Skip to content

Commit 5ca2bd0

Browse files
Pierre-vhPierre-vh
committed
(cherry-pick) [AMDGPU] Do not allow M0 as v_readlane_b32 dst (llvm#128867)
See llvm#128851 - this is the same patch, but for v_readlane_b32. This instruction is used much less often so there were less changes required.
1 parent 98e63d4 commit 5ca2bd0

14 files changed

+56
-50
lines changed

‎llvm/lib/Target/AMDGPU/SIISelLowering.cpp‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4917,7 +4917,8 @@ static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
49174917
Register NewActiveBitsReg = MRI.createVirtualRegister(WaveMaskRegClass);
49184918

49194919
Register FF1Reg = MRI.createVirtualRegister(DstRegClass);
4920-
Register LaneValueReg = MRI.createVirtualRegister(DstRegClass);
4920+
Register LaneValueReg =
4921+
MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
49214922

49224923
bool IsWave32 = ST.isWave32();
49234924
unsigned MovOpc = IsWave32 ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;

‎llvm/lib/Target/AMDGPU/SIInstrInfo.cpp‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2230,6 +2230,8 @@ bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
22302230

22312231
case AMDGPU::SI_RESTORE_S32_FROM_VGPR:
22322232
MI.setDesc(get(AMDGPU::V_READLANE_B32));
2233+
MI.getMF()->getRegInfo().constrainRegClass(MI.getOperand(0).getReg(),
2234+
&AMDGPU::SReg_32_XM0RegClass);
22332235
break;
22342236

22352237
case AMDGPU::V_MOV_B64_PSEUDO: {

‎llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2126,6 +2126,8 @@ bool SIRegisterInfo::spillEmergencySGPR(MachineBasicBlock::iterator MI,
21262126
// Don't need to write VGPR out.
21272127
}
21282128

2129+
MachineRegisterInfo &MRI = MI->getMF()->getRegInfo();
2130+
21292131
// Restore clobbered registers in the specified restore block.
21302132
MI = RestoreMBB.end();
21312133
SB.setMI(&RestoreMBB, MI);
@@ -2140,6 +2142,7 @@ bool SIRegisterInfo::spillEmergencySGPR(MachineBasicBlock::iterator MI,
21402142
SB.NumSubRegs == 1
21412143
? SB.SuperReg
21422144
: Register(getSubReg(SB.SuperReg, SB.SplitParts[i]));
2145+
MRI.constrainRegClass(SubReg, &AMDGPU::SReg_32_XM0RegClass);
21432146
bool LastSubReg = (i + 1 == e);
21442147
auto MIB = BuildMI(*SB.MBB, MI, SB.DL, SB.TII.get(AMDGPU::V_READLANE_B32),
21452148
SubReg)

‎llvm/lib/Target/AMDGPU/VOP2Instructions.td‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -685,7 +685,7 @@ def VOP2e_I16_I16_I16_I1_fake16 : VOP2e_SGPR<[i16, i16, i16, i1]> {
685685
}
686686

687687
def VOP_READLANE : VOPProfile<[i32, i32, i32, untyped]> {
688-
let Outs32 = (outs SReg_32:$vdst);
688+
let Outs32 = (outs SReg_32_XM0:$vdst);
689689
let Outs64 = Outs32;
690690
let Ins32 = (ins VRegOrLdsSrc_32:$src0, SCSrc_b32:$src1);
691691
let Ins64 = Ins32;

‎llvm/test/CodeGen/AMDGPU/GlobalISel/global-atomic-fadd.f32-no-rtn.ll‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -191,7 +191,7 @@ define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_atomicrmw(ptr addrspa
191191
; GFX90A_GFX940-NEXT: [[V_MOV_B32_dpp5:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY17]], [[V_ADD_F32_e64_4]], 323, 12, 15, 0, implicit $exec
192192
; GFX90A_GFX940-NEXT: [[V_ADD_F32_e64_5:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_4]], 0, [[V_MOV_B32_dpp5]], 0, 0, implicit $mode, implicit $exec
193193
; GFX90A_GFX940-NEXT: [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 63
194-
; GFX90A_GFX940-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[V_ADD_F32_e64_5]], [[S_MOV_B32_3]]
194+
; GFX90A_GFX940-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[V_ADD_F32_e64_5]], [[S_MOV_B32_3]]
195195
; GFX90A_GFX940-NEXT: [[COPY18:%[0-9]+]]:vgpr_32 = COPY [[V_READLANE_B32_]]
196196
; GFX90A_GFX940-NEXT: [[STRICT_WWM:%[0-9]+]]:vgpr_32 = STRICT_WWM [[COPY18]], implicit $exec
197197
; GFX90A_GFX940-NEXT: [[COPY19:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]

