[InstCombine] prematurely promoting floating point negation

[llvmbot]

Bugzilla Link	21914
Version	trunk
OS	All
Attachments	source code of the example, the resultant LLVM IR
Reporter	LLVM Bugzilla Contributor
CC	@RKSimon,@rotateright

Extended Description

The optimization that transforms (-x * y) to -(x * y) for floating pointer numbers seems premature.

http://llvm.org/doxygen/InstCombineMulDivRem_8cpp_source.html#l00602

00602 if (Opnd0->hasOneUse()) {
00603 // -X * Y => -(X*Y) (Promote negation as high as possible)
00604 Value *T = Builder->CreateFMul(N0, Opnd1);
00605 Value *Neg = Builder->CreateFNeg(T);
00606 Neg->takeName(&I);
00607 return ReplaceInstUsesWith(I, Neg);
00608 }

While I understand the motivation of this transformation, it prevents potential LICM if X is a loop invariant and Y is a loop variant. For instance,

int bar(int, int);

void foo(int n, float *output, float beta) {
float sum = 0;
for (int i = 0; i < n; ++i)
sum += bar(i, i * (-beta));
*output = sum;
}

With i*(-beta) transformed to -(i*beta), the LICM passes after instcombine is unable to hoist -beta outside the loop. As a result, "opt -O3" leaves a fsub instruction inside the loop:

define void @​_Z3fooiPff(i32 %n, float* nocapture %output, float %beta) #​0 {
entry:
%cmp7 = icmp sgt i32 %n, 0
br i1 %cmp7, label %for.body.lr.ph, label %for.end

for.body.lr.ph: ; preds = %entry
%0 = add i32 %n, -1
br label %for.body

for.body: ; preds = %for.body, %for.body.lr.ph
%i.09 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
%sum.08 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add, %for.body ]
%conv = sitofp i32 %i.09 to float
%1 = fmul float %conv, %beta
%mul = fsub float -0.000000e+00, %1
%conv1 = fptosi float %mul to i32
%call = tail call i32 @​_Z3barii(i32 %i.09, i32 %conv1)
%conv2 = sitofp i32 %call to float
%add = fadd float %sum.08, %conv2
%inc = add nsw i32 %i.09, 1
%exitcond = icmp eq i32 %i.09, %0
br i1 %exitcond, label %for.end.loopexit, label %for.body

for.end.loopexit: ; preds = %for.body
%add.lcssa = phi float [ %add, %for.body ]
br label %for.end

for.end: ; preds = %for.end.loopexit, %entry
%sum.0.lcssa = phi float [ 0.000000e+00, %entry ], [ %add.lcssa, %for.end.loopexit ]
store float %sum.0.lcssa, float* %output, align 4, !tbaa !​1
ret void
}

Jingyue

[llvmbot]

The transformation mentioned above was introduced in r172551. Shuxin, do you have any idea on how to solve this issue? Thanks!

Jingyue

[rotateright]

On x86, this is extra bad because we also fail to remove a 16-byte constant load out of the loop:
xorps LCPI0_0(%rip), %xmm1

[rotateright]

This could expose other problems, but it solves the example shown here:
https://reviews.llvm.org/D65399