[llvm] [AMDGPU] In VectorLegalizer::Expand, if UnrollVectorOp returns Load, … (PR #88475)
via llvm-commits
llvm-commits at lists.llvm.org
Thu Apr 11 23:19:47 PDT 2024
https://github.com/choikwa created https://github.com/llvm/llvm-project/pull/88475
…return only Load since other output is chain.
Added testcase that showed mismatched expected arity when Load and chain were returned as separate items after 003b58f65bdd5d9c7d0c1b355566c9ef430c0e7d
>From 70167e3d11d8f8796fd7a18ef219cd6e61719f5c Mon Sep 17 00:00:00 2001
From: Kevin Choi <kevin.choi at amd.com>
Date: Fri, 12 Apr 2024 00:57:26 -0500
Subject: [PATCH] [AMDGPU] In VectorLegalizer::Expand, if UnrollVectorOp
returns Load, return only Load since other output is chain.
Added testcase that showed mismatched expected arity when Load and chain were returned as separate items after 003b58f65bdd5d9c7d0c1b355566c9ef430c0e7d
---
.../SelectionDAG/LegalizeVectorOps.cpp | 8 +-
llvm/test/CodeGen/AMDGPU/build_vector.ll | 98 +++++++++++++++++++
2 files changed, 104 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
index 57a3f6a65e002c..9b268dac484df1 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
@@ -1159,8 +1159,12 @@ void VectorLegalizer::Expand(SDNode *Node, SmallVectorImpl<SDValue> &Results) {
}
SDValue Unrolled = DAG.UnrollVectorOp(Node);
- for (unsigned I = 0, E = Unrolled->getNumValues(); I != E; ++I)
- Results.push_back(Unrolled.getValue(I));
+ const LoadSDNode *Ld = dyn_cast<LoadSDNode>(Unrolled.getNode());
+ if (Ld)
+ Results.push_back(Unrolled);
+ else
+ for (unsigned I = 0, E = Unrolled->getNumValues(); I != E; ++I)
+ Results.push_back(Unrolled.getValue(I));
}
SDValue VectorLegalizer::ExpandSELECT(SDNode *Node) {
diff --git a/llvm/test/CodeGen/AMDGPU/build_vector.ll b/llvm/test/CodeGen/AMDGPU/build_vector.ll
index 37412ac3aa5418..5b52e3aa70ab60 100644
--- a/llvm/test/CodeGen/AMDGPU/build_vector.ll
+++ b/llvm/test/CodeGen/AMDGPU/build_vector.ll
@@ -3,6 +3,7 @@
; RUN: llc < %s -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs | FileCheck %s --check-prefixes=GFX8,GFX678,ALL
; RUN: llc < %s -mtriple=amdgcn-amd-amdpal -mcpu=gfx1030 -verify-machineinstrs | FileCheck %s --check-prefixes=GFX10,GFX1011,ALL
; RUN: llc < %s -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs | FileCheck %s --check-prefixes=GFX11,GFX1011,ALL
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx940 | FileCheck %s --check-prefixes=GFX940,ALL
; ALL-LABEL: {{^}}build_vector2:
; R600: MOV
@@ -96,3 +97,100 @@ define amdgpu_kernel void @build_vector_v2i16_trunc (ptr addrspace(1) %out, i32
store <2 x i16> %ins.1, ptr addrspace(1) %out
ret void
}
+
+; R600-LABEL: build_v2i32_from_v4i16_shuffle:
+; R600: ; %bb.0: ; %entry
+; R600-NEXT: ALU 0, @10, KC0[], KC1[]
+; R600-NEXT: TEX 1 @6
+; R600-NEXT: ALU 4, @11, KC0[CB0:0-32], KC1[]
+; R600-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
+; R600-NEXT: CF_END
+; R600-NEXT: PAD
+; R600-NEXT: Fetch clause starting at 6:
+; R600-NEXT: VTX_READ_16 T1.X, T0.X, 48, #3
+; R600-NEXT: VTX_READ_16 T0.X, T0.X, 44, #3
