[llvm] [AMDGPU] Elide bitcast fold i64 imm to build_vector (PR #154115)
Janek van Oirschot via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 22 07:11:38 PDT 2025
https://github.com/JanekvO updated https://github.com/llvm/llvm-project/pull/154115
>From 60147e8677ac43a84144e4264d7378b110788193 Mon Sep 17 00:00:00 2001
From: Janek van Oirschot <janek.vanoirschot at amd.com>
Date: Fri, 22 Aug 2025 15:06:54 +0100
Subject: [PATCH] [AMDGPU] Elide bitcast combine to build_vector in case i64
constant can be materialized
---
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 11 ++++
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h | 3 +
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 28 ++++++++-
.../AMDGPU/av-split-dead-valno-crash.ll | 59 ++++++++++---------
llvm/test/CodeGen/AMDGPU/flat-scratch.ll | 6 +-
llvm/test/CodeGen/AMDGPU/imm.ll | 3 +-
6 files changed, 74 insertions(+), 36 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index c048371b11d77..d80479784a7ea 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -5296,6 +5296,13 @@ SDValue AMDGPUTargetLowering::performRcpCombine(SDNode *N,
return DCI.DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
}
+bool AMDGPUTargetLowering::canMov64bImm(uint64_t Val, SelectionDAG &DAG) const {
+ if (!Subtarget->isGCN())
+ return false;
+ auto &ST = DAG.getSubtarget<GCNSubtarget>();
+ return ST.hasMovB64() && (ST.has64BitLiterals() || isUInt<32>(Val));
+}
+
SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
@@ -5346,6 +5353,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) {
SDLoc SL(N);
uint64_t CVal = C->getZExtValue();
+ if (canMov64bImm(CVal, DAG))
+ break;
SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
@@ -5356,6 +5365,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
const APInt &Val = C->getValueAPF().bitcastToAPInt();
SDLoc SL(N);
uint64_t CVal = Val.getZExtValue();
+ if (canMov64bImm(CVal, DAG))
+ break;
SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 78394ac9cd2dd..abf1d9a841a4d 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -103,6 +103,9 @@ class AMDGPUTargetLowering : public TargetLowering {
SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
protected:
+ /// Check whether value Val can be supported by v_mov_b64, for the current
+ /// target.
+ bool canMov64bImm(uint64_t Val, SelectionDAG &DAG) const;
bool shouldCombineMemoryType(EVT VT) const;
SDValue performLoadCombine(SDNode *N, DAGCombinerInfo &DCI) const;
SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const;
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 66c1dfc71c2f5..5fafb7ca2fa0b 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -14584,13 +14584,39 @@ SITargetLowering::performExtractVectorEltCombine(SDNode *N,
return V;
}
+ // EXTRACT_VECTOR_ELT (v2i32 bitcast (i64/f64:k), Idx)
+ // =>
+ // i32:Lo(k) if Idx == 0, or
+ // i32:Hi(k) if Idx == 1
+ auto *Idx = dyn_cast<ConstantSDNode>(N->getOperand(1));
+ if (Vec.getOpcode() == ISD::BITCAST && VecVT == MVT::v2i32 && Idx) {
+ SDLoc SL(N);
+ SDValue PeekThrough = peekThroughBitcasts(Vec);
+ auto *KImm = dyn_cast<ConstantSDNode>(PeekThrough);
+ if (KImm && KImm->getValueType(0).getSizeInBits() == 64) {
+ uint64_t KImmValue = KImm->getZExtValue();
+ if (Idx->getZExtValue() == 0)
+ return DAG.getConstant(Lo_32(KImmValue), SL, MVT::i32);
+ else
+ return DAG.getConstant(Hi_32(KImmValue), SL, MVT::i32);
+ }
+ auto *KFPImm = dyn_cast<ConstantFPSDNode>(PeekThrough);
+ if (KFPImm && KFPImm->getValueType(0).getSizeInBits() == 64) {
+ uint64_t KFPImmValue =
+ KFPImm->getValueAPF().bitcastToAPInt().getZExtValue();
+ if (Idx->getZExtValue() == 0)
+ return DAG.getConstant(Lo_32(KFPImmValue), SL, MVT::i32);
+ else
+ return DAG.getConstant(Hi_32(KFPImmValue), SL, MVT::i32);
+ }
+ }
+
if (!DCI.isBeforeLegalize())
return SDValue();
// Try to turn sub-dword accesses of vectors into accesses of the same 32-bit
// elements. This exposes more load reduction opportunities by replacing
// multiple small extract_vector_elements with a single 32-bit extract.