‎llvm/test/CodeGen/AMDGPU/GlobalISel/global-atomic-fadd.f32-rtn.ll‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -203,7 +203,7 @@ define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_atomicrmw(ptr addrspace
203203
; GFX90A-NEXT: [[COPY18:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_2]]
204204
; GFX90A-NEXT: [[V_MOV_B32_dpp6:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY18]], [[V_ADD_F32_e64_5]], 312, 15, 15, 0, implicit $exec
205205
; GFX90A-NEXT: [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 63
206-
; GFX90A-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[V_ADD_F32_e64_5]], [[S_MOV_B32_3]]
206+
; GFX90A-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[V_ADD_F32_e64_5]], [[S_MOV_B32_3]]
207207
; GFX90A-NEXT: [[COPY19:%[0-9]+]]:vgpr_32 = COPY [[V_READLANE_B32_]]
208208
; GFX90A-NEXT: [[STRICT_WWM:%[0-9]+]]:vgpr_32 = STRICT_WWM [[COPY19]], implicit $exec
209209
; GFX90A-NEXT: [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
@@ -296,7 +296,7 @@ define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_atomicrmw(ptr addrspace
296296
; GFX940-NEXT: [[COPY18:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_2]]
297297
; GFX940-NEXT: [[V_MOV_B32_dpp6:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY18]], [[V_ADD_F32_e64_5]], 312, 15, 15, 0, implicit $exec
298298
; GFX940-NEXT: [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 63
299-
; GFX940-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[V_ADD_F32_e64_5]], [[S_MOV_B32_3]]
299+
; GFX940-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[V_ADD_F32_e64_5]], [[S_MOV_B32_3]]
300300
; GFX940-NEXT: [[COPY19:%[0-9]+]]:vgpr_32 = COPY [[V_READLANE_B32_]]
301301
; GFX940-NEXT: [[STRICT_WWM:%[0-9]+]]:vgpr_32 = STRICT_WWM [[COPY19]], implicit $exec
302302
; GFX940-NEXT: [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
@@ -380,11 +380,11 @@ define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_atomicrmw(ptr addrspace
380380
; GFX11-NEXT: [[COPY12:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
381381
; GFX11-NEXT: [[V_MOV_B32_dpp5:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY12]], [[V_ADD_F32_e64_4]], 273, 15, 15, 0, implicit $exec
382382
; GFX11-NEXT: [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 15
383-
; GFX11-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[V_ADD_F32_e64_4]], [[S_MOV_B32_3]]
383+
; GFX11-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[V_ADD_F32_e64_4]], [[S_MOV_B32_3]]
384384
; GFX11-NEXT: [[S_MOV_B32_4:%[0-9]+]]:sreg_32 = S_MOV_B32 16
385385
; GFX11-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 [[V_READLANE_B32_]], [[S_MOV_B32_4]], [[V_MOV_B32_dpp5]]
386386
; GFX11-NEXT: [[S_MOV_B32_5:%[0-9]+]]:sreg_32 = S_MOV_B32 31
387-
; GFX11-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[V_ADD_F32_e64_4]], [[S_MOV_B32_5]]
387+
; GFX11-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[V_ADD_F32_e64_4]], [[S_MOV_B32_5]]
388388
; GFX11-NEXT: [[COPY13:%[0-9]+]]:vgpr_32 = COPY [[V_READLANE_B32_1]]
389389
; GFX11-NEXT: [[STRICT_WWM:%[0-9]+]]:vgpr_32 = STRICT_WWM [[COPY13]], implicit $exec
390390
; GFX11-NEXT: [[COPY14:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]

