This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Relax check against commuting XOR and ASHIFTRT in combine.c

Rainer Orth wrote:
However, as a quick first step, does adding the ilp32 / lp64 (and keeping
the architectures list for now) solve the immediate problem? Patch
attached, OK for trunk?

No, as I said this is wrong for biarch targets like sparc and i386.

When you say no this does not solve the immediate problem, are you saying that you are (still) seeing test failures with the require-effective-target patch applied? Or is the issue that this would not execute the tests as widely as might be possible? In principle I'm quite happy to relax the target patterns, although have been having issues with sparc (below)...

Re. "what the architectures have in common" is largely that these are the primary/secondary archs on which I've checked the test passes! I can now add mips and microblaze to this list, however I'm nervous of dropping the target entirely given the very large number of target architectures gcc supports; and e.g. IA64 (in ILP32 mode) generates an ashiftrt:DI by 31 places, not ashiftrt:SI, which does not match the simplification criteria in combine.c.

This should be something like
  { target aarch64*-*-* i?86-*-* powerpc*-*-* sparc*-*-* x86_64-*-* }

E.g. sparc-sun-solaris2.11 with -m64 is lp64, but would be excluded by
your target list.  Keep the list sorted alphabetically and best add an
explanation so others know what those targets have in common.

So I've built a stage-1 compiler with --target=sparc-sun-solaris2.11, and I find

* without -m64, my "dg-require-effective-target ilp32" causes the 32-bit test to execute, and pass; "dg-require-effective-target lp64" prevents execution of the 64-bit test (which would fail) - so all as expected and desired.

  * with -lp64, behaviour is as previous (this is probably expected)

* with -m64, "dg-require-effective-target ilp32" still causes the test to execute (but it fails, as the RTL now has an ashiftrt:DI by 31 places, which doesn't meet the simplification criteria in combine.c - this is pretty much as expected). "dg-require-effective-target lp64" stops the 64-bit test from executing however (despite that it would now pass).

Can you clarify what I should be doing on sparc, therefore?

Thanks for your help!


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]