- auto *Idx = dyn_cast<ConstantSDNode>(N->getOperand(1));
if (isa<MemSDNode>(Vec) && VecEltSize <= 16 && VecEltVT.isByteSized() &&
VecSize > 32 && VecSize % 32 == 0 && Idx) {
EVT NewVT = getEquivalentMemType(*DAG.getContext(), VecVT);
diff --git a/llvm/test/CodeGen/AMDGPU/av-split-dead-valno-crash.ll b/llvm/test/CodeGen/AMDGPU/av-split-dead-valno-crash.ll
index 614b1e38a530f..d301f16512a60 100644
--- a/llvm/test/CodeGen/AMDGPU/av-split-dead-valno-crash.ll
+++ b/llvm/test/CodeGen/AMDGPU/av-split-dead-valno-crash.ll
@@ -7,8 +7,8 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
; CHECK-NEXT: s_load_dword s0, s[4:5], 0x8
; CHECK-NEXT: s_load_dwordx2 s[10:11], s[4:5], 0x0
; CHECK-NEXT: s_load_dwordx4 s[12:15], s[4:5], 0x10
-; CHECK-NEXT: v_mov_b32_e32 v1, 0x3e21eeb6
-; CHECK-NEXT: v_mov_b32_e32 v2, 0xa17f65f6
+; CHECK-NEXT: v_mov_b32_e32 v30, 0x9037ab78
+; CHECK-NEXT: v_mov_b32_e32 v31, 0x3e21eeb6
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
; CHECK-NEXT: s_bitcmp1_b32 s0, 0
; CHECK-NEXT: s_cselect_b64 s[16:17], -1, 0
@@ -16,12 +16,10 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
; CHECK-NEXT: s_bitcmp1_b32 s0, 8
; CHECK-NEXT: s_cselect_b64 s[2:3], -1, 0
; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[2:3]
-; CHECK-NEXT: v_cmp_ne_u32_e64 s[0:1], 1, v0
-; CHECK-NEXT: v_mov_b32_e32 v0, 0x9037ab78
-; CHECK-NEXT: v_accvgpr_write_b32 a3, v1
; CHECK-NEXT: s_xor_b64 s[20:21], s[2:3], -1
+; CHECK-NEXT: v_cmp_ne_u32_e64 s[0:1], 1, v0
; CHECK-NEXT: s_and_b64 s[2:3], exec, s[2:3]
-; CHECK-NEXT: v_accvgpr_write_b32 a2, v0
+; CHECK-NEXT: v_mov_b32_e32 v2, 0xa17f65f6
; CHECK-NEXT: v_mov_b32_e32 v3, 0xbe927e4f
; CHECK-NEXT: v_mov_b32_e32 v4, 0x19f4ec90
; CHECK-NEXT: v_mov_b32_e32 v5, 0x3efa01a0
@@ -36,14 +34,15 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
; CHECK-NEXT: v_mov_b32_e32 v14, 0x8427b883
; CHECK-NEXT: v_mov_b32_e32 v15, 0x3fae1bb4
; CHECK-NEXT: s_mov_b64 s[22:23], 0
-; CHECK-NEXT: v_mov_b32_e32 v0, 0x57b87036
-; CHECK-NEXT: v_mov_b32_e32 v1, 0x3fb3b136
+; CHECK-NEXT: v_mov_b32_e32 v20, 0x57b87036
+; CHECK-NEXT: v_mov_b32_e32 v21, 0x3fb3b136
; CHECK-NEXT: s_and_b64 s[4:5], exec, s[16:17]
; CHECK-NEXT: v_mov_b32_e32 v18, 0x55555523
; CHECK-NEXT: v_mov_b32_e32 v19, 0xbfd55555
; CHECK-NEXT: s_and_b64 s[6:7], exec, s[18:19]
-; CHECK-NEXT: v_mov_b32_e32 v20, 0
-; CHECK-NEXT: ; implicit-def: $vgpr30_vgpr31
+; CHECK-NEXT: v_mov_b32_e32 v0, 0
+; CHECK-NEXT: v_mov_b64_e32 v[16:17], 0
+; CHECK-NEXT: ; implicit-def: $agpr0_agpr1
; CHECK-NEXT: ; implicit-def: $vgpr22_vgpr23
; CHECK-NEXT: s_branch .LBB0_2
; CHECK-NEXT: .LBB0_1: ; %Flow9
@@ -63,12 +62,11 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
; CHECK-NEXT: v_mov_b64_e32 v[24:25], s[14:15]
; CHECK-NEXT: flat_load_dwordx2 v[24:25], v[24:25]
-; CHECK-NEXT: v_accvgpr_read_b32 v27, a3
-; CHECK-NEXT: v_accvgpr_read_b32 v26, a2
+; CHECK-NEXT: v_mov_b64_e32 v[26:27], v[30:31]
