[PATCH] D13417: [MachineCombiner] make slack optional in critical path cost calculation (PR25016)

[Gerolf Hoflehner via llvm-commits]

Gerolf added a comment.

Perhaps there is another change in your tree.  We are looking at different code and/or I don’t know how your reference sequence for your test case. With fast-math and reassociation turned off I’m getting:

  vmovsd  8(%rsp), %xmm1          # 8-byte Reload
                                  # xmm1 = mem[0],zero
  vaddsd  16(%rsp), %xmm1, %xmm1  # 8-byte Folded Reload
  vaddsd  (%rsp), %xmm0, %xmm0    # 8-byte Folded Reload
  vaddsd  %xmm0, %xmm1, %xmm0

Which is happens to be the same code I get w/o fast-math.

That should not be faster than fast-math with reassociation.

I don’t understand why the equation *holds* although you measure a performance loss. I very much prefer a single equation that holds for all pattern.

-Gerolf

PS:
reassoc.ll

declare double @bar()
define double @foo4_reassociated() {

  %a = call double @bar()
  %b = call double @bar()
  %c = call double @bar()
  %d = call double @bar()
  %t0 = fadd double %a, %b 
  %t1 = fadd double %c, %d 
  %t2 = fadd double %t0, %t1 
  ret double %t2

}


http://reviews.llvm.org/D13417