On 7/6/23 6:28 PM, Segher Boessenkool wrote: > On Thu, Jul 06, 2023 at 02:48:19PM -0500, Peter Bergner wrote: >> On 7/6/23 12:33 PM, Segher Boessenkool wrote: >>> On Wed, Jul 05, 2023 at 05:21:18PM +0530, P Jeevitha wrote: >>>> --- a/gcc/config/rs6000/rs6000.cc >>>> +++ b/gcc/config/rs6000/rs6000.cc >>>> @@ -9894,6 +9894,8 @@ rs6000_legitimate_address_p (machine_mode mode, rtx >>>> x, bool reg_ok_strict) >>>> >>>> /* Handle unaligned altivec lvx/stvx type addresses. */ >>>> if (VECTOR_MEM_ALTIVEC_OR_VSX_P (mode) >>>> + && mode != OOmode >>>> + && mode != XOmode >>>> && GET_CODE (x) == AND >>>> && CONST_INT_P (XEXP (x, 1)) >>>> && INTVAL (XEXP (x, 1)) == -16) >>> >>> Why do we need this for OOmode and XOmode here, but not for the other >>> modes that are equally not allowed? That makes no sense. >> >> VECTOR_MEM_ALTIVEC_OR_VSX_P (mode) already filters those modes out >> (eg, SImode, DFmode, etc.), just not OOmode and XOmode, since those both >> are modes used in/with VSX registers. > > It does not filter anything out, no. That simply checks if a datum of > that mode can be loaded into vector registers or not. For example > SImode could very well be loaded into vector registers! (It just is not > such a great idea).
I spent some time looking at how the compiler fixes this up in the -mno-block-ops-vector-pair case and I see the constraints used in the vsx_mov<mode>_64bit pattern for loads and stores disallows these types of addresses, so LRA fixes them up for us. Clearly movoo should do the same and that is enough to fix the ICE. I'll work with Jeevitha on submitting a patch using that solution. That said, I think it would be good to modify rs6000_legitimate_address_p to disallow these altivec style addresses for OOmode and XOmode, since we know early-on that they're not going to be valid, but that would be a different patch. > dg-do compile *does* invoke the assembler, btw. As it should. There is dg-do "preprocess", "compile", "assemble", "link" and "run" (ignoring "precompile" and "repo"). Dg-do compile produces an assembly file, but doesn't actually call the assembler, which we don't strictly need for a test case that checks whether GCC ICEs or not. If you want to run the assembler too and then stop, then you'd want dg-do assemble. Peter