; CHECK-NEXT: v_mov_b64_e32 v[28:29], v[2:3]
-; CHECK-NEXT: v_mov_b64_e32 v[16:17], v[0:1]
-; CHECK-NEXT: v_accvgpr_write_b32 a0, 0
-; CHECK-NEXT: v_accvgpr_write_b32 a1, 0
+; CHECK-NEXT: v_mov_b64_e32 v[16:17], v[20:21]
+; CHECK-NEXT: v_accvgpr_write_b32 a2, 0
+; CHECK-NEXT: v_accvgpr_write_b32 a3, 0
; CHECK-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; CHECK-NEXT: v_fmac_f64_e32 v[26:27], 0, v[24:25]
; CHECK-NEXT: v_fmac_f64_e32 v[28:29], 0, v[26:27]
@@ -95,30 +93,32 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
; CHECK-NEXT: .LBB0_6: ; %.preheader1855.i.i.i3329
; CHECK-NEXT: ; Parent Loop BB0_2 Depth=1
; CHECK-NEXT: ; => This Inner Loop Header: Depth=2
-; CHECK-NEXT: v_accvgpr_read_b32 v29, a1
-; CHECK-NEXT: v_accvgpr_read_b32 v28, a0
+; CHECK-NEXT: v_accvgpr_read_b32 v29, a3
+; CHECK-NEXT: v_accvgpr_read_b32 v28, a2
; CHECK-NEXT: s_mov_b64 s[24:25], -1
; CHECK-NEXT: s_mov_b64 s[8:9], -1
; CHECK-NEXT: s_mov_b64 vcc, s[2:3]
-; CHECK-NEXT: ; implicit-def: $agpr0_agpr1
+; CHECK-NEXT: ; implicit-def: $agpr2_agpr3
; CHECK-NEXT: s_cbranch_vccz .LBB0_5
; CHECK-NEXT: ; %bb.7: ; %.lr.ph2070.i.i.i3291
; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=2
-; CHECK-NEXT: v_accvgpr_write_b32 a0, v30
-; CHECK-NEXT: v_accvgpr_write_b32 a1, v31
+; CHECK-NEXT: v_accvgpr_mov_b32 a3, a1
+; CHECK-NEXT: v_accvgpr_mov_b32 a2, a0
; CHECK-NEXT: s_mov_b64 s[8:9], s[18:19]
; CHECK-NEXT: s_mov_b64 vcc, s[6:7]
; CHECK-NEXT: s_cbranch_vccz .LBB0_5
; CHECK-NEXT: ; %bb.8: ; %.preheader1856.preheader.i.i.i3325
; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=2
-; CHECK-NEXT: v_accvgpr_write_b32 a0, v26
+; CHECK-NEXT: v_accvgpr_write_b32 a2, v26
; CHECK-NEXT: s_mov_b64 s[24:25], 0
-; CHECK-NEXT: v_accvgpr_write_b32 a1, v27
+; CHECK-NEXT: v_accvgpr_write_b32 a3, v27
; CHECK-NEXT: s_mov_b64 s[8:9], 0
; CHECK-NEXT: s_branch .LBB0_5
; CHECK-NEXT: .LBB0_9: ; in Loop: Header=BB0_2 Depth=1
+; CHECK-NEXT: v_mov_b64_e32 v[24:25], s[10:11]
+; CHECK-NEXT: v_accvgpr_write_b32 a0, v24
; CHECK-NEXT: s_mov_b64 s[22:23], 0
-; CHECK-NEXT: v_mov_b64_e32 v[30:31], s[10:11]
+; CHECK-NEXT: v_accvgpr_write_b32 a1, v25
; CHECK-NEXT: s_mov_b64 s[8:9], s[20:21]
; CHECK-NEXT: s_branch .LBB0_15
; CHECK-NEXT: .LBB0_10: ; in Loop: Header=BB0_2 Depth=1
@@ -135,19 +135,21 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
; CHECK-NEXT: v_cndmask_b32_e64 v23, v23, 0, s[16:17]
; CHECK-NEXT: v_cndmask_b32_e64 v22, v22, 0, s[16:17]
; CHECK-NEXT: v_cndmask_b32_e64 v16, 0, 1, s[8:9]
-; CHECK-NEXT: v_mov_b32_e32 v17, v16
; CHECK-NEXT: s_and_b64 s[8:9], exec, s[16:17]
-; CHECK-NEXT: global_store_dwordx2 v20, v[16:17], s[12:13]
+; CHECK-NEXT: v_mov_b32_e32 v17, v16
; CHECK-NEXT: s_cselect_b32 s23, s23, 0
; CHECK-NEXT: s_cselect_b32 s22, s22, 0
; CHECK-NEXT: s_mov_b64 s[8:9], -1
+; CHECK-NEXT: global_store_dwordx2 v0, v[16:17], s[12:13]
; CHECK-NEXT: s_branch .LBB0_14
; CHECK-NEXT: .LBB0_13: ; in Loop: Header=BB0_2 Depth=1
