[llvm] r190871 - SLPVectorizer: Don't vectorize phi nodes that use invoke values

[Arnold Schwaighofer]

Nothing wrong in principle. The cost of doing it must be justified by an expected benefit. Without further analysis, it is not clear to me that there is going to be benefit.


On Sep 17, 2013, at 12:15 PM, Quentin Colombet <qcolombet at apple.com> wrote:

> Hi Arnold,
> 
> I am curious, what is wrong with splitting critical edges?
> 
> Is it related to the potential cost it will incurved or is it just something we do not want to do?
> 
> Thanks,
> -Quentin
> 
> On Sep 17, 2013, at 10:03 AM, Arnold Schwaighofer <aschwaighofer at apple.com> wrote:
> 
>> Author: arnolds
>> Date: Tue Sep 17 12:03:29 2013
>> New Revision: 190871
>> 
>> URL: http://llvm.org/viewvc/llvm-project?rev=190871&view=rev
>> Log:
>> SLPVectorizer: Don't vectorize phi nodes that use invoke values
>> 
>> We can't insert an insertelement after an invoke. We would have to split a
>> critical edge. So when we see a phi node that uses an invoke we just give up.
>> 
>> radar://14990770
>> 
>> Modified:
>>    llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
>>    llvm/trunk/test/Transforms/SLPVectorizer/X86/ordering.ll
>> 
>> Modified: llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp?rev=190871&r1=190870&r2=190871&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp (original)
>> +++ llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp Tue Sep 17 12:03:29 2013
>> @@ -639,6 +639,18 @@ void BoUpSLP::buildTree_rec(ArrayRef<Val
>>   switch (Opcode) {
>>     case Instruction::PHI: {
>>       PHINode *PH = dyn_cast<PHINode>(VL0);
>> +
>> +      // Check for terminator values (e.g. invoke).
>> +      for (unsigned j = 0; j < VL.size(); ++j)
>> +        for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
>> +          TerminatorInst *Term = dyn_cast<TerminatorInst>(cast<PHINode>(VL[j])->getIncomingValue(i));
>> +          if (Term) {
>> +            DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
>> +            newTreeEntry(VL, false);
>> +            return;
>> +          }
>> +        }
>> +
>>       newTreeEntry(VL, true);
>>       DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
>> 
>> 
>> Modified: llvm/trunk/test/Transforms/SLPVectorizer/X86/ordering.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/SLPVectorizer/X86/ordering.ll?rev=190871&r1=190870&r2=190871&view=diff
>> ==============================================================================
>> --- llvm/trunk/test/Transforms/SLPVectorizer/X86/ordering.ll (original)
>> +++ llvm/trunk/test/Transforms/SLPVectorizer/X86/ordering.ll Tue Sep 17 12:03:29 2013
>> @@ -17,3 +17,65 @@ entry:
>>   %cmp11 = fcmp olt double %add, 0.000000e+00
>>   ret void
>> }
>> +
>> +declare i8* @objc_msgSend(i8*, i8*, ...)
>> +declare i32 @personality_v0(...)
>> +
>> +define void @invoketest() {
>> +entry:
>> +  br i1 undef, label %cond.true, label %cond.false
>> +
>> +cond.true:
>> +  %call49 = invoke double bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to double (i8*, i8*)*)(i8* undef, i8* undef) 
>> +          to label %cond.true54 unwind label %lpad
>> +
>> +cond.false:
>> +  %call51 = invoke double bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to double (i8*, i8*)*)(i8* undef, i8* undef)
>> +          to label %cond.false57 unwind label %lpad
>> +
>> +cond.true54:
>> +  %call56 = invoke double bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to double (i8*, i8*)*)(i8* undef, i8* undef) 
>> +          to label %cond.end60 unwind label %lpad
>> +
>> +cond.false57:
>> +  %call59 = invoke double bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to double (i8*, i8*)*)(i8* undef, i8* undef)
>> +          to label %cond.end60 unwind label %lpad
>> +
>> +; Make sure we don't vectorize these phis - they have invokes as inputs.
>> +
>> +; RUN: opt < %s -slp-vectorizer -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7 | FileCheck %s
>> +
>> +; CHECK-LABEL: invoketest
>> +
>> +; CHECK-LABEL: cond.end60
>> +; CHECK-NEXT-NOT: phi <2 x double>
>> +; CHECK: insertelement
>> +; CHECK-LABEL: if.then63
>> +
>> +cond.end60:
>> +  %cond126 = phi double [ %call49, %cond.true54 ], [ %call51, %cond.false57 ]
>> +  %cond61 = phi double [ %call56, %cond.true54 ], [ %call59, %cond.false57 ]
>> +  br i1 undef, label %if.end98, label %if.then63
>> +
>> +if.then63:
>> +  %conv69 = fptrunc double undef to float
>> +  %conv70 = fpext float %conv69 to double
>> +  %div71 = fdiv double %cond126, %conv70
>> +  %conv78 = fptrunc double undef to float
>> +  %conv79 = fpext float %conv78 to double
>> +  %div80 = fdiv double %cond61, %conv79
>> +  br label %if.end98
>> +
>> +lpad:
>> +  %l = landingpad { i8*, i32 } personality i8* bitcast (i32 (...)* @personality_v0 to i8*)
>> +          cleanup
>> +  resume { i8*, i32 } %l
>> +
>> +if.end98:
>> +  %dimensionsResult.sroa.0.0 = phi double [ %div71, %if.then63 ], [ %cond126, %cond.end60 ]
>> +  %dimensionsResult.sroa.6.0 = phi double [ %div80, %if.then63 ], [ %cond61, %cond.end60 ]
>> +  br label %if.end99
>> +
>> +if.end99:
>> +  ret void
>> +}
>> 
>> 
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>