<div dir="ltr"><div dir="ltr"><br></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Wed, Sep 9, 2020 at 3:50 AM Roman Lebedev <<a href="mailto:lebedev.ri@gmail.com">lebedev.ri@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">On Wed, Sep 9, 2020 at 6:19 AM Eric Christopher <<a href="mailto:echristo@gmail.com" target="_blank">echristo@gmail.com</a>> wrote:<br>
><br>
> Hi Roman,<br>
Hi.<br>
<br>
> This seems like a pretty good idea,<br>
<br>
> but the change to EarlyCSE seems to be necessary before disabling a bunch of optimization?<br>
> Or am I misreading the comment in the commit?<br>
I really don't see what you mean.<br>
Nothing in that commit message talks about the need for an EarlyCSE change,<br>
or about "disabling a bunch of optimization". What it does talk about,<br>
is that we may want an InstCombine-based PHI CSE, and we got that<br>
in <a href="https://reviews.llvm.org/rGbf21ce7b908e16db76f5704714ee83e61a9da95b" rel="noreferrer" target="_blank">https://reviews.llvm.org/rGbf21ce7b908e16db76f5704714ee83e61a9da95b</a><br>
<br>
> I've managed to find a couple of regressions in tcmalloc with this and was wondering if you'd seen performance or just compile time improvements?<br>
Note this patch isn't a standalone "let's have less `extractvalue`e",<br>
but it belongs to the series of folds designed to ensure that as many<br>
`invoke`s as possible are no longer `invoke`s, but are converted into `call`s,<br>
which, i think you'll agree, generally allows much more optimizations.<br>
<br></blockquote><div><br></div><div>Agreed, but only if you have invokes :) Most of my code is compiled with -fno-exceptions.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
That end goal is almost a correctness thing, while the patch is just<br>
an exploratory step. As i've since established with certainty<br>
(<a href="https://reviews.llvm.org/D86530#2247674" rel="noreferrer" target="_blank">https://reviews.llvm.org/D86530#2247674</a>), i'll need a more<br>
general approach, so *in principle* this particular<br>
change could go away afterwards.<br>
<br>
But no, i haven't observed any notable (above the noiselevel)<br>
performance changes in this patch.<br>
<br></blockquote><div><br></div><div>OK. It was this set of comments that concerned me:</div><div><br></div><div>>> 1. After this fold, as it can be seen in tests, we may (will) end up with trivially redundant PHI nodes.<br>
>> We don't CSE them in InstCombine presently, which means that EarlyCSE needs to run and then InstCombine rerun.<br>
>> 2. But then, EarlyCSE not only manages to fold such redundant PHI's,<br>
>> it also sees that the extract-insert chain recreates the original aggregate,<br>
>> and replaces it with the original aggregate.<br></div><div><br></div><div>I am seeing some performance regressions, but I'll investigate and see if I can't get some program counters and a bit more here for you.</div><div><br></div><div>Thanks!</div><div><br></div><div>-eric</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
> -eric<br>
<br>
<br>
Roman<br>
<br>
> On Wed, Aug 26, 2020 at 2:58 AM Roman Lebedev via llvm-commits <<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a>> wrote:<br>
>><br>
>><br>
