Revert "[DependenceAnalysis] Dependecies for loads marked with "ivnariant.load" should not be shared with general accesses. Fix for https://bugs.llvm.org/show_bug.cgi?id=42151"

[george.karpenkov]

Summary

Revert "[DependenceAnalysis] Dependecies for loads marked with "ivnariant.load" should not be shared with general accesses. Fix for https://bugs.llvm.org/show_bug.cgi?id=42151"

This reverts commit 5f026b6d9e882941fde9b7e5dc0a2d807f7f24f5.

We're (tensorflow.org/xla team) seeing some misscompiles with the new change, only at -O3, with fast math disabled.

I'm still trying to come up with a useful/small/external example, but for now, the following IR:

; ModuleID = '__compute_module'
source_filename = "__compute_module"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-grtev4-linux-gnu"

@0 = private unnamed_addr constant [4 x i8] c"\DB\0F\C9@"
@1 = private unnamed_addr constant [4 x i8] c"\00\00\00?"

; Function Attrs: uwtable
define void @jit_wrapped_fun.31(i8* %retval, i8* noalias %run_options, i8** noalias %params, i8** noalias %buffer_table, i64* noalias %prof_counters) #0 {
entry:
  %fusion.invar_address.dim.2 = alloca i64
  %fusion.invar_address.dim.1 = alloca i64
  %fusion.invar_address.dim.0 = alloca i64
  %fusion.1.invar_address.dim.2 = alloca i64
  %fusion.1.invar_address.dim.1 = alloca i64
  %fusion.1.invar_address.dim.0 = alloca i64
  %0 = getelementptr inbounds i8*, i8** %buffer_table, i64 1
  %1 = load i8*, i8** %0, !invariant.load !0, !dereferenceable !1, !align !2
  %parameter.3 = bitcast i8* %1 to [2 x [1 x [4 x float]]]*
  %2 = getelementptr inbounds i8*, i8** %buffer_table, i64 5
  %3 = load i8*, i8** %2, !invariant.load !0, !dereferenceable !1, !align !2
  %fusion.1 = bitcast i8* %3 to [2 x [1 x [4 x float]]]*
  store i64 0, i64* %fusion.1.invar_address.dim.0
  br label %fusion.1.loop_header.dim.0

fusion.1.loop_header.dim.0:                       ; preds = %fusion.1.loop_exit.dim.1, %entry
  %fusion.1.indvar.dim.0 = load i64, i64* %fusion.1.invar_address.dim.0
  %4 = icmp uge i64 %fusion.1.indvar.dim.0, 2
  br i1 %4, label %fusion.1.loop_exit.dim.0, label %fusion.1.loop_body.dim.0

fusion.1.loop_body.dim.0:                         ; preds = %fusion.1.loop_header.dim.0
  store i64 0, i64* %fusion.1.invar_address.dim.1
  br label %fusion.1.loop_header.dim.1

fusion.1.loop_header.dim.1:                       ; preds = %fusion.1.loop_exit.dim.2, %fusion.1.loop_body.dim.0
  %fusion.1.indvar.dim.1 = load i64, i64* %fusion.1.invar_address.dim.1
  %5 = icmp uge i64 %fusion.1.indvar.dim.1, 1
  br i1 %5, label %fusion.1.loop_exit.dim.1, label %fusion.1.loop_body.dim.1

fusion.1.loop_body.dim.1:                         ; preds = %fusion.1.loop_header.dim.1
  store i64 0, i64* %fusion.1.invar_address.dim.2
  br label %fusion.1.loop_header.dim.2

fusion.1.loop_header.dim.2:                       ; preds = %fusion.1.loop_body.dim.2, %fusion.1.loop_body.dim.1
  %fusion.1.indvar.dim.2 = load i64, i64* %fusion.1.invar_address.dim.2
  %6 = icmp uge i64 %fusion.1.indvar.dim.2, 4
  br i1 %6, label %fusion.1.loop_exit.dim.2, label %fusion.1.loop_body.dim.2

fusion.1.loop_body.dim.2:                         ; preds = %fusion.1.loop_header.dim.2
  %7 = load float, float* bitcast ([4 x i8]* @0 to float*)
  %8 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %parameter.3, i64 0, i64 %fusion.1.indvar.dim.0, i64 0, i64 %fusion.1.indvar.dim.2
  %9 = load float, float* %8, !invariant.load !0, !noalias !3
  %10 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %parameter.3, i64 0, i64 %fusion.1.indvar.dim.0, i64 0, i64 %fusion.1.indvar.dim.2
  %11 = load float, float* %10, !invariant.load !0, !noalias !3
  %12 = fmul float %9, %11
  %13 = fmul float %7, %12
  %14 = call float @llvm.log.f32(float %13)
  %15 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %fusion.1, i64 0, i64 %fusion.1.indvar.dim.0, i64 0, i64 %fusion.1.indvar.dim.2
  store float %14, float* %15, !alias.scope !7, !noalias !8
  %invar.inc2 = add nuw nsw i64 %fusion.1.indvar.dim.2, 1
  store i64 %invar.inc2, i64* %fusion.1.invar_address.dim.2
  br label %fusion.1.loop_header.dim.2

