[llvm] AMDGPU/GlobalISel: RegBankLegalize rules for mfma_scale (PR #203279)
Petar Avramovic via llvm-commits
llvm-commits at lists.llvm.org
Fri Jun 12 05:04:50 PDT 2026
https://github.com/petar-avramovic updated https://github.com/llvm/llvm-project/pull/203279
>From 67444b66423bd0b3fab9d10dd272fb275e22239a Mon Sep 17 00:00:00 2001
From: Petar Avramovic <Petar.Avramovic at amd.com>
Date: Thu, 11 Jun 2026 15:44:46 +0200
Subject: [PATCH] AMDGPU/GlobalISel: RegBankLegalize rules for mfma_scale
---
.../AMDGPU/AMDGPURegBankLegalizeRules.cpp | 7 +
....amdgcn.mfma.scale.f32.16x16x128.f8f6f4.ll | 4207 +++++++-
...m.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.ll | 9288 ++++++++++++++++-
3 files changed, 13500 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
index ece2552ef32ae..8db23307d4ef6 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
@@ -2642,6 +2642,13 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
{{VgprOrAgprAnyTy},
{IntrId, VgprAnyTy, VgprAnyTy, VgprOrAgprAnyTy, VgprAnyTy}}});
+ addRulesForIOpcs({amdgcn_mfma_scale_f32_32x32x64_f8f6f4,
+ amdgcn_mfma_scale_f32_16x16x128_f8f6f4})
+ .Any({{DivAnyTy},
+ {{VgprOrAgprAnyTy},
+ {IntrId, VgprOrAgprAnyTy, VgprOrAgprAnyTy, VgprOrAgprAnyTy, Imm,
+ Imm, Imm, Vgpr32, Imm, Vgpr32}}});
+
// WMMA/SWMMAC intrinsics: all register operands map to VGPR.
addRulesForIOpcs(
{// WMMA GFX11+
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.ll
index 20f6f5021d5ef..7e935e5ad483a 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.ll
@@ -1,6 +1,8 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=0 < %s | FileCheck -check-prefixes=GCN,SDAG %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=1 < %s | FileCheck -check-prefixes=GCN,GISEL %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=1 -new-reg-bank-select < %s | FileCheck -check-prefixes=GCN,GISEL %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=0 --amdgpu-mfma-vgpr-form=0 < %s | FileCheck -check-prefixes=AGPR,AGPR-SDAG %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=1 -new-reg-bank-select --amdgpu-mfma-vgpr-form=0 < %s | FileCheck -check-prefixes=AGPR,AGPR-GISEL %s
; 0 = fp8
; 1 = bf8
@@ -19,6 +21,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp0(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -32,6 +82,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_1_1__cbsz1__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[0,0,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_1_1__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel:[1,1,0] op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_1_1__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -45,6 +143,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_2_2__cbsz1__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[1,1,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_2_2__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_2_2__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -58,6 +204,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_3_3__cbsz1__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[1,1,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_3_3__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_3_3__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -71,6 +265,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_3__cbsz1__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[0,1,0] op_sel_hi:[0,1,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_3__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel:[0,1,0] op_sel_hi:[0,1,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_3__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel:[0,1,0] op_sel_hi:[0,1,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -84,6 +326,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_3_0__cbsz1__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,0,0] op_sel_hi:[1,0,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_3_0__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel:[1,0,0] op_sel_hi:[1,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_3_0__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel:[1,0,0] op_sel_hi:[1,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -97,6 +387,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_2_3__cbsz1__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[0,1,0] op_sel_hi:[1,1,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_2_3__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel:[0,1,0] op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_2_3__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel:[0,1,0] op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -110,6 +448,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_3_2__cbsz1__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,0,0] op_sel_hi:[1,1,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_3_2__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel:[1,0,0] op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_3_2__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel:[1,0,0] op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -124,6 +510,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp0__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -138,6 +572,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 1, ; blgp
@@ -152,6 +634,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp1__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19] blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3] blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19] blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 1, ; blgp
@@ -166,6 +696,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp2(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17], v18, v19 op_sel_hi:[0,0,0] blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 2, ; blgp
@@ -180,6 +756,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp2__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17] blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp2__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3] blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp2__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17] blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 2, ; blgp
@@ -194,6 +816,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp3(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17], v18, v19 op_sel_hi:[0,0,0] blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 3, ; blgp
@@ -208,6 +876,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp3__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17] blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3] blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17] blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 3, ; blgp
@@ -222,6 +936,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp4(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:11], v[12:15], v16, v17 op_sel_hi:[0,0,0] blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:11], a[0:3], v16, v17 op_sel_hi:[0,0,0] blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:11], a[12:15], v16, v17 op_sel_hi:[0,0,0] blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v4i32(<8 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 4, ; blgp
@@ -236,6 +994,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp4__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:11], v[12:15] blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:11], a[0:3] blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz0__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:11], a[12:15] blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v4i32(<8 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 4, ; blgp
@@ -250,6 +1052,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp0(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] cbsz:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 0, ; blgp
@@ -264,6 +1114,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp0__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19] cbsz:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3] cbsz:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19] cbsz:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 0, ; blgp
@@ -278,6 +1176,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp1(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:1 blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] cbsz:1 blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:1 blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 1, ; blgp
@@ -293,6 +1239,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp1__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19] cbsz:1 blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3] cbsz:1 blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19] cbsz:1 blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 1, ; blgp
@@ -307,6 +1301,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp2(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:1 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] cbsz:1 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:1 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 2, ; blgp
@@ -320,6 +1360,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp2__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17] cbsz:1 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp2__constant_scale_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3] cbsz:1 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp2__constant_scale_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17] cbsz:1 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 2, ; blgp
@@ -334,6 +1420,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp3(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:1 blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] cbsz:1 blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:1 blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 3, ; blgp
@@ -348,6 +1480,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp3__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17] cbsz:1 blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3] cbsz:1 blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17] cbsz:1 blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 3, ; blgp
@@ -362,6 +1540,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp4(<8 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:11], v[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:1 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:11], a[0:3], v16, v17 op_sel_hi:[0,0,0] cbsz:1 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:11], a[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:1 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v4i32(<8 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 4, ; blgp
@@ -376,6 +1598,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp4__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:11], v[12:15] cbsz:1 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:11], a[0:3] cbsz:1 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz1__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:11], a[12:15] cbsz:1 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v4i32(<8 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 1, ; cbsz
i32 4, ; blgp
@@ -390,6 +1656,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp0(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] cbsz:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 0, ; blgp
@@ -404,6 +1716,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp0__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17] cbsz:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3] cbsz:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17] cbsz:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 0, ; blgp
@@ -418,6 +1776,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp1(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:2 blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] cbsz:2 blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:2 blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 1, ; blgp
@@ -432,6 +1836,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp1__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17] cbsz:2 blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3] cbsz:2 blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17] cbsz:2 blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 1, ; blgp
@@ -446,6 +1896,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp2(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:11], v[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:2 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:11], a[0:3], v16, v17 op_sel_hi:[0,0,0] cbsz:2 blgp:2
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:11], a[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:2 blgp:2
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 2, ; blgp
@@ -460,6 +1954,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp2__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:11], v[12:15] cbsz:2 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp2__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:11], a[0:3] cbsz:2 blgp:2
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp2__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:11], a[12:15] cbsz:2 blgp:2
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 2, ; blgp
@@ -474,6 +2012,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp3(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:11], v[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:2 blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:11], a[0:3], v16, v17 op_sel_hi:[0,0,0] cbsz:2 blgp:3
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:11], a[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:2 blgp:3
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 3, ; blgp
@@ -488,6 +2070,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp3__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:11], v[12:15] cbsz:2 blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:11], a[0:3] cbsz:2 blgp:3
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:11], a[12:15] cbsz:2 blgp:3
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 3, ; blgp
@@ -503,6 +2129,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp0(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] cbsz:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 0, ; blgp
@@ -517,6 +2189,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp0__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17] cbsz:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3] cbsz:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17] cbsz:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 0, ; blgp
@@ -531,6 +2249,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp1(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:3 blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] cbsz:3 blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:3 blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 1, ; blgp
@@ -545,6 +2309,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp1__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17] cbsz:3 blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3] cbsz:3 blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17] cbsz:3 blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 1, ; blgp
@@ -559,6 +2369,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp2(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:11], v[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:3 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:11], a[0:3], v16, v17 op_sel_hi:[0,0,0] cbsz:3 blgp:2
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:11], a[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:3 blgp:2
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 2, ; blgp