>> Author: Roman Lebedev<br>
>> Date: 2020-08-26T09:57:50+03:00<br>
>> New Revision: 1f90d45b9eb4550a4143ad846e2623fc1b7c05ab<br>
>><br>
>> URL: <a href="https://github.com/llvm/llvm-project/commit/1f90d45b9eb4550a4143ad846e2623fc1b7c05ab" rel="noreferrer" target="_blank">https://github.com/llvm/llvm-project/commit/1f90d45b9eb4550a4143ad846e2623fc1b7c05ab</a><br>
>> DIFF: <a href="https://github.com/llvm/llvm-project/commit/1f90d45b9eb4550a4143ad846e2623fc1b7c05ab.diff" rel="noreferrer" target="_blank">https://github.com/llvm/llvm-project/commit/1f90d45b9eb4550a4143ad846e2623fc1b7c05ab.diff</a><br>
>><br>
>> LOG: [InstCombine] PHI-of-extractvalues -> extractvalue-of-PHI, aka invokes are bad<br>
>><br>
>> While since D86306 we do it's sibling fold for `insertvalue`,<br>
>> we should also do this for `extractvalue`'s.<br>
>><br>
>> And unlike that one, the results here are, quite honestly, shocking,<br>
>> as it can be observed here on vanilla llvm test-suite + RawSpeed results:<br>
>><br>
>> ```<br>
>> | statistic name | baseline | proposed | Δ | % | |%| |<br>
>> |----------------------------------------------------|-----------|-----------|--------:|--------:|-------:|<br>
>> | asm-printer.EmittedInsts | 7945095 | 7942507 | -2588 | -0.03% | 0.03% |<br>
>> | assembler.ObjectBytes | 273209920 | 273069800 | -140120 | -0.05% | 0.05% |<br>
>> | early-cse.NumCSE | 2183363 | 2183398 | 35 | 0.00% | 0.00% |<br>
>> | early-cse.NumSimplify | 541847 | 550017 | 8170 | 1.51% | 1.51% |<br>
>> | instcombine.NumAggregateReconstructionsSimplified | 2139 | 108 | -2031 | -94.95% | 94.95% |<br>
>> | instcombine.NumCombined | 3601364 | 3635448 | 34084 | 0.95% | 0.95% |<br>
>> | instcombine.NumConstProp | 27153 | 27157 | 4 | 0.01% | 0.01% |<br>
>> | instcombine.NumDeadInst | 1694521 | 1765022 | 70501 | 4.16% | 4.16% |<br>
>> | instcombine.NumPHIsOfExtractValues | 0 | 37546 | 37546 | 0.00% | 0.00% |<br>
>> | instcombine.NumSunkInst | 63158 | 63686 | 528 | 0.84% | 0.84% |<br>
>> | instcount.NumBrInst | 874304 | 871857 | -2447 | -0.28% | 0.28% |<br>
>> | instcount.NumCallInst | 1757657 | 1758402 | 745 | 0.04% | 0.04% |<br>
>> | instcount.NumExtractValueInst | 45623 | 11483 | -34140 | -74.83% | 74.83% |<br>
>> | instcount.NumInsertValueInst | 4983 | 580 | -4403 | -88.36% | 88.36% |<br>
>> | instcount.NumInvokeInst | 61018 | 59478 | -1540 | -2.52% | 2.52% |<br>
>> | instcount.NumLandingPadInst | 35334 | 34215 | -1119 | -3.17% | 3.17% |<br>
>> | instcount.NumPHIInst | 344428 | 331116 | -13312 | -3.86% | 3.86% |<br>
>> | instcount.NumRetInst | 100773 | 100772 | -1 | 0.00% | 0.00% |<br>
>> | instcount.TotalBlocks | 1081154 | 1077166 | -3988 | -0.37% | 0.37% |<br>
>> | instcount.TotalFuncs | 101443 | 101442 | -1 | 0.00% | 0.00% |<br>
>> | instcount.TotalInsts | 8890201 | 8833747 | -56454 | -0.64% | 0.64% |<br>
>> | instsimplify.NumSimplified | 75822 | 75707 | -115 | -0.15% | 0.15% |<br>
>> | simplifycfg.NumHoistCommonCode | 24203 | 24197 | -6 | -0.02% | 0.02% |<br>
>> | simplifycfg.NumHoistCommonInstrs | 48201 | 48195 | -6 | -0.01% | 0.01% |<br>
>> | simplifycfg.NumInvokes | 2785 | 4298 | 1513 | 54.33% | 54.33% |<br>
>> | simplifycfg.NumSimpl | 997332 | 1018189 | 20857 | 2.09% | 2.09% |<br>