fusion.1.loop_exit.dim.2:                         ; preds = %fusion.1.loop_header.dim.2
  %invar.inc1 = add nuw nsw i64 %fusion.1.indvar.dim.1, 1
  store i64 %invar.inc1, i64* %fusion.1.invar_address.dim.1
  br label %fusion.1.loop_header.dim.1

fusion.1.loop_exit.dim.1:                         ; preds = %fusion.1.loop_header.dim.1
  %invar.inc = add nuw nsw i64 %fusion.1.indvar.dim.0, 1
  store i64 %invar.inc, i64* %fusion.1.invar_address.dim.0
  br label %fusion.1.loop_header.dim.0

fusion.1.loop_exit.dim.0:                         ; preds = %fusion.1.loop_header.dim.0
  %16 = getelementptr inbounds i8*, i8** %buffer_table, i64 4
  %17 = load i8*, i8** %16, !invariant.load !0, !dereferenceable !9, !align !2
  %parameter.1 = bitcast i8* %17 to float*
  %18 = getelementptr inbounds i8*, i8** %buffer_table, i64 2
  %19 = load i8*, i8** %18, !invariant.load !0, !dereferenceable !10, !align !2
  %parameter.2 = bitcast i8* %19 to [3 x [1 x float]]*
  %20 = getelementptr inbounds i8*, i8** %buffer_table, i64 0
  %21 = load i8*, i8** %20, !invariant.load !0, !dereferenceable !11, !align !2
  %fusion = bitcast i8* %21 to [2 x [3 x [4 x float]]]*
  store i64 0, i64* %fusion.invar_address.dim.0
  br label %fusion.loop_header.dim.0

fusion.loop_header.dim.0:                         ; preds = %fusion.loop_exit.dim.1, %fusion.1.loop_exit.dim.0
  %fusion.indvar.dim.0 = load i64, i64* %fusion.invar_address.dim.0
  %22 = icmp uge i64 %fusion.indvar.dim.0, 2
  br i1 %22, label %fusion.loop_exit.dim.0, label %fusion.loop_body.dim.0

fusion.loop_body.dim.0:                           ; preds = %fusion.loop_header.dim.0
  store i64 0, i64* %fusion.invar_address.dim.1
  br label %fusion.loop_header.dim.1

fusion.loop_header.dim.1:                         ; preds = %fusion.loop_exit.dim.2, %fusion.loop_body.dim.0
  %fusion.indvar.dim.1 = load i64, i64* %fusion.invar_address.dim.1
  %23 = icmp uge i64 %fusion.indvar.dim.1, 3
  br i1 %23, label %fusion.loop_exit.dim.1, label %fusion.loop_body.dim.1

fusion.loop_body.dim.1:                           ; preds = %fusion.loop_header.dim.1
  store i64 0, i64* %fusion.invar_address.dim.2
  br label %fusion.loop_header.dim.2

fusion.loop_header.dim.2:                         ; preds = %fusion.loop_body.dim.2, %fusion.loop_body.dim.1
  %fusion.indvar.dim.2 = load i64, i64* %fusion.invar_address.dim.2
  %24 = icmp uge i64 %fusion.indvar.dim.2, 4
  br i1 %24, label %fusion.loop_exit.dim.2, label %fusion.loop_body.dim.2