‎llvm/test/CodeGen/AMDGPU/GlobalISel/global-atomic-fadd.f64.ll‎

Lines changed: 16 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -428,9 +428,9 @@ define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_atomicrmw(ptr addrspa
428428
; GFX90A_ITERATIVE-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[V_FFBL_B32_e64_1]], [[V_MOV_B32_e32_1]], 0, implicit $exec
429429
; GFX90A_ITERATIVE-NEXT: [[V_MIN_U32_e64_:%[0-9]+]]:vgpr_32 = V_MIN_U32_e64 [[V_FFBL_B32_e64_]], [[V_ADD_U32_e64_]], implicit $exec
430430
; GFX90A_ITERATIVE-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[V_MIN_U32_e64_]], implicit $exec
431-
; GFX90A_ITERATIVE-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY2]], [[V_READFIRSTLANE_B32_]]
431+
; GFX90A_ITERATIVE-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY2]], [[V_READFIRSTLANE_B32_]]
432432
; GFX90A_ITERATIVE-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[V_MIN_U32_e64_]], implicit $exec
433-
; GFX90A_ITERATIVE-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY3]], [[V_READFIRSTLANE_B32_1]]
433+
; GFX90A_ITERATIVE-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY3]], [[V_READFIRSTLANE_B32_1]]
434434
; GFX90A_ITERATIVE-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READLANE_B32_]], %subreg.sub0, [[V_READLANE_B32_1]], %subreg.sub1
435435
; GFX90A_ITERATIVE-NEXT: [[COPY7:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
436436
; GFX90A_ITERATIVE-NEXT: [[V_ADD_F64_e64_:%[0-9]+]]:vreg_64_align2 = nofpexcept V_ADD_F64_e64 0, [[PHI]], 0, [[COPY7]], 0, 0, implicit $mode, implicit $exec
@@ -529,8 +529,8 @@ define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_atomicrmw(ptr addrspa
529529
; GFX90A_DPP-NEXT: [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 63
530530
; GFX90A_DPP-NEXT: [[COPY19:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_5]].sub0
531531
; GFX90A_DPP-NEXT: [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_5]].sub1
532-
; GFX90A_DPP-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY19]], [[S_MOV_B32_2]]
533-
; GFX90A_DPP-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY20]], [[S_MOV_B32_2]]
532+
; GFX90A_DPP-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY19]], [[S_MOV_B32_2]]
533+
; GFX90A_DPP-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY20]], [[S_MOV_B32_2]]
534534
; GFX90A_DPP-NEXT: [[REG_SEQUENCE3:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READLANE_B32_]], %subreg.sub0, [[V_READLANE_B32_1]], %subreg.sub1
535535
; GFX90A_DPP-NEXT: [[COPY21:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE3]]
536536
; GFX90A_DPP-NEXT: [[STRICT_WWM:%[0-9]+]]:vreg_64_align2 = STRICT_WWM [[COPY21]], implicit $exec
@@ -609,9 +609,9 @@ define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_atomicrmw(ptr addrspa
609609
; GFX940_ITERATIVE-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[V_FFBL_B32_e64_1]], [[V_MOV_B32_e32_1]], 0, implicit $exec
610610
; GFX940_ITERATIVE-NEXT: [[V_MIN_U32_e64_:%[0-9]+]]:vgpr_32 = V_MIN_U32_e64 [[V_FFBL_B32_e64_]], [[V_ADD_U32_e64_]], implicit $exec
611611
; GFX940_ITERATIVE-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[V_MIN_U32_e64_]], implicit $exec
612-
; GFX940_ITERATIVE-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY2]], [[V_READFIRSTLANE_B32_]]
612+
; GFX940_ITERATIVE-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY2]], [[V_READFIRSTLANE_B32_]]
613613
; GFX940_ITERATIVE-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[V_MIN_U32_e64_]], implicit $exec
614-
; GFX940_ITERATIVE-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY3]], [[V_READFIRSTLANE_B32_1]]
614+
; GFX940_ITERATIVE-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY3]], [[V_READFIRSTLANE_B32_1]]
615615
; GFX940_ITERATIVE-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READLANE_B32_]], %subreg.sub0, [[V_READLANE_B32_1]], %subreg.sub1
616616
; GFX940_ITERATIVE-NEXT: [[COPY7:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
617617
; GFX940_ITERATIVE-NEXT: [[V_ADD_F64_e64_:%[0-9]+]]:vreg_64_align2 = nofpexcept V_ADD_F64_e64 0, [[PHI]], 0, [[COPY7]], 0, 0, implicit $mode, implicit $exec
@@ -710,8 +710,8 @@ define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_atomicrmw(ptr addrspa
710710
; GFX940_DPP-NEXT: [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 63
711711
; GFX940_DPP-NEXT: [[COPY19:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_5]].sub0
712712
; GFX940_DPP-NEXT: [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_5]].sub1
713-
; GFX940_DPP-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY19]], [[S_MOV_B32_2]]
714-
; GFX940_DPP-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY20]], [[S_MOV_B32_2]]
713+
; GFX940_DPP-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY19]], [[S_MOV_B32_2]]
714+
; GFX940_DPP-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY20]], [[S_MOV_B32_2]]
715715
; GFX940_DPP-NEXT: [[REG_SEQUENCE3:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READLANE_B32_]], %subreg.sub0, [[V_READLANE_B32_1]], %subreg.sub1
716716
; GFX940_DPP-NEXT: [[COPY21:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE3]]
717717