>> | simplifycfg.NumSinkCommonCode | 7088 | 6464 | -624 | -8.80% | 8.80% |<br>
>> | simplifycfg.NumSinkCommonInstrs | 15117 | 14021 | -1096 | -7.25% | 7.25% |<br>
>> ```<br>
>> ... which tells us that this new fold fires whopping 38k times,<br>
>> increasing the amount of SimplifyCFG's `invoke`->`call` transforms by +54% (+1513) (again, D85787 did that last time),<br>
>> decreasing total instruction count by -0.64% (-56454),<br>
>> and sharply decreasing count of `insertvalue`'s (-88.36%, i.e. 9 times less)<br>
>> and `extractvalue`'s (-74.83%, i.e. four times less).<br>
>><br>
>> This causes geomean -0.01% binary size decrease<br>
>> <a href="http://llvm-compile-time-tracker.com/compare.php?from=4d5ca22b8adfb6643466e4e9f48ba14bb48938bc&to=97dacca0111cb2ae678204e52a3cee00e3a69208&stat=size-text" rel="noreferrer" target="_blank">http://llvm-compile-time-tracker.com/compare.php?from=4d5ca22b8adfb6643466e4e9f48ba14bb48938bc&to=97dacca0111cb2ae678204e52a3cee00e3a69208&stat=size-text</a><br>
>> and, ignoring `O0-g`, is a geomean -0.01%..-0.05% compile-time improvement<br>
>> <a href="http://llvm-compile-time-tracker.com/compare.php?from=4d5ca22b8adfb6643466e4e9f48ba14bb48938bc&to=97dacca0111cb2ae678204e52a3cee00e3a69208&stat=instructions" rel="noreferrer" target="_blank">http://llvm-compile-time-tracker.com/compare.php?from=4d5ca22b8adfb6643466e4e9f48ba14bb48938bc&to=97dacca0111cb2ae678204e52a3cee00e3a69208&stat=instructions</a><br>
>><br>
>> The other thing that tells is, is that while this is a massive win for `invoke`->`call` transform<br>
>> `InstCombinerImpl::foldAggregateConstructionIntoAggregateReuse()` fold,<br>
>> which is supposed to be dealing with such aggregate reconstructions,<br>
>> fires a lot less now. There are two reasons why:<br>
>> 1. After this fold, as it can be seen in tests, we may (will) end up with trivially redundant PHI nodes.<br>
>> We don't CSE them in InstCombine presently, which means that EarlyCSE needs to run and then InstCombine rerun.<br>
>> 2. But then, EarlyCSE not only manages to fold such redundant PHI's,<br>
>> it also sees that the extract-insert chain recreates the original aggregate,<br>
>> and replaces it with the original aggregate.<br>
>><br>
>> The take-aways are<br>
>> 1. We maybe should do most trivial, same-BB PHI CSE in InstCombine<br>
>> 2. I need to check if what other patterns remain, and how they can be resolved.<br>
>> (i.e. i wonder if `foldAggregateConstructionIntoAggregateReuse()` might go away)<br>
>><br>
>> This is a reland of the original commit fcb51d8c2460faa23b71e06abb7e826243887dd6,<br>
>> because originally i forgot to ensure that the base aggregate types match.<br>
>><br>
>> Reviewed By: spatel<br>
>><br>
>> Differential Revision: <a href="https://reviews.llvm.org/D86530" rel="noreferrer" target="_blank">https://reviews.llvm.org/D86530</a><br>
>><br>
>> Added:<br>
>><br>
>><br>
>> Modified:<br>
>> llvm/lib/Transforms/InstCombine/InstCombineInternal.h<br>
>> llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp<br>
>> llvm/test/Transforms/InstCombine/phi-aware-aggregate-reconstruction.ll<br>
>> llvm/test/Transforms/InstCombine/phi-of-extractvalues.ll<br>
>><br>
>> Removed:<br>
>><br>
