[Bug rtl-optimization/49095] Horrible code generation for trivial decrement with test

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49095

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-05-27 16:33:33 UTC ---
.text sizes before/after the patch (in each case on identical sources, for
cc1plus I've reverted the patch afterwards and did make -j64 cc1plus in gcc/
subdir), the size changes aren't that big, but perhaps other projects use it
more often and/or something is hidden due to alignment.

        32-bit before    32-bit after    64-bit before    64-bit after
libstdc++.so.6    0x71708        0x716e8        0x67ee6        0x67ec6
libgcj.so.12    0xa3ec58    0xa3eb98    0x90cd68    0x90cce8
cc1plus        0xe8a29c    0xe8a25c    0xdccf98    0xdccf08

In 64-bit cc1plus it only made a difference in:
c-decl.o
cfg.o
dbxout.o
ebitmap.o
final.o
ira-color.o
jump.o
regcprop.o
reg-stack.o
reload.o
tree-ssa-dom.o
tree-ssa-loop-ivopts.o

Anyway, I'll post the patch now and see what Uros says about it.