[llvm-branch-commits] [llvm] [AMDGPU] Patterns for <2 x bfloat> fneg (fabs) (PR #142911)
Stanislav Mekhanoshin via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Jun 4 23:50:44 PDT 2025
https://github.com/rampitec created https://github.com/llvm/llvm-project/pull/142911
None
>From 44a9017e98eff94456889a528a166d6aabca842d Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Wed, 4 Jun 2025 23:49:43 -0700
Subject: [PATCH] [AMDGPU] Patterns for <2 x bfloat> fneg (fabs)
---
llvm/lib/Target/AMDGPU/SIInstructions.td | 11 +++----
llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll | 38 +++++++++-------------
2 files changed, 21 insertions(+), 28 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index a0285e3512a08..360fd05cb3d96 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -1840,22 +1840,21 @@ def : GCNPat <
(UniformUnaryFrag<fabs> (v2fp16vt SReg_32:$src)),
(S_AND_B32 SReg_32:$src, (S_MOV_B32 (i32 0x7fff7fff)))
>;
-}
// This is really (fneg (fabs v2f16:$src))
//
// fabs is not reported as free because there is modifier for it in
// VOP3P instructions, so it is turned into the bit op.
def : GCNPat <
- (UniformUnaryFrag<fneg> (v2f16 (bitconvert (and_oneuse (i32 SReg_32:$src), 0x7fff7fff)))),
+ (UniformUnaryFrag<fneg> (v2fp16vt (bitconvert (and_oneuse (i32 SReg_32:$src), 0x7fff7fff)))),
(S_OR_B32 SReg_32:$src, (S_MOV_B32 (i32 0x80008000))) // Set sign bit
>;
def : GCNPat <
- (UniformUnaryFrag<fneg> (v2f16 (fabs SReg_32:$src))),
+ (UniformUnaryFrag<fneg> (v2fp16vt (fabs SReg_32:$src))),
(S_OR_B32 SReg_32:$src, (S_MOV_B32 (i32 0x80008000))) // Set sign bit
>;
-
+}
// COPY_TO_REGCLASS is needed to avoid using SCC from S_XOR_B32 instead
// of the real value.
@@ -1986,12 +1985,12 @@ def : GCNPat <
(fabs (v2fp16vt VGPR_32:$src)),
(V_AND_B32_e64 (S_MOV_B32 (i32 0x7fff7fff)), VGPR_32:$src)
>;
-}
def : GCNPat <
- (fneg (v2f16 (fabs VGPR_32:$src))),
+ (fneg (v2fp16vt (fabs VGPR_32:$src))),
(V_OR_B32_e64 (S_MOV_B32 (i32 0x80008000)), VGPR_32:$src)
>;
+}
def : GCNPat <
(fabs (f64 VReg_64:$src)),
diff --git a/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll b/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll
index 243469d39cc11..d189b6d4c1e83 100644
--- a/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll
@@ -523,8 +523,7 @@ define amdgpu_kernel void @s_fneg_fabs_v2bf16_non_bc_src(ptr addrspace(1) %out,
; VI-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT: v_alignbit_b32 v0, v1, v0, 16
-; VI-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0
-; VI-NEXT: v_xor_b32_e32 v2, 0x80008000, v0
+; VI-NEXT: v_or_b32_e32 v2, 0x80008000, v0
; VI-NEXT: v_mov_b32_e32 v0, s0
; VI-NEXT: v_mov_b32_e32 v1, s1
; VI-NEXT: flat_store_dword v[0:1], v2
@@ -556,8 +555,7 @@ define amdgpu_kernel void @s_fneg_fabs_v2bf16_non_bc_src(ptr addrspace(1) %out,
; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_and_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT: v_lshl_or_b32 v1, v1, 16, v2
-; GFX9-NEXT: v_and_b32_e32 v1, 0x7fff7fff, v1
-; GFX9-NEXT: v_xor_b32_e32 v1, 0x80008000, v1
+; GFX9-NEXT: v_or_b32_e32 v1, 0x80008000, v1
; GFX9-NEXT: global_store_dword v0, v1, s[0:1]
; GFX9-NEXT: s_endpgm
;
@@ -590,9 +588,9 @@ define amdgpu_kernel void @s_fneg_fabs_v2bf16_non_bc_src(ptr addrspace(1) %out,
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_and_b32 v0, 0x7fff7fff, v0