@@ -573,6 +2427,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp2__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:11], v[12:15] cbsz:3 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp2__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:11], a[0:3] cbsz:3 blgp:2
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp2__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:11], a[12:15] cbsz:3 blgp:2
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 2, ; blgp
@@ -587,6 +2485,48 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp4(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:9], v[10:13], v14, v15 op_sel_hi:[0,0,0] cbsz:3 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:9], a[0:3], v14, v15 op_sel_hi:[0,0,0] cbsz:3 blgp:4
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:9], a[10:13], v14, v15 op_sel_hi:[0,0,0] cbsz:3 blgp:4
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v4i32(<6 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 4, ; blgp
@@ -601,6 +2541,48 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp4__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:9], v[10:13] cbsz:3 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:9], a[0:3] cbsz:3 blgp:4
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:9], a[10:13] cbsz:3 blgp:4
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v4i32(<6 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 4, ; blgp
@@ -615,6 +2597,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp3(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:11], v[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:3 blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:11], a[0:3], v16, v17 op_sel_hi:[0,0,0] cbsz:3 blgp:3
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:11], a[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:3 blgp:3
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 3, ; blgp
@@ -629,6 +2655,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp3__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:11], v[12:15] cbsz:3 blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:11], a[0:3] cbsz:3 blgp:3
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz3__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:11], a[12:15] cbsz:3 blgp:3
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 3, ; cbsz
i32 3, ; blgp
@@ -643,6 +2713,48 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp4(<6 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:9], v[10:13], v14, v15 op_sel_hi:[0,0,0] cbsz:2 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:9], a[0:3], v14, v15 op_sel_hi:[0,0,0] cbsz:2 blgp:4
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:9], a[10:13], v14, v15 op_sel_hi:[0,0,0] cbsz:2 blgp:4
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v4i32(<6 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 4, ; blgp
@@ -657,6 +2769,48 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp4__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:9], v[10:13] cbsz:2 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:9], a[0:3] cbsz:2 blgp:4
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz2__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:5], a[6:9], a[10:13] cbsz:2 blgp:4
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v4i32(<6 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 4, ; blgp
@@ -671,6 +2825,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp0(<4 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:11], v[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:11], a[0:3], v16, v17 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[8:11], a[0:7], a[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v8i32(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 0, ; blgp
@@ -685,6 +2883,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp0__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:11], v[12:15] cbsz:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:11], a[0:3] cbsz:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[8:11], a[0:7], a[12:15] cbsz:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v8i32(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 0, ; blgp
@@ -699,6 +2941,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp1(<4 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:11], v[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:4 blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:11], a[0:3], v16, v17 op_sel_hi:[0,0,0] cbsz:4 blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[8:11], a[0:7], a[12:15], v16, v17 op_sel_hi:[0,0,0] cbsz:4 blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v8i32(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 1, ; blgp
@@ -713,6 +2999,50 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp1__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:11], v[12:15] cbsz:4 blgp:1
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:11], a[0:3] cbsz:4 blgp:1
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[8:11], a[0:7], a[12:15] cbsz:4 blgp:1
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v8i32(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 1, ; blgp
@@ -727,6 +3057,48 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp2(<4 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:9], v[10:13], v14, v15 op_sel_hi:[0,0,0] cbsz:4 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:9], a[0:3], v14, v15 op_sel_hi:[0,0,0] cbsz:4 blgp:2
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[6:9], a[0:5], a[10:13], v14, v15 op_sel_hi:[0,0,0] cbsz:4 blgp:2
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v6i32(<4 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 2, ; blgp
@@ -741,6 +3113,48 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp2__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:9], v[10:13] cbsz:4 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp2__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:9], a[0:3] cbsz:4 blgp:2
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp2__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[6:9], a[0:5], a[10:13] cbsz:4 blgp:2
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v6i32(<4 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 2, ; blgp
@@ -755,6 +3169,48 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp3(<4 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:9], v[10:13], v14, v15 op_sel_hi:[0,0,0] cbsz:4 blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:9], a[0:3], v14, v15 op_sel_hi:[0,0,0] cbsz:4 blgp:3
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[6:9], a[0:5], a[10:13], v14, v15 op_sel_hi:[0,0,0] cbsz:4 blgp:3
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v6i32(<4 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 3, ; blgp
@@ -769,6 +3225,48 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp3__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:9], v[10:13] cbsz:4 blgp:3
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:9], a[0:3] cbsz:4 blgp:3
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[6:9], a[0:5], a[10:13] cbsz:4 blgp:3
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v6i32(<4 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 3, ; blgp
@@ -783,6 +3281,46 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp4(<4 x
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:7], v[8:11], v12, v13 op_sel_hi:[0,0,0] cbsz:4 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:7], a[0:3], v12, v13 op_sel_hi:[0,0,0] cbsz:4 blgp:4
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:3], a[4:7], a[8:11], v12, v13 op_sel_hi:[0,0,0] cbsz:4 blgp:4
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v4i32(<4 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 4, ; blgp
@@ -797,6 +3335,46 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp4__cons
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:3], v[4:7], v[8:11] cbsz:4 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:3], v[4:7], a[0:3] cbsz:4 blgp:4
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__cbsz4__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:3], a[4:7], a[8:11] cbsz:4 blgp:4
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v4i32.v4i32(<4 x i32> %arg0, <4 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 4, ; blgp
@@ -817,6 +3395,58 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__sgpr_scaleA__sgpr_
; GCN-NEXT: s_nop 1
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__sgpr_scaleA__sgpr_scaleB:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s1
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v16, v17 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__sgpr_scaleA__sgpr_scaleB:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, s1
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v0, v1 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -829,6 +3459,56 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__sgpr_scaleA__vgpr_
; GCN-NEXT: s_nop 1
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v21, v20 op_sel_hi:[0,0,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__sgpr_scaleA__vgpr_scaleB:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s0
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v16, v20 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__sgpr_scaleA__vgpr_scaleB:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s0
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v0, v20 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -841,6 +3521,56 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__vgpr_scaleA__sgpr_
; GCN-NEXT: s_nop 1
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__vgpr_scaleA__sgpr_scaleB:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s0
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v16 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__vgpr_scaleA__sgpr_scaleB:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s0
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -908,6 +3638,74 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgprs(<8 x i32> inr
; GISEL-NEXT: s_mov_b64 exec, s[0:1]
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgprs:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s19
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s0, v1
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s1, v0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s23
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s24
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s25
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s26
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, s28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, s29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, s1
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, s0
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[12:19], v[4:11], a[0:3], v2, v3 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgprs:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s19
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s4, v0
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s5, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, s20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, s21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, s22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, s23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, s24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, s25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, s26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, s27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s5
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v2, v3 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -944,6 +3742,64 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_vgpr__sgp
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[14:21], v[0:7], v[8:11], v13, v12 op_sel_hi:[0,0,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_vgpr__sgpr_vgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s20
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[14:21], v[0:7], a[0:3], v8, v12 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_vgpr__sgpr_vgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v11
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s20
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v0, v12 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -980,6 +3836,64 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_vgpr__vgp
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[14:21], v[0:7], v[8:11], v12, v13 op_sel_hi:[0,0,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_vgpr__vgpr_sgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s20
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[14:21], v[0:7], a[0:3], v12, v8 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_vgpr__vgpr_sgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v11
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s20
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v12, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -1016,6 +3930,64 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0_vgpr_sgpr_vgpr__vgp
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[14:21], v[8:11], v12, v13 op_sel_hi:[0,0,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_vgpr_sgpr_vgpr__vgpr_sgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s20
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[14:21], a[0:3], v12, v8 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_vgpr_sgpr_vgpr__vgpr_sgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v11
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s20
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v12, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -1042,6 +4014,56 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0_vgpr_vgpr_sgpr__vgp
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[18:21], v16, v17 op_sel_hi:[0,0,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_vgpr_vgpr_sgpr__vgpr_sgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, s0
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, s1
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, s2
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s16
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v16, v17 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_vgpr_vgpr_sgpr__vgpr_sgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s3
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s16
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v16, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -1084,6 +4106,64 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_sgpr__vgp
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[10:17], v[0:7], v[18:21], v8, v9 op_sel_hi:[0,0,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_sgpr__vgpr_sgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, s20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, s21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, s22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, s23
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s24
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[10:17], v[0:7], a[0:3], v8, v9 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0_sgpr_vgpr_sgpr__vgpr_sgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s23
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s24
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v8, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <4 x float> %result
}
@@ -1106,6 +4186,58 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_inlineimm__
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[1,1,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_inlineimm__scaleB_inlineimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, -2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, 33
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v21, v20 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_inlineimm__scaleB_inlineimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 33
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, -2
