AMDGPU/GlobalISel: Don't assume instruction can be erased when selecting exts

The G_ANYEXT handling can end up reaching selectCOPY, which mutates
the instruction in place.

llvm-svn: 366915

2 files changed, 27 insertions, 15 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index a68817d5f52..5db2e6fb4e6 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -1006,6 +1006,7 @@ bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
     BuildMI(MBB, I, DL, TII.get(Opcode), DstReg)
       .addImm(0)
       .addImm(Signed ? -1 : 1);
+    I.eraseFromParent();
     return RBI.constrainGenericRegister(DstReg, *DstRC, MRI);
   }
 
@@ -1020,6 +1021,7 @@ bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
       .addImm(0)               // src1_modifiers
       .addImm(Signed ? -1 : 1) // src1
       .addUse(SrcReg);
+    I.eraseFromParent();
     return constrainSelectedInstRegOperands(*ExtI, TII, TRI, RBI);
   }
 
@@ -1036,6 +1038,7 @@ bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
       BuildMI(MBB, I, DL, TII.get(AMDGPU::V_AND_B32_e32), DstReg)
         .addImm(Mask)
         .addReg(SrcReg);
+      I.eraseFromParent();
       return constrainSelectedInstRegOperands(*ExtI, TII, TRI, RBI);
     }
 
@@ -1045,6 +1048,7 @@ bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
       .addReg(SrcReg)
       .addImm(0) // Offset
       .addImm(SrcSize); // Width
+    I.eraseFromParent();
     return constrainSelectedInstRegOperands(*ExtI, TII, TRI, RBI);
   }
 
@@ -1057,6 +1061,7 @@ bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
         AMDGPU::S_SEXT_I32_I8 : AMDGPU::S_SEXT_I32_I16;
       BuildMI(MBB, I, DL, TII.get(SextOpc), DstReg)
         .addReg(SrcReg);
+      I.eraseFromParent();
       return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_32RegClass, MRI);
     }
 
@@ -1081,6 +1086,7 @@ bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
         .addReg(ExtReg)
         .addImm(SrcSize << 16);
 
+      I.eraseFromParent();
       return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_64RegClass, MRI);
     }
 
@@ -1095,6 +1101,7 @@ bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
         .addImm(SrcSize << 16);
     }
 
+    I.eraseFromParent();
     return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_32RegClass, MRI);
   }
 
@@ -1369,12 +1376,7 @@ bool AMDGPUInstructionSelector::select(MachineInstr &I,
   case TargetOpcode::G_SEXT:
   case TargetOpcode::G_ZEXT:
   case TargetOpcode::G_ANYEXT:
-    if (selectG_SZA_EXT(I)) {
-      I.eraseFromParent();
-      return true;
-    }
-
-    return false;
+    return selectG_SZA_EXT(I);
   case TargetOpcode::G_BRCOND:
     return selectG_BRCOND(I);
   case TargetOpcode::G_FRAME_INDEX:
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-anyext.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-anyext.mir
index a2a0d5f85c9..39f64b26690 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-anyext.mir
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-anyext.mir
@@ -1,5 +1,5 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -march=amdgcn -run-pass=instruction-select -verify-machineinstrs -global-isel %s -o - | FileCheck %s -check-prefixes=GCN
+# RUN: llc -march=amdgcn -run-pass=instruction-select -verify-machineinstrs %s -o - | FileCheck %s -check-prefixes=GCN
 
 ---
 
@@ -55,7 +55,8 @@ body: |
     liveins: $sgpr0
 
     ; GCN-LABEL: name: anyext_sgpr_s1_to_sgpr_s32
-    ; GCN: $sgpr0 = COPY %2:sreg_32_xm0
+    ; GCN: [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr0
+    ; GCN: $sgpr0 = COPY [[COPY]]
     %0:sgpr(s32) = COPY $sgpr0
     %1:sgpr(s1) = G_TRUNC %0
     %2:sgpr(s32) = G_ANYEXT %1
@@ -72,7 +73,9 @@ body: |
     liveins: $sgpr0
 
     ; GCN-LABEL: name: anyext_sgpr_s1_to_sgpr_s64
-    ; GCN: $sgpr0_sgpr1 = COPY %2:sreg_64_xexec
+    ; GCN: [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr0
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY [[COPY]]
+    ; GCN: $sgpr0_sgpr1 = COPY [[COPY1]]
     %0:sgpr(s32) = COPY $sgpr0
     %1:sgpr(s1) = G_TRUNC %0
     %2:sgpr(s64) = G_ANYEXT %1
@@ -89,7 +92,8 @@ body: |
     liveins: $sgpr0
 
     ; GCN-LABEL: name: anyext_sgpr_s8_to_sgpr_s32
-    ; GCN: $sgpr0 = COPY %2:sreg_32_xm0
+    ; GCN: [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr0
+    ; GCN: $sgpr0 = COPY [[COPY]]
     %0:sgpr(s32) = COPY $sgpr0
     %1:sgpr(s8) = G_TRUNC %0
     %2:sgpr(s32) = G_ANYEXT %1
@@ -107,7 +111,8 @@ body: |
     liveins: $sgpr0
 
     ; GCN-LABEL: name: anyext_sgpr_s16_to_sgpr_s32
-    ; GCN: $sgpr0 = COPY %2:sreg_32_xm0
+    ; GCN: [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr0
+    ; GCN: $sgpr0 = COPY [[COPY]]
     %0:sgpr(s32) = COPY $sgpr0
     %1:sgpr(s16) = G_TRUNC %0
     %2:sgpr(s32) = G_ANYEXT %1
@@ -125,7 +130,9 @@ body: |
     liveins: $sgpr0
 
     ; GCN-LABEL: name: anyext_sgpr_s16_to_sgpr_s64
-    ; GCN: $sgpr0_sgpr1 = COPY %2:sreg_64_xexec
+    ; GCN: [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr0
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY [[COPY]]
+    ; GCN: $sgpr0_sgpr1 = COPY [[COPY1]]
     %0:sgpr(s32) = COPY $sgpr0
     %1:sgpr(s16) = G_TRUNC %0
     %2:sgpr(s64) = G_ANYEXT %1
@@ -163,7 +170,8 @@ body: |
     liveins: $vgpr0
 
     ; GCN-LABEL: name: anyext_vgpr_s1_to_vgpr_s32
-    ; GCN: $vgpr0 = COPY %2:vgpr_32
+    ; GCN: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GCN: $vgpr0 = COPY [[COPY]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s1) = G_TRUNC %0
     %2:vgpr(s32) = G_ANYEXT %1
@@ -180,7 +188,8 @@ body: |
     liveins: $vgpr0
 
     ; GCN-LABEL: name: anyext_vgpr_s8_to_vgpr_s32
-    ; GCN: $vgpr0 = COPY %2:vgpr_32
+    ; GCN: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GCN: $vgpr0 = COPY [[COPY]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s8) = G_TRUNC %0
     %2:vgpr(s32) = G_ANYEXT %1
@@ -198,7 +207,8 @@ body: |
     liveins: $vgpr0
 
     ; GCN-LABEL: name: anyext_vgpr_s16_to_vgpr_s32
-    ; GCN: $vgpr0 = COPY %2:vgpr_32
+    ; GCN: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GCN: $vgpr0 = COPY [[COPY]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s16) = G_TRUNC %0
     %2:vgpr(s32) = G_ANYEXT %1