[SimplifyCFG] Added condition assumption for unreachable blocks

[xbolva00]

Summary

PR41688

[xbolva00]

Update old tests

[xbolva00]

This comment was removed by xbolva00.

[lebedev.ri]

Please precommit the check-line regeneration

[nikic]

This is a good idea in theory, but I'm somewhat concerned that this will backfire in practice because assumes often pessimize optimization. For example LLVM has a really bad offender where it inserts assumes for alignment annotations during inlining -- thankfully there's an option to disable it, because it's a major perf regression otherwise.

This feels like it should be less problematic though.

[lebedev.ri]

In D61409#1487366, @nikic wrote:

This is a good idea in theory, but I'm somewhat concerned that this will backfire in practice because assumes often pessimize optimization. For example LLVM has a really bad offender where it inserts assumes for alignment annotations during inlining -- thankfully there's an option to disable it, because it's a major perf regression otherwise.

Pessimization of the final produced code, or of the llvm itself?

This feels like it should be less problematic though.

[nikic]

In D61409#1487372, @lebedev.ri wrote:

In D61409#1487366, @nikic wrote:

This is a good idea in theory, but I'm somewhat concerned that this will backfire in practice because assumes often pessimize optimization. For example LLVM has a really bad offender where it inserts assumes for alignment annotations during inlining -- thankfully there's an option to disable it, because it's a major perf regression otherwise.

Pessimization of the final produced code, or of the llvm itself?

Of the final produced code. I think it's mainly because assumes cause one-use checks to fail.

[lebedev.ri]

In D61409#1487375, @nikic wrote:

In D61409#1487372, @lebedev.ri wrote:

In D61409#1487366, @nikic wrote:

This is a good idea in theory, but I'm somewhat concerned that this will backfire in practice because assumes often pessimize optimization. For example LLVM has a really bad offender where it inserts assumes for alignment annotations during inlining -- thankfully there's an option to disable it, because it's a major perf regression otherwise.

Pessimization of the final produced code, or of the llvm itself?

Of the final produced code. I think it's mainly because assumes cause one-use checks to fail.

Oh, good point. It would make sense to not count @llvm.assume() during use-checking,
but then this isn't just about the call void @llvm.assume(), but about all the instructions
that would be dropped when the call is dropped during back-end lowering...

[xbolva00]

Since llvm.assume is not a really true use, oneUse check fails? Then we should really fix it to not count it.

[xbolva00]

It would be good if anybody familiar with llvm patternmatch could fix m_OneUse to ignore llvm.assume.

Then this patch will not cause any problems or regressions..

@spatel @RKSimon

[lebedev.ri]

In D61409#1488239, @xbolva00 wrote:

It would be good if anybody familiar with llvm patternmatch could fix m_OneUse to ignore llvm.assume.

In D61409#1487377, @lebedev.ri wrote:

In D61409#1487375, @nikic wrote:

In D61409#1487372, @lebedev.ri wrote:

In D61409#1487366, @nikic wrote:

This is a good idea in theory, but I'm somewhat concerned that this will backfire in practice because assumes often pessimize optimization. For example LLVM has a really bad offender where it inserts assumes for alignment annotations during inlining -- thankfully there's an option to disable it, because it's a major perf regression otherwise.

Pessimization of the final produced code, or of the llvm itself?

Of the final produced code. I think it's mainly because assumes cause one-use checks to fail.

Oh, good point. It would make sense to not count @llvm.assume() during use-checking,
but then this isn't just about the call void @llvm.assume(), but about all the instructions
that would be dropped when the call is dropped during back-end lowering...

I.e. just ignoring only the @llvm.assume() itself is only tip of the iceberg.

In D61409#1488239, @xbolva00 wrote:

Then this patch will not cause any problems or regressions..