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 2, i32 33, i32 2, i32 -2)
ret <4 x float> %result
}
@@ -1128,6 +4260,58 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_kimm__scale
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[1,1,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_kimm__scaleB_inlineimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, -2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, 0x41
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v21, v20 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_kimm__scaleB_inlineimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 0x41
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, -2
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 2, i32 65, i32 2, i32 -2)
ret <4 x float> %result
}
@@ -1150,6 +4334,58 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_kimm__scale
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[1,1,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_kimm__scaleB_kimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, 0x4d
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, 0x41
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v21, v20 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4_0_0__scaleA_kimm__scaleB_kimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 0x41
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 0x4d
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 2, i32 65, i32 2, i32 77)
ret <4 x float> %result
}
@@ -1213,6 +4449,65 @@ define amdgpu_kernel void @test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd(<8 x i32
; GISEL-NEXT: s_nop 10
; GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[30:31]
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, 0
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v0, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v1, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s15
+; AGPR-SDAG-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x40
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s23
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, s13
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v21, v22 op_sel:[1,1,0] op_sel_hi:[1,0,0] blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: global_store_dwordx4 v20, v[0:3], s[14:15]
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx8 s[24:31], s[4:5], 0x40
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[24:25]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[26:27]
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v20, s28
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v21, s29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0] blgp:2
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v4, 0
+; AGPR-GISEL-NEXT: s_nop 10
+; AGPR-GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[30:31]
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 2, i32 3, i32 %scale0, i32 1, i32 %scale1)
store <4 x float> %result, ptr addrspace(1) %ptr, align 16
ret void
@@ -1276,6 +4571,64 @@ define amdgpu_kernel void @test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA
; GISEL-NEXT: s_nop 10
; GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA_kimm__scaleB__inlineimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x40
+; AGPR-SDAG-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x50
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, -2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, 0x41
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, 0
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v0, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v1, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s15
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[18:19], s[2:3]
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[16:17], s[0:1]
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v22, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: global_store_dwordx4 v20, v[0:3], s[6:7]
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA_kimm__scaleB__inlineimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x40
+; AGPR-GISEL-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x50
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v20, 0x41
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v21, -2
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[2:3]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[0:1]
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0]
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v4, 0
+; AGPR-GISEL-NEXT: s_nop 10
+; AGPR-GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 3, i32 65, i32 1, i32 -2)
store <4 x float> %result, ptr addrspace(1) %ptr, align 16
ret void
@@ -1339,6 +4692,64 @@ define amdgpu_kernel void @test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA
; GISEL-NEXT: s_nop 10
; GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA_kimm__scaleB__FP_literal:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x40
+; AGPR-SDAG-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x50
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, 1.0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, 0x41
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, 0
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v0, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v1, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s15
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[18:19], s[2:3]
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[16:17], s[0:1]
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v22, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: global_store_dwordx4 v20, v[0:3], s[6:7]
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA_kimm__scaleB__FP_literal:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x40
+; AGPR-GISEL-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x50
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v20, 0x41
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v21, 1.0
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[2:3]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[0:1]
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0]
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v4, 0
+; AGPR-GISEL-NEXT: s_nop 10
+; AGPR-GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 3, i32 65, i32 1, i32 1065353216)
store <4 x float> %result, ptr addrspace(1) %ptr, align 16
ret void
@@ -1402,6 +4813,64 @@ define amdgpu_kernel void @test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA
; GISEL-NEXT: s_nop 10
; GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA_FP_literal__scaleB__inline_imm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x40
+; AGPR-SDAG-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x50
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, -2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, 1.0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, 0
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v0, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v1, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s15
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[18:19], s[2:3]
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[16:17], s[0:1]
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v22, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: global_store_dwordx4 v20, v[0:3], s[6:7]
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA_FP_literal__scaleB__inline_imm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x40
+; AGPR-GISEL-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x50
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v20, 1.0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v21, -2
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[2:3]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[0:1]
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0]
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v4, 0
+; AGPR-GISEL-NEXT: s_nop 10
+; AGPR-GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 3, i32 1065353216, i32 1, i32 -2)
store <4 x float> %result, ptr addrspace(1) %ptr, align 16
ret void
@@ -1465,6 +4934,64 @@ define amdgpu_kernel void @test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA
; GISEL-NEXT: s_nop 10
; GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA_FP_literal__scaleB__FP_literal:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x40
+; AGPR-SDAG-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x50
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, 0.15915494
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, 1.0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, 0
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v0, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v1, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s15
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[18:19], s[2:3]
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[16:17], s[0:1]
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v22, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: global_store_dwordx4 v20, v[0:3], s[6:7]
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4__vgprcd___scaleA_FP_literal__scaleB__FP_literal:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x40
+; AGPR-GISEL-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x50
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v20, 1.0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v21, 0.15915494
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[2:3]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[0:1]
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel:[1,1,0] op_sel_hi:[1,0,0]
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v4, 0
+; AGPR-GISEL-NEXT: s_nop 10
+; AGPR-GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 3, i32 1065353216, i32 1, i32 1042479491)
store <4 x float> %result, ptr addrspace(1) %ptr, align 16
ret void
@@ -1477,6 +5004,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_0_a(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_0_a:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_0_a:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)
ret <4 x float> %result
}
@@ -1488,6 +5063,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_0_b(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19]
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_0_b:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_0_b:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 3, i32 0, i32 1, i32 0)
ret <4 x float> %result
}
@@ -1510,6 +5133,58 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_1(<8
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, 1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, 0
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v21, v20 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_0_1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 1
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v0, v1 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 0, i32 0, i32 1)
ret <4 x float> %result
}
@@ -1532,6 +5207,58 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_1_0_a(
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_1_0_a:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, 0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, 1
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v21, v20 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___constant_scale_1_0_a:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 1
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 0
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v0, v1 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 0, i32 0, i32 0, i32 1, i32 0, i32 0)
ret <4 x float> %result
}
@@ -1546,6 +5273,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v8i32_fp6(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v8i32_fp6:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v8i32_fp6:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 2, ; blgp
@@ -1559,6 +5334,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp8(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp8:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] cbsz:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp8:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 0, ; blgp
@@ -1572,6 +5395,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp6(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:2 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp6:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] cbsz:2 blgp:2
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp6:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:2 blgp:2
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 2, ; blgp
@@ -1585,6 +5456,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp6_
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19] cbsz:2 blgp:2
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp6__0_scale:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3] cbsz:2 blgp:2
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp6__v8i32_fp6__0_scale:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19] cbsz:2 blgp:2
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 2, ; cbsz
i32 2, ; blgp
@@ -1598,6 +5517,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v8i32_fp4(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v8i32_fp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v8i32_fp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 4, ; blgp
@@ -1611,6 +5578,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp8(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp8:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp8:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 0, ; blgp
@@ -1624,6 +5639,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v6i32_fp4(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v6i32_fp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp8__v6i32_fp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:13], a[14:17], v18, v19 op_sel_hi:[0,0,0] blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <4 x float> %arg2,
i32 0, ; cbsz
i32 4, ; blgp
@@ -1637,6 +5698,52 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v6i32_fp4__v8i32_fp8(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:5], v[6:13], v[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v6i32_fp4__v8i32_fp8:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:5], v[6:13], a[0:3], v18, v19 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v6i32_fp4__v8i32_fp8:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[8:13], a[0:7], a[14:17], v18, v19 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 0, ; blgp
@@ -1650,6 +5757,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp4(
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:4 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3], v20, v21 op_sel_hi:[0,0,0] cbsz:4 blgp:4
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19], v20, v21 op_sel_hi:[0,0,0] cbsz:4 blgp:4
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 4, ; blgp
@@ -1663,6 +5818,54 @@ define <4 x float> @test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp4_
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[0:7], v[8:15], v[16:19] cbsz:4 blgp:4
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp4__0_scale:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], v[0:7], v[8:15], a[0:3] cbsz:4 blgp:4
+; AGPR-SDAG-NEXT: s_nop 7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_16x16x128_f8f6f4___v8i32_fp4__v8i32_fp4__0_scale:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v19
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_16x16x128_f8f6f4 a[0:3], a[0:7], a[8:15], a[16:19] cbsz:4 blgp:4
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2,
i32 4, ; cbsz
i32 4, ; blgp
@@ -1683,3 +5886,5 @@ declare <4 x float> @llvm.amdgcn.mfma.scale.f32.16x16x128.f8f6f4.v8i32.v6i32(<8
attributes #0 = { "amdgpu-flat-work-group-size"="512,512" "amdgpu-agpr-alloc"="0,0" nounwind }