; GFX940_DPP-NEXT: [[STRICT_WWM:%[0-9]+]]:vreg_64_align2 = STRICT_WWM [[COPY21]], implicit $exec
@@ -818,9 +818,9 @@ define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_atomicrmw(ptr addrspac
818818
; GFX90A_ITERATIVE-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[V_FFBL_B32_e64_1]], [[V_MOV_B32_e32_1]], 0, implicit $exec
819819
; GFX90A_ITERATIVE-NEXT: [[V_MIN_U32_e64_:%[0-9]+]]:vgpr_32 = V_MIN_U32_e64 [[V_FFBL_B32_e64_]], [[V_ADD_U32_e64_]], implicit $exec
820820
; GFX90A_ITERATIVE-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[V_MIN_U32_e64_]], implicit $exec
821-
; GFX90A_ITERATIVE-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY2]], [[V_READFIRSTLANE_B32_4]]
821+
; GFX90A_ITERATIVE-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY2]], [[V_READFIRSTLANE_B32_4]]
822822
; GFX90A_ITERATIVE-NEXT: [[V_READFIRSTLANE_B32_5:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[V_MIN_U32_e64_]], implicit $exec
823-
; GFX90A_ITERATIVE-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY3]], [[V_READFIRSTLANE_B32_5]]
823+
; GFX90A_ITERATIVE-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY3]], [[V_READFIRSTLANE_B32_5]]
824824
; GFX90A_ITERATIVE-NEXT: [[REG_SEQUENCE3:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READLANE_B32_]], %subreg.sub0, [[V_READLANE_B32_1]], %subreg.sub1
825825
; GFX90A_ITERATIVE-NEXT: [[COPY16:%[0-9]+]]:vgpr_32 = COPY [[PHI2]].sub0
826826
; GFX90A_ITERATIVE-NEXT: [[COPY17:%[0-9]+]]:vgpr_32 = COPY [[PHI2]].sub1
@@ -936,8 +936,8 @@ define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_atomicrmw(ptr addrspac
936936
; GFX90A_DPP-NEXT: [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 63
937937
; GFX90A_DPP-NEXT: [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_5]].sub0
938938
; GFX90A_DPP-NEXT: [[COPY21:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_5]].sub1
939-
; GFX90A_DPP-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY20]], [[S_MOV_B32_2]]
940-
; GFX90A_DPP-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY21]], [[S_MOV_B32_2]]
939+
; GFX90A_DPP-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY20]], [[S_MOV_B32_2]]
940+
; GFX90A_DPP-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY21]], [[S_MOV_B32_2]]
941941
; GFX90A_DPP-NEXT: [[REG_SEQUENCE3:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READLANE_B32_]], %subreg.sub0, [[V_READLANE_B32_1]], %subreg.sub1
942942
; GFX90A_DPP-NEXT: [[COPY22:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE3]]
943943
; GFX90A_DPP-NEXT: [[STRICT_WWM:%[0-9]+]]:vreg_64_align2 = STRICT_WWM [[COPY22]], implicit $exec
@@ -1070,9 +1070,9 @@ define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_atomicrmw(ptr addrspac
10701070
; GFX940_ITERATIVE-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[V_FFBL_B32_e64_1]], [[V_MOV_B32_e32_1]], 0, implicit $exec
10711071
; GFX940_ITERATIVE-NEXT: [[V_MIN_U32_e64_:%[0-9]+]]:vgpr_32 = V_MIN_U32_e64 [[V_FFBL_B32_e64_]], [[V_ADD_U32_e64_]], implicit $exec
10721072
; GFX940_ITERATIVE-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[V_MIN_U32_e64_]], implicit $exec
1073-
; GFX940_ITERATIVE-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY2]], [[V_READFIRSTLANE_B32_4]]
1073+
; GFX940_ITERATIVE-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY2]], [[V_READFIRSTLANE_B32_4]]
10741074
; GFX940_ITERATIVE-NEXT: [[V_READFIRSTLANE_B32_5:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[V_MIN_U32_e64_]], implicit $exec
1075-
; GFX940_ITERATIVE-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY3]], [[V_READFIRSTLANE_B32_5]]
1075+
; GFX940_ITERATIVE-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY3]], [[V_READFIRSTLANE_B32_5]]
10761076
; GFX940_ITERATIVE-NEXT: [[REG_SEQUENCE3:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READLANE_B32_]], %subreg.sub0, [[V_READLANE_B32_1]], %subreg.sub1
10771077
; GFX940_ITERATIVE-NEXT: [[COPY16:%[0-9]+]]:vgpr_32 = COPY [[PHI2]].sub0
10781078
; GFX940_ITERATIVE-NEXT: [[COPY17:%[0-9]+]]:vgpr_32 = COPY [[PHI2]].sub1
@@ -1188,8 +1188,8 @@ define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_atomicrmw(ptr addrspac
11881188
; GFX940_DPP-NEXT: [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 63
11891189
; GFX940_DPP-NEXT: [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_5]].sub0
11901190
; GFX940_DPP-NEXT: [[COPY21:%[0-9]+]]:vgpr_32 = COPY [[V_ADD_F64_e64_5]].sub1
1191-
; GFX940_DPP-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY20]], [[S_MOV_B32_2]]
1192-
; GFX940_DPP-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[COPY21]], [[S_MOV_B32_2]]
1191+
; GFX940_DPP-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY20]], [[S_MOV_B32_2]]
1192+
; GFX940_DPP-NEXT: [[V_READLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 [[COPY21]], [[S_MOV_B32_2]]
11931193
; GFX940_DPP-NEXT: [[REG_SEQUENCE3:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READLANE_B32_]], %subreg.sub0, [[V_READLANE_B32_1]], %subreg.sub1
11941194
; GFX940_DPP-NEXT: [[COPY22:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE3]]
11951195
; GFX940_DPP-NEXT: [[STRICT_WWM:%[0-9]+]]:vreg_64_align2 = STRICT_WWM [[COPY22]], implicit $exec

