Skip to content

Commit 341cdbc

Browse files
JanekvOJanekvO
authored
[AMDGPU] Elide bitcast fold i64 imm to build_vector (llvm#154115)
Elide bitcast combine to build_vector in case of i64 immediate that can be materialized through 64b mov
1 parent 3ef066f commit 341cdbc

File tree

6 files changed

+99
-59
lines changed

6 files changed

+99
-59
lines changed

‎llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp‎

Lines changed: 28 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -5287,6 +5287,30 @@ SDValue AMDGPUTargetLowering::performRcpCombine(SDNode *N,
52875287
return DCI.DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
52885288
}
52895289

5290+
bool AMDGPUTargetLowering::isInt64ImmLegal(SDNode *N, SelectionDAG &DAG) const {
5291+
if (!Subtarget->isGCN())
5292+
return false;
5293+
5294+
ConstantSDNode *SDConstant = dyn_cast<ConstantSDNode>(N);
5295+
ConstantFPSDNode *SDFPConstant = dyn_cast<ConstantFPSDNode>(N);
5296+
auto &ST = DAG.getSubtarget<GCNSubtarget>();
5297+
const auto *TII = ST.getInstrInfo();
5298+
5299+
if (!ST.hasMovB64() || (!SDConstant && !SDFPConstant))
5300+
return false;
5301+
5302+
if (ST.has64BitLiterals())
5303+
return true;
5304+
5305+
if (SDConstant) {
5306+
const APInt &APVal = SDConstant->getAPIntValue();
5307+
return isUInt<32>(APVal.getZExtValue()) || TII->isInlineConstant(APVal);
5308+
}
5309+
5310+
APInt Val = SDFPConstant->getValueAPF().bitcastToAPInt();
5311+
return isUInt<32>(Val.getZExtValue()) || TII->isInlineConstant(Val);
5312+
}
5313+
52905314
SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
52915315
DAGCombinerInfo &DCI) const {
52925316
SelectionDAG &DAG = DCI.DAG;
@@ -5336,6 +5360,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
53365360
SDValue Src = N->getOperand(0);
53375361
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) {
53385362
SDLoc SL(N);
5363+
if (isInt64ImmLegal(C, DAG))
5364+
break;
53395365
uint64_t CVal = C->getZExtValue();
53405366
SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
53415367
DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
@@ -5346,6 +5372,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
53465372
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Src)) {
53475373
const APInt &Val = C->getValueAPF().bitcastToAPInt();
53485374
SDLoc SL(N);
5375+
if (isInt64ImmLegal(C, DAG))
5376+
break;
53495377
uint64_t CVal = Val.getZExtValue();
53505378
SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
53515379
DAG.getConstant(Lo_32(CVal), SL, MVT::i32),

‎llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -103,6 +103,9 @@ class AMDGPUTargetLowering : public TargetLowering {
103103
SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
104104

105105
protected:
106+
/// Check whether value Val can be supported by v_mov_b64, for the current
107+
/// target.
108+
bool isInt64ImmLegal(SDNode *Val, SelectionDAG &DAG) const;
106109
bool shouldCombineMemoryType(EVT VT) const;
107110
SDValue performLoadCombine(SDNode *N, DAGCombinerInfo &DCI) const;
108111
SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const;

‎llvm/lib/Target/AMDGPU/SIISelLowering.cpp‎

Lines changed: 24 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -15202,13 +15202,36 @@ SITargetLowering::performExtractVectorEltCombine(SDNode *N,
1520215202
return V;
1520315203
}
1520415204

15205+
// EXTRACT_VECTOR_ELT (v2i32 bitcast (i64/f64:k), Idx)
15206+
// =>
15207+
// i32:Lo(k) if Idx == 0, or
15208+
// i32:Hi(k) if Idx == 1
15209+
auto *Idx = dyn_cast<ConstantSDNode>(N->getOperand(1));
15210+
if (Vec.getOpcode() == ISD::BITCAST && VecVT == MVT::v2i32 && Idx) {
15211+
SDLoc SL(N);
15212+
SDValue PeekThrough = Vec.getOperand(0);
15213+
auto *KImm = dyn_cast<ConstantSDNode>(PeekThrough);
15214+
if (KImm && KImm->getValueType(0).getSizeInBits() == 64) {
15215+
uint64_t KImmValue = KImm->getZExtValue();
15216+
return DAG.getConstant(
15217+
(KImmValue >> (32 * Idx->getZExtValue())) & 0xffffffff, SL, MVT::i32);
15218+
}
15219+
auto *KFPImm = dyn_cast<ConstantFPSDNode>(PeekThrough);
15220+
if (KFPImm && KFPImm->getValueType(0).getSizeInBits() == 64) {
15221+
uint64_t KFPImmValue =
15222+
KFPImm->getValueAPF().bitcastToAPInt().getZExtValue();
15223+
return DAG.getConstant((KFPImmValue >> (32 * Idx->getZExtValue())) &
15224+
0xffffffff,
15225+
SL, MVT::i32);
15226+
}
15227+
}
15228+
1520515229
if (!DCI.isBeforeLegalize())
1520615230
return SDValue();
1520715231