fusion.loop_body.dim.2:                           ; preds = %fusion.loop_header.dim.2
  %25 = mul nuw nsw i64 %fusion.indvar.dim.2, 1
  %26 = add nuw nsw i64 0, %25
  %27 = udiv i64 %26, 4
  %28 = mul nuw nsw i64 %fusion.indvar.dim.0, 1
  %29 = add nuw nsw i64 0, %28
  %30 = udiv i64 %29, 2
  %31 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %fusion.1, i64 0, i64 %29, i64 0, i64 %26
  %32 = load float, float* %31, !alias.scope !7, !noalias !8
  %33 = mul nuw nsw i64 %fusion.indvar.dim.1, 1
  %34 = add nuw nsw i64 0, %33
  %35 = udiv i64 %34, 3
  %36 = load float, float* %parameter.1, !invariant.load !0, !noalias !3
  %37 = getelementptr inbounds [3 x [1 x float]], [3 x [1 x float]]* %parameter.2, i64 0, i64 %34, i64 0
  %38 = load float, float* %37, !invariant.load !0, !noalias !3
  %39 = fsub float %36, %38
  %40 = fmul float %39, %39
  %41 = mul nuw nsw i64 %fusion.indvar.dim.2, 1
  %42 = add nuw nsw i64 0, %41
  %43 = udiv i64 %42, 4
  %44 = mul nuw nsw i64 %fusion.indvar.dim.0, 1
  %45 = add nuw nsw i64 0, %44
  %46 = udiv i64 %45, 2
  %47 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %parameter.3, i64 0, i64 %45, i64 0, i64 %42
  %48 = load float, float* %47, !invariant.load !0, !noalias !3
  %49 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %parameter.3, i64 0, i64 %45, i64 0, i64 %42
  %50 = load float, float* %49, !invariant.load !0, !noalias !3
  %51 = fmul float %48, %50
  %52 = fdiv float %40, %51
  %53 = fadd float %32, %52
  %54 = fneg float %53
  %55 = load float, float* bitcast ([4 x i8]* @1 to float*)
  %56 = fmul float %54, %55
  %57 = getelementptr inbounds [2 x [3 x [4 x float]]], [2 x [3 x [4 x float]]]* %fusion, i64 0, i64 %fusion.indvar.dim.0, i64 %fusion.indvar.dim.1, i64 %fusion.indvar.dim.2
  store float %56, float* %57, !alias.scope !8, !noalias !12
  %invar.inc5 = add nuw nsw i64 %fusion.indvar.dim.2, 1
  store i64 %invar.inc5, i64* %fusion.invar_address.dim.2
  br label %fusion.loop_header.dim.2

fusion.loop_exit.dim.2:                           ; preds = %fusion.loop_header.dim.2
  %invar.inc4 = add nuw nsw i64 %fusion.indvar.dim.1, 1
  store i64 %invar.inc4, i64* %fusion.invar_address.dim.1
  br label %fusion.loop_header.dim.1

fusion.loop_exit.dim.1:                           ; preds = %fusion.loop_header.dim.1
  %invar.inc3 = add nuw nsw i64 %fusion.indvar.dim.0, 1
  store i64 %invar.inc3, i64* %fusion.invar_address.dim.0
  br label %fusion.loop_header.dim.0

fusion.loop_exit.dim.0:                           ; preds = %fusion.loop_header.dim.0
  %58 = getelementptr inbounds i8*, i8** %buffer_table, i64 3
  %59 = load i8*, i8** %58, !invariant.load !0, !dereferenceable !2, !align !2
  %tuple.30 = bitcast i8* %59 to [1 x i8*]*
  %60 = bitcast [2 x [3 x [4 x float]]]* %fusion to i8*
  %61 = getelementptr inbounds [1 x i8*], [1 x i8*]* %tuple.30, i64 0, i64 0
  store i8* %60, i8** %61, !alias.scope !14, !noalias !8
  ret void
}

; Function Attrs: nounwind readnone speculatable willreturn
declare float @llvm.log.f32(float) #1

attributes #0 = { uwtable "no-frame-pointer-elim"="false" }
attributes #1 = { nounwind readnone speculatable willreturn }

!0 = !{}
!1 = !{i64 32}
!2 = !{i64 8}
!3 = !{!4, !6}
!4 = !{!"buffer: {index:0, offset:0, size:96}", !5}
!5 = !{!"XLA global AA domain"}
!6 = !{!"buffer: {index:5, offset:0, size:32}", !5}
!7 = !{!6}
!8 = !{!4}
!9 = !{i64 4}
!10 = !{i64 12}
!11 = !{i64 96}
!12 = !{!13, !6}
!13 = !{!"buffer: {index:3, offset:0, size:8}", !5}
!14 = !{!13}

gets (correctly) optimized to the one below without the change:

; ModuleID = '__compute_module'
source_filename = "__compute_module"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-grtev4-linux-gnu"