@spatel @RKSimon

[spatel]

In D61409#1488242, @lebedev.ri wrote:

In D61409#1488239, @xbolva00 wrote:

It would be good if anybody familiar with llvm patternmatch could fix m_OneUse to ignore llvm.assume.

In D61409#1487377, @lebedev.ri wrote:

In D61409#1487375, @nikic wrote:

In D61409#1487372, @lebedev.ri wrote:

In D61409#1487366, @nikic wrote:

This is a good idea in theory, but I'm somewhat concerned that this will backfire in practice because assumes often pessimize optimization. For example LLVM has a really bad offender where it inserts assumes for alignment annotations during inlining -- thankfully there's an option to disable it, because it's a major perf regression otherwise.

Pessimization of the final produced code, or of the llvm itself?

Of the final produced code. I think it's mainly because assumes cause one-use checks to fail.

Oh, good point. It would make sense to not count @llvm.assume() during use-checking,
but then this isn't just about the call void @llvm.assume(), but about all the instructions
that would be dropped when the call is dropped during back-end lowering...

I.e. just ignoring only the @llvm.assume() itself is only tip of the iceberg.

In D61409#1488239, @xbolva00 wrote:

Then this patch will not cause any problems or regressions..

@spatel @RKSimon

I agree that this is a (much) bigger than fixing the pattern matcher. I have no instinct or data to say how much inserted assumes can/do harm optimization, but this is a long-standing problem with that implementation. @hfinkel or @reames may have a better idea.

[xbolva00]

Any advices about next steps for this patch?

if inserted llvm.assume is a big problem and “fixing” oneUse match check is not proper solution, then this patch is dead..

[hfinkel]

In D61409#1490577, @xbolva00 wrote:

Any advices about next steps for this patch?

if inserted llvm.assume is a big problem and “fixing” oneUse match check is not proper solution, then this patch is dead..

In general, automatically llvm.assume much be done with extreme care, as it can pessimize optimizations. We try do this only in cases where the user as specifically indicated that there is some information relevant to optimization that is worth preserving. In this case, it might be worth doing, because the if (x) unreachable is the recommended "assume" pattern for use with GCC.

[lebedev.ri]

In D61409#1490638, @hfinkel wrote:

In D61409#1490577, @xbolva00 wrote:

Any advices about next steps for this patch?

if inserted llvm.assume is a big problem and “fixing” oneUse match check is not proper solution, then this patch is dead..

In general, automatically llvm.assume much be done with extreme care, as it can pessimize optimizations. We try do this only in cases where the user as specifically indicated that there is some information relevant to optimization that is worth preserving.

In D61409#1490638, @hfinkel wrote:

In this case, it might be worth doing, because the if (x) unreachable is the recommended "assume" pattern for use with GCC.

It looked this patch was moving in the right direction?

[xbolva00]

Possibly. But more fixes are needed as preconditions to enable this, as you mentioned here (fix oneUse check). Without them, we cannot push this futher.

[nikic]

Just to be clear, I didn't mean to say that we shouldn't do this, just that it is not guaranteed to be a win in the end. If it's possible to do some end-to-end performance testing with this patch to make sure that it doesn't come with unexpected regressions, I see no reason not to move forward with this change. The problem with assumes is a long-standing one and I don't think it has any simple solution.

[lebedev.ri]

In D61409#1503626, @nikic wrote:

Just to be clear, I didn't mean to say that we shouldn't do this, just that it is not guaranteed to be a win in the end. If it's possible to do some end-to-end performance testing with this patch to make sure that it doesn't come with unexpected regressions, I see no reason not to move forward with this change. The problem with assumes is a long-standing one and I don't think it has any simple solution.

That was my take, too.

