On Thu, Oct 8, 2015 at 12:36 AM, Alexandre Oliva <aol...@redhat.com> wrote: > On Sep 29, 2015, Szabolcs Nagy <szabolcs.n...@arm.com> wrote: > >> this commit > >> commit 33cc9081157a8c90460e4c0bdda2ac461a3822cc >> Author: aoliva <aoliva@138bc75d-0d04-0410-961f-82ee72b054a4> >> Date: 2015-09-27 09:02:00 +0000 > >> revert to assign_parms assignments using default defs >> ... > >> introduced a test failure on arm-none-eabi (using newlib, compiling >> with -mthumb -march=armv8-a -mfpu=crypto-neon-fp-armv8 -mfloat-abi=hard ): > >> FAIL: gcc.target/arm/pr43920-2.c scan-assembler-times pop 2 > > Thanks for the report. > > The problem here is that we don't allocate the pseudo assigned to > <retval> to r0. That's because we coalesce <retval> versions with > another variable that crosses a function call. We do that because > uncprop brings these unrelated variables, that happen to contain the > same -1 value we want to return, into the PHI node with the final > <retval> value. > > We can't coalesce both start and end with <retval>, because start and > end conflict, but by chance we try start first, and that succeeds. If > we tried end first (e.g., by giving it a higher coalesce priority, > because fewer calls are crossed by its value in the path to the relevant > edge), we could have got the coalesced variable assigned to r0, and that > would enable us to optimize out the copy to r0 before return, and so > merge the return-only basic block with other blocks. But ATM we don't > take the definition point or path to the edge into account when > computing coalesce costs, so we can't deterministically do better for > this testcase, and I'm not sure using these additional information would > make it better overall. > > Compiling with -fno-tree-dominator-opts skips uncprop so that we don't > even try to coalesce other variables with <retval>, so we get the code > expected by the testcase. But we obviously don't want to disable this > optimization in general. > > Any other thoughts, anyone?
Bad luck? Add some heuristics that always help? Ok, that wasn't really useful :/ Richard. > > -- > Alexandre Oliva, freedom fighter http://FSFLA.org/~lxoliva/ > You must be the change you wish to see in the world. -- Gandhi > Be Free! -- http://FSFLA.org/ FSF Latin America board member > Free Software Evangelist|Red Hat Brasil GNU Toolchain Engineer