; CHECK-NEXT: s_mov_b64 s[8:9], 0
; CHECK-NEXT: v_mov_b64_e32 v[22:23], 0
-; CHECK-NEXT: .LBB0_14: ; %Flow6
+; CHECK-NEXT: .LBB0_14: ; %Flow8
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
-; CHECK-NEXT: v_mov_b64_e32 v[30:31], v[24:25]
+; CHECK-NEXT: v_accvgpr_write_b32 a0, v24
+; CHECK-NEXT: v_mov_b64_e32 v[16:17], 0
+; CHECK-NEXT: v_accvgpr_write_b32 a1, v25
; CHECK-NEXT: .LBB0_15: ; %Flow6
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
; CHECK-NEXT: s_mov_b64 s[24:25], -1
@@ -155,9 +157,8 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
; CHECK-NEXT: s_cbranch_vccz .LBB0_1
; CHECK-NEXT: ; %bb.16: ; %._crit_edge2105.i.i.i2330
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
-; CHECK-NEXT: v_mov_b32_e32 v21, v20
; CHECK-NEXT: s_mov_b64 s[24:25], 0
-; CHECK-NEXT: global_store_dwordx2 v20, v[20:21], s[12:13]
+; CHECK-NEXT: global_store_dwordx2 v0, v[16:17], s[12:13]
; CHECK-NEXT: s_branch .LBB0_1
; CHECK-NEXT: .LBB0_17: ; %DummyReturnBlock
; CHECK-NEXT: s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/flat-scratch.ll b/llvm/test/CodeGen/AMDGPU/flat-scratch.ll
index fc8883924dfbc..4eaa1965c66f1 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-scratch.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-scratch.ll
@@ -4152,8 +4152,7 @@ define void @store_load_i64_aligned(ptr addrspace(5) nocapture %arg) {
; GFX942-LABEL: store_load_i64_aligned:
; GFX942: ; %bb.0: ; %bb
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942-NEXT: v_mov_b32_e32 v2, 15
-; GFX942-NEXT: v_mov_b32_e32 v3, 0
+; GFX942-NEXT: v_mov_b64_e32 v[2:3], 15
; GFX942-NEXT: scratch_store_dwordx2 v0, v[2:3], off sc0 sc1
; GFX942-NEXT: s_waitcnt vmcnt(0)
; GFX942-NEXT: scratch_load_dwordx2 v[0:1], v0, off sc0 sc1
@@ -4263,8 +4262,7 @@ define void @store_load_i64_unaligned(ptr addrspace(5) nocapture %arg) {
; GFX942-LABEL: store_load_i64_unaligned:
; GFX942: ; %bb.0: ; %bb
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942-NEXT: v_mov_b32_e32 v2, 15
-; GFX942-NEXT: v_mov_b32_e32 v3, 0
+; GFX942-NEXT: v_mov_b64_e32 v[2:3], 15
; GFX942-NEXT: scratch_store_dwordx2 v0, v[2:3], off sc0 sc1
; GFX942-NEXT: s_waitcnt vmcnt(0)
; GFX942-NEXT: scratch_load_dwordx2 v[0:1], v0, off sc0 sc1
diff --git a/llvm/test/CodeGen/AMDGPU/imm.ll b/llvm/test/CodeGen/AMDGPU/imm.ll
index 21390003ee565..3735796b65a19 100644
--- a/llvm/test/CodeGen/AMDGPU/imm.ll
+++ b/llvm/test/CodeGen/AMDGPU/imm.ll
@@ -2163,10 +2163,9 @@ define amdgpu_kernel void @store_inline_imm_0.0_f64(ptr addrspace(1) %out) {
; GFX942-LABEL: store_inline_imm_0.0_f64:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
-; GFX942-NEXT: v_mov_b32_e32 v0, 0
; GFX942-NEXT: s_mov_b32 s3, 0xf000
; GFX942-NEXT: s_mov_b32 s2, -1
-; GFX942-NEXT: v_mov_b32_e32 v1, v0
+; GFX942-NEXT: v_mov_b64_e32 v[0:1], 0
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GFX942-NEXT: s_endpgm
More information about the llvm-commits
mailing list