[xbolva00]

$ ../test-suite/utils/compare.py baseline.json
Tests: 1163
Metric: exec_time

Program baseline
LCALS/Subs...aw.test:BM_MAT_X_MAT_RAW/44217 868459.97
LCALS/Subs...test:BM_MAT_X_MAT_LAMBDA/44217 563058.65
ImageProce...t:BENCHMARK_GAUSSIAN_BLUR/1024 99965.67
ImageProce...HMARK_ANISTROPIC_DIFFUSION/256 77470.40
harris/har...est:BENCHMARK_HARRIS/2048/2048 52498.97
ImageProce...MARK_BICUBIC_INTERPOLATION/256 44359.86
LCALS/Subs...Raw.test:BM_MAT_X_MAT_RAW/5001 25499.45
ImageProce...st:BENCHMARK_GAUSSIAN_BLUR/512 24021.84
LCALS/Subs....test:BM_MAT_X_MAT_LAMBDA/5001 20025.00
ImageProce...HMARK_ANISTROPIC_DIFFUSION/128 18171.17
harris/har...est:BENCHMARK_HARRIS/1024/1024 15093.33
ImageProce...ate.test:BENCHMARK_DILATE/1024 12321.46
ImageProce...ARK_BILINEAR_INTERPOLATION/256 11589.59
ImageProce...MARK_BICUBIC_INTERPOLATION/128 10496.46
LCALS/Subs...Raw.test:BM_HYDRO_2D_RAW/44217 8448.96

baseline

count 1147.000000
mean 1727.619647
std 30862.929090
min 0.000000
25% 0.004000
50% 0.008800
75% 3.645378
max 868459.973000
$ ../test-suite/utils/compare.py patch.json
Tests: 1163
Metric: exec_time

Program patch
LCALS/Subs...test:BM_MAT_X_MAT_LAMBDA/44217 648771.10
LCALS/Subs...aw.test:BM_MAT_X_MAT_RAW/44217 549594.39
ImageProce...t:BENCHMARK_GAUSSIAN_BLUR/1024 112422.28
ImageProce...HMARK_ANISTROPIC_DIFFUSION/256 91073.89
harris/har...est:BENCHMARK_HARRIS/2048/2048 52924.20
ImageProce...MARK_BICUBIC_INTERPOLATION/256 50989.07
ImageProce...st:BENCHMARK_GAUSSIAN_BLUR/512 26065.95
LCALS/Subs...Raw.test:BM_MAT_X_MAT_RAW/5001 23588.53
ImageProce...HMARK_ANISTROPIC_DIFFUSION/128 21120.28
LCALS/Subs....test:BM_MAT_X_MAT_LAMBDA/5001 20901.71
ImageProce...ARK_BILINEAR_INTERPOLATION/256 15573.54
harris/har...est:BENCHMARK_HARRIS/1024/1024 14383.34
ImageProce...ate.test:BENCHMARK_DILATE/1024 12928.17
ImageProce...MARK_BICUBIC_INTERPOLATION/128 10768.69
LCALS/Subs...Raw.test:BM_HYDRO_2D_RAW/44217 9542.02

patch

count 1147.000000
mean 1571.500067
std 25580.275818
min 0.000000
25% 0.004000
50% 0.008900
75% 3.815259
max 648771.096000

[fhahn]

You can use something like the command below to get a comparison between 2 sets of runs. Also, doing single runs only will probably result in quite noisy results.

./test-suite/utils/compare.py -m exec_time patch.json vs master.json

[xbolva00]

In D61409#1507829, @fhahn wrote:

You can use something like the command below to get a comparison between 2 sets of runs. Also, doing single runs only will probably result in quite noisy results.

./test-suite/utils/compare.py -m exec_time patch.json vs master.json

Not sure why this crashes for me:
/home/xbolva00/.local/lib/python2.7/site-packages/scipy/stats/stats.py:308: RuntimeWarning: divide by zero encountered in log

log_a = np.log(np.array(a, dtype=dtype))

/home/xbolva00/.local/lib/python2.7/site-packages/numpy/core/_methods.py:75: RuntimeWarning: invalid value encountered in reduce

ret = umr_sum(arr, axis, dtype, out, keepdims)

Anway, I used --filter-short.. I ran and compare it 3x..
https://pastebin.com/kjFBv2ZP

Yes, noise a bit but I think there are some stable improvements..

