On Wed, Jan 16, 2019 at 10:26 AM Martin Liška <mli...@suse.cz> wrote: > > And there's patch with Richi's validation check that he provided. > It fails on following 2 tests in test-suite: > > $ ./xgcc -B. > /home/marxin/Programming/gcc/gcc/testsuite/gcc.target/i386/pr68674.c > DECL_MODE BLK vs TYPE_MODE V8SI [V8SI]: a > ... > > $ ./xgcc -B. > /home/marxin/Programming/gcc/gcc/testsuite/gcc.target/i386/pr80583.c -c > DECL_MODE BLK vs TYPE_MODE V8SI [V8SI]: a > ... > > In both cases we access a global variable from a function with a different > target > attributes. I guess Honza has seen that in inliner. > > What to do with these, can it be potentially dangerous?
I guess so. There's not much we can do about this other than making DECL_MODE dynamic the same way as TYPE_MODE. I still believe this is the wrong direction and instead RTL expansion should properly adjust DECL_RTL and generally on GIMPLE we shouldn't need to commit to modes early at all... [but then we eventually want to expose more ABI details to GIMPLE] Richard. > > Thanks, > Martin >