1520815232
// Try to turn sub-dword accesses of vectors into accesses of the same 32-bit
1520915233
// elements. This exposes more load reduction opportunities by replacing
1521015234
// multiple small extract_vector_elements with a single 32-bit extract.
15211-
auto *Idx = dyn_cast<ConstantSDNode>(N->getOperand(1));
1521215235
if (isa<MemSDNode>(Vec) && VecEltSize <= 16 && VecEltVT.isByteSized() &&
1521315236
VecSize > 32 && VecSize % 32 == 0 && Idx) {
1521415237
EVT NewVT = getEquivalentMemType(*DAG.getContext(), VecVT);

‎llvm/test/CodeGen/AMDGPU/av-split-dead-valno-crash.ll‎

Lines changed: 29 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -7,21 +7,18 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
77
; CHECK-NEXT: s_load_dword s0, s[4:5], 0x8
88
; CHECK-NEXT: s_load_dwordx2 s[10:11], s[4:5], 0x0
99
; CHECK-NEXT: s_load_dwordx4 s[12:15], s[4:5], 0x10
10-
; CHECK-NEXT: v_mov_b32_e32 v1, 0x3e21eeb6
11-
; CHECK-NEXT: v_mov_b32_e32 v20, 0
10+
; CHECK-NEXT: v_mov_b32_e32 v30, 0x9037ab78
11+
; CHECK-NEXT: v_mov_b32_e32 v31, 0x3e21eeb6
1212
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
1313
; CHECK-NEXT: s_bitcmp1_b32 s0, 0
1414
; CHECK-NEXT: s_cselect_b64 s[16:17], -1, 0
1515
; CHECK-NEXT: s_xor_b64 s[18:19], s[16:17], -1
1616
; CHECK-NEXT: s_bitcmp1_b32 s0, 8
1717
; CHECK-NEXT: s_cselect_b64 s[2:3], -1, 0
1818
; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[2:3]
19-
; CHECK-NEXT: v_cmp_ne_u32_e64 s[0:1], 1, v0
20-
; CHECK-NEXT: v_mov_b32_e32 v0, 0x9037ab78
21-
; CHECK-NEXT: v_accvgpr_write_b32 a3, v1
2219
; CHECK-NEXT: s_xor_b64 s[20:21], s[2:3], -1
20+
; CHECK-NEXT: v_cmp_ne_u32_e64 s[0:1], 1, v0
2321
; CHECK-NEXT: s_and_b64 s[2:3], exec, s[2:3]
24-
; CHECK-NEXT: v_accvgpr_write_b32 a2, v0
2522
; CHECK-NEXT: v_mov_b32_e32 v2, 0xa17f65f6
2623
; CHECK-NEXT: v_mov_b32_e32 v3, 0xbe927e4f
2724
; CHECK-NEXT: v_mov_b32_e32 v4, 0x19f4ec90
@@ -37,14 +34,15 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
3734
; CHECK-NEXT: v_mov_b32_e32 v14, 0x8427b883
3835
; CHECK-NEXT: v_mov_b32_e32 v15, 0x3fae1bb4
3936
; CHECK-NEXT: s_mov_b64 s[22:23], 0
40-
; CHECK-NEXT: v_mov_b32_e32 v0, 0x57b87036
41-
; CHECK-NEXT: v_mov_b32_e32 v1, 0x3fb3b136
37+
; CHECK-NEXT: v_mov_b32_e32 v20, 0x57b87036
38+
; CHECK-NEXT: v_mov_b32_e32 v21, 0x3fb3b136
4239
; CHECK-NEXT: s_and_b64 s[4:5], exec, s[16:17]
4340
; CHECK-NEXT: v_mov_b32_e32 v18, 0x55555523
4441
; CHECK-NEXT: v_mov_b32_e32 v19, 0xbfd55555
4542
; CHECK-NEXT: s_and_b64 s[6:7], exec, s[18:19]
46-
; CHECK-NEXT: v_mov_b32_e32 v21, v20
47-
; CHECK-NEXT: ; implicit-def: $vgpr30_vgpr31
43+
; CHECK-NEXT: v_mov_b32_e32 v0, 0
44+
; CHECK-NEXT: v_mov_b64_e32 v[16:17], 0
45+
; CHECK-NEXT: ; implicit-def: $agpr0_agpr1
4846
; CHECK-NEXT: ; implicit-def: $vgpr22_vgpr23
4947
; CHECK-NEXT: s_branch .LBB0_2
5048
; CHECK-NEXT: .LBB0_1: ; %Flow9
@@ -64,12 +62,11 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
6462
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
6563
; CHECK-NEXT: v_mov_b64_e32 v[24:25], s[14:15]
6664
; CHECK-NEXT: flat_load_dwordx2 v[24:25], v[24:25]
67-
; CHECK-NEXT: v_accvgpr_read_b32 v27, a3
68-
; CHECK-NEXT: v_accvgpr_read_b32 v26, a2
65+
; CHECK-NEXT: v_mov_b64_e32 v[26:27], v[30:31]
6966
; CHECK-NEXT: v_mov_b64_e32 v[28:29], v[2:3]
70-
; CHECK-NEXT: v_mov_b64_e32 v[16:17], v[0:1]
71-
; CHECK-NEXT: v_accvgpr_write_b32 a0, 0
72-
; CHECK-NEXT: v_accvgpr_write_b32 a1, 0
67+
; CHECK-NEXT: v_mov_b64_e32 v[16:17], v[20:21]
68+
; CHECK-NEXT: v_accvgpr_write_b32 a2, 0
69+
; CHECK-NEXT: v_accvgpr_write_b32 a3, 0
7370
; CHECK-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
7471
; CHECK-NEXT: v_fmac_f64_e32 v[26:27], 0, v[24:25]
7572
; CHECK-NEXT: v_fmac_f64_e32 v[28:29], 0, v[26:27]
@@ -96,30 +93,32 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