>><br>
>><br>
>> ################################################################################<br>
>> diff --git a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h<br>
>> index e5fe99367b7b..86b0bfe24d28 100644<br>
>> --- a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h<br>
>> +++ b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h<br>
>> @@ -618,6 +618,7 @@ class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final<br>
>> Instruction *foldPHIArgOpIntoPHI(PHINode &PN);<br>
>> Instruction *foldPHIArgBinOpIntoPHI(PHINode &PN);<br>
>> Instruction *foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN);<br>
>> + Instruction *foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN);<br>
>> Instruction *foldPHIArgGEPIntoPHI(PHINode &PN);<br>
>> Instruction *foldPHIArgLoadIntoPHI(PHINode &PN);<br>
>> Instruction *foldPHIArgZextsIntoPHI(PHINode &PN);<br>
>><br>
>> diff --git a/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp b/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp<br>
>> index 269b4d550e43..4c43e835186a 100644<br>
>> --- a/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp<br>
>> +++ b/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp<br>
>> @@ -32,6 +32,8 @@ MaxNumPhis("instcombine-max-num-phis", cl::init(512),<br>
>><br>
>> STATISTIC(NumPHIsOfInsertValues,<br>
>> "Number of phi-of-insertvalue turned into insertvalue-of-phis");<br>
>> +STATISTIC(NumPHIsOfExtractValues,<br>
>> + "Number of phi-of-extractvalue turned into extractvalue-of-phi");<br>
>><br>
>> /// The PHI arguments will be folded into a single operation with a PHI node<br>
>> /// as input. The debug location of the single operation will be the merged<br>
>> @@ -336,6 +338,43 @@ InstCombinerImpl::foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN) {<br>
>> return NewIVI;<br>
>> }<br>
>><br>
>> +/// If we have something like phi [extractvalue(a,0), extractvalue(b,0)],<br>
>> +/// turn this into a phi[a,b] and a single extractvalue.<br>
>> +Instruction *<br>
>> +InstCombinerImpl::foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN) {<br>
>> + auto *FirstEVI = cast<ExtractValueInst>(PN.getIncomingValue(0));<br>
>> +<br>
>> + // Scan to see if all operands are `extractvalue`'s with the same indicies,<br>
>> + // and all have a single use.<br>
>> + for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {<br>
>> + auto *I = dyn_cast<ExtractValueInst>(PN.getIncomingValue(i));<br>
>> + if (!I || !I->hasOneUse() || I->getIndices() != FirstEVI->getIndices() ||<br>
>> + I->getAggregateOperand()->getType() !=<br>
>> + FirstEVI->getAggregateOperand()->getType())<br>
>> + return nullptr;<br>
>> + }<br>
>> +<br>
>> + // Create a new PHI node to receive the values the aggregate operand has<br>
>> + // in each incoming basic block.<br>
>> + auto *NewAggregateOperand = PHINode::Create(<br>
>> + FirstEVI->getAggregateOperand()->getType(), PN.getNumIncomingValues(),<br>
>> + FirstEVI->getAggregateOperand()->getName() + ".pn");<br>
>> + // And populate the PHI with said values.<br>
>> + for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))<br>
>> + NewAggregateOperand->addIncoming(<br>
>> + cast<ExtractValueInst>(std::get<1>(Incoming))->getAggregateOperand(),<br>
