Missed optimization in math expression: sqrt(pow(a, 2.0)) == |a|

[zamazan4ik]

Bugzilla Link	35601
Resolution	FIXED
Resolved on	Dec 10, 2017 09:38
Version	trunk
OS	All
Blocks	#34959
CC	@hfinkel,@rotateright

Extended Description

clang(trunk) with '--std=c++17 -O3 -march=native -ffast-math' flags for this code:

#include <cmath>

double test(double a, double b, double c, double d)
{
    return sqrt(pow(a, 2.0));
}

generates this assembly:

test(double, double, double, double): # @test(double, double, double, double)
  vmulsd xmm0, xmm0, xmm0
  vsqrtsd xmm0, xmm0, xmm0
  ret

gcc(trunk) with '--std=c++17 -O3 -march=native -ffast-math' flags generates this:

test(double, double, double, double):
        vandpd  xmm0, xmm0, XMMWORD PTR .LC0[rip]
        ret
.LC0:
        .long   4294967295
        .long   2147483647
        .long   0
        .long   0

[zamazan4ik]

assigned to @rotateright

[hfinkel]

Could be in InstCombine.

[rotateright]

I thought we already had this one via:
// We're looking for a repeated factor in a multiplication tree,
// so we can do this fold: sqrt(x * x) -> fabs(x);

Let me see how it escaped.

[rotateright]

The fmul lost its 'fast' FMF:

define double @test(double) local_unnamed_addr #0 {
  %2 = fmul double %0, %0
  %3 = tail call fast double @llvm.sqrt.f64(double %2)
  ret double %3
}

[zamazan4ik]

See a lot of optimizations here: https://github.com/gcc-mirror/gcc/blob/07b69d3f1cd3dd8ebb0af1fbff95914daee477d2/gcc/match.pd

[rotateright]

Should be fixed after:
https://reviews.llvm.org/rL320309
https://reviews.llvm.org/rL320310

Note that there's a sibling bug for pow(x, -1.0) that is still alive (I marked it with a 'FIXME' comment in the source).

Also, we should be getting these folds with vector types too, but I suspect we're missing a bunch of those cases.

[rotateright]

Note that there's a sibling bug for pow(x, -1.0) that is still alive (I
marked it with a 'FIXME' comment in the source).

https://reviews.llvm.org/rL320312

I don't think that's a factor in any of the related bugs so far, but feel free to double-check. Thanks for filing these!

[RKSimon]

mentioned in issue #34959