On 12/09/13 16:18, Eric Botcazou wrote:
It's the pessimization introduced on the ARM (and other RISC targets) by the
distribution of truncations in simplify_truncation. Further information at:
http://gcc.gnu.org/ml/gcc/2013-12/msg00019.html
The change started as a simple address reassociation trick for x32 and evolved
into a general distribution of truncations for common operations. But the
outcome doesn't really play nice with other transformations that help on RISC
targets so I think that we should restrict it. I'm not sure whether it really
helps on CISC targets either, but it's at least neutral and it's too late to
backpedal anyway since further changes were installed in the x86 back-end.
Hence the proposed minimal patch based on WORD_REGISTER_OPERATIONS. Not clear
whether it's the best long term solution, but I think that we want to fix the
regression on the 4.8 branch as well.
Tested on x86-64/Linux, PowerPC/Linux and IA-64/Linux. As reported by Kyrill,
this fixes gcc.target/arm/unsigned-extend-1.c on the ARM.
Comments?
2013-12-09 Eric Botcazou <ebotca...@adacore.com>
PR rtl-optimization/58295
* simplify-rtx.c (simplify_truncation): Restrict the distribution for
WORD_REGISTER_OPERATIONS targets.
Seems like a reasonable way to disable until we sort how best to deal
with this in the long term.
jeff