9693
; CHECK-NEXT: .LBB0_6: ; %.preheader1855.i.i.i3329
9794
; CHECK-NEXT: ; Parent Loop BB0_2 Depth=1
9895
; CHECK-NEXT: ; => This Inner Loop Header: Depth=2
99-
; CHECK-NEXT: v_accvgpr_read_b32 v29, a1
100-
; CHECK-NEXT: v_accvgpr_read_b32 v28, a0
96+
; CHECK-NEXT: v_accvgpr_read_b32 v29, a3
97+
; CHECK-NEXT: v_accvgpr_read_b32 v28, a2
10198
; CHECK-NEXT: s_mov_b64 s[24:25], -1
10299
; CHECK-NEXT: s_mov_b64 s[8:9], -1
103100
; CHECK-NEXT: s_mov_b64 vcc, s[2:3]
104-
; CHECK-NEXT: ; implicit-def: $agpr0_agpr1
101+
; CHECK-NEXT: ; implicit-def: $agpr2_agpr3
105102
; CHECK-NEXT: s_cbranch_vccz .LBB0_5
106103
; CHECK-NEXT: ; %bb.7: ; %.lr.ph2070.i.i.i3291
107104
; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=2
108-
; CHECK-NEXT: v_accvgpr_write_b32 a0, v30
109-
; CHECK-NEXT: v_accvgpr_write_b32 a1, v31
105+
; CHECK-NEXT: v_accvgpr_mov_b32 a3, a1
106+
; CHECK-NEXT: v_accvgpr_mov_b32 a2, a0
110107
; CHECK-NEXT: s_mov_b64 s[8:9], s[18:19]
111108
; CHECK-NEXT: s_mov_b64 vcc, s[6:7]
112109
; CHECK-NEXT: s_cbranch_vccz .LBB0_5
113110
; CHECK-NEXT: ; %bb.8: ; %.preheader1856.preheader.i.i.i3325
114111
; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=2
115-
; CHECK-NEXT: v_accvgpr_write_b32 a0, v26
112+
; CHECK-NEXT: v_accvgpr_write_b32 a2, v26
116113
; CHECK-NEXT: s_mov_b64 s[24:25], 0
117-
; CHECK-NEXT: v_accvgpr_write_b32 a1, v27
114+
; CHECK-NEXT: v_accvgpr_write_b32 a3, v27
118115
; CHECK-NEXT: s_mov_b64 s[8:9], 0
119116
; CHECK-NEXT: s_branch .LBB0_5
120117
; CHECK-NEXT: .LBB0_9: ; in Loop: Header=BB0_2 Depth=1
118+
; CHECK-NEXT: v_mov_b64_e32 v[24:25], s[10:11]
119+
; CHECK-NEXT: v_accvgpr_write_b32 a0, v24
121120
; CHECK-NEXT: s_mov_b64 s[22:23], 0
122-
; CHECK-NEXT: v_mov_b64_e32 v[30:31], s[10:11]
121+
; CHECK-NEXT: v_accvgpr_write_b32 a1, v25
123122
; CHECK-NEXT: s_mov_b64 s[8:9], s[20:21]
124123
; CHECK-NEXT: s_branch .LBB0_15
125124
; CHECK-NEXT: .LBB0_10: ; in Loop: Header=BB0_2 Depth=1
@@ -136,19 +135,21 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
136135
; CHECK-NEXT: v_cndmask_b32_e64 v23, v23, 0, s[16:17]
137136
; CHECK-NEXT: v_cndmask_b32_e64 v22, v22, 0, s[16:17]
138137
; CHECK-NEXT: v_cndmask_b32_e64 v16, 0, 1, s[8:9]
139-
; CHECK-NEXT: v_mov_b32_e32 v17, v16
140138
; CHECK-NEXT: s_and_b64 s[8:9], exec, s[16:17]
141-
; CHECK-NEXT: global_store_dwordx2 v20, v[16:17], s[12:13]
139+
; CHECK-NEXT: v_mov_b32_e32 v17, v16
142140
; CHECK-NEXT: s_cselect_b32 s23, s23, 0
143141
; CHECK-NEXT: s_cselect_b32 s22, s22, 0
144142
; CHECK-NEXT: s_mov_b64 s[8:9], -1
143+
; CHECK-NEXT: global_store_dwordx2 v0, v[16:17], s[12:13]
145144
; CHECK-NEXT: s_branch .LBB0_14
146145
; CHECK-NEXT: .LBB0_13: ; in Loop: Header=BB0_2 Depth=1
147146
; CHECK-NEXT: s_mov_b64 s[8:9], 0
148147
; CHECK-NEXT: v_mov_b64_e32 v[22:23], 0
149-
; CHECK-NEXT: .LBB0_14: ; %Flow6
148+
; CHECK-NEXT: .LBB0_14: ; %Flow8
150149
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
151-
; CHECK-NEXT: v_mov_b64_e32 v[30:31], v[24:25]
150+
; CHECK-NEXT: v_accvgpr_write_b32 a0, v24
151+
; CHECK-NEXT: v_mov_b64_e32 v[16:17], 0
152+
; CHECK-NEXT: v_accvgpr_write_b32 a1, v25
152153
; CHECK-NEXT: .LBB0_15: ; %Flow6
153154
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
154155
; CHECK-NEXT: s_mov_b64 s[24:25], -1
@@ -157,7 +158,7 @@ define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2,
157158
; CHECK-NEXT: ; %bb.16: ; %._crit_edge2105.i.i.i2330
158159
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
159160
; CHECK-NEXT: s_mov_b64 s[24:25], 0
160-
; CHECK-NEXT: global_store_dwordx2 v20, v[20:21], s[12:13]
161+
; CHECK-NEXT: global_store_dwordx2 v0, v[16:17], s[12:13]
161162
; CHECK-NEXT: s_branch .LBB0_1
162163
; CHECK-NEXT: .LBB0_17: ; %DummyReturnBlock
163164
; CHECK-NEXT: s_endpgm