>> + std::get<0>(Incoming));<br>
>> + InsertNewInstBefore(NewAggregateOperand, PN);<br>
>> +<br>
>> + // And finally, create `extractvalue` over the newly-formed PHI nodes.<br>
>> + auto *NewEVI = ExtractValueInst::Create(NewAggregateOperand,<br>
>> + FirstEVI->getIndices(), PN.getName());<br>
>> +<br>
>> + PHIArgMergedDebugLoc(NewEVI, PN);<br>
>> + ++NumPHIsOfExtractValues;<br>
>> + return NewEVI;<br>
>> +}<br>
>> +<br>
>> /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the<br>
>> /// adds all have a single use, turn this into a phi and a single binop.<br>
>> Instruction *InstCombinerImpl::foldPHIArgBinOpIntoPHI(PHINode &PN) {<br>
>> @@ -788,6 +827,8 @@ Instruction *InstCombinerImpl::foldPHIArgOpIntoPHI(PHINode &PN) {<br>
>> return foldPHIArgLoadIntoPHI(PN);<br>
>> if (isa<InsertValueInst>(FirstInst))<br>
>> return foldPHIArgInsertValueInstructionIntoPHI(PN);<br>
>> + if (isa<ExtractValueInst>(FirstInst))<br>
>> + return foldPHIArgExtractValueInstructionIntoPHI(PN);<br>
>><br>
>> // Scan the instruction, looking for input operations that can be folded away.<br>
>> // If all input operands to the phi are the same instruction (e.g. a cast from<br>
>><br>
>> diff --git a/llvm/test/Transforms/InstCombine/phi-aware-aggregate-reconstruction.ll b/llvm/test/Transforms/InstCombine/phi-aware-aggregate-reconstruction.ll<br>
>> index b97b40827ccd..939126bd2e34 100644<br>
>> --- a/llvm/test/Transforms/InstCombine/phi-aware-aggregate-reconstruction.ll<br>
>> +++ b/llvm/test/Transforms/InstCombine/phi-aware-aggregate-reconstruction.ll<br>
>> @@ -24,9 +24,14 @@ define { i32, i32 } @test0({ i32, i32 } %agg_left, { i32, i32 } %agg_right, i1 %<br>
>> ; CHECK-NEXT: call void @bar()<br>
>> ; CHECK-NEXT: br label [[END]]<br>
>> ; CHECK: end:<br>
>> -; CHECK-NEXT: [[I8_MERGED:%.*]] = phi { i32, i32 } [ [[AGG_RIGHT:%.*]], [[RIGHT]] ], [ [[AGG_LEFT:%.*]], [[LEFT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN:%.*]] = phi { i32, i32 } [ [[AGG_LEFT:%.*]], [[LEFT]] ], [ [[AGG_RIGHT:%.*]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN1:%.*]] = phi { i32, i32 } [ [[AGG_LEFT]], [[LEFT]] ], [ [[AGG_RIGHT]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[I6:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN1]], 1<br>
>> +; CHECK-NEXT: [[I5:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN]], 0<br>
>> ; CHECK-NEXT: call void @baz()<br>
>> -; CHECK-NEXT: ret { i32, i32 } [[I8_MERGED]]<br>
>> +; CHECK-NEXT: [[I7:%.*]] = insertvalue { i32, i32 } undef, i32 [[I5]], 0<br>
>> +; CHECK-NEXT: [[I8:%.*]] = insertvalue { i32, i32 } [[I7]], i32 [[I6]], 1<br>
>> +; CHECK-NEXT: ret { i32, i32 } [[I8]]<br>
>> ;<br>
>> entry:<br>
>> br i1 %c, label %left, label %right<br>
>> @@ -58,18 +63,16 @@ define { i32, i32 } @negative_test1({ i32, i32 } %agg_left, { i32, i32 } %agg_ri<br>
>> ; CHECK-NEXT: entry:<br>
>> ; CHECK-NEXT: br i1 [[C:%.*]], label [[LEFT:%.*]], label [[RIGHT:%.*]]<br>
>> ; CHECK: left:<br>
>> -; CHECK-NEXT: [[I4:%.*]] = extractvalue { i32, i32 } [[AGG_RIGHT:%.*]], 1<br>
>> -; CHECK-NEXT: [[I0:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT:%.*]], 0<br>
>> ; CHECK-NEXT: call void @foo()<br>
>> ; CHECK-NEXT: br label [[END:%.*]]<br>