attributes #1 = { convergent nocallback nofree nosync nounwind willreturn memory(none) }
attributes #2 = { nounwind }
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; AGPR: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.ll
index f9332d752482d..e2838438f6cce 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.ll
@@ -1,6 +1,8 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=0 < %s | FileCheck -check-prefixes=GCN,SDAG %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=1 < %s | FileCheck -check-prefixes=GCN,GISEL %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=1 -new-reg-bank-select < %s | FileCheck -check-prefixes=GCN,GISEL %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=0 --amdgpu-mfma-vgpr-form=0 < %s | FileCheck -check-prefixes=AGPR,AGPR-SDAG %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -global-isel=1 -new-reg-bank-select --amdgpu-mfma-vgpr-form=0 < %s | FileCheck -check-prefixes=AGPR,AGPR-GISEL %s
; 0 = fp8
; 1 = bf8
@@ -69,6 +71,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp0(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -132,6 +238,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_1_1__cbsz1__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_1_1__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel:[1,1,0] op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_1_1__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel:[1,1,0] op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -195,6 +405,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_2_2__cbsz1__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_2_2__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_2_2__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -258,6 +572,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_3_3__cbsz1__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_3_3__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_3_3__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -321,6 +739,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_3__cbsz1__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_3__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel:[0,1,0] op_sel_hi:[0,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_3__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel:[0,1,0] op_sel_hi:[0,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -384,6 +906,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_3_0__cbsz1__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_3_0__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel:[1,0,0] op_sel_hi:[1,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_3_0__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel:[1,0,0] op_sel_hi:[1,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -447,6 +1073,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_2_3__cbsz1__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_2_3__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel:[0,1,0] op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_2_3__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel:[0,1,0] op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -510,6 +1240,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_3_2__cbsz1__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_3_2__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel:[1,0,0] op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_3_2__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel:[1,0,0] op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -544,6 +1378,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp0__cons
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 0, ; blgp
@@ -608,6 +1542,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 1, ; blgp
@@ -641,6 +1679,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp1__cons
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15] blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15] blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 1, ; blgp
@@ -675,6 +1813,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp2(<8 x
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] blgp:2
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15], v30, v31 op_sel_hi:[0,0,0] blgp:2
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 2, ; blgp
@@ -706,6 +1944,102 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp2__cons
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp2__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15] blgp:2
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp2__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15] blgp:2
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 2, ; blgp
@@ -740,6 +2074,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp3(<8 x
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] blgp:3
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15], v30, v31 op_sel_hi:[0,0,0] blgp:3
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 3, ; blgp
@@ -771,6 +2205,102 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp3__cons
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15] blgp:3
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15] blgp:3
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 3, ; blgp
@@ -803,6 +2333,100 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp4(<8 x
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:11], a[0:15], v28, v29 op_sel_hi:[0,0,0] blgp:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:27], a[0:15], v28, v29 op_sel_hi:[0,0,0] blgp:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v4i32(<8 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 4, ; blgp
@@ -834,6 +2458,100 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp4__cons
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:11], a[0:15] blgp:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz0__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:27], a[0:15] blgp:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v4i32(<8 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 4, ; blgp
@@ -898,6 +2616,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp0(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] cbsz:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] cbsz:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 0, ; blgp
@@ -931,6 +2753,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp0__cons
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15] cbsz:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15] cbsz:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 0, ; blgp
@@ -995,6 +2917,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp1(<8 x
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] cbsz:1 blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] cbsz:1 blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 1, ; blgp
@@ -1029,6 +3055,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp1__cons
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15] cbsz:1 blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15] cbsz:1 blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 1, ; blgp
@@ -1063,6 +3189,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp2(<8 x
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:1 blgp:2
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:1 blgp:2
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 2, ; blgp
@@ -1094,6 +3320,102 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp2__cons
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp2__constant_scale_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15] cbsz:1 blgp:2
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp2__constant_scale_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15] cbsz:1 blgp:2
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 2, ; blgp
@@ -1128,6 +3450,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp3(<8 x
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:1 blgp:3
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:1 blgp:3
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 3, ; blgp
@@ -1159,6 +3581,102 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp3__cons
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15] cbsz:1 blgp:3
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15] cbsz:1 blgp:3
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 3, ; blgp
@@ -1191,6 +3709,100 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp4(<8 x
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:11], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:1 blgp:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:27], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:1 blgp:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v4i32(<8 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 4, ; blgp
@@ -1222,6 +3834,100 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp4__cons
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:11], a[0:15] cbsz:1 blgp:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz1__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:27], a[0:15] cbsz:1 blgp:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v4i32(<8 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 1, ; cbsz
i32 4, ; blgp
@@ -1256,6 +3962,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp0(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:2
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:2
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 0, ; blgp
@@ -1287,6 +4093,102 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp0__cons
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15] cbsz:2
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15] cbsz:2
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 0, ; blgp
@@ -1321,6 +4223,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp1(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:2 blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:2 blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 1, ; blgp
@@ -1352,6 +4354,102 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp1__cons
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15] cbsz:2 blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15] cbsz:2 blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 1, ; blgp
@@ -1383,6 +4481,98 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp2(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:11], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:2 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:27], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:2 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 2, ; blgp
@@ -1413,6 +4603,98 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp2__cons
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp2__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:11], a[0:15] cbsz:2 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp2__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:27], a[0:15] cbsz:2 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 2, ; blgp
@@ -1444,6 +4726,98 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp3(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:11], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:2 blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:27], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:2 blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 3, ; blgp
@@ -1474,6 +4848,98 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp3__cons
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:11], a[0:15] cbsz:2 blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:27], a[0:15] cbsz:2 blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 3, ; blgp
@@ -1509,6 +4975,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp0(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:3
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:3
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 0, ; blgp
@@ -1540,6 +5106,102 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp0__cons
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15] cbsz:3
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15] cbsz:3
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 0, ; blgp
@@ -1574,6 +5236,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp1(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:3 blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:3 blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 1, ; blgp
@@ -1605,6 +5367,102 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp1__cons
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15] cbsz:3 blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15] cbsz:3 blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 1, ; blgp
@@ -1636,6 +5494,98 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp2(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:11], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:3 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:27], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:3 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 2, ; blgp
@@ -1666,6 +5616,98 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp2__cons
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp2__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:11], a[0:15] cbsz:3 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp2__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:27], a[0:15] cbsz:3 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 2, ; blgp
@@ -1697,6 +5739,96 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp4(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v24
; GCN-NEXT: v_mov_b32_e32 v15, v25
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:9], a[0:15], v26, v27 op_sel_hi:[0,0,0] cbsz:3 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:25], a[0:15], v26, v27 op_sel_hi:[0,0,0] cbsz:3 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v4i32(<6 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 4, ; blgp
@@ -1727,6 +5859,96 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp4__cons
; GCN-NEXT: v_mov_b32_e32 v14, v24
; GCN-NEXT: v_mov_b32_e32 v15, v25
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:9], a[0:15] cbsz:3 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:25], a[0:15] cbsz:3 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v4i32(<6 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 4, ; blgp
@@ -1758,6 +5980,98 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp3(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:11], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:3 blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:27], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:3 blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 3, ; blgp
@@ -1788,6 +6102,98 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp3__cons
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:11], a[0:15] cbsz:3 blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz3__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:27], a[0:15] cbsz:3 blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v6i32(<6 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 3, ; cbsz
i32 3, ; blgp
@@ -1819,6 +6225,96 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp4(<6 x
; GCN-NEXT: v_mov_b32_e32 v14, v24
; GCN-NEXT: v_mov_b32_e32 v15, v25
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:9], a[0:15], v26, v27 op_sel_hi:[0,0,0] cbsz:2 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:25], a[0:15], v26, v27 op_sel_hi:[0,0,0] cbsz:2 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v4i32(<6 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 4, ; blgp
@@ -1849,6 +6345,96 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp4__cons
; GCN-NEXT: v_mov_b32_e32 v14, v24
; GCN-NEXT: v_mov_b32_e32 v15, v25