‎llvm/test/CodeGen/AMDGPU/global-atomic-fadd.f32-no-rtn.ll‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -185,7 +185,7 @@ define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_atomicrmw(ptr addrspa
185185
; GFX908-NEXT: [[V_MOV_B32_dpp5:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[V_MOV_B32_e32_]], [[V_ADD_F32_e64_4]], 323, 12, 15, 0, implicit $exec
186186
; GFX908-NEXT: [[V_ADD_F32_e64_5:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_4]], 0, killed [[V_MOV_B32_dpp5]], 0, 0, implicit $mode, implicit $exec
187187
; GFX908-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 63
188-
; GFX908-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 killed [[V_ADD_F32_e64_5]], killed [[S_MOV_B32_1]]
188+
; GFX908-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 killed [[V_ADD_F32_e64_5]], killed [[S_MOV_B32_1]]
189189
; GFX908-NEXT: early-clobber %1:sgpr_32 = STRICT_WWM killed [[V_READLANE_B32_]], implicit $exec
190190
; GFX908-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 killed [[V_MBCNT_HI_U32_B32_e64_]], [[S_MOV_B32_]], implicit $exec
191191
; GFX908-NEXT: [[SI_IF1:%[0-9]+]]:sreg_64 = SI_IF killed [[V_CMP_EQ_U32_e64_]], %bb.3, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
@@ -246,7 +246,7 @@ define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_atomicrmw(ptr addrspa
246246
; GFX90A_GFX940-NEXT: [[V_MOV_B32_dpp5:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[V_MOV_B32_e32_]], [[V_ADD_F32_e64_4]], 323, 12, 15, 0, implicit $exec
247247
; GFX90A_GFX940-NEXT: [[V_ADD_F32_e64_5:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_4]], 0, killed [[V_MOV_B32_dpp5]], 0, 0, implicit $mode, implicit $exec
248248
; GFX90A_GFX940-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 63
249-
; GFX90A_GFX940-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 killed [[V_ADD_F32_e64_5]], killed [[S_MOV_B32_1]]
249+
; GFX90A_GFX940-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READLANE_B32 killed [[V_ADD_F32_e64_5]], killed [[S_MOV_B32_1]]
250250
; GFX90A_GFX940-NEXT: early-clobber %1:sgpr_32 = STRICT_WWM killed [[V_READLANE_B32_]], implicit $exec
251251
; GFX90A_GFX940-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 killed [[V_MBCNT_HI_U32_B32_e64_]], [[S_MOV_B32_]], implicit $exec
252252
; GFX90A_GFX940-NEXT: [[SI_IF1:%[0-9]+]]:sreg_64 = SI_IF killed [[V_CMP_EQ_U32_e64_]], %bb.3, implicit-def dead $exec, implicit-def dead $scc, implicit $exec

0 commit comments

Comments
 (0)