>> ; CHECK: right:<br>
>> -; CHECK-NEXT: [[I3:%.*]] = extractvalue { i32, i32 } [[AGG_RIGHT]], 0<br>
>> -; CHECK-NEXT: [[I2:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT]], 1<br>
>> ; CHECK-NEXT: call void @bar()<br>
>> ; CHECK-NEXT: br label [[END]]<br>
>> ; CHECK: end:<br>
>> -; CHECK-NEXT: [[I5:%.*]] = phi i32 [ [[I0]], [[LEFT]] ], [ [[I3]], [[RIGHT]] ]<br>
>> -; CHECK-NEXT: [[I6:%.*]] = phi i32 [ [[I4]], [[LEFT]] ], [ [[I2]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN:%.*]] = phi { i32, i32 } [ [[AGG_LEFT:%.*]], [[LEFT]] ], [ [[AGG_RIGHT:%.*]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[AGG_RIGHT_PN:%.*]] = phi { i32, i32 } [ [[AGG_RIGHT]], [[LEFT]] ], [ [[AGG_LEFT]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[I6:%.*]] = extractvalue { i32, i32 } [[AGG_RIGHT_PN]], 1<br>
>> +; CHECK-NEXT: [[I5:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN]], 0<br>
>> ; CHECK-NEXT: call void @baz()<br>
>> ; CHECK-NEXT: [[I7:%.*]] = insertvalue { i32, i32 } undef, i32 [[I5]], 0<br>
>> ; CHECK-NEXT: [[I8:%.*]] = insertvalue { i32, i32 } [[I7]], i32 [[I6]], 1<br>
>> @@ -154,24 +157,20 @@ define { i32, i32 } @test3({ i32, i32 } %agg_00, { i32, i32 } %agg_01, { i32, i3<br>
>> ; CHECK: bb0.dispatch:<br>
>> ; CHECK-NEXT: br i1 [[C1:%.*]], label [[BB00:%.*]], label [[BB01:%.*]]<br>
>> ; CHECK: bb00:<br>
>> -; CHECK-NEXT: [[I0:%.*]] = extractvalue { i32, i32 } [[AGG_00:%.*]], 0<br>
>> -; CHECK-NEXT: [[I1:%.*]] = extractvalue { i32, i32 } [[AGG_00]], 1<br>
>> ; CHECK-NEXT: br label [[BB0_MERGE:%.*]]<br>
>> ; CHECK: bb01:<br>
>> -; CHECK-NEXT: [[I2:%.*]] = extractvalue { i32, i32 } [[AGG_01:%.*]], 0<br>
>> -; CHECK-NEXT: [[I3:%.*]] = extractvalue { i32, i32 } [[AGG_01]], 1<br>
>> ; CHECK-NEXT: br label [[BB0_MERGE]]<br>
>> ; CHECK: bb0.merge:<br>
>> -; CHECK-NEXT: [[I4:%.*]] = phi i32 [ [[I0]], [[BB00]] ], [ [[I2]], [[BB01]] ]<br>
>> -; CHECK-NEXT: [[I5:%.*]] = phi i32 [ [[I1]], [[BB00]] ], [ [[I3]], [[BB01]] ]<br>
>> +; CHECK-NEXT: [[AGG_00_PN:%.*]] = phi { i32, i32 } [ [[AGG_00:%.*]], [[BB00]] ], [ [[AGG_01:%.*]], [[BB01]] ]<br>
>> +; CHECK-NEXT: [[AGG_00_PN1:%.*]] = phi { i32, i32 } [ [[AGG_00]], [[BB00]] ], [ [[AGG_01]], [[BB01]] ]<br>
>> ; CHECK-NEXT: br label [[END:%.*]]<br>
>> ; CHECK: bb10:<br>
>> -; CHECK-NEXT: [[I6:%.*]] = extractvalue { i32, i32 } [[AGG_10:%.*]], 0<br>
>> -; CHECK-NEXT: [[I7:%.*]] = extractvalue { i32, i32 } [[AGG_10]], 1<br>
>> ; CHECK-NEXT: br label [[END]]<br>
>> ; CHECK: end:<br>
>> -; CHECK-NEXT: [[I8:%.*]] = phi i32 [ [[I4]], [[BB0_MERGE]] ], [ [[I6]], [[BB10]] ]<br>
>> -; CHECK-NEXT: [[I9:%.*]] = phi i32 [ [[I5]], [[BB0_MERGE]] ], [ [[I7]], [[BB10]] ]<br>
>> +; CHECK-NEXT: [[AGG_00_PN_PN:%.*]] = phi { i32, i32 } [ [[AGG_00_PN]], [[BB0_MERGE]] ], [ [[AGG_10:%.*]], [[BB10]] ]<br>
>> +; CHECK-NEXT: [[AGG_00_PN1_PN:%.*]] = phi { i32, i32 } [ [[AGG_00_PN1]], [[BB0_MERGE]] ], [ [[AGG_10]], [[BB10]] ]<br>
>> +; CHECK-NEXT: [[I9:%.*]] = extractvalue { i32, i32 } [[AGG_00_PN1_PN]], 1<br>
>> +; CHECK-NEXT: [[I8:%.*]] = extractvalue { i32, i32 } [[AGG_00_PN_PN]], 0<br>
>> ; CHECK-NEXT: call void @baz()<br>
>> ; CHECK-NEXT: [[I10:%.*]] = insertvalue { i32, i32 } undef, i32 [[I8]], 0<br>