-; GFX11-NEXT: v_xor_b32_e32 v0, 0x80008000, v0
+; GFX11-NEXT: v_mov_b32_e32 v1, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_or_b32_e32 v0, 0x80008000, v0
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: global_store_b32 v1, v0, s[0:1]
; GFX11-NEXT: s_endpgm
@@ -634,8 +632,7 @@ define amdgpu_kernel void @s_fneg_fabs_v2bf16_bc_src(ptr addrspace(1) %out, <2 x
; VI-NEXT: s_mov_b32 flat_scratch_lo, s13
; VI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8
; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: s_and_b32 s2, s2, 0x7fff7fff
-; VI-NEXT: s_xor_b32 s2, s2, 0x80008000
+; VI-NEXT: s_or_b32 s2, s2, 0x80008000
; VI-NEXT: v_mov_b32_e32 v0, s0
; VI-NEXT: v_mov_b32_e32 v1, s1
; VI-NEXT: v_mov_b32_e32 v2, s2
@@ -648,8 +645,7 @@ define amdgpu_kernel void @s_fneg_fabs_v2bf16_bc_src(ptr addrspace(1) %out, <2 x
; GFX9-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x0
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_and_b32 s2, s2, 0x7fff7fff
-; GFX9-NEXT: s_xor_b32 s2, s2, 0x80008000
+; GFX9-NEXT: s_or_b32 s2, s2, 0x80008000
; GFX9-NEXT: v_mov_b32_e32 v1, s2
; GFX9-NEXT: global_store_dword v0, v1, s[0:1]
; GFX9-NEXT: s_endpgm
@@ -660,9 +656,8 @@ define amdgpu_kernel void @s_fneg_fabs_v2bf16_bc_src(ptr addrspace(1) %out, <2 x
; GFX11-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX11-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_and_b32 s2, s2, 0x7fff7fff
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT: s_xor_b32 s2, s2, 0x80008000
+; GFX11-NEXT: s_or_b32 s2, s2, 0x80008000
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
; GFX11-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX11-NEXT: s_endpgm
@@ -977,7 +972,7 @@ define amdgpu_kernel void @s_fneg_multi_use_fabs_v2bf16(ptr addrspace(1) %out0,
; VI-NEXT: v_mov_b32_e32 v0, s0
; VI-NEXT: s_and_b32 s0, s4, 0x7fff7fff
; VI-NEXT: v_mov_b32_e32 v1, s1
-; VI-NEXT: s_xor_b32 s1, s0, 0x80008000
+; VI-NEXT: s_or_b32 s1, s4, 0x80008000
; VI-NEXT: v_mov_b32_e32 v4, s0
; VI-NEXT: v_mov_b32_e32 v2, s2
; VI-NEXT: v_mov_b32_e32 v3, s3
@@ -992,12 +987,12 @@ define amdgpu_kernel void @s_fneg_multi_use_fabs_v2bf16(ptr addrspace(1) %out0,
; GFX9-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_and_b32 s4, s4, 0x7fff7fff
-; GFX9-NEXT: s_xor_b32 s5, s4, 0x80008000
-; GFX9-NEXT: v_mov_b32_e32 v1, s4
-; GFX9-NEXT: global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT: s_and_b32 s5, s4, 0x7fff7fff
+; GFX9-NEXT: s_or_b32 s4, s4, 0x80008000
; GFX9-NEXT: v_mov_b32_e32 v1, s5
-; GFX9-NEXT: global_store_dword v0, v1, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT: global_store_dword v0, v2, s[2:3]
; GFX9-NEXT: s_endpgm
;
; GFX11-LABEL: s_fneg_multi_use_fabs_v2bf16:
@@ -1007,9 +1002,8 @@ define amdgpu_kernel void @s_fneg_multi_use_fabs_v2bf16(ptr addrspace(1) %out0,
; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x0
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_and_b32 s4, s6, 0x7fff7fff
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_or_b32 s5, s6, 0x80008000
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s4
-; GFX11-NEXT: s_xor_b32 s5, s4, 0x80008000
; GFX11-NEXT: v_mov_b32_e32 v2, s5
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_store_b32 v0, v1, s[0:1]
More information about the llvm-branch-commits
mailing list