Missed optimization in math expression: missing cube of the sum formula

[zamazan4ik]

Bugzilla Link	35597
Version	trunk
OS	Linux
Blocks	#34959
CC	@apolukhin,@RKSimon,@rotateright

Extended Description

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

double test(double a, double b)
{
    return a*a*a + 3.0*a*a*b + 3.0*a*b*b + b*b*b;
}

generates this assembly:

.LCPI0_0:
  .quad 4613937818241073152 # double 3
test(double, double): # @test(double, double)
  vmulsd xmm2, xmm0, qword ptr [rip + .LCPI0_0]
  vmulsd xmm3, xmm0, xmm0
  vmulsd xmm4, xmm1, xmm1
  vaddsd xmm5, xmm2, xmm1
  vmulsd xmm4, xmm4, xmm5
  vfmadd231sd xmm3, xmm2, xmm1
  vfmadd213sd xmm0, xmm3, xmm4
  ret

But there is formula: aaa + 3.0aab + 3.0abb + bbb == (a + b)^3. And it can be compiled in faster way.

[rotateright]

I think this one is different than most/all of the related fast-math bugs that were recently filed. This seems like a shortcoming in -reassociate rather than LibCallSimplifier / InstCombine.

[RKSimon]

The simpler square of the sum case is not handled either:

https://godbolt.org/g/8wPQSa

double test2x(double a, double b)
{
    return a*a + 2.0*a*b + b*b; // (a+b)*(a+b)
}

.LCPI1_0:
  .quad 4611686018427387904 # double 2
_Z6test2xdd: # @_Z6test2xdd
  vmulsd %xmm0, %xmm0, %xmm2
  vfmaddsd %xmm1, .LCPI1_0(%rip), %xmm0, %xmm0
  vfmaddsd %xmm2, %xmm1, %xmm0, %xmm0
  retq

[RKSimon]

mentioned in issue #34959