>> ; CHECK-NEXT: [[I11:%.*]] = insertvalue { i32, i32 } [[I10]], i32 [[I9]], 1<br>
>> @@ -278,12 +277,17 @@ define { i32, i32 } @test5({ i32, i32 } %agg_left, { i32, i32 } %agg_right, i1 %<br>
>> ; CHECK-NEXT: call void @bar()<br>
>> ; CHECK-NEXT: br label [[MIDDLE]]<br>
>> ; CHECK: middle:<br>
>> -; CHECK-NEXT: [[I8_MERGED:%.*]] = phi { i32, i32 } [ [[I8_MERGED]], [[MIDDLE]] ], [ [[AGG_RIGHT:%.*]], [[RIGHT]] ], [ [[AGG_LEFT:%.*]], [[LEFT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN:%.*]] = phi { i32, i32 } [ [[AGG_LEFT:%.*]], [[LEFT]] ], [ [[AGG_RIGHT:%.*]], [[RIGHT]] ], [ [[I8:%.*]], [[MIDDLE]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN1:%.*]] = phi { i32, i32 } [ [[AGG_LEFT]], [[LEFT]] ], [ [[AGG_RIGHT]], [[RIGHT]] ], [ [[I8]], [[MIDDLE]] ]<br>
>> +; CHECK-NEXT: [[I6:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN1]], 1<br>
>> +; CHECK-NEXT: [[I5:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN]], 0<br>
>> ; CHECK-NEXT: call void @baz()<br>
>> +; CHECK-NEXT: [[I7:%.*]] = insertvalue { i32, i32 } undef, i32 [[I5]], 0<br>
>> +; CHECK-NEXT: [[I8]] = insertvalue { i32, i32 } [[I7]], i32 [[I6]], 1<br>
>> ; CHECK-NEXT: [[C1:%.*]] = call i1 @geni1()<br>
>> ; CHECK-NEXT: br i1 [[C1]], label [[END:%.*]], label [[MIDDLE]]<br>
>> ; CHECK: end:<br>
>> -; CHECK-NEXT: ret { i32, i32 } [[I8_MERGED]]<br>
>> +; CHECK-NEXT: ret { i32, i32 } [[I8]]<br>
>> ;<br>
>> entry:<br>
>> br i1 %c0, label %left, label %right<br>
>> @@ -327,15 +331,20 @@ define { i32, i32 } @test6({ i32, i32 } %agg_left, { i32, i32 } %agg_right, i1 %<br>
>> ; CHECK-NEXT: call void @bar()<br>
>> ; CHECK-NEXT: br label [[MERGE]]<br>
>> ; CHECK: merge:<br>
>> -; CHECK-NEXT: [[I8_MERGED:%.*]] = phi { i32, i32 } [ [[AGG_RIGHT:%.*]], [[RIGHT]] ], [ [[AGG_LEFT:%.*]], [[LEFT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN:%.*]] = phi { i32, i32 } [ [[AGG_LEFT:%.*]], [[LEFT]] ], [ [[AGG_RIGHT:%.*]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN1:%.*]] = phi { i32, i32 } [ [[AGG_LEFT]], [[LEFT]] ], [ [[AGG_RIGHT]], [[RIGHT]] ]<br>
>> ; CHECK-NEXT: call void @baz()<br>
>> ; CHECK-NEXT: br i1 [[C1:%.*]], label [[END:%.*]], label [[PASSTHROUGH:%.*]]<br>
>> ; CHECK: passthrough:<br>
>> ; CHECK-NEXT: call void @qux()<br>
>> ; CHECK-NEXT: br label [[END]]<br>
>> ; CHECK: end:<br>
>> +; CHECK-NEXT: [[I6:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN1]], 1<br>
>> +; CHECK-NEXT: [[I5:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN]], 0<br>
>> ; CHECK-NEXT: call void @quux()<br>
>> -; CHECK-NEXT: ret { i32, i32 } [[I8_MERGED]]<br>
>> +; CHECK-NEXT: [[I7:%.*]] = insertvalue { i32, i32 } undef, i32 [[I5]], 0<br>
>> +; CHECK-NEXT: [[I8:%.*]] = insertvalue { i32, i32 } [[I7]], i32 [[I6]], 1<br>
>> +; CHECK-NEXT: ret { i32, i32 } [[I8]]<br>
>> ;<br>
>> entry:<br>
>> br i1 %c0, label %left, label %right<br>
>> @@ -490,9 +499,14 @@ define { i32, i32 } @test9({ i32, i32 } %agg_left, { i32, i32 } %agg_right, i1 %<br>
>> ; CHECK-NEXT: call void @bar()<br>
>> ; CHECK-NEXT: br label [[END]]<br>
>> ; CHECK: end:<br>
