http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54317
--- Comment #2 from dave.anglin at bell dot net 2012-08-20 16:23:57 UTC --- On 8/20/2012 11:46 AM, glisse at gcc dot gnu.org wrote: > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54317 > > --- Comment #1 from Marc Glisse <glisse at gcc dot gnu.org> 2012-08-20 > 15:46:58 UTC --- > Wow, I didn't expect that patch to break a multiplication test... > It sounds like you have before and after compilers. Do you have tree-vrp dumps > from both? (I would ask if a stage1 compiler fails too, to rule out > miscompilation of the compiler, but I have no idea how ada works...) At the moment, I have no idea as to which module is broken but I believe the problem is in the Ada runtime library. I compared the .o files for the first test of the four failing tests using before and after compilers and they were identical. The ada tests are not compiled with debugging enabled, so they are a bit difficult to debug. The tests invoke a lot of arithmetic operations... > Is hwint 32 bits on this platform? I am looking for acats testresults from > other such platforms in gcc-testresults but can't find them... > Yes, hwint is 32 bits. I just fixed a couple of issues with expand_mult (PR middle-end/53823). The synth_mult code can't handle multiplication by negative coefficients when the mode is larger than a hwint. One possibility might be that this code is being invoked by another path.