[xbolva00]

Ping

I thinks the results are okay..

[nikic]

These results look pretty noisy, so it's hard to say. Maybe someone who already has a good benchmarking setup (maybe @dmgreen?) could run some tests?

[nikic]

Independently of that, could you please precommit your new test, as well as the regenerated test checks? Especially for the LoopVectorize/if-pred-stores.ll test it's not clear what has actually changed.

[xbolva00]

Rebased on updated test checks..

Some tests are failing
Failing Tests (4):

LLVM :: CodeGen/Hexagon/bit-visit-flowq.ll
LLVM :: CodeGen/Hexagon/rdf-ignore-undef.ll
LLVM :: CodeGen/Hexagon/reg-scavengebug.ll
LLVM :: CodeGen/Hexagon/regalloc-block-overlap.ll

The root cause: tests contain "br undef" and this is now optimized away. Not sure how to progress here..

[nikic]

The root cause: tests contain "br undef" and this is now optimized away. Not sure how to progress here..

This is a common problem with bugpoint reduced tests. You'll have to update these tests (as well as if-pred-stores.ll) either by replacing unreachables with something else (say a dummy return), or replacing the undef branch with a proper condition (e.g. pass in a bool as arg), while trying to preserve the original intention of the test.

[xbolva00]

Rebased

[hfinkel]

This LGTM. Please watch the LNT bots carefully and revert if we find performance regressions. A fear of performance regressions is specifically why we didn't do this when the assumption facility was first introduced.

[xbolva00]

Thanks!

I will wait a few days in case @dmgreen wants to do more benchmarks.

[dmgreen]

Hello. Looks fine from the tests I ran. For codesize too, which can be a good way to get low-noise indication of how things change.

[xbolva00]

Thank you!

Commited.

Please also see https://bugs.llvm.org/show_bug.cgi?id=42019 if you have any ideas how to handle intrinsics in use-checking code.

lib/Transforms/Utils/SimplifyCFG.cpp

for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
     if (auto *BI = dyn_cast<BranchInst>(TI)) {
       if (BI->isUnconditional()) {
         if (BI->getSuccessor(0) == BB) {
           new UnreachableInst(TI->getContext(), TI);
           TI->eraseFromParent();
           Changed = true;
         }
       } else {
+        Value* Cond = BI->getCondition();
         if (BI->getSuccessor(0) == BB) {
+          Builder.CreateAssumption(Builder.CreateNot(Cond));
           Builder.CreateBr(BI->getSuccessor(1));
           EraseTerminatorAndDCECond(BI);
         } else if (BI->getSuccessor(1) == BB) {
+          Builder.CreateAssumption(Cond);
           Builder.CreateBr(BI->getSuccessor(0));
           EraseTerminatorAndDCECond(BI);
           Changed = true;
         }
       }
     } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
       for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
         if (i->getCaseSuccessor() != BB) {

test/Analysis/ValueTracking/select-pattern.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -simplifycfg < %s -S | FileCheck %s
 
 ; The dead code would cause a select that had itself
 ; as an operand to be analyzed. This would then cause
 ; infinite recursion and eventual crash.
 
 define void @PR36045(i1 %t, i32* %b) {
 ; CHECK-LABEL: @PR36045(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[T:%.*]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br i1 %t, label %if, label %end
 
 if:
   br i1 %t, label %unreach, label %pre
 

test/CodeGen/ARM/crash-greedy.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -regalloc=greedy -mcpu=cortex-a8 -relocation-model=pic -frame-pointer=all -verify-machineinstrs | FileCheck %s
 ;
 ; ARM tests that crash or fail with the greedy register allocator.
 