‎llvm/test/CodeGen/AMDGPU/flat-scratch.ll‎

Lines changed: 2 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -4152,8 +4152,7 @@ define void @store_load_i64_aligned(ptr addrspace(5) nocapture %arg) {
41524152
; GFX942-LABEL: store_load_i64_aligned:
41534153
; GFX942: ; %bb.0: ; %bb
41544154
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
4155-
; GFX942-NEXT: v_mov_b32_e32 v2, 15
4156-
; GFX942-NEXT: v_mov_b32_e32 v3, 0
4155+
; GFX942-NEXT: v_mov_b64_e32 v[2:3], 15
41574156
; GFX942-NEXT: scratch_store_dwordx2 v0, v[2:3], off sc0 sc1
41584157
; GFX942-NEXT: s_waitcnt vmcnt(0)
41594158
; 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) {
42634262
; GFX942-LABEL: store_load_i64_unaligned:
42644263
; GFX942: ; %bb.0: ; %bb
42654264
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
4266-
; GFX942-NEXT: v_mov_b32_e32 v2, 15
4267-
; GFX942-NEXT: v_mov_b32_e32 v3, 0
4265+
; GFX942-NEXT: v_mov_b64_e32 v[2:3], 15
42684266
; GFX942-NEXT: scratch_store_dwordx2 v0, v[2:3], off sc0 sc1
42694267
; GFX942-NEXT: s_waitcnt vmcnt(0)
42704268
; GFX942-NEXT: scratch_load_dwordx2 v[0:1], v0, off sc0 sc1

