Missed optimization in math expression: squashing sqrt functions

[zamazan4ik]

Bugzilla Link	35592
Resolution	FIXED
Resolved on	Apr 15, 2018 20:04
Version	trunk
OS	All
Blocks	#34959
CC	@hfinkel,@jryans,@RKSimon,@rotateright

Extended Description

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

#include <cmath>

double f(double a, double b, double c, double d)
{
    return sqrt(a)*sqrt(b)*sqrt(c)*sqrt(d);
}

generates this assembly:

f(double, double, double, double): # @f(double, double, double, double)
  vsqrtsd xmm0, xmm0, xmm0
  vsqrtsd xmm1, xmm1, xmm1
  vmulsd xmm0, xmm1, xmm0
  vsqrtsd xmm1, xmm2, xmm2
  vsqrtsd xmm2, xmm3, xmm3
  vmulsd xmm1, xmm1, xmm2
  vmulsd xmm0, xmm0, xmm1
  ret

GCC (trunk) with '--std=c++17 -O3 -march=native -ffast-math' flags generates more optimal code:

f(double, double, double, double):
        vmulsd  xmm0, xmm0, xmm1
        vmulsd  xmm2, xmm2, xmm3
        vmulsd  xmm2, xmm2, xmm0
        vsqrtsd xmm0, xmm0, xmm2
        ret

[hfinkel]

It will certainly make sense to have sqrt(a)sqrt(b) -> sqrt(ab) in InstCombine.

[zamazan4ik]

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

[hfinkel]

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

You might want to create a "meta bug" about all of these, and then link the individual bug reports to that as dependencies.

Also, there are other sources of these kinds of things that we should look at. For example, in https://github.com/sympy/sympy/blob/master/sympy/functions/elementary/trigonometric.py we have:

class cos(TrigonometricFunction):
  ...
    def eval(cls, arg):
  ...
        if isinstance(arg, acos):
            return arg.args[0]

        if isinstance(arg, atan):
            x = arg.args[0]
            return 1 / sqrt(1 + x**2)

        if isinstance(arg, atan2):
            y, x = arg.args
            return x / sqrt(x**2 + y**2)

        if isinstance(arg, asin):
            x = arg.args[0]
            return sqrt(1 - x ** 2)

        if isinstance(arg, acot):
            x = arg.args[0]
            return 1 / sqrt(1 + 1 / x**2)

        if isinstance(arg, acsc):
            x = arg.args[0]
            return sqrt(1 - 1 / x**2)

        if isinstance(arg, asec):
            x = arg.args[0]
            return 1 / x

[jryans]

It appears this was reviewed in https://reviews.llvm.org/D41322 and fixed in r321637.

[RKSimon]

mentioned in issue #34959