 target triple = "thumbv7-apple-darwin"
 
 declare double @exp(double)
 

 
 for.end:                                          ; preds = %cond.end
   %conv78 = sitofp i32 %z to double
   %conv83 = fpext float %c to double
   %mul84 = fmul double %mul17, %conv83
   %call85 = tail call double @exp(double %mul84) nounwind
   %mul86 = fmul double %conv78, %call85
   %add88 = fadd double 0.000000e+00, %mul86
-; CHECK: bl _exp
   %call100 = tail call double @exp(double %mul84) nounwind
   %mul101 = fmul double undef, %call100
   %add103 = fadd double %add46, %mul101
   %mul111 = fmul double undef, %conv83
   %mul119 = fmul double %mul111, undef
   %add121 = fadd double undef, %mul119
   %div = fdiv double 1.000000e+00, %conv16
   %div126 = fdiv double %add, %add75
   %sub = fsub double %div, %div126
   %div129 = fdiv double %add103, %add88
   %add130 = fadd double %sub, %div129
   %conv131 = fptrunc double %add130 to float
   store float %conv131, float* %ret_f, align 4
   %mul139 = fmul double %div129, %div129
   %div142 = fdiv double %add121, %add88
   %sub143 = fsub double %mul139, %div142
 ; %lambda is passed on the stack, and the stack slot load is rematerialized.
 ; The rematted load of a float constrains the D register used for the mul.
-; CHECK: vldr
   %mul146 = fmul float %lambda, %lambda
   %conv147 = fpext float %mul146 to double
   %div148 = fdiv double 1.000000e+00, %conv147
   %sub149 = fsub double %sub143, %div148
   %conv150 = fptrunc double %sub149 to float
   store float %conv150, float* %ret_df, align 4
   ret void
 }

test/Transforms/CallSiteSplitting/split-loop.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -S -callsite-splitting -simplifycfg < %s | FileCheck %s
 
 define i16 @test1() {
 ; CHECK-LABEL: @test1(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[SPEC_SELECT:%.*]] = select i1 undef, i16 1, i16 0
+; CHECK-NEXT:    [[TOBOOL18:%.*]] = icmp ne i16 [[SPEC_SELECT]], 0
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[TOBOOL18]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    br label [[FOR_COND12:%.*]]
 ; CHECK:       for.cond12:
 ; CHECK-NEXT:    call void @callee(i16 [[SPEC_SELECT]])
 ; CHECK-NEXT:    br label [[FOR_COND12]]
 ;
 entry:
   %spec.select = select i1 undef, i16 1, i16 0
   %tobool18 = icmp ne i16 %spec.select, 0
   br i1 %tobool18, label %for.cond12.us, label %for.cond12
 
 for.cond12.us:
   unreachable
 
 for.cond12:
   call void @callee(i16 %spec.select)
   br label %for.cond12
 }
 
 define i16 @test2() {
 ; CHECK-LABEL: @test2(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[S:%.*]] = select i1 undef, i16 1, i16 0
+; CHECK-NEXT:    [[TOBOOL18:%.*]] = icmp ne i16 [[S]], 0
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[TOBOOL18]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    br label [[FOR_COND12:%.*]]
 ; CHECK:       for.cond12:
 ; CHECK-NEXT:    call void @callee(i16 [[S]])
 ; CHECK-NEXT:    [[ADD:%.*]] = add i16 [[S]], 10
 ; CHECK-NEXT:    [[ADD2:%.*]] = add i16 [[S]], 10
 ; CHECK-NEXT:    br label [[FOR_COND12]]
 ;
 entry:
for.cond12:
   %add2 = add i16 %s, 10
   br label %for.cond12
 }
 