; Function Attrs: nofree nounwind uwtable
define void @jit_wrapped_fun.31(i8* nocapture readnone %retval, i8* noalias nocapture readnone %run_options, i8** noalias nocapture readnone %params, i8** noalias nocapture readonly %buffer_table, i64* noalias nocapture readnone %prof_counters) local_unnamed_addr #0 {
entry:
  %0 = getelementptr inbounds i8*, i8** %buffer_table, i64 1
  %1 = bitcast i8** %0 to [2 x [1 x [4 x float]]]**
  %2 = load [2 x [1 x [4 x float]]]*, [2 x [1 x [4 x float]]]** %1, align 8, !invariant.load !0, !dereferenceable !1, !align !2
  %3 = getelementptr inbounds i8*, i8** %buffer_table, i64 5
  %4 = bitcast i8** %3 to [2 x [1 x [4 x float]]]**
  %5 = load [2 x [1 x [4 x float]]]*, [2 x [1 x [4 x float]]]** %4, align 8, !invariant.load !0, !dereferenceable !1, !align !2
  %6 = bitcast [2 x [1 x [4 x float]]]* %2 to <4 x float>*
  %7 = load <4 x float>, <4 x float>* %6, align 8, !invariant.load !0, !noalias !3
  %8 = fmul <4 x float> %7, %7
  %9 = fmul <4 x float> %8, <float 0x401921FB60000000, float 0x401921FB60000000, float 0x401921FB60000000, float 0x401921FB60000000>
  %10 = call <4 x float> @llvm.log.v4f32(<4 x float> %9)
  %11 = bitcast [2 x [1 x [4 x float]]]* %5 to <4 x float>*
  store <4 x float> %10, <4 x float>* %11, align 8, !alias.scope !7, !noalias !8
  %12 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %2, i64 0, i64 1, i64 0, i64 0
  %13 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %5, i64 0, i64 1, i64 0, i64 0
  %14 = bitcast float* %12 to <4 x float>*
  %15 = load <4 x float>, <4 x float>* %14, align 8, !invariant.load !0, !noalias !3
  %16 = fmul <4 x float> %15, %15
  %17 = fmul <4 x float> %16, <float 0x401921FB60000000, float 0x401921FB60000000, float 0x401921FB60000000, float 0x401921FB60000000>
  %18 = call <4 x float> @llvm.log.v4f32(<4 x float> %17)
  %19 = bitcast float* %13 to <4 x float>*
  store <4 x float> %18, <4 x float>* %19, align 8, !alias.scope !7, !noalias !8
  %20 = getelementptr inbounds i8*, i8** %buffer_table, i64 4
  %21 = bitcast i8** %20 to float**
  %22 = load float*, float** %21, align 8, !invariant.load !0, !dereferenceable !9, !align !2
  %23 = getelementptr inbounds i8*, i8** %buffer_table, i64 2
  %24 = bitcast i8** %23 to [3 x [1 x float]]**
  %25 = load [3 x [1 x float]]*, [3 x [1 x float]]** %24, align 8, !invariant.load !0, !dereferenceable !10, !align !2
  %26 = load i8*, i8** %buffer_table, align 8, !invariant.load !0, !dereferenceable !11, !align !2
  %27 = load float, float* %22, align 8, !invariant.load !0, !noalias !3
  %.phi.trans.insert28 = getelementptr inbounds [3 x [1 x float]], [3 x [1 x float]]* %25, i64 0, i64 2, i64 0
  %.pre29 = load float, float* %.phi.trans.insert28, align 8, !invariant.load !0, !noalias !3
  %28 = bitcast [3 x [1 x float]]* %25 to <2 x float>*
  %29 = load <2 x float>, <2 x float>* %28, align 8, !invariant.load !0, !noalias !3
  %30 = insertelement <2 x float> undef, float %27, i32 0
  %31 = shufflevector <2 x float> %30, <2 x float> undef, <2 x i32> zeroinitializer
  %32 = fsub <2 x float> %31, %29
  %33 = fmul <2 x float> %32, %32
  %shuffle30 = shufflevector <2 x float> %33, <2 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 1, i32 1, i32 1, i32 1>
  %34 = fsub float %27, %.pre29
  %35 = fmul float %34, %34
  %36 = insertelement <4 x float> undef, float %35, i32 0
  %37 = shufflevector <4 x float> %36, <4 x float> undef, <4 x i32> zeroinitializer
  %shuffle = shufflevector <4 x float> %10, <4 x float> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
  %38 = fmul <4 x float> %7, %7
  %shuffle31 = shufflevector <4 x float> %38, <4 x float> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
  %39 = fdiv <8 x float> %shuffle30, %shuffle31
  %40 = fadd <8 x float> %shuffle, %39
  %41 = fmul <8 x float> %40, <float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01>
  %42 = bitcast i8* %26 to <8 x float>*
  store <8 x float> %41, <8 x float>* %42, align 8, !alias.scope !8, !noalias !12
  %43 = getelementptr inbounds i8, i8* %26, i64 32
  %44 = fdiv <4 x float> %37, %38
  %45 = fadd <4 x float> %10, %44
  %46 = fmul <4 x float> %45, <float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01>
  %47 = bitcast i8* %43 to <4 x float>*
  store <4 x float> %46, <4 x float>* %47, align 8, !alias.scope !8, !noalias !12
  %.phi.trans.insert = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %5, i64 0, i64 1, i64 0, i64 0
  %.phi.trans.insert12 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %2, i64 0, i64 1, i64 0, i64 0
  %48 = bitcast float* %.phi.trans.insert to <4 x float>*
  %49 = load <4 x float>, <4 x float>* %48, align 8, !alias.scope !7, !noalias !8
  %50 = bitcast float* %.phi.trans.insert12 to <4 x float>*
  %51 = load <4 x float>, <4 x float>* %50, align 8, !invariant.load !0, !noalias !3
  %shuffle.1 = shufflevector <4 x float> %49, <4 x float> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
  %52 = getelementptr inbounds i8, i8* %26, i64 48
  %53 = fmul <4 x float> %51, %51
  %shuffle31.1 = shufflevector <4 x float> %53, <4 x float> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
  %54 = fdiv <8 x float> %shuffle30, %shuffle31.1
  %55 = fadd <8 x float> %shuffle.1, %54
  %56 = fmul <8 x float> %55, <float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01>
  %57 = bitcast i8* %52 to <8 x float>*
  store <8 x float> %56, <8 x float>* %57, align 8, !alias.scope !8, !noalias !12
  %58 = getelementptr inbounds i8, i8* %26, i64 80
  %59 = fdiv <4 x float> %37, %53
  %60 = fadd <4 x float> %49, %59
  %61 = fmul <4 x float> %60, <float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01>
  %62 = bitcast i8* %58 to <4 x float>*
  store <4 x float> %61, <4 x float>* %62, align 8, !alias.scope !8, !noalias !12
  %63 = getelementptr inbounds i8*, i8** %buffer_table, i64 3
  %64 = bitcast i8** %63 to [1 x i8*]**
  %65 = load [1 x i8*]*, [1 x i8*]** %64, align 8, !invariant.load !0, !dereferenceable !2, !align !2
  %66 = getelementptr inbounds [1 x i8*], [1 x i8*]* %65, i64 0, i64 0
  store i8* %26, i8** %66, align 8, !alias.scope !14, !noalias !8
  ret void
}

