llvm11/trunk/amdgpu-avoid-an-illegal-operand-in-si-shrink-instr.patch

   1 commit b08a140a8fe8d0b0d16a93042b4952d6e34ab913
   2 Author: Piotr Sobczak <Piotr.Sobczak@amd.com>
   3 Date:   Wed Jan 27 16:02:49 2021 +0100
   4
   5     [AMDGPU] Avoid an illegal operand in si-shrink-instructions
   6
   7     Before the patch it was possible to trigger a constant bus
   8     violation when folding immediates into a shrunk instruction.
   9
  10     The patch adds a check to enforce the legality of the new operand.
  11
  12     Differential Revision: https://reviews.llvm.org/D95527
  13
  14 diff --git a/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp b/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
  15 index 9c6833a7dab6..6c1b16eddc84 100644
  16 --- a/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
  17 +++ b/llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp
  18 @@ -84,21 +84,23 @@ static bool foldImmediates(MachineInstr &MI, const SIInstrInfo *TII,
  19          MachineOperand &MovSrc = Def->getOperand(1);
  20          bool ConstantFolded = false;
  21
  22 -        if (MovSrc.isImm() && (isInt<32>(MovSrc.getImm()) ||
  23 -                               isUInt<32>(MovSrc.getImm()))) {
  24 -          // It's possible to have only one component of a super-reg defined by
  25 -          // a single mov, so we need to clear any subregister flag.
  26 -          Src0.setSubReg(0);
  27 -          Src0.ChangeToImmediate(MovSrc.getImm());
  28 -          ConstantFolded = true;
  29 -        } else if (MovSrc.isFI()) {
  30 -          Src0.setSubReg(0);
  31 -          Src0.ChangeToFrameIndex(MovSrc.getIndex());
  32 -          ConstantFolded = true;
  33 -        } else if (MovSrc.isGlobal()) {
  34 -          Src0.ChangeToGA(MovSrc.getGlobal(), MovSrc.getOffset(),
  35 -                          MovSrc.getTargetFlags());
  36 -          ConstantFolded = true;
  37 +        if (TII->isOperandLegal(MI, Src0Idx, &MovSrc)) {
  38 +          if (MovSrc.isImm() &&
  39 +              (isInt<32>(MovSrc.getImm()) || isUInt<32>(MovSrc.getImm()))) {
  40 +            // It's possible to have only one component of a super-reg defined
  41 +            // by a single mov, so we need to clear any subregister flag.
  42 +            Src0.setSubReg(0);
  43 +            Src0.ChangeToImmediate(MovSrc.getImm());
  44 +            ConstantFolded = true;
  45 +          } else if (MovSrc.isFI()) {
  46 +            Src0.setSubReg(0);
  47 +            Src0.ChangeToFrameIndex(MovSrc.getIndex());
  48 +            ConstantFolded = true;
  49 +          } else if (MovSrc.isGlobal()) {
  50 +            Src0.ChangeToGA(MovSrc.getGlobal(), MovSrc.getOffset(),
  51 +                            MovSrc.getTargetFlags());
  52 +            ConstantFolded = true;
  53 +          }
  54          }
  55
  56          if (ConstantFolded) {
  57 diff --git a/llvm/test/CodeGen/AMDGPU/shrink-instructions-illegal-fold.mir b/llvm/test/CodeGen/AMDGPU/shrink-instructions-illegal-fold.mir
  58 new file mode 100644
  59 index 000000000000..7889f437facf
  60 --- /dev/null
  61 +++ b/llvm/test/CodeGen/AMDGPU/shrink-instructions-illegal-fold.mir
  62 @@ -0,0 +1,23 @@
  63 +# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass=si-shrink-instructions --verify-machineinstrs %s -o - | FileCheck %s
  64 +
  65 +# Make sure immediate folding into V_CNDMASK respects constant bus restrictions.
  66 +---
  67 +
  68 +name:            shrink_cndmask_illegal_imm_folding
  69 +tracksRegLiveness: true
  70 +body:             |
  71 +  bb.0:
  72 +    liveins: $vgpr0, $vgpr1
  73 +    ; CHECK-LABEL: name: shrink_cndmask_illegal_imm_folding
  74 +    ; CHECK: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
  75 +    ; CHECK: [[MOV:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 32768, implicit $exec
  76 +    ; CHECK: V_CMP_EQ_U32_e32 0, [[COPY]], implicit-def $vcc, implicit $exec
  77 +    ; CHECK: V_CNDMASK_B32_e32 [[MOV]], killed [[COPY]], implicit $vcc, implicit $exec
  78 +
  79 +    %0:vgpr_32 = COPY $vgpr0
  80 +    %1:vgpr_32 = V_MOV_B32_e32 32768, implicit $exec
  81 +    V_CMP_EQ_U32_e32 0, %0:vgpr_32, implicit-def $vcc, implicit $exec
  82 +    %2:vgpr_32 = V_CNDMASK_B32_e64 0, %1:vgpr_32, 0, killed %0:vgpr_32, $vcc, implicit $exec
  83 +    S_NOP 0
  84 +
  85 +...