 define i16 @test3() {
 ; CHECK-LABEL: @test3(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[S:%.*]] = select i1 undef, i16 1, i16 0
+; CHECK-NEXT:    [[TOBOOL18:%.*]] = icmp ne i16 [[S]], 0
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[TOBOOL18]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    br label [[FOR_COND12:%.*]]
 ; CHECK:       for.cond12:
 ; CHECK-NEXT:    call void @callee(i16 [[S]])
 ; CHECK-NEXT:    [[ADD:%.*]] = add i16 [[S]], 10
 ; CHECK-NEXT:    [[ADD2:%.*]] = add i16 [[ADD]], 10
 ; CHECK-NEXT:    br i1 undef, label [[FOR_COND12]], label [[EXIT:%.*]]
 ; CHECK:       exit:
 ; CHECK-NEXT:    ret i16 [[ADD2]]

test/Transforms/LoopVectorize/if-pred-stores.ll

 ; Track basic blocks when unrolling conditional blocks. This code used to assert
 ; because we did not update the phi nodes with the proper predecessor in the
 ; vectorized loop body.
 ; PR18724
 
 define void @bug18724() {
 ; UNROLL-LABEL: @bug18724(
 ; UNROLL-NEXT:  entry:
-; UNROLL-NEXT:    br label [[FOR_BODY14:%.*]]
-; UNROLL:       for.body14:
-; UNROLL-NEXT:    [[INDVARS_IV3:%.*]] = phi i64 [ [[INDVARS_IV_NEXT4:%.*]], [[FOR_INC23:%.*]] ], [ undef, [[ENTRY:%.*]] ]
-; UNROLL-NEXT:    [[INEWCHUNKS_120:%.*]] = phi i32 [ [[INEWCHUNKS_2:%.*]], [[FOR_INC23]] ], [ undef, [[ENTRY]] ]
-; UNROLL-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds [768 x i32], [768 x i32]* undef, i64 0, i64 [[INDVARS_IV3]]
-; UNROLL-NEXT:    [[TMP:%.*]] = load i32, i32* [[ARRAYIDX16]], align 4
-; UNROLL-NEXT:    br i1 undef, label [[IF_THEN18:%.*]], label [[FOR_INC23]]
-; UNROLL:       if.then18:
-; UNROLL-NEXT:    store i32 2, i32* [[ARRAYIDX16]], align 4
-; UNROLL-NEXT:    [[INC21:%.*]] = add nsw i32 [[INEWCHUNKS_120]], 1
-; UNROLL-NEXT:    br label [[FOR_INC23]]
-; UNROLL:       for.inc23:
-; UNROLL-NEXT:    [[INEWCHUNKS_2]] = phi i32 [ [[INC21]], [[IF_THEN18]] ], [ [[INEWCHUNKS_120]], [[FOR_BODY14]] ]
-; UNROLL-NEXT:    [[INDVARS_IV_NEXT4]] = add nsw i64 [[INDVARS_IV3]], 1
-; UNROLL-NEXT:    br label [[FOR_BODY14]]
+; UNROLL-NEXT:    unreachable
 ;
 ; UNROLL-NOSIMPLIFY-LABEL: @bug18724(
 ; UNROLL-NOSIMPLIFY-NEXT:  entry:
 ; UNROLL-NOSIMPLIFY-NEXT:    br label [[FOR_BODY9:%.*]]
 ; UNROLL-NOSIMPLIFY:       for.body9:
 ; UNROLL-NOSIMPLIFY-NEXT:    br i1 undef, label [[FOR_INC26:%.*]], label [[FOR_BODY14_PREHEADER:%.*]]
 ; UNROLL-NOSIMPLIFY:       for.body14.preheader:
 ; UNROLL-NOSIMPLIFY-NEXT:    br i1 true, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]