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:9], a[0:15] cbsz:2 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz2__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:21], a[22:25], a[0:15] cbsz:2 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v4i32(<6 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 4, ; blgp
@@ -1881,6 +6467,100 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp0(<4 x
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:11], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[24:27], a[16:23], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v8i32(<4 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 0, ; blgp
@@ -1912,6 +6592,100 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp0__cons
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp0__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:11], a[0:15] cbsz:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp0__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[24:27], a[16:23], a[0:15] cbsz:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v8i32(<4 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 0, ; blgp
@@ -1944,6 +6718,100 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp1(<4 x
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:11], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:4 blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[24:27], a[16:23], a[0:15], v28, v29 op_sel_hi:[0,0,0] cbsz:4 blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v8i32(<4 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 1, ; blgp
@@ -1975,6 +6843,100 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp1__cons
; GCN-NEXT: v_mov_b32_e32 v14, v26
; GCN-NEXT: v_mov_b32_e32 v15, v27
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp1__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:11], a[0:15] cbsz:4 blgp:1
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp1__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v27
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[24:27], a[16:23], a[0:15] cbsz:4 blgp:1
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v8i32(<4 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 1, ; blgp
@@ -2006,6 +6968,96 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp2(<4 x
; GCN-NEXT: v_mov_b32_e32 v14, v24
; GCN-NEXT: v_mov_b32_e32 v15, v25
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp2:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:9], a[0:15], v26, v27 op_sel_hi:[0,0,0] cbsz:4 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp2:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[22:25], a[16:21], a[0:15], v26, v27 op_sel_hi:[0,0,0] cbsz:4 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v6i32(<4 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 2, ; blgp
@@ -2036,6 +7088,96 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp2__cons
; GCN-NEXT: v_mov_b32_e32 v14, v24
; GCN-NEXT: v_mov_b32_e32 v15, v25
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp2__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:9], a[0:15] cbsz:4 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp2__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[22:25], a[16:21], a[0:15] cbsz:4 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v6i32(<4 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 2, ; blgp
@@ -2067,6 +7209,96 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp3(<4 x
; GCN-NEXT: v_mov_b32_e32 v14, v24
; GCN-NEXT: v_mov_b32_e32 v15, v25
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp3:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:9], a[0:15], v26, v27 op_sel_hi:[0,0,0] cbsz:4 blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp3:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[22:25], a[16:21], a[0:15], v26, v27 op_sel_hi:[0,0,0] cbsz:4 blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v6i32(<4 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 3, ; blgp
@@ -2097,6 +7329,96 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp3__cons
; GCN-NEXT: v_mov_b32_e32 v14, v24
; GCN-NEXT: v_mov_b32_e32 v15, v25
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp3__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:9], a[0:15] cbsz:4 blgp:3
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp3__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v25
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[22:25], a[16:21], a[0:15] cbsz:4 blgp:3
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v6i32(<4 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 3, ; blgp
@@ -2128,6 +7450,94 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp4(<4 x
; GCN-NEXT: v_mov_b32_e32 v14, v22
; GCN-NEXT: v_mov_b32_e32 v15, v23
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:7], a[0:15], v24, v25 op_sel_hi:[0,0,0] cbsz:4 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:19], a[20:23], a[0:15], v24, v25 op_sel_hi:[0,0,0] cbsz:4 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v4i32(<4 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 4, ; blgp
@@ -2158,6 +7568,94 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp4__cons
; GCN-NEXT: v_mov_b32_e32 v14, v22
; GCN-NEXT: v_mov_b32_e32 v15, v23
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp4__constant_scale_0_0:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:3], v[4:7], a[0:15] cbsz:4 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__cbsz4__blgp4__constant_scale_0_0:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:19], a[20:23], a[0:15] cbsz:4 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v4i32.v4i32(<4 x i32> %arg0, <4 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 4, ; blgp
@@ -2198,6 +7696,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__sgpr_scaleA__sgpr_
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__sgpr_scaleA__sgpr_scaleB:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v16, v17 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__sgpr_scaleA__sgpr_scaleB:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, s1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2231,6 +7833,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__sgpr_scaleA__vgpr_
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__sgpr_scaleA__vgpr_scaleB:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s0
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v16, v31 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__sgpr_scaleA__vgpr_scaleB:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s0
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v31 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2264,6 +7970,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__vgpr_scaleA__sgpr_
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__vgpr_scaleA__sgpr_scaleB:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s0
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v31, v16 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__vgpr_scaleA__sgpr_scaleB:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s0
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2390,6 +8200,148 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgprs(<8 x i32> inr
; GISEL-NEXT: s_mov_b64 exec, s[0:1]
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgprs:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v23, s19
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s4, v13
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s5, v12
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s6, v11
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s7, v10
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s8, v9
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s9, v8
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s10, v7
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s11, v6
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s12, v5
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s13, v4
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s14, v3
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s15, v2
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s40, v1
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s41, v0
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, s28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, s29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, s41
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, s40
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, s15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, s14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, s13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, s12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, s11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, s10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, s9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, s8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, s7
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, s6
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, s5
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, s4
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v0, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v1, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s23
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s24
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s25
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s26
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s27
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[16:23], v[0:7], a[0:15], v14, v15 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgprs:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, s19
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s4, v0
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s5, v1
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s6, v2
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s7, v3
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s8, v4
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s9, v5
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s10, v6
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s11, v7
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s12, v8
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s13, v9
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s14, v10
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s15, v11
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s40, v12
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s41, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, s20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, s21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, s22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, s23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, s24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, s25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, s26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, s27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, s10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, s11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, s12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, s13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, s14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, s15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, s40
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, s41
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v14, v15 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2462,6 +8414,114 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_vgpr__sgp
; GISEL-NEXT: v_mov_b32_e32 v14, v22
; GISEL-NEXT: v_mov_b32_e32 v15, v23
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_vgpr__sgpr_vgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v26, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v27, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v28, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v29, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v30, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v33, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s20
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[26:33], v[0:7], a[0:15], v8, v24 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_vgpr__sgpr_vgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s20
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v24 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2534,6 +8594,114 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_vgpr__vgp
; GISEL-NEXT: v_mov_b32_e32 v14, v22
; GISEL-NEXT: v_mov_b32_e32 v15, v23
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_vgpr__vgpr_sgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v26, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v27, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v28, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v29, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v30, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v33, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s20
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[26:33], v[0:7], a[0:15], v24, v8 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_vgpr__vgpr_sgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s20
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v24, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2606,6 +8774,114 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0_vgpr_sgpr_vgpr__vgp
; GISEL-NEXT: v_mov_b32_e32 v14, v22
; GISEL-NEXT: v_mov_b32_e32 v15, v23
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_vgpr_sgpr_vgpr__vgpr_sgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v26, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v27, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v28, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v29, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v30, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v33, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s20
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[26:33], a[0:15], v24, v8 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_vgpr_sgpr_vgpr__vgpr_sgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v23
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s20
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v24, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2686,6 +8962,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0_vgpr_vgpr_sgpr__vgp
; GISEL-NEXT: s_nop 1
; GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[18:25], v[26:33], v[0:15], v16, v17 op_sel_hi:[0,0,0]
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_vgpr_vgpr_sgpr__vgpr_sgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, s0
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, s1
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, s2
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, s3
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, s16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, s17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, s18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, s20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, s21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, s22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, s23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, s24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, s25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, s26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, s27
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s28
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v16, v17 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_vgpr_vgpr_sgpr__vgpr_sgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, s20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, s21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, s22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, s23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, s24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, s25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, s26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, s27
