Crash in SIInstrInfo::isSafeToSink

[jayfoad]

New test case test/CodeGen/AMDGPU/machine-sink-crash.mir:

# RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -run-pass=machine-sink %s | FileCheck %s

---
name: f
body: |
  bb.0:
    %81:sreg_32_xm0_xexec = IMPLICIT_DEF
    S_CBRANCH_SCC1 %bb.2, implicit undef $scc
    S_BRANCH %bb.1
  
  bb.1:
    S_BRANCH %bb.3
  
  bb.2:
    S_BRANCH %bb.6
  
  bb.3:
    successors: %bb.4, %bb.5
  
    %10:sreg_32_xm0_xexec = SI_IF undef %81, %bb.5, implicit-def $exec, implicit-def $scc, implicit $exec
    S_BRANCH %bb.4
  
  bb.4:
  
  bb.5:
    successors: %bb.2(0x00000000), %bb.3(0x80000000)
  
    %61:sreg_32 = COPY undef %10
    S_CBRANCH_VCCNZ %bb.3, implicit undef $vcc
    S_BRANCH %bb.2
  
  bb.6:
    ADJCALLSTACKUP 0, 0, implicit-def $scc
    ADJCALLSTACKDOWN 0, 0, implicit-def $scc
...

I get:

llc: lib/Target/AMDGPU/SIInstrInfo.cpp:220: virtual bool llvm::SIInstrInfo::isSafeToSink(llvm::MachineInstr &, llvm::MachineBasicBlock *, llvm::MachineCycleInfo *) const: Assertion `ExitBlocks[0]->getSinglePredecessor()' failed.

SIInstrInfo::isSafeToSink was introduced by #67456.

Cc @sameerds

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Jay Foad (jayfoad)

New test case `test/CodeGen/AMDGPU/machine-sink-crash.mir`: ``` # RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -run-pass=machine-sink %s | FileCheck %s

---

name: f
body: |
bb.0:
%81:sreg_32_xm0_xexec = IMPLICIT_DEF
S_CBRANCH_SCC1 %bb.2, implicit undef $scc
S_BRANCH %bb.1

bb.1:
S_BRANCH %bb.3

bb.2:
S_BRANCH %bb.6

bb.3:
successors: %bb.4, %bb.5

%10:sreg_32_xm0_xexec = SI_IF undef %81, %bb.5, implicit-def $exec, implicit-def $scc, implicit $exec
S_BRANCH %bb.4

bb.4:

bb.5:
successors: %bb.2(0x00000000), %bb.3(0x80000000)

%61:sreg_32 = COPY undef %10
S_CBRANCH_VCCNZ %bb.3, implicit undef $vcc
S_BRANCH %bb.2

bb.6:
ADJCALLSTACKUP 0, 0, implicit-def $scc
ADJCALLSTACKDOWN 0, 0, implicit-def $scc
...

I get:

llc: lib/Target/AMDGPU/SIInstrInfo.cpp:220: virtual bool llvm::SIInstrInfo::isSafeToSink(llvm::MachineInstr &, llvm::MachineBasicBlock *, llvm::MachineCycleInfo *) const: Assertion `ExitBlocks[0]->getSinglePredecessor()' failed.

`SIInstrInfo::isSafeToSink` was introduced by #<!-- -->67456.

Cc @<!-- -->sameerds
</details>

[petar-avramovic]

I am aware of this, after structurize-cfg, cycle exit block should have exactly one predecessor, but again it could be removed in early-tailduplication I don't know if that block deletion is safe but again since it is late in the pipeline it is probably fine?
The quick fix is to search for predecessor that is in the cycle.

[petar-avramovic]

Can you also attach llvm-ir just to check if block was deleted in early-tailduplication

[jayfoad]

Here's the IR I used:

source_filename = "reduced.ll"
target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128-p9:192:256:256:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7:8:9"
target triple = "amdgcn--amdpal"

define amdgpu_ps void @_amdgpu_ps_main(i1 %arg) {
bb:
  %i = call i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32> zeroinitializer, i32 0, i32 0)
  %i1 = icmp slt i32 %i, 0
  br i1 %i1, label %bb2, label %bb12

bb2:                                              ; preds = %bb6, %bb
  %i3 = phi i1 [ %i9, %bb6 ], [ false, %bb ]
  %i4 = phi i32 [ %i10, %bb6 ], [ 0, %bb ]
  br i1 %arg, label %bb5, label %bb6

bb5:                                              ; preds = %bb2
  br label %bb6

bb6:                                              ; preds = %bb5, %bb2
  %i7 = phi i32 [ 0, %bb2 ], [ 1, %bb5 ]
  %i8 = icmp ne i32 %i7, 0
  %i9 = select i1 %i8, i1 %i3, i1 false
  %i10 = or i32 %i4, 1
  %i11 = icmp slt i32 %i4, 0
  br i1 %i11, label %bb2, label %bb12

bb12:                                             ; preds = %bb6, %bb
  %i13 = phi i1 [ false, %bb ], [ %i9, %bb6 ]
  %i14 = select i1 %i13, float 0.000000e+00, float 1.000000e+00
  %i15 = insertelement <4 x float> zeroinitializer, float %i14, i64 0
  call amdgpu_gfx addrspace(4) void null(<4 x float> %i15, i32 0)
  unreachable
}

; Function Attrs: nocallback nofree nosync nounwind willreturn memory(none)
declare i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32>, i32, i32 immarg) #0

attributes #0 = { nocallback nofree nosync nounwind willreturn memory(none) }

[petar-avramovic]

it was early tail duplication, it deleted bb.7.Flow
I already have fix for this, will open a pr

[sameerds]

@ssahasra

[jayfoad]

@ssahasra

Sorry. The problem is that when I type @ and then sameer, this id does not appear in the list of suggested completions.