‎llvm/test/CodeGen/AMDGPU/imm.ll‎

Lines changed: 13 additions & 26 deletions
Original file line numberDiff line numberDiff line change
@@ -1969,10 +1969,9 @@ define amdgpu_kernel void @add_inline_imm_neg_1_f64(ptr addrspace(1) %out, [8 x
19691969
; GFX942-LABEL: add_inline_imm_neg_1_f64:
19701970
; GFX942: ; %bb.0:
19711971
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
1972-
; GFX942-NEXT: v_mov_b32_e32 v0, -1
19731972
; GFX942-NEXT: s_mov_b32 s3, 0xf000
19741973
; GFX942-NEXT: s_mov_b32 s2, -1
1975-
; GFX942-NEXT: v_mov_b32_e32 v1, v0
1974+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], -1
19761975
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
19771976
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
19781977
; GFX942-NEXT: s_endpgm
@@ -2009,8 +2008,7 @@ define amdgpu_kernel void @add_inline_imm_neg_2_f64(ptr addrspace(1) %out, [8 x
20092008
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
20102009
; GFX942-NEXT: s_mov_b32 s3, 0xf000
20112010
; GFX942-NEXT: s_mov_b32 s2, -1
2012-
; GFX942-NEXT: v_mov_b32_e32 v0, -2
2013-
; GFX942-NEXT: v_mov_b32_e32 v1, -1
2011+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], -2
20142012
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
20152013
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
20162014
; GFX942-NEXT: s_endpgm
@@ -2047,8 +2045,7 @@ define amdgpu_kernel void @add_inline_imm_neg_16_f64(ptr addrspace(1) %out, [8 x
20472045
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
20482046
; GFX942-NEXT: s_mov_b32 s3, 0xf000
20492047
; GFX942-NEXT: s_mov_b32 s2, -1
2050-
; GFX942-NEXT: v_mov_b32_e32 v0, -16
2051-
; GFX942-NEXT: v_mov_b32_e32 v1, -1
2048+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], -16
20522049
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
20532050
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
20542051
; GFX942-NEXT: s_endpgm
@@ -2163,10 +2160,9 @@ define amdgpu_kernel void @store_inline_imm_0.0_f64(ptr addrspace(1) %out) {
21632160
; GFX942-LABEL: store_inline_imm_0.0_f64:
21642161
; GFX942: ; %bb.0:
21652162
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
2166-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
21672163
; GFX942-NEXT: s_mov_b32 s3, 0xf000
21682164
; GFX942-NEXT: s_mov_b32 s2, -1
2169-
; GFX942-NEXT: v_mov_b32_e32 v1, v0
2165+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], 0
21702166
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
21712167
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
21722168
; GFX942-NEXT: s_endpgm
@@ -2239,8 +2235,7 @@ define amdgpu_kernel void @store_inline_imm_0.5_f64(ptr addrspace(1) %out) {
22392235
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
22402236
; GFX942-NEXT: s_mov_b32 s3, 0xf000
22412237
; GFX942-NEXT: s_mov_b32 s2, -1
2242-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
2243-
; GFX942-NEXT: v_mov_b32_e32 v1, 0x3fe00000
2238+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], 0.5
22442239
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
22452240
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
22462241
; GFX942-NEXT: s_endpgm
@@ -2276,8 +2271,7 @@ define amdgpu_kernel void @store_inline_imm_m_0.5_f64(ptr addrspace(1) %out) {
22762271
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
22772272
; GFX942-NEXT: s_mov_b32 s3, 0xf000
22782273
; GFX942-NEXT: s_mov_b32 s2, -1
2279-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
2280-
; GFX942-NEXT: v_mov_b32_e32 v1, 0xbfe00000
2274+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], -0.5
22812275
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
22822276
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
22832277
; GFX942-NEXT: s_endpgm
@@ -2313,8 +2307,7 @@ define amdgpu_kernel void @store_inline_imm_1.0_f64(ptr addrspace(1) %out) {
23132307
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
23142308