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s28
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v16, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2811,6 +9187,128 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_sgpr__vgp
; GISEL-NEXT: s_mov_b64 exec, s[0:1]
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_sgpr__vgpr_sgpr:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s3
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v23, s19
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s4, v13
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s5, v12
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s6, v11
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s7, v10
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s8, v9
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s9, v8
+; AGPR-SDAG-NEXT: v_readfirstlane_b32 s0, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, s20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, s21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, s22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, s23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, s24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, s25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, s26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, s27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, s28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, s29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, s9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, s8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, s7
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, s6
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, s5
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, s4
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s0
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[16:23], v[0:7], a[0:15], v14, v8 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0_sgpr_vgpr_sgpr__vgpr_sgpr:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, s19
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s4, v8
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s5, v9
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s6, v10
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s7, v11
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s8, v12
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s9, v13
+; AGPR-GISEL-NEXT: v_readfirstlane_b32 s10, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, s28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, s29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, s4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, s5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, s6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, s7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, s8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, s9
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s10
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v14, v0 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
ret <16 x float> %result
}
@@ -2873,6 +9371,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_inlineimm__
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_inlineimm__scaleB_inlineimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, -2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 33
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_inlineimm__scaleB_inlineimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 33
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, -2
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 2, i32 33, i32 2, i32 -2)
ret <16 x float> %result
}
@@ -2935,6 +9537,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scale
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scaleB_inlineimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, -2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 0x41
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scaleB_inlineimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 0x41
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, -2
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 2, i32 65, i32 2, i32 -2)
ret <16 x float> %result
}
@@ -2997,6 +9703,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scale
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scaleB_FP_literal:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, 1.0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 0x41
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scaleB_FP_literal:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 0x41
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 1.0
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 2, i32 65, i32 2, i32 1065353216)
ret <16 x float> %result
}
@@ -3059,6 +9869,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_FP_literal_
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_FP_literal__scaleB_inlineimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, -2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 1.0
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_FP_literal__scaleB_inlineimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 1.0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, -2
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 2, i32 1065353216, i32 2, i32 -2)
ret <16 x float> %result
}
@@ -3121,6 +10035,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_FP_literal_
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_FP_literal__scaleB_FP_literal:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, 1.0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 0.15915494
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_FP_literal__scaleB_FP_literal:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 0.15915494
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 1.0
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 2, i32 1042479491, i32 2, i32 1065353216)
ret <16 x float> %result
}
@@ -3183,6 +10201,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scale
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scaleB_kimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, 0x4d
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 0x41
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[1,1,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__scaleA_kimm__scaleB_kimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 0x41
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 0x4d
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[1,1,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 2, i32 65, i32 2, i32 77)
ret <16 x float> %result
}
@@ -3265,6 +10387,84 @@ define amdgpu_kernel void @test_mfma_scale_f32_32x32x64_f8f6f4__vgprcd(<8 x i32>
; GISEL-NEXT: global_store_dwordx4 v16, v[8:11], s[2:3] offset:32
; GISEL-NEXT: global_store_dwordx4 v16, v[12:15], s[2:3] offset:48
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4__vgprcd:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x40
+; AGPR-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x80
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[0:1], s[36:37]
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v23, s15
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v24, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v25, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v26, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v27, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v28, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v29, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v30, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[2:3], s[38:39]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[4:5], s[40:41]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[6:7], s[42:43]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[8:9], s[44:45]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[10:11], s[46:47]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[12:13], s[48:49]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[14:15], s[50:51]
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v33, s1
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[16:23], v[24:31], v[0:15], v32, v33 op_sel:[1,1,0] op_sel_hi:[1,0,0] blgp:2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, 0
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 2
+; AGPR-SDAG-NEXT: global_store_dwordx4 v16, v[12:15], s[2:3] offset:48
+; AGPR-SDAG-NEXT: global_store_dwordx4 v16, v[8:11], s[2:3] offset:32
+; AGPR-SDAG-NEXT: global_store_dwordx4 v16, v[4:7], s[2:3] offset:16
+; AGPR-SDAG-NEXT: global_store_dwordx4 v16, v[0:3], s[2:3]
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4__vgprcd:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x40
+; AGPR-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x80
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[22:23], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[20:21], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[30:31], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[36:37]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[28:29], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[26:27], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[24:25], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[38:39]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[40:41]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[42:43]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[44:45]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[46:47]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[48:49]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[50:51]
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v32, s0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v33, s1
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[16:23], v[24:31], v[0:15], v32, v33 op_sel:[1,1,0] op_sel_hi:[1,0,0] blgp:2
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v16, 0
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 2
+; AGPR-GISEL-NEXT: global_store_dwordx4 v16, v[0:3], s[2:3]
+; AGPR-GISEL-NEXT: global_store_dwordx4 v16, v[4:7], s[2:3] offset:16
+; AGPR-GISEL-NEXT: global_store_dwordx4 v16, v[8:11], s[2:3] offset:32
+; AGPR-GISEL-NEXT: global_store_dwordx4 v16, v[12:15], s[2:3] offset:48
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 2, i32 3, i32 %scale0, i32 1, i32 %scale1)
store <16 x float> %result, ptr addrspace(1) %ptr, align 64
ret void
@@ -3348,6 +10548,84 @@ define amdgpu_kernel void @test_mfma_scale_f32_32x32x64_f8f6f4__vgprcd___scaleA_
; GISEL-NEXT: global_store_dwordx4 v16, v[8:11], s[0:1] offset:32
; GISEL-NEXT: global_store_dwordx4 v16, v[12:15], s[0:1] offset:48
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4__vgprcd___scaleA_kimm__scaleB__inlineimm:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x40
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, -2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v33, 0x41
+; AGPR-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x80
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v23, s15
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[0:1], s[36:37]
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v24, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v25, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v26, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v27, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v28, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v29, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v30, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[2:3], s[38:39]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[4:5], s[40:41]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[6:7], s[42:43]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[8:9], s[44:45]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[10:11], s[46:47]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[12:13], s[48:49]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[14:15], s[50:51]
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[16:23], v[24:31], v[0:15], v33, v32 op_sel:[1,1,0] op_sel_hi:[1,0,0] blgp:2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, 0
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 2
+; AGPR-SDAG-NEXT: global_store_dwordx4 v16, v[12:15], s[0:1] offset:48
+; AGPR-SDAG-NEXT: global_store_dwordx4 v16, v[8:11], s[0:1] offset:32
+; AGPR-SDAG-NEXT: global_store_dwordx4 v16, v[4:7], s[0:1] offset:16
+; AGPR-SDAG-NEXT: global_store_dwordx4 v16, v[0:3], s[0:1]
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4__vgprcd___scaleA_kimm__scaleB__inlineimm:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x40
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v32, 0x41
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v33, -2
+; AGPR-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x80
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[22:23], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[20:21], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[30:31], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[36:37]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[28:29], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[26:27], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[24:25], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[38:39]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[40:41]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[42:43]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[44:45]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[46:47]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[48:49]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[50:51]
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[16:23], v[24:31], v[0:15], v32, v33 op_sel:[1,1,0] op_sel_hi:[1,0,0] blgp:2
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v16, 0
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 2
+; AGPR-GISEL-NEXT: global_store_dwordx4 v16, v[0:3], s[0:1]
+; AGPR-GISEL-NEXT: global_store_dwordx4 v16, v[4:7], s[0:1] offset:16
+; AGPR-GISEL-NEXT: global_store_dwordx4 v16, v[8:11], s[0:1] offset:32
+; AGPR-GISEL-NEXT: global_store_dwordx4 v16, v[12:15], s[0:1] offset:48
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 2, i32 3, i32 65, i32 1, i32 -2)
store <16 x float> %result, ptr addrspace(1) %ptr, align 64
ret void
@@ -3483,6 +10761,175 @@ define amdgpu_kernel void @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__nonmac(<8 x
; GISEL-NEXT: global_store_dwordx4 v[38:39], v[12:15], off sc0 sc1
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__nonmac:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[12:27], s[4:5], 0x0
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s15
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s23
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
+; AGPR-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x80
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s24
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s25
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s26
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, s9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, s10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, s11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, s12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, s13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, s14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, s15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, s16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, s17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, s18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, s20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, s21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, s22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, s23