>> -; CHECK-NEXT: [[I8_MERGED:%.*]] = phi { i32, i32 } [ [[AGG_RIGHT:%.*]], [[RIGHT]] ], [ [[AGG_LEFT:%.*]], [[LEFT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN:%.*]] = phi { i32, i32 } [ [[AGG_LEFT:%.*]], [[LEFT]] ], [ [[AGG_RIGHT:%.*]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN1:%.*]] = phi { i32, i32 } [ [[AGG_LEFT]], [[LEFT]] ], [ [[AGG_RIGHT]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[I7:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN1]], 1<br>
>> +; CHECK-NEXT: [[I0_PN:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN]], 0<br>
>> +; CHECK-NEXT: [[I6:%.*]] = insertvalue { i32, i32 } undef, i32 [[I0_PN]], 0<br>
>> ; CHECK-NEXT: call void @baz()<br>
>> -; CHECK-NEXT: ret { i32, i32 } [[I8_MERGED]]<br>
>> +; CHECK-NEXT: [[I8:%.*]] = insertvalue { i32, i32 } [[I6]], i32 [[I7]], 1<br>
>> +; CHECK-NEXT: ret { i32, i32 } [[I8]]<br>
>> ;<br>
>> entry:<br>
>> br i1 %c, label %left, label %right<br>
>><br>
>> diff --git a/llvm/test/Transforms/InstCombine/phi-of-extractvalues.ll b/llvm/test/Transforms/InstCombine/phi-of-extractvalues.ll<br>
>> index 78ea09f7e636..46cbd123f5fb 100644<br>
>> --- a/llvm/test/Transforms/InstCombine/phi-of-extractvalues.ll<br>
>> +++ b/llvm/test/Transforms/InstCombine/phi-of-extractvalues.ll<br>
>> @@ -10,13 +10,12 @@ define i32 @test0({ i32, i32 } %agg_left, { i32, i32 } %agg_right, i1 %c) {<br>
>> ; CHECK-NEXT: entry:<br>
>> ; CHECK-NEXT: br i1 [[C:%.*]], label [[LEFT:%.*]], label [[RIGHT:%.*]]<br>
>> ; CHECK: left:<br>
>> -; CHECK-NEXT: [[I0:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT:%.*]], 0<br>
>> ; CHECK-NEXT: br label [[END:%.*]]<br>
>> ; CHECK: right:<br>
>> -; CHECK-NEXT: [[I1:%.*]] = extractvalue { i32, i32 } [[AGG_RIGHT:%.*]], 0<br>
>> ; CHECK-NEXT: br label [[END]]<br>
>> ; CHECK: end:<br>
>> -; CHECK-NEXT: [[R:%.*]] = phi i32 [ [[I0]], [[LEFT]] ], [ [[I1]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN:%.*]] = phi { i32, i32 } [ [[AGG_LEFT:%.*]], [[LEFT]] ], [ [[AGG_RIGHT:%.*]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[R:%.*]] = extractvalue { i32, i32 } [[AGG_LEFT_PN]], 0<br>
>> ; CHECK-NEXT: ret i32 [[R]]<br>
>> ;<br>
>> entry:<br>
>> @@ -169,13 +168,12 @@ define i32 @test5({{ i32, i32 }, { i32, i32 }} %agg_left, {{ i32, i32 }, { i32,<br>
>> ; CHECK-NEXT: entry:<br>
>> ; CHECK-NEXT: br i1 [[C:%.*]], label [[LEFT:%.*]], label [[RIGHT:%.*]]<br>
>> ; CHECK: left:<br>
>> -; CHECK-NEXT: [[I0:%.*]] = extractvalue { { i32, i32 }, { i32, i32 } } [[AGG_LEFT:%.*]], 0, 0<br>
>> ; CHECK-NEXT: br label [[END:%.*]]<br>
>> ; CHECK: right:<br>
>> -; CHECK-NEXT: [[I1:%.*]] = extractvalue { { i32, i32 }, { i32, i32 } } [[AGG_RIGHT:%.*]], 0, 0<br>
>> ; CHECK-NEXT: br label [[END]]<br>
>> ; CHECK: end:<br>
>> -; CHECK-NEXT: [[R:%.*]] = phi i32 [ [[I0]], [[LEFT]] ], [ [[I1]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[AGG_LEFT_PN:%.*]] = phi { { i32, i32 }, { i32, i32 } } [ [[AGG_LEFT:%.*]], [[LEFT]] ], [ [[AGG_RIGHT:%.*]], [[RIGHT]] ]<br>
>> +; CHECK-NEXT: [[R:%.*]] = extractvalue { { i32, i32 }, { i32, i32 } } [[AGG_LEFT_PN]], 0, 0<br>
>> ; CHECK-NEXT: ret i32 [[R]]<br>
>> ;<br>
>> entry:<br>
>><br>
>><br>
>><br>
>> _______________________________________________<br>
>> llvm-commits mailing list<br>
>> <a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
>> <a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
</blockquote></div></div>