Skip to content

Commit b215a26

Browse files
ruilingruiling
committed
[AMDGPU] Update LiveVariables in convertToThreeAddress()
This can fix an asan failure like below. ==15856==ERROR: AddressSanitizer: use-after-poison on address ... READ of size 8 at 0x6210001a3cb0 thread T0 #0 llvm::MachineInstr::getParent() #1 llvm::LiveVariables::VarInfo::findKill() rust-lang#2 TwoAddressInstructionPass::rescheduleMIBelowKill() rust-lang#3 TwoAddressInstructionPass::tryInstructionTransform() rust-lang#4 TwoAddressInstructionPass::runOnMachineFunction() We need to update the Kills if we replace instructions. The Kills may be later accessed within TwoAddressInstruction pass. Differential Revision: https://reviews.llvm.org/D89092
1 parent 7f8dc34 commit b215a26

File tree

2 files changed

+99
-38
lines changed

2 files changed

+99
-38
lines changed

‎llvm/lib/Target/AMDGPU/SIInstrInfo.cpp‎

Lines changed: 63 additions & 38 deletions
Original file line numberDiff line numberDiff line change
@@ -15,10 +15,10 @@
1515
#include "AMDGPU.h"
1616
#include "AMDGPUSubtarget.h"
1717
#include "GCNHazardRecognizer.h"
18+
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
1819
#include "SIDefines.h"
1920
#include "SIMachineFunctionInfo.h"
2021
#include "SIRegisterInfo.h"
21-
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
2222
#include "Utils/AMDGPUBaseInfo.h"
2323
#include "llvm/ADT/APInt.h"
2424
#include "llvm/ADT/ArrayRef.h"
@@ -28,6 +28,7 @@
2828
#include "llvm/Analysis/AliasAnalysis.h"
2929
#include "llvm/Analysis/MemoryLocation.h"
3030
#include "llvm/Analysis/ValueTracking.h"
31+
#include "llvm/CodeGen/LiveVariables.h"
3132
#include "llvm/CodeGen/MachineBasicBlock.h"
3233
#include "llvm/CodeGen/MachineDominators.h"
3334
#include "llvm/CodeGen/MachineFrameInfo.h"
@@ -2841,6 +2842,18 @@ static int64_t getFoldableImm(const MachineOperand* MO) {
28412842
return AMDGPU::NoRegister;
28422843
}
28432844

2845+
static void updateLiveVariables(LiveVariables *LV, MachineInstr &MI,
2846+
MachineInstr &NewMI) {
2847+
if (LV) {
2848+
unsigned NumOps = MI.getNumOperands();
2849+
for (unsigned I = 1; I < NumOps; ++I) {
2850+
MachineOperand &Op = MI.getOperand(I);
2851+
if (Op.isReg() && Op.isKill())
2852+
LV->replaceKillInstruction(Op.getReg(), MI, NewMI);
2853+
}
2854+
}
2855+
}
2856+
28442857
MachineInstr *SIInstrInfo::convertToThreeAddress(MachineFunction::iterator &MBB,
28452858
MachineInstr &MI,
28462859
LiveVariables *LV) const {
@@ -2888,43 +2901,53 @@ MachineInstr *SIInstrInfo::convertToThreeAddress(MachineFunction::iterator &MBB,
28882901
const MachineOperand *Src2 = getNamedOperand(MI, AMDGPU::OpName::src2);
28892902
const MachineOperand *Clamp = getNamedOperand(MI, AMDGPU::OpName::clamp);
28902903
const MachineOperand *Omod = getNamedOperand(MI, AMDGPU::OpName::omod);
2904+
MachineInstrBuilder MIB;
28912905

