x86 PIC regressions

Robert Lipe robertl@dgii.com
Fri Sep 4 14:22:00 GMT 1998

Richard Henderson wrote:
> On Wed, Sep 02, 1998 at 02:07:15PM -0500, Robert Lipe wrote:
> > It now fails to bootstrap for me on either i686-pc-linux-gnu or
> > i686-pc-sco3.2v5.0.5.   Both targets abort in the same place.  Reverting
> > Richard's patch from this morning allows the bootstrap to finish.
> What a pain in the ass.  Push the bubble down here, and it pops
> up there.  With this one, I don't see any regressions; Uli is
> running glibc tests for me now -- we'll see what happens there.

Sorry. I'm just the messenger. :-) If it helps, I do appreciate your work.

> 	* i386.h (PREFERRED_RELOAD_CLASS): i387 constants get FP_TOP_REG.
> 	* i386.md (movsf, movdf, movxf): Call validize_mem to clean up
> 	force_const_mem output for -fpic.  Do not try to emit REG_NOTES.
> 	During reload, fix up not-quite-correct constant pool addresses.

With this installed, the compiler will indeed bootstrap.  The GCC
testsuite gives only believable failures on OpenServer in all of
{coff,elf,elf+pic} modes.

There is something bizarre going on in g++ land, but my gut tells me
this isn't related to your patch. I think this is just more dwarf-isms
becuase the same code on Linux works.  I probably won't get to
investigate this today, but will send that under a different subject
to better catch the attention of the g++ crew.

Thanx, Richard!


More information about the Gcc mailing list