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v0, s0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v1, s1
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[2:9], v[10:17], a[0:15], v0, v1 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[0:1], 48
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s17
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[2:3], 32
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s15
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s13
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[4:5], 16
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[4:5], v[6:9], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s9
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[6:7], 0
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[6:7], v[8:11], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[2:3], a[8:11], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[0:1], a[12:15], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[6:7], a[0:3], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[4:5], a[4:7], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__nonmac:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
+; AGPR-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x80
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], 0
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], 16
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s36
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s37
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s38
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s39
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, s40
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, s41
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, s42
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, s43
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, s44
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, s45
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, s46
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, s47
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, s48
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, s49
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, s50
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, s51
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, s20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, s21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, s22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, s23
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, s0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, s1
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[20:21], 32
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[22:23], 48
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[16:17], v[0:3], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[18:19], v[4:7], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[20:21], v[8:11], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[22:23], v[12:15], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[16:17], v[0:3], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[18:19], v[4:7], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[20:21], v[8:11], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[22:23], v[12:15], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 %scale0, i32 0, i32 %scale1)
store volatile <16 x float> %arg2, ptr addrspace(1) null, align 64
store volatile <16 x float> %result, ptr addrspace(1) null, align 64
@@ -3616,6 +11063,173 @@ define amdgpu_kernel void @test_mfma_scale_f32_32x32x64_f8f6f4_25_42__nonmac(<8
; GISEL-NEXT: global_store_dwordx4 v[38:39], v[12:15], off sc0 sc1
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_25_42__nonmac:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[12:27], s[4:5], 0x0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v0, 42
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v1, 25
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s15
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v12, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v13, s23
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v14, s24
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v15, s25
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s26
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s27
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, s8
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, s9
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, s10
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, s11
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, s12
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, s13
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, s14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, s15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, s16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, s17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, s18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, s19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, s20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, s21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, s22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, s23
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[2:9], v[10:17], a[0:15], v1, v0 op_sel_hi:[0,0,0] blgp:2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v2, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v3, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[0:1], 48
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v4, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v5, s17
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[2:3], 32
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s15
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v6, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v7, s13
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[4:5], 16
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[4:5], v[6:9], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v10, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v11, s11
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v8, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v9, s9
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[6:7], 0
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[6:7], v[8:11], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[2:3], a[8:11], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[0:1], a[12:15], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[6:7], a[0:3], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[4:5], a[4:7], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_25_42__nonmac:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 25
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 42
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], 0
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, s36
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, s37
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, s38
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, s39
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, s40
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, s41
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, s42
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, s43
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, s44
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, s8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, s45
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, s46
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, s47
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, s48
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, s49
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, s50
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, s51
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, s9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, s10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, s11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, s12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, s13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, s14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, s15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, s16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, s17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, s18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, s19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, s20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, s21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, s22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, s23
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[0,0,0] blgp:2
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], 16
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[20:21], 32
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[22:23], 48
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[16:17], v[0:3], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[18:19], v[4:7], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[20:21], v[8:11], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[22:23], v[12:15], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[16:17], v[0:3], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[18:19], v[4:7], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[20:21], v[8:11], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[22:23], v[12:15], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 2, i32 0, i32 25, i32 0, i32 42)
store volatile <16 x float> %arg2, ptr addrspace(1) null, align 64
store volatile <16 x float> %result, ptr addrspace(1) null, align 64
@@ -3740,6 +11354,124 @@ define amdgpu_kernel void @test_mfma_scale_f32_32x32x64_f8f6f4_0_0__vgprcd_nonma
; GISEL-NEXT: global_store_dwordx4 v[38:39], v[12:15], off sc0 sc1
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__vgprcd_nonmac:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[12:27], s[4:5], 0x0
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v33, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v34, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v35, s15
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v36, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v37, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v38, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v39, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v40, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v41, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v42, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v43, s23
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v44, s24
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v45, s25
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v46, s26
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v47, s27
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[30:31], s[22:23]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[28:29], s[20:21]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[26:27], s[18:19]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[24:25], s[16:17]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[22:23], s[14:15]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[20:21], s[12:13]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[18:19], s[10:11]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[16:17], s[8:9]
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[32:39], v[40:47], v[16:31] blgp:2
+; AGPR-SDAG-NEXT: s_nop 14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[20:21], 48
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[20:21], v[16:19], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[22:23], 32
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[24:25], 16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s19
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[22:23], v[16:19], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[26:27], 0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s15
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[24:25], v[16:19], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s11
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[26:27], v[16:19], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[22:23], v[8:11], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[20:21], v[12:15], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[26:27], v[0:3], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[24:25], v[4:7], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_0_0__vgprcd_nonmac:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[32:33], s[36:37]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[34:35], s[38:39]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[36:37], s[40:41]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[38:39], s[42:43]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[40:41], s[44:45]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[30:31], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[42:43], s[46:47]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[44:45], s[48:49]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[46:47], s[50:51]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[28:29], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[26:27], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[24:25], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[22:23], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[20:21], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[8:9]
+; AGPR-GISEL-NEXT: s_nop 1
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[32:39], v[40:47], v[16:31] blgp:2
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[32:33], 0
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[34:35], 16
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[36:37], 32
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[38:39], 48
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[32:33], v[16:19], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[34:35], v[20:23], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[36:37], v[24:27], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[38:39], v[28:31], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 7
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[32:33], v[0:3], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[34:35], v[4:7], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[36:37], v[8:11], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[38:39], v[12:15], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 2, i32 0, i32 0, i32 0, i32 0)
store volatile <16 x float> %arg2, ptr addrspace(1) null, align 64
store volatile <16 x float> %result, ptr addrspace(1) null, align 64
@@ -3874,6 +11606,134 @@ define amdgpu_kernel void @test_mfma_scale_f32_32x32x64_f8f6f4_25_42__vgprcd_non
; GISEL-NEXT: global_store_dwordx4 v[38:39], v[12:15], off sc0 sc1
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_endpgm
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_25_42__vgprcd_nonmac:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[12:27], s[4:5], 0x0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 42
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v33, 25
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s15
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v20, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v21, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v22, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v23, s19
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v24, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v25, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v26, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v27, s23
