[Bug target/56944] New: Better isfinite in some cases?

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56944

             Bug #: 56944
           Summary: Better isfinite in some cases?
    Classification: Unclassified
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: glisse@gcc.gnu.org
            Target: x86_64-linux-gnu


Hello,

for isfinite, gcc typically generates this sequence:

    movsd    .LC0(%rip), %xmm1
    andpd    %xmm1, %xmm0
    movsd    .LC1(%rip), %xmm1
    ucomisd    %xmm0, %xmm1
    setae    %al

With -fno-trapping-math, I tried this shorter sequence instead, which should be
valid:

    subsd    %xmm0, %xmm0
    ucomisd    %xmm0, %xmm0
    setnp    %al

Depending on the tests, it seemed to be either the same speed or 15% faster,
whether the argument is normal, infinite or nan. For a denormal argument, it is
15% slower (but then both codes take 100 times as long as the normal case). The
results might also be different on a more recent processor.

I don't know if we want to try and generate this code when -fno-trapping-math
is present.

(related to PR 30652)