; Function Attrs: nounwind readnone speculatable willreturn
declare <4 x float> @llvm.log.v4f32(<4 x float>) #1

attributes #0 = { nofree nounwind uwtable "no-frame-pointer-elim"="false" }
attributes #1 = { nounwind readnone speculatable willreturn }

!0 = !{}
!1 = !{i64 32}
!2 = !{i64 8}
!3 = !{!4, !6}
!4 = !{!"buffer: {index:0, offset:0, size:96}", !5}
!5 = !{!"XLA global AA domain"}
!6 = !{!"buffer: {index:5, offset:0, size:32}", !5}
!7 = !{!6}
!8 = !{!4}
!9 = !{i64 4}
!10 = !{i64 12}
!11 = !{i64 96}
!12 = !{!13, !6}
!13 = !{!"buffer: {index:3, offset:0, size:8}", !5}
!14 = !{!13}

and (incorrectly) optimized to the one below with the change:

; ModuleID = '__compute_module'
source_filename = "__compute_module"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-grtev4-linux-gnu"

; Function Attrs: nofree nounwind uwtable
define void @jit_wrapped_fun.31(i8* nocapture readnone %retval, i8* noalias nocapture readnone %run_options, i8** noalias nocapture readnone %params, i8** noalias nocapture readonly %buffer_table, i64* noalias nocapture readnone %prof_counters) local_unnamed_addr #0 {
entry:
  %0 = getelementptr inbounds i8*, i8** %buffer_table, i64 1
  %1 = bitcast i8** %0 to [2 x [1 x [4 x float]]]**
  %2 = load [2 x [1 x [4 x float]]]*, [2 x [1 x [4 x float]]]** %1, align 8, !invariant.load !0, !dereferenceable !1, !align !2
  %3 = getelementptr inbounds i8*, i8** %buffer_table, i64 5
  %4 = bitcast i8** %3 to [2 x [1 x [4 x float]]]**
  %5 = load [2 x [1 x [4 x float]]]*, [2 x [1 x [4 x float]]]** %4, align 8, !invariant.load !0, !dereferenceable !1, !align !2
  %6 = bitcast [2 x [1 x [4 x float]]]* %2 to <4 x float>*
  %7 = load <4 x float>, <4 x float>* %6, align 8, !invariant.load !0, !noalias !3
  %8 = fmul <4 x float> %7, %7
  %9 = fmul <4 x float> %8, <float 0x401921FB60000000, float 0x401921FB60000000, float 0x401921FB60000000, float 0x401921FB60000000>
  %10 = call <4 x float> @llvm.log.v4f32(<4 x float> %9)
  %11 = bitcast [2 x [1 x [4 x float]]]* %5 to <4 x float>*
  store <4 x float> %10, <4 x float>* %11, align 8, !alias.scope !7, !noalias !8
  %12 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %2, i64 0, i64 1, i64 0, i64 0
  %13 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %5, i64 0, i64 1, i64 0, i64 0
  %14 = bitcast float* %12 to <4 x float>*
  %15 = load <4 x float>, <4 x float>* %14, align 8, !invariant.load !0, !noalias !3
  %16 = fmul <4 x float> %15, %15
  %17 = fmul <4 x float> %16, <float 0x401921FB60000000, float 0x401921FB60000000, float 0x401921FB60000000, float 0x401921FB60000000>
  %18 = call <4 x float> @llvm.log.v4f32(<4 x float> %17)
  %19 = bitcast float* %13 to <4 x float>*
  store <4 x float> %18, <4 x float>* %19, align 8, !alias.scope !7, !noalias !8
  %20 = getelementptr inbounds i8*, i8** %buffer_table, i64 4
  %21 = bitcast i8** %20 to float**
  %22 = load float*, float** %21, align 8, !invariant.load !0, !dereferenceable !9, !align !2
  %23 = getelementptr inbounds i8*, i8** %buffer_table, i64 2
  %24 = bitcast i8** %23 to [3 x [1 x float]]**
  %25 = load [3 x [1 x float]]*, [3 x [1 x float]]** %24, align 8, !invariant.load !0, !dereferenceable !10, !align !2
  %26 = load i8*, i8** %buffer_table, align 8, !invariant.load !0, !dereferenceable !11, !align !2
  %27 = load float, float* %22, align 8, !invariant.load !0, !noalias !3
  %.phi.trans.insert28 = getelementptr inbounds [3 x [1 x float]], [3 x [1 x float]]* %25, i64 0, i64 2, i64 0
  %.pre29 = load float, float* %.phi.trans.insert28, align 8, !invariant.load !0, !noalias !3
  %28 = bitcast [3 x [1 x float]]* %25 to <2 x float>*
  %29 = load <2 x float>, <2 x float>* %28, align 8, !invariant.load !0, !noalias !3
  %30 = insertelement <2 x float> undef, float %27, i32 0
  %31 = shufflevector <2 x float> %30, <2 x float> undef, <2 x i32> zeroinitializer
  %32 = fsub <2 x float> %31, %29
  %33 = fmul <2 x float> %32, %32
  %shuffle32 = shufflevector <2 x float> %33, <2 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 1, i32 1, i32 1, i32 1>
  %34 = fsub float %27, %.pre29
  %35 = fmul float %34, %34
  %36 = insertelement <4 x float> undef, float %35, i32 0
  %37 = shufflevector <4 x float> %36, <4 x float> undef, <4 x i32> zeroinitializer
  %shuffle = shufflevector <4 x float> %10, <4 x float> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