+; AGPR-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v28, s24
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v29, s25
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v30, s26
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, s27
+; AGPR-SDAG-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-SDAG-NEXT: s_nop 1
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[16:23], v[24:31], v[0:15], v33, v32 op_sel_hi:[0,0,0] blgp:2
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s20
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s21
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s22
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s23
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[20:21], 48
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[20:21], v[16:19], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[22:23], 32
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[24:25], 16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s16
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s17
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s18
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s19
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[22:23], v[16:19], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: v_mov_b64_e32 v[26:27], 0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s12
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s13
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s14
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s15
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[24:25], v[16:19], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v16, s8
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v17, s9
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v18, s10
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v19, s11
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[26:27], v[16:19], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[22:23], v[8:11], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[20:21], v[12:15], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[26:27], v[0:3], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: global_store_dwordx4 v[24:25], v[4:7], off sc0 sc1
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_endpgm
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4_25_42__vgprcd_nonmac:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[36:51], s[4:5], 0x0
+; AGPR-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x40
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v32, 25
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v33, 42
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[34:35], 16
+; AGPR-GISEL-NEXT: s_waitcnt lgkmcnt(0)
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[36:37]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[38:39]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[20:21], s[40:41]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[22:23], s[42:43]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[24:25], s[44:45]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[26:27], s[46:47]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[28:29], s[48:49]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[30:31], s[50:51]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[20:21]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[36:37], 32
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[38:39], 48
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[16:23], v[24:31], v[0:15], v32, v33 op_sel_hi:[0,0,0] blgp:2
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[10:11]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[32:33], 0
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[8:9]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[22:23], s[14:15]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[26:27], s[18:19]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[30:31], s[22:23]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[20:21], s[12:13]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[24:25], s[16:17]
+; AGPR-GISEL-NEXT: v_mov_b64_e32 v[28:29], s[20:21]
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[32:33], v[16:19], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[34:35], v[20:23], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[36:37], v[24:27], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[38:39], v[28:31], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 2
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[32:33], v[0:3], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[34:35], v[4:7], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[36:37], v[8:11], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: global_store_dwordx4 v[38:39], v[12:15], off sc0 sc1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_endpgm
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 2, i32 0, i32 25, i32 0, i32 42)
store volatile <16 x float> %arg2, ptr addrspace(1) null, align 64
store volatile <16 x float> %result, ptr addrspace(1) null, align 64
@@ -3906,6 +11766,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_0_a(
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_0_a:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_0_a:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)
ret <16 x float> %result
}
@@ -3936,6 +11896,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_0_b(
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_0_b:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_0_b:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 3, i32 0, i32 1, i32 0)
ret <16 x float> %result
}
@@ -3998,6 +12058,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_1(<8
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_1:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, 1
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 0
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_0_1:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 0
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 1
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 0, i32 0, i32 1)
ret <16 x float> %result
}
@@ -4060,6 +12224,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_1_0_a(
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_1_0_a:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v31, 0
+; AGPR-SDAG-NEXT: v_mov_b32_e32 v32, 1
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0]
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___constant_scale_1_0_a:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v0, 1
+; AGPR-GISEL-NEXT: v_mov_b32_e32 v1, 0
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v0, v1 op_sel_hi:[0,0,0]
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 0, i32 0, i32 0, i32 1, i32 0, i32 0)
ret <16 x float> %result
}
@@ -4124,6 +12392,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v8i32_fp6(
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v8i32_fp6:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] blgp:2
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v8i32_fp6:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] blgp:2
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 2, ; blgp
@@ -4187,6 +12559,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp8(
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp8:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] cbsz:2
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp8:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] cbsz:2
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 0, ; blgp
@@ -4248,6 +12724,108 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp6(
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp6:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] cbsz:2 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp6:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] cbsz:2 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 2, ; blgp
@@ -4280,6 +12858,104 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp6_
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp6__0_scale:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15] cbsz:2 blgp:2
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp6__v8i32_fp6__0_scale:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15] cbsz:2 blgp:2
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 2, ; cbsz
i32 2, ; blgp
@@ -4343,6 +13019,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v8i32_fp4(
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v8i32_fp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] blgp:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v8i32_fp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] blgp:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 4, ; blgp
@@ -4406,6 +13186,110 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp8(
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp8:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp8:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 0, ; blgp
@@ -4439,6 +13323,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v6i32_fp4(
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v6i32_fp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] blgp:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp8__v6i32_fp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:29], a[0:15], v30, v31 op_sel_hi:[0,0,0] blgp:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8 x i32> %arg0, <6 x i32> %arg1, <16 x float> %arg2,
i32 0, ; cbsz
i32 4, ; blgp
@@ -4472,6 +13456,106 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v6i32_fp4__v8i32_fp8(
; GCN-NEXT: v_mov_b32_e32 v14, v28
; GCN-NEXT: v_mov_b32_e32 v15, v29
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v6i32_fp4__v8i32_fp8:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:5], v[6:13], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-SDAG-NEXT: s_nop 15
+; AGPR-SDAG-NEXT: s_nop 3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v6i32_fp4__v8i32_fp8:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a15, v29
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[24:29], a[16:23], a[0:15], v30, v31 op_sel_hi:[0,0,0] cbsz:4
+; AGPR-GISEL-NEXT: s_nop 15
+; AGPR-GISEL-NEXT: s_nop 3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v6i32.v8i32(<6 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 0, ; blgp
@@ -4533,6 +13617,108 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp4(
; GISEL-NEXT: v_mov_b32_e32 v14, v30
; GISEL-NEXT: v_mov_b32_e32 v15, v31
; GISEL-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp4:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: scratch_load_dword v31, off, s32 offset:8
+; AGPR-SDAG-NEXT: scratch_load_dword v32, off, s32 offset:4
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15], v32, v31 op_sel_hi:[0,0,0] cbsz:4 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp4:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: scratch_load_dword v31, off, s32 offset:4
+; AGPR-GISEL-NEXT: scratch_load_dword v32, off, s32 offset:8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_scale_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15], v31, v32 op_sel_hi:[0,0,0] cbsz:4 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 4, ; blgp
@@ -4565,6 +13751,104 @@ define <16 x float> @test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp4_
; GCN-NEXT: v_mov_b32_e32 v14, v30
; GCN-NEXT: v_mov_b32_e32 v15, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-SDAG-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp4__0_scale:
+; AGPR-SDAG: ; %bb.0:
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-SDAG-NEXT: scratch_load_dword a15, off, s32
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-SDAG-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-SDAG-NEXT: s_waitcnt vmcnt(0)
+; AGPR-SDAG-NEXT: s_nop 0
+; AGPR-SDAG-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], v[0:7], v[8:15], a[0:15] cbsz:4 blgp:4
+; AGPR-SDAG-NEXT: s_nop 11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-SDAG-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-SDAG-NEXT: s_setpc_b64 s[30:31]
+;
+; AGPR-GISEL-LABEL: test_mfma_scale_f32_32x32x64_f8f6f4___v8i32_fp4__v8i32_fp4__0_scale:
+; AGPR-GISEL: ; %bb.0:
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; AGPR-GISEL-NEXT: scratch_load_dword a15, off, s32
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a16, v0
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a17, v1
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a18, v2
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a19, v3
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a20, v4
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a21, v5
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a22, v6
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a23, v7
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a24, v8
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a25, v9
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a26, v10
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a27, v11
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a28, v12
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a29, v13
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a30, v14
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a31, v15
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a0, v16
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a1, v17
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a2, v18
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a3, v19
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a4, v20
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a5, v21
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a6, v22
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a7, v23
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a8, v24
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a9, v25
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a10, v26
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a11, v27
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a12, v28
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a13, v29
+; AGPR-GISEL-NEXT: v_accvgpr_write_b32 a14, v30
+; AGPR-GISEL-NEXT: s_waitcnt vmcnt(0)
+; AGPR-GISEL-NEXT: s_nop 0
+; AGPR-GISEL-NEXT: v_mfma_f32_32x32x64_f8f6f4 a[0:15], a[16:23], a[24:31], a[0:15] cbsz:4 blgp:4
+; AGPR-GISEL-NEXT: s_nop 11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v0, a0
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v1, a1
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v2, a2
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v3, a3
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v4, a4
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v5, a5
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v6, a6
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v7, a7
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v8, a8
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v9, a9
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v10, a10
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v11, a11
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v12, a12
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v13, a13
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v14, a14
+; AGPR-GISEL-NEXT: v_accvgpr_read_b32 v15, a15
+; AGPR-GISEL-NEXT: s_setpc_b64 s[30:31]
%result = call <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v8i32(<8 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2,
i32 4, ; cbsz
i32 4, ; blgp
@@ -4585,3 +13869,5 @@ declare <16 x float> @llvm.amdgcn.mfma.scale.f32.32x32x64.f8f6f4.v8i32.v6i32(<8
attributes #0 = { "amdgpu-flat-work-group-size"="512,512" "amdgpu-agpr-alloc"="0,0" nounwind }
attributes #1 = { "amdgpu-flat-work-group-size"="128,128" nounwind }
attributes #2 = { convergent nocallback nofree nosync nounwind willreturn memory(none) }
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; AGPR: {{.*}}
More information about the llvm-commits
mailing list