Missed optimization in math expression: aggressive optimization with std::pow

[zamazan4ik]

Bugzilla Link	35595
Version	trunk
OS	All
Blocks	#34959
CC	@davidbolvansky,@efriedma-quic,@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 x)
{
    return pow(a, x) * a * a * a * a;
}

generates this assembly:

test(double, double): # @test(double, double)
  push rax
  vmovsd qword ptr [rsp], xmm0 # 8-byte Spill
  call pow
  vmovsd xmm1, qword ptr [rsp] # 8-byte Reload
  vmulsd xmm1, xmm1, xmm1
  vmulsd xmm1, xmm1, xmm1
  vmulsd xmm0, xmm1, xmm0
  pop rax
  ret

As you can see, me can simplify it by adding 4 to 'x' variable and after call std::pow.

[hfinkel]

Another missing InstCombine.

[zamazan4ik]

One more case:

#include <cmath>

double test(double a, double b, double c)
{
    return pow(a,b) * pow(c, b);
}

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

test(double, double, double): # @test(double, double, double)
  sub rsp, 24
  vmovsd qword ptr [rsp + 16], xmm2 # 8-byte Spill
  vmovsd qword ptr [rsp], xmm1 # 8-byte Spill
  call pow
  vmovsd qword ptr [rsp + 8], xmm0 # 8-byte Spill
  vmovsd xmm0, qword ptr [rsp + 16] # 8-byte Reload
  vmovsd xmm1, qword ptr [rsp] # 8-byte Reload
  call pow
  vmulsd xmm0, xmm0, qword ptr [rsp + 8] # 8-byte Folded Reload
  add rsp, 24
  ret

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

test(double, double, double):
        vmulsd  xmm0, xmm0, xmm2
        jmp     __pow_finite

[zamazan4ik]

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

[llvmbot]

Same for opposite case

double f2(double a, double b, double c, double d)
{
    return pow(a,b)*pow(a,c)*pow(a,d)*pow(a,-2);
}

can be simplified to pow(a,b+c+d-2)

[llvmbot]

See https://reviews.llvm.org/D41608

[RKSimon]

mentioned in issue #34959

[rotateright]

All of the examples here should be optimized now. If there's still something missing, please open a new issue.