https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68130
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok. The FRE2 difference is interesting but ok. We CSE
_7 = __builtin_signbit (1.2999999523162841796875e+0);
if (_7 != 0)
goto <bb 4>;
else
goto <bb 3>;
<bb 3>:
<bb 4>:
# iftmp.9_8 = PHI <-1.0e+0(2), 1.0e+0(3)>
...
<bb 14>:
if (_7 != 0)
goto <bb 16>;
else
goto <bb 15>;
<bb 15>:
<bb 16>:
# iftmp.13_15 = PHI <-1.0e+0(14), 1.0e+0(15)>
_16 = __builtin_copysignf (-1.2999999523162841796875e+0, iftmp.13_15);
to use iftmp.9_8 in the last copysignf statement, removing the PHI node
preceeding it as well as the redundant test of _7 != 0. We do this for
five cases.
The PRE dump difference is pretty unreadable already. So I wonder if
the issue reproduces with -fno-tree-pre added. If not I wonder
if it reproduces with -fdisable-tree-fre2 or at least -fno-tree-tail-merge.
AFAIK x86_64-darwin uses SSE math even with -m32, correct? On x86_64-linux
I can't fully get rid of the use of the FP stack (library calls).
