[PATCH] FIx bug in x87 FP stackifier pass
bob.wilson at apple.com
Fri Jul 18 15:04:47 PDT 2014
Ping (on behalf of Akira).
I’ve looked over the patch and it seems like a very nice improvement to the way we handle x87 registers. Unfortunately I have very little experience in this area and don’t feel qualified to review it.
Can someone else more familiar with x87 stuff please review it?
> On Jul 1, 2014, at 4:55 PM, Akira Hatanaka <ahatanak at gmail.com> wrote:
> The x87 FP stackifier pass has a bug where the liveness of an ST register defined by an inline-asm instruction is not tracked precisely.
> This causes the assertion on line 1365 to fail when an ST register defined by an inline-asm is live across another inline-asm, as shown in the following example:
> INLINEASM <es:frndint> [sideeffect] [attdialect], $0:[regdef], %ST0<imp-def,tied5>, $1:[reguse tiedto:$0], %ST0<tied3>, $2:[clobber], %EFLAGS<earlyclobber,imp-def,dead>
> INLINEASM <es:fldcw $0> [sideeffect] [mayload] [attdialect], $0:[mem], %EAX<undef>, 1, %noreg, 0, %noreg, $1:[clobber], %EFLAGS<earlyclobber,imp-def,dead>
> %FP0<def> = COPY %ST0 // %ST0 is defined by the first inline-asm instruction.
> In order to fix this bug, the attached patch uses FP registers for function returns and inline-asm instructions where previously ST registers were being used and adds kill flags that are missing in FP register operands.
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-commits