Commit fbe05633 authored by Hans Wennborg's avatar Hans Wennborg
Browse files

Merging r341416: 

------------------------------------------------------------------------
r341416 | annat | 2018-09-05 00:12:23 +0200 (Wed, 05 Sep 2018) | 11 lines

[LV] First order recurrence phis should not be treated as uniform

This is fix for PR38786.
First order recurrence phis were incorrectly treated as uniform,
which caused them to be vectorized as uniform instructions.

Patch by Ayal Zaks and Orivej Desh!

Reviewed by: Anna

Differential Revision: https://reviews.llvm.org/D51639
------------------------------------------------------------------------

llvm-svn: 341523
parent 3f83c2fa

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

+7 −0
Original line number Diff line number Diff line
@@ -4510,6 +4510,13 @@ void LoopVectorizationCostModel::collectLoopUniforms(unsigned VF) { 
    for (auto OV : I->operand_values()) {

      if (isOutOfScope(OV))

        continue;

      // First order recurrence Phi's should typically be considered

      // non-uniform.

      auto *OP = dyn_cast<PHINode>(OV);

      if (OP && Legal->isFirstOrderRecurrence(OP))

        continue;

      // If all the users of the operand are uniform, then add the

      // operand into the uniform worklist.

      auto *OI = cast<Instruction>(OV);

      if (llvm::all_of(OI->users(), [&](User *U) -> bool {

            auto *J = cast<Instruction>(U);

llvm/test/Transforms/LoopVectorize/X86/uniform-phi.ll

+22 −0
Original line number Diff line number Diff line
@@ -75,3 +75,25 @@ for.end: ; preds = %for.body 
  ret i64 %retval

}



; CHECK-LABEL: PR38786

; Check that first order recurrence phis (%phi32 and %phi64) are not uniform.

; CHECK-NOT: LV: Found uniform instruction:   %phi

define void @PR38786(double* %y, double* %x, i64 %n) {

entry:

  br label %for.body



for.body:

  %phi32 = phi i32 [ 0, %entry ], [ %i32next, %for.body ]

  %phi64 = phi i64 [ 0, %entry ], [ %i64next, %for.body ]

  %i32next = add i32 %phi32, 1

  %i64next = zext i32 %i32next to i64

  %xip = getelementptr inbounds double, double* %x, i64 %i64next

  %yip = getelementptr inbounds double, double* %y, i64 %phi64

  %xi = load double, double* %xip, align 8

  store double %xi, double* %yip, align 8

  %cmp = icmp slt i64 %i64next, %n

  br i1 %cmp, label %for.body, label %for.end



for.end:

  ret void

}