bug fixes

Author: akadutta

llvm/lib/Target/AMDGPU/SIPreEmitPeephole.cpp

@@ -713,8 +713,8 @@ MachineInstrBuilder SIPreEmitPeephole::createUnpackedMI(MachineInstr &I,
                                                         bool IsHiBits) {
   MachineBasicBlock &MBB = *I.getParent();
   const DebugLoc &DL = I.getDebugLoc();
-  const MachineOperand *SrcMO1 = TII->getNamedOperand(I, AMDGPU::OpName::src0);
-  const MachineOperand *SrcMO2 = TII->getNamedOperand(I, AMDGPU::OpName::src1);
+  const MachineOperand *SrcMO0 = TII->getNamedOperand(I, AMDGPU::OpName::src0);
+  const MachineOperand *SrcMO1 = TII->getNamedOperand(I, AMDGPU::OpName::src1);
   Register DstReg = I.getOperand(0).getReg();
   unsigned OpCode = I.getOpcode();
   Register UnpackedDstReg = IsHiBits ? TRI->getSubReg(DstReg, AMDGPU::sub1)
@@ -730,11 +730,12 @@ MachineInstrBuilder SIPreEmitPeephole::createUnpackedMI(MachineInstr &I,
   NewMI.addDef(UnpackedDstReg); // vdst
   if (AMDGPU::hasNamedOperand(UnpackedOpcode, AMDGPU::OpName::src0) &&
       AMDGPU::hasNamedOperand(UnpackedOpcode, AMDGPU::OpName::src1)) {
-    addOperandAndMods(NewMI, Src0Mods, IsHiBits, *SrcMO1);
-    addOperandAndMods(NewMI, Src1Mods, IsHiBits, *SrcMO2);
+    addOperandAndMods(NewMI, Src0Mods, IsHiBits, *SrcMO0);
+    addOperandAndMods(NewMI, Src1Mods, IsHiBits, *SrcMO1);
   } else {
-    const MachineOperand *SrcMO = IsHiBits ? SrcMO2 : SrcMO1;
-    addOperandAndMods(NewMI, Src1Mods, IsHiBits, *SrcMO);
+    const MachineOperand *SrcMO = IsHiBits ? SrcMO1 : SrcMO0;
+    unsigned SrcMods = IsHiBits ? Src1Mods : Src0Mods;
+    addOperandAndMods(NewMI, SrcMods, IsHiBits, *SrcMO);
   }
 
   if (AMDGPU::hasNamedOperand(OpCode, AMDGPU::OpName::src2)) {

llvm/test/CodeGen/AMDGPU/unpack-non-coissue-insts-post-ra-scheduler.mir

@@ -1687,7 +1687,7 @@ body:             |
     ; GFX950-NEXT: $vgpr10 = V_MOV_B32_e32 killed $sgpr6, implicit $exec, implicit $exec
     ; GFX950-NEXT: $vgpr11 = V_MOV_B32_e32 killed $sgpr7, implicit $exec, implicit $exec
     ; GFX950-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec
-    ; GFX950-NEXT: $vgpr4 = V_MOV_B32_e32 killed $vgpr0, implicit $exec
+    ; GFX950-NEXT: $vgpr4 = V_MOV_B32_e32 killed $vgpr1, implicit $exec
     ; GFX950-NEXT: $vgpr5 = V_MOV_B32_e32 killed $vgpr2, implicit $exec
     ; GFX950-NEXT: S_ENDPGM 0
     ;
@@ -1707,7 +1707,7 @@ body:             |
     ; GFX942-NEXT: $vgpr10 = V_MOV_B32_e32 killed $sgpr6, implicit $exec, implicit $exec
     ; GFX942-NEXT: $vgpr11 = V_MOV_B32_e32 killed $sgpr7, implicit $exec, implicit $exec
     ; GFX942-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec
-    ; GFX942-NEXT: $vgpr4 = V_MOV_B32_e32 killed $vgpr0, implicit $exec
+    ; GFX942-NEXT: $vgpr4 = V_MOV_B32_e32 killed $vgpr1, implicit $exec
     ; GFX942-NEXT: $vgpr5 = V_MOV_B32_e32 killed $vgpr2, implicit $exec
     ; GFX942-NEXT: S_ENDPGM 0
     ;