  %38 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %5, i64 0, i64 0, i64 0, i64 3
  %39 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %2, i64 0, i64 0, i64 0, i64 3
  %40 = fmul <4 x float> %7, %7
  %41 = shufflevector <4 x float> %40, <4 x float> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
  %42 = fdiv <8 x float> %shuffle32, %41
  %43 = fadd <8 x float> %shuffle, %42
  %44 = fmul <8 x float> %43, <float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01>
  %45 = bitcast i8* %26 to <8 x float>*
  store <8 x float> %44, <8 x float>* %45, align 8, !alias.scope !8, !noalias !12
  %46 = extractelement <4 x float> %10, i32 0
  %47 = getelementptr inbounds i8, i8* %26, i64 32
  %48 = extractelement <4 x float> %10, i32 1
  %49 = extractelement <4 x float> %10, i32 2
  %50 = load float, float* %38, align 4, !alias.scope !7, !noalias !8
  %51 = load float, float* %39, align 4, !invariant.load !0, !noalias !3
  %52 = fmul float %51, %51
  %53 = insertelement <4 x float> undef, float %52, i32 3
  %54 = fdiv <4 x float> %37, %53
  %55 = insertelement <4 x float> undef, float %46, i32 0
  %56 = insertelement <4 x float> %55, float %48, i32 1
  %57 = insertelement <4 x float> %56, float %49, i32 2
  %58 = insertelement <4 x float> %57, float %50, i32 3
  %59 = fadd <4 x float> %58, %54
  %60 = fmul <4 x float> %59, <float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01>
  %61 = bitcast i8* %47 to <4 x float>*
  store <4 x float> %60, <4 x float>* %61, align 8, !alias.scope !8, !noalias !12
  %.phi.trans.insert = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %5, i64 0, i64 1, i64 0, i64 0
  %.phi.trans.insert12 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %2, i64 0, i64 1, i64 0, i64 0
  %62 = bitcast float* %.phi.trans.insert to <4 x float>*
  %63 = load <4 x float>, <4 x float>* %62, align 8, !alias.scope !7, !noalias !8
  %64 = bitcast float* %.phi.trans.insert12 to <4 x float>*
  %65 = load <4 x float>, <4 x float>* %64, align 8, !invariant.load !0, !noalias !3
  %shuffle.1 = shufflevector <4 x float> %63, <4 x float> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
  %66 = getelementptr inbounds i8, i8* %26, i64 48
  %67 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %5, i64 0, i64 1, i64 0, i64 3
  %68 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* %2, i64 0, i64 1, i64 0, i64 3
  %69 = fmul <4 x float> %65, %65
  %70 = shufflevector <4 x float> %69, <4 x float> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
  %71 = fdiv <8 x float> %shuffle32, %70
  %72 = fadd <8 x float> %shuffle.1, %71
  %73 = fmul <8 x float> %72, <float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01>
  %74 = bitcast i8* %66 to <8 x float>*
  store <8 x float> %73, <8 x float>* %74, align 8, !alias.scope !8, !noalias !12
  %75 = extractelement <4 x float> %69, i32 0
  %76 = extractelement <4 x float> %63, i32 0
  %77 = getelementptr inbounds i8, i8* %26, i64 80
  %78 = extractelement <4 x float> %69, i32 1
  %79 = extractelement <4 x float> %63, i32 1
  %80 = extractelement <4 x float> %69, i32 2
  %81 = extractelement <4 x float> %63, i32 2
  %82 = load float, float* %67, align 4, !alias.scope !7, !noalias !8
  %83 = load float, float* %68, align 4, !invariant.load !0, !noalias !3
  %84 = fmul float %83, %83
  %85 = insertelement <4 x float> undef, float %75, i32 0
  %86 = insertelement <4 x float> %85, float %78, i32 1
  %87 = insertelement <4 x float> %86, float %80, i32 2
  %88 = insertelement <4 x float> %87, float %84, i32 3
  %89 = fdiv <4 x float> %37, %88
  %90 = insertelement <4 x float> undef, float %76, i32 0
  %91 = insertelement <4 x float> %90, float %79, i32 1
  %92 = insertelement <4 x float> %91, float %81, i32 2
  %93 = insertelement <4 x float> %92, float %82, i32 3
  %94 = fadd <4 x float> %93, %89
  %95 = fmul <4 x float> %94, <float -5.000000e-01, float -5.000000e-01, float -5.000000e-01, float -5.000000e-01>
  %96 = bitcast i8* %77 to <4 x float>*
  store <4 x float> %95, <4 x float>* %96, align 8, !alias.scope !8, !noalias !12
  %97 = getelementptr inbounds i8*, i8** %buffer_table, i64 3
  %98 = bitcast i8** %97 to [1 x i8*]**
  %99 = load [1 x i8*]*, [1 x i8*]** %98, align 8, !invariant.load !0, !dereferenceable !2, !align !2
  %100 = getelementptr inbounds [1 x i8*], [1 x i8*]* %99, i64 0, i64 0
  store i8* %26, i8** %100, align 8, !alias.scope !14, !noalias !8
  ret void
}