28922906
if (!Src0Mods && !Src1Mods && !Clamp && !Omod &&
28932907
// If we have an SGPR input, we will violate the constant bus restriction.
2894-
(ST.getConstantBusLimit(Opc) > 1 ||
2895-
!Src0->isReg() ||
2908+
(ST.getConstantBusLimit(Opc) > 1 || !Src0->isReg() ||
28962909
!RI.isSGPRReg(MBB->getParent()->getRegInfo(), Src0->getReg()))) {
28972910
if (auto Imm = getFoldableImm(Src2)) {
28982911
unsigned NewOpc =
2899-
IsFMA ? (IsF16 ? AMDGPU::V_FMAAK_F16 : AMDGPU::V_FMAAK_F32)
2900-
: (IsF16 ? AMDGPU::V_MADAK_F16 : AMDGPU::V_MADAK_F32);
2901-
if (pseudoToMCOpcode(NewOpc) != -1)
2902-
return BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpc))
2903-
.add(*Dst)
2904-
.add(*Src0)
2905-
.add(*Src1)
2906-
.addImm(Imm);
2907-
}
2908-
unsigned NewOpc =
2909-
IsFMA ? (IsF16 ? AMDGPU::V_FMAMK_F16 : AMDGPU::V_FMAMK_F32)
2910-
: (IsF16 ? AMDGPU::V_MADMK_F16 : AMDGPU::V_MADMK_F32);
2912+
IsFMA ? (IsF16 ? AMDGPU::V_FMAAK_F16 : AMDGPU::V_FMAAK_F32)
2913+
: (IsF16 ? AMDGPU::V_MADAK_F16 : AMDGPU::V_MADAK_F32);
2914+
if (pseudoToMCOpcode(NewOpc) != -1) {
2915+
MIB = BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpc))
2916+
.add(*Dst)
2917+
.add(*Src0)
2918+
.add(*Src1)
2919+
.addImm(Imm);
2920+
updateLiveVariables(LV, MI, *MIB);
2921+
return MIB;
2922+
}
2923+
}
2924+
unsigned NewOpc = IsFMA
2925+
? (IsF16 ? AMDGPU::V_FMAMK_F16 : AMDGPU::V_FMAMK_F32)
2926+
: (IsF16 ? AMDGPU::V_MADMK_F16 : AMDGPU::V_MADMK_F32);
29112927
if (auto Imm = getFoldableImm(Src1)) {
2912-
if (pseudoToMCOpcode(NewOpc) != -1)
2913-
return BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpc))
2914-
.add(*Dst)
2915-
.add(*Src0)
2916-
.addImm(Imm)
2917-
.add(*Src2);
2928+
if (pseudoToMCOpcode(NewOpc) != -1) {
2929+
MIB = BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpc))
2930+
.add(*Dst)
2931+
.add(*Src0)
2932+
.addImm(Imm)
2933+
.add(*Src2);
2934+
updateLiveVariables(LV, MI, *MIB);
2935+
return MIB;
2936+
}
29182937
}
29192938
if (auto Imm = getFoldableImm(Src0)) {
29202939
if (pseudoToMCOpcode(NewOpc) != -1 &&
2921-
isOperandLegal(MI, AMDGPU::getNamedOperandIdx(NewOpc,
2922-
AMDGPU::OpName::src0), Src1))
2923-
return BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpc))
2924-
.add(*Dst)
2925-
.add(*Src1)
2926-
.addImm(Imm)
2927-
.add(*Src2);
2940+
isOperandLegal(
2941+
MI, AMDGPU::getNamedOperandIdx(NewOpc, AMDGPU::OpName::src0),
2942+
Src1)) {
2943+
MIB = BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpc))
2944+
.add(*Dst)
2945+
.add(*Src1)
2946+
.addImm(Imm)
2947+
.add(*Src2);
2948+
updateLiveVariables(LV, MI, *MIB);
2949+
return MIB;
2950+
}
29282951
}
29292952
}
29302953

@@ -2933,16 +2956,18 @@ MachineInstr *SIInstrInfo::convertToThreeAddress(MachineFunction::iterator &MBB,
29332956
if (pseudoToMCOpcode(NewOpc) == -1)
29342957
return nullptr;
29352958

2936-
return BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpc))
2937-
.add(*Dst)
2938-
.addImm(Src0Mods ? Src0Mods->getImm() : 0)
2939-
.add(*Src0)
2940-
.addImm(Src1Mods ? Src1Mods->getImm() : 0)
2941-
.add(*Src1)
2942-
.addImm(0) // Src mods
2943-
.add(*Src2)
2944-
.addImm(Clamp ? Clamp->getImm() : 0)
2945-
.addImm(Omod ? Omod->getImm() : 0);
2959+
MIB = BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpc))
2960+
.add(*Dst)
2961+
.addImm(Src0Mods ? Src0Mods->getImm() : 0)
2962+
.add(*Src0)
2963+
.addImm(Src1Mods ? Src1Mods->getImm() : 0)
2964+
.add(*Src1)
2965+
.addImm(0) // Src mods
2966+
.add(*Src2)
2967+
.addImm(Clamp ? Clamp->getImm() : 0)
2968+
.addImm(Omod ? Omod->getImm() : 0);
2969+
updateLiveVariables(LV, MI, *MIB);
2970+
return MIB;
29462971
}
29472972

29482973
// It's not generally safe to move VALU instructions across these since it will

‎llvm/test/CodeGen/AMDGPU/stale-livevar-in-twoaddr-pass.mir‎

Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass=livevars,phi-node-elimination,twoaddressinstruction -verify-machineinstrs -o - %s | FileCheck %s
2+
# This used to fail under ASAN enabled build because we didn't update LiveVariables in SIInstrInfo::convertToThreeAddress()
3+
# CHECK: _amdgpu_ps_main
4+
5+
---
6+
name: _amdgpu_ps_main
7+
alignment: 1
8+
tracksRegLiveness: true
9+
body: |
10+
bb.0:
11+
liveins: $sgpr2, $vgpr2, $vgpr3
12+
13+
%0:vgpr_32 = COPY $vgpr3
14+
%1:vgpr_32 = COPY $vgpr2
15+
S_BRANCH %bb.3
16+
17+
bb.1:
18+
%2:vgpr_32 = V_MAC_F32_e32 0, %0, %1, implicit $mode, implicit $exec
19+
%3:vgpr_32 = V_MED3_F32 0, %1, 0, %2, 0, %2, 0, 0, implicit $mode, implicit $exec
20+
21+
bb.2:
22+
%4:vgpr_32 = PHI %5, %bb.3, %3, %bb.1
23+
SI_END_CF %6, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
24+
EXP_DONE 0, %4, %4, %4, %4, -1, 0, 15, implicit $exec
25+
S_ENDPGM 0
26+
27+
bb.3:
28+
successors: %bb.1, %bb.2
29+
30+
%5:vgpr_32 = V_MAC_F32_e32 0, %1, %0, implicit $mode, implicit $exec
31+
%7:vgpr_32 = V_CVT_I32_F32_e32 %5, implicit $mode, implicit $exec
32+
%8:sreg_64 = V_CMP_NE_U32_e64 1, %7, implicit $exec
33+
%6:sreg_64 = SI_IF %8, %bb.2, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
34+
S_BRANCH %bb.1
35+
36+
...

0 commit comments

Comments
 (0)