 ; UNROLL-NOSIMPLIFY-NEXT:    [[INEWCHUNKS_2_LCSSA:%.*]] = phi i32 [ [[INEWCHUNKS_2]], [[FOR_INC23]] ], [ [[BIN_RDX]], [[MIDDLE_BLOCK]] ]
 ; UNROLL-NOSIMPLIFY-NEXT:    br label [[FOR_INC26]]
 ; UNROLL-NOSIMPLIFY:       for.inc26:
 ; UNROLL-NOSIMPLIFY-NEXT:    [[INEWCHUNKS_1_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY9]] ], [ [[INEWCHUNKS_2_LCSSA]], [[FOR_INC26_LOOPEXIT]] ]
 ; UNROLL-NOSIMPLIFY-NEXT:    unreachable
 ;
 ; VEC-LABEL: @bug18724(
 ; VEC-NEXT:  entry:
-; VEC-NEXT:    br label [[FOR_BODY14:%.*]]
-; VEC:       for.body14:
-; VEC-NEXT:    [[INDVARS_IV3:%.*]] = phi i64 [ [[INDVARS_IV_NEXT4:%.*]], [[FOR_INC23:%.*]] ], [ undef, [[ENTRY:%.*]] ]
-; VEC-NEXT:    [[INEWCHUNKS_120:%.*]] = phi i32 [ [[INEWCHUNKS_2:%.*]], [[FOR_INC23]] ], [ undef, [[ENTRY]] ]
-; VEC-NEXT:    [[ARRAYIDX16:%.*]] = getelementptr inbounds [768 x i32], [768 x i32]* undef, i64 0, i64 [[INDVARS_IV3]]
-; VEC-NEXT:    [[TMP:%.*]] = load i32, i32* [[ARRAYIDX16]], align 4
-; VEC-NEXT:    br i1 undef, label [[IF_THEN18:%.*]], label [[FOR_INC23]]
-; VEC:       if.then18:
-; VEC-NEXT:    store i32 2, i32* [[ARRAYIDX16]], align 4
-; VEC-NEXT:    [[INC21:%.*]] = add nsw i32 [[INEWCHUNKS_120]], 1
-; VEC-NEXT:    br label [[FOR_INC23]]
-; VEC:       for.inc23:
-; VEC-NEXT:    [[INEWCHUNKS_2]] = phi i32 [ [[INC21]], [[IF_THEN18]] ], [ [[INEWCHUNKS_120]], [[FOR_BODY14]] ]
-; VEC-NEXT:    [[INDVARS_IV_NEXT4]] = add nsw i64 [[INDVARS_IV3]], 1
-; VEC-NEXT:    br label [[FOR_BODY14]]
+; VEC-NEXT:    unreachable
 ;
 entry:
   br label %for.body9
 
 for.body9:
   br i1 undef, label %for.inc26, label %for.body14
 
 for.body14:

test/Transforms/SimplifyCFG/PR30210.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -S -simplifycfg < %s | FileCheck %s
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 declare i32* @fn1(i32* returned)
 
 define i32 @test1(i1 %B) {
 ; CHECK-LABEL: @test1(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br label [[FOR_COND_US:%.*]]
 ; CHECK:       for.cond.us:
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[B:%.*]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    br label [[FOR_COND_US]]
 ;
 entry:
   br label %for.cond.us
 
 for.cond.us:                                      ; preds = %for.cond.us, %entry
   br i1 %B, label %for.cond4.preheader, label %for.cond.us
 
for.cond4:                                        ; preds = %for.end, %for.cond4.preheader
   br label %for.cond5
 
 for.cond5:                                        ; preds = %for.cond5, %for.cond4
   br i1 %B, label %for.cond5, label %for.end
 
 for.end:                                          ; preds = %for.cond5
   %load = load i32, i32* %call, align 4
   br label %for.cond4
-}
+}
 

test/Transforms/SimplifyCFG/UnreachableEliminate.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -simplifycfg -S | FileCheck %s
 
 define void @test1(i1 %C, i1* %BP) {
 ; CHECK-LABEL: @test1(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[C:%.*]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br i1 %C, label %T, label %F
 T:
   store i1 %C, i1* %BP
   unreachable
 F:

