Created attachment 2916 [details] Testcase (bitcode) Testcase attached; to reproduce, use the following command, at least on my computer: llc x2.bc -o - | fgrep "REGSA: %eax %eax" I'm on Ubuntu Linux on x86-32 using a Core Duo (hopefully, that doesn't make a difference). The relevant inline asm is as follows: %0 = call i32 asm sideeffect "REGSA: $0 $1 $2", "=&{ax},r,r,~{cx},~{bx},~{dx},~{si},~{memory},~{dirflag},~{fpsr},~{flags}"(i8* %arrayidx.i.i, %struct.CABACContext* %tmp3.i.i) nounwind ; <i32> [#uses=1] Unfortunately, it's difficult to simplify a testcase like this; the construct in question isn't generally broken, just not working in this particular case. Testcase reduced from libavcodec/h264.c in ffmpeg.
This is totally being ignored by me. :-( Dale, do you have a bit of time to dig into this?
Not now, I'm going on vacation as of tomorrow, remember?
Ok, when you get back then. :-) Thx.
Dale, any update? This may be related to PR 4964 where two outputs are getting assigned to the same register.
No, I've been busy with Apple bugs. Maybe next week if Evan thinks this is more important than the Apple bugs in my queue.
This is not the same as 4964; I fixed that one, which was simpler than this one.
*** Bug 6539 has been marked as a duplicate of this bug. ***
Any update? This has been affecting ffmpeg for ages.
106816 should fix this.