; Function Attrs: nounwind readnone speculatable willreturn
declare <4 x float> @llvm.log.v4f32(<4 x float>) #1

attributes #0 = { nofree nounwind uwtable "no-frame-pointer-elim"="false" }
attributes #1 = { nounwind readnone speculatable willreturn }

!0 = !{}
!1 = !{i64 32}
!2 = !{i64 8}
!3 = !{!4, !6}
!4 = !{!"buffer: {index:0, offset:0, size:96}", !5}
!5 = !{!"XLA global AA domain"}
!6 = !{!"buffer: {index:5, offset:0, size:32}", !5}
!7 = !{!6}
!8 = !{!4}
!9 = !{i64 4}
!10 = !{i64 12}
!11 = !{i64 96}
!12 = !{!13, !6}
!13 = !{!"buffer: {index:3, offset:0, size:8}", !5}
!14 = !{!13}

This results in bad numerical answers when used through XLA.
Again, it's not that easy to give a small fully-reproducible example, but the misscompare is:

Expected literal:
(
f32[2,3,4] {
{
  { nan, -inf, -3181.35, -inf },
  { nan, -inf, -28.2577019, -inf },
  { nan, -inf, -28.2577019, -inf }
},
{
  { -inf, -inf, -inf, -inf },
  { -6.60753046e+28, -1.47314833e+23, -inf, -inf },
  { -2.43504347e+30, -5.42892693e+24, -inf, -inf }
}
}
)

Actual literal:
(
f32[2,3,4] {
{
  { nan, -inf, -3181.35, -inf },
  { nan, -inf, -inf, -inf },
  { inf, -inf, -28.2577019, -inf }
},
{
  { -inf, -inf, -inf, -inf },
  { -6.60753046e+28, -1.47314833e+23, -inf, -inf },
  { -2.43504347e+30, -5.42892693e+24, -inf, -inf }
}
}
)

[george.karpenkov]

Would you be OK speculatively reverting this change?

[reames]

In D70516#1754275, @george.karpenkov wrote:

Would you be OK speculatively reverting this change?

If you're reasonable sure this is miscompiling, definitely revert. Help narrowing it down would be appreciated though.