 
 ;; We can either convert the following control-flow to a select or remove the
 ;; unreachable control flow because of the undef store of null. Make sure we do
 ;; the latter.
 
 define void @test5(i1 %cond, i8* %ptr) {
 ; CHECK-LABEL: @test5(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[COND:%.*]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    store i8 2, i8* [[PTR:%.*]], align 8
 ; CHECK-NEXT:    ret void
 ;
 
 entry:
   br i1 %cond, label %bb1, label %bb3
 
 bb3:
bb2:
   %ptr.2 = phi i8* [ %ptr, %bb3 ], [ null, %bb1 ]
   store i8 2, i8* %ptr.2, align 8
   ret void
 }
 
 define void @test6(i1 %cond, i8* %ptr) {
 ; CHECK-LABEL: @test6(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[COND:%.*]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    store i8 2, i8* [[PTR:%.*]], align 8
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br i1 %cond, label %bb1, label %bb2
 
 bb1:
   br label %bb2
bb2:
   store i8 2, i8* %ptr.2, align 8
   ret void
 }
 
 
 define i32 @test7(i1 %X) {
 ; CHECK-LABEL: @test7(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[X:%.*]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    ret i32 0
 ;
 entry:
   br i1 %X, label %if, label %else
 
 if:
   call void undef()
   br label %else
 
 else:
   %phi = phi i32 [ 0, %entry ], [ 1, %if ]
   ret i32 %phi
 }
 
 define void @test8(i1 %X, void ()* %Y) {
 ; CHECK-LABEL: @test8(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[X:%.*]], true
+; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP0]])
 ; CHECK-NEXT:    call void [[Y:%.*]]()
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br i1 %X, label %if, label %else
 
 if:
   br label %else
if:
   br label %else
 
 else:
   %phi = phi void ()* [ %Y, %entry ], [ null, %if ]
   call void %phi()
   ret void
 }
 
-attributes #0 = { "null-pointer-is-valid"="true" }
+attributes #0 = { "null-pointer-is-valid"="true" }
 

test/Transforms/SimplifyCFG/unreachable_assume.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt %s -simplifycfg -instcombine -S | FileCheck %s
 
-; TODO: ABS call should be optimized away
 define i32 @assume1(i32 %p) {
 ; CHECK-LABEL: @assume1(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[TMP0:%.*]] = icmp slt i32 [[P:%.*]], 0
-; CHECK-NEXT:    [[NEG:%.*]] = sub nsw i32 0, [[P]]
-; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[TMP0]], i32 [[NEG]], i32 [[P]]
-; CHECK-NEXT:    ret i32 [[TMP1]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[P:%.*]], 0
+; CHECK-NEXT:    call void @llvm.assume(i1 [[CMP]])
+; CHECK-NEXT:    ret i32 [[P]]
 ;
 entry:
   %cmp = icmp sle i32 %p, 0
   br i1 %cmp, label %if.then, label %if.end
 
 if.then:
   unreachable
 
 if.end:
   %call = call i32 @abs(i32 %p)
   ret i32 %call
 }
 
 
 define i32 @assume2(i32 %p) {
 ; CHECK-LABEL: @assume2(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[TMP0:%.*]] = icmp slt i32 [[P:%.*]], 0
-; CHECK-NEXT:    [[NEG:%.*]] = sub nsw i32 0, [[P]]
-; CHECK-NEXT:    [[TMP1:%.*]] = select i1 [[TMP0]], i32 [[NEG]], i32 [[P]]
-; CHECK-NEXT:    ret i32 [[TMP1]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[P:%.*]], 0
+; CHECK-NEXT:    call void @llvm.assume(i1 [[CMP]])
+; CHECK-NEXT:    ret i32 [[P]]
 ;
 entry:
   %cmp = icmp sgt i32 %p, 0
   br i1 %cmp, label %if.then, label %if.else
 
 if.then:
   br label %if.end