+; R600-NEXT: ALU clause starting at 10:
+; R600-NEXT: MOV * T0.X, 0.0,
+; R600-NEXT: ALU clause starting at 11:
+; R600-NEXT: LSHL * T0.Y, T1.X, literal.x,
+; R600-NEXT: 16(2.242078e-44), 0(0.000000e+00)
+; R600-NEXT: LSHL T0.X, T0.X, literal.x,
+; R600-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
+; R600-NEXT: 16(2.242078e-44), 2(2.802597e-45)
+;
+; GFX6-LABEL: build_v2i32_from_v4i16_shuffle:
+; GFX6: ; %bb.0: ; %entry
+; GFX6-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: s_lshl_b32 s3, s3, 16
+; GFX6-NEXT: s_lshl_b32 s2, s2, 16
+; GFX6-NEXT: s_mov_b32 s6, -1
+; GFX6-NEXT: s_mov_b32 s4, s0
+; GFX6-NEXT: s_mov_b32 s5, s1
+; GFX6-NEXT: v_mov_b32_e32 v0, s2
+; GFX6-NEXT: v_mov_b32_e32 v1, s3
+; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GFX6-NEXT: s_endpgm
+;
+; GFX8-LABEL: build_v2i32_from_v4i16_shuffle:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX8-NEXT: s_mov_b32 s7, 0xf000
+; GFX8-NEXT: s_mov_b32 s6, -1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 s4, s0
+; GFX8-NEXT: s_mov_b32 s5, s1
+; GFX8-NEXT: s_lshl_b32 s0, s3, 16
+; GFX8-NEXT: s_lshl_b32 s1, s2, 16
+; GFX8-NEXT: v_mov_b32_e32 v0, s1
+; GFX8-NEXT: v_mov_b32_e32 v1, s0
+; GFX8-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GFX8-NEXT: s_endpgm
+;
+; GFX10-LABEL: build_v2i32_from_v4i16_shuffle:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x0
+; GFX10-NEXT: v_mov_b32_e32 v2, 0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_lshl_b32 s2, s2, 16
+; GFX10-NEXT: s_lshl_b32 s3, s3, 16
+; GFX10-NEXT: v_mov_b32_e32 v0, s2
+; GFX10-NEXT: v_mov_b32_e32 v1, s3
+; GFX10-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1]
+; GFX10-NEXT: s_endpgm
+;
+; GFX11-LABEL: build_v2i32_from_v4i16_shuffle:
+; GFX11: ; %bb.0: ; %entry
+; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x0
+; GFX11-NEXT: v_mov_b32_e32 v2, 0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_lshl_b32 s2, s2, 16
+; GFX11-NEXT: s_lshl_b32 s3, s3, 16
+; GFX11-NEXT: v_mov_b32_e32 v0, s2
+; GFX11-NEXT: v_mov_b32_e32 v1, s3
+; GFX11-NEXT: global_store_b64 v2, v[0:1], s[0:1]
+; GFX11-NEXT: s_nop 0
+; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-NEXT: s_endpgm
+;
+; GFX940-LABEL: build_v2i32_from_v4i16_shuffle:
+; GFX940: ; %bb.0: ; %entry
+; GFX940-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX940-NEXT: v_mov_b32_e32 v2, 0
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_lshl_b32 s3, s3, 16
+; GFX940-NEXT: s_lshl_b32 s2, s2, 16
+; GFX940-NEXT: v_mov_b32_e32 v0, s2
+; GFX940-NEXT: v_mov_b32_e32 v1, s3
+; GFX940-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1] sc0 sc1
+; GFX940-NEXT: s_endpgm
+define amdgpu_kernel void @build_v2i32_from_v4i16_shuffle(ptr addrspace(1) %out, <4 x i16> %in) {
+entry:
+ %shuf = shufflevector <4 x i16> %in, <4 x i16> zeroinitializer, <2 x i32> <i32 0, i32 2>
+ %zextended = zext <2 x i16> %shuf to <2 x i32>
+ %shifted = shl <2 x i32> %zextended, <i32 16, i32 16>
+ store <2 x i32> %shifted, ptr addrspace(1) %out
+ ret void
+}
+
More information about the llvm-commits
mailing list