; GFX942-NEXT: s_mov_b32 s3, 0xf000
23152309
; GFX942-NEXT: s_mov_b32 s2, -1
2316-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
2317-
; GFX942-NEXT: v_mov_b32_e32 v1, 0x3ff00000
2310+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], 1.0
23182311
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
23192312
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
23202313
; GFX942-NEXT: s_endpgm
@@ -2350,8 +2343,7 @@ define amdgpu_kernel void @store_inline_imm_m_1.0_f64(ptr addrspace(1) %out) {
23502343
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
23512344
; GFX942-NEXT: s_mov_b32 s3, 0xf000
23522345
; GFX942-NEXT: s_mov_b32 s2, -1
2353-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
2354-
; GFX942-NEXT: v_mov_b32_e32 v1, 0xbff00000
2346+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], -1.0
23552347
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
23562348
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
23572349
; GFX942-NEXT: s_endpgm
@@ -2387,8 +2379,7 @@ define amdgpu_kernel void @store_inline_imm_2.0_f64(ptr addrspace(1) %out) {
23872379
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
23882380
; GFX942-NEXT: s_mov_b32 s3, 0xf000
23892381
; GFX942-NEXT: s_mov_b32 s2, -1
2390-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
2391-
; GFX942-NEXT: v_mov_b32_e32 v1, 2.0
2382+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], 2.0
23922383
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
23932384
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
23942385
; GFX942-NEXT: s_endpgm
@@ -2424,8 +2415,7 @@ define amdgpu_kernel void @store_inline_imm_m_2.0_f64(ptr addrspace(1) %out) {
24242415
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
24252416
; GFX942-NEXT: s_mov_b32 s3, 0xf000
24262417
; GFX942-NEXT: s_mov_b32 s2, -1
2427-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
2428-
; GFX942-NEXT: v_mov_b32_e32 v1, -2.0
2418+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], -2.0
24292419
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
24302420
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
24312421
; GFX942-NEXT: s_endpgm
@@ -2461,8 +2451,7 @@ define amdgpu_kernel void @store_inline_imm_4.0_f64(ptr addrspace(1) %out) {
24612451
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
24622452
; GFX942-NEXT: s_mov_b32 s3, 0xf000
24632453
; GFX942-NEXT: s_mov_b32 s2, -1
2464-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
2465-
; GFX942-NEXT: v_mov_b32_e32 v1, 0x40100000
2454+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], 4.0
24662455
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
24672456
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
24682457
; GFX942-NEXT: s_endpgm
@@ -2498,8 +2487,7 @@ define amdgpu_kernel void @store_inline_imm_m_4.0_f64(ptr addrspace(1) %out) {
24982487
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
24992488
; GFX942-NEXT: s_mov_b32 s3, 0xf000
25002489
; GFX942-NEXT: s_mov_b32 s2, -1
2501-
; GFX942-NEXT: v_mov_b32_e32 v0, 0
2502-
; GFX942-NEXT: v_mov_b32_e32 v1, 0xc0100000
2490+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], -4.0
25032491
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
25042492
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
25052493
; GFX942-NEXT: s_endpgm
@@ -2535,8 +2523,7 @@ define amdgpu_kernel void @store_inv_2pi_f64(ptr addrspace(1) %out) {
25352523
; GFX942-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
25362524
; GFX942-NEXT: s_mov_b32 s3, 0xf000
25372525
; GFX942-NEXT: s_mov_b32 s2, -1
2538-
; GFX942-NEXT: v_mov_b32_e32 v0, 0x6dc9c882
2539-
; GFX942-NEXT: v_mov_b32_e32 v1, 0x3fc45f30
2526+
; GFX942-NEXT: v_mov_b64_e32 v[0:1], 0.15915494309189532
25402527
; GFX942-NEXT: s_waitcnt lgkmcnt(0)
25412528
; GFX942-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
25422529
; GFX942-NEXT: s_endpgm

0 commit comments

Comments
 (0)