[george.karpenkov]

If you're reasonable sure this is miscompiling, definitely revert. Help narrowing it down would be appreciated though.

Yes, we're seeing multiple different failures from this commit. Let me try to narrow it down to a smaller IR input.

[ebrevnov]

Looking at the provided IRs I can see what happened but I can't understand why. Provided IR is fine for me to continue investigation. I just need to know command line to reproduce. Interestingly, if I apply instcombine to initial IR (results are identical with and without my change) it is simplified to the form which I believe can't result in the provided "incorrect" version of IR.

[bkramer]

Pushed this in 446acafb82b5c116b6c94c11d4ac4db7641fa58d. The difference is visible with opt -O3, right? Or do you need an executable test (that's hard)

[ebrevnov]

In D70516#1754782, @bkramer wrote:

Pushed this in 446acafb82b5c116b6c94c11d4ac4db7641fa58d. The difference is visible with opt -O3, right? Or do you need an executable test (that's hard)

Yes, I was able to reproduce. Thanks!

llvm/lib/Analysis/MemoryDependenceAnalysis.cpp

 /// cached information in Cache or by doing a lookup (which may use dirty cache
 /// info if available).
 ///
 /// If we do a lookup, add the result to the cache.
 MemDepResult MemoryDependenceResults::GetNonLocalInfoForBlock(
     Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad,
     BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
 
-  bool isInvariantLoad = false;
-
-  if (LoadInst *LI = dyn_cast_or_null<LoadInst>(QueryInst))
-    isInvariantLoad = LI->getMetadata(LLVMContext::MD_invariant_load);
-
   // Do a binary search to see if we already have an entry for this block in
   // the cache set.  If so, find it.
   NonLocalDepInfo::iterator Entry = std::upper_bound(
       Cache->begin(), Cache->begin() + NumSortedEntries, NonLocalDepEntry(BB));
   if (Entry != Cache->begin() && (Entry - 1)->getBB() == BB)
     --Entry;
 
   NonLocalDepEntry *ExistingResult = nullptr;
   if (Entry != Cache->begin() + NumSortedEntries && Entry->getBB() == BB)
     ExistingResult = &*Entry;
 
-  // Use cached result for invariant load only if there is no dependency for non
-  // invariant load. In this case invariant load can not have any dependency as
-  // well.
-  if (ExistingResult && isInvariantLoad &&
-      !ExistingResult->getResult().isNonFuncLocal())
-    ExistingResult = nullptr;
-
   // If we have a cached entry, and it is non-dirty, use it as the value for
   // this dependency.
   if (ExistingResult && !ExistingResult->getResult().isDirty()) {
     ++NumCacheNonLocalPtr;
     return ExistingResult->getResult();
   }
 
   // Otherwise, we have to scan for the value.  If we have a dirty cache
MemDepResult MemoryDependenceResults::GetNonLocalInfoForBlock(
   } else {
     ++NumUncacheNonLocalPtr;
   }
 
   // Scan the block for the dependency.
   MemDepResult Dep =
       getPointerDependencyFrom(Loc, isLoad, ScanPos, BB, QueryInst);
 
-  // Don't cache results for invariant load.
-  if (isInvariantLoad)
-    return Dep;
-
   // If we had a dirty entry for the block, update it.  Otherwise, just add
   // a new entry.
   if (ExistingResult)
     ExistingResult->setResult(Dep);
   else
     Cache->push_back(NonLocalDepEntry(BB, Dep));
 
   // If the block has a dependency (i.e. it isn't completely transparent to
PredTranslationFailure:
     // If *nothing* works, mark the pointer as unknown.
     //
     // If this is the magic first block, return this as a clobber of the whole
     // incoming value.  Since we can't phi translate to one of the predecessors,
     // we have to bail out.
     if (SkipFirstBlock)
       return false;
 
+    bool foundBlock = false;
     for (NonLocalDepEntry &I : llvm::reverse(*Cache)) {
       if (I.getBB() != BB)
         continue;
 
       assert((GotWorklistLimit || I.getResult().isNonLocal() ||
               !DT.isReachableFromEntry(BB)) &&
              "Should only be here with transparent block");
+      foundBlock = true;
       I.setResult(MemDepResult::getUnknown());
+      Result.push_back(
+          NonLocalDepResult(I.getBB(), I.getResult(), Pointer.getAddr()));
       break;
     }
-    // Go ahead and report unknown dependence.
-    Result.push_back(
-        NonLocalDepResult(BB, MemDepResult::getUnknown(), Pointer.getAddr()));
+    (void)foundBlock; (void)GotWorklistLimit;
+    assert((foundBlock || GotWorklistLimit) && "Current block not in cache?");
   }
 
   // Okay, we're done now.  If we added new values to the cache, re-sort it.
   SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
   LLVM_DEBUG(AssertSorted(*Cache));
   return true;
 }