Created attachment 3272 [details] Reduced testcase Per summary, incorrect error; to reproduce, run llc (with no special arguments) over attached. Seems to work with "-regalloc=local". Testcase reduced from ffmpeg (the source of all fun inline asm issues).
This works for me (not my doing). I've put the example in the testsuite to keep it from regressing; this is a fragile area.
The original is still breaking; I'll attach it in a moment.
Created attachment 3295 [details] Re-reduced testcase
It seems to think the EDX result from the first asm, which is never used, is still live across the second asm.
The reduced test case is going to fail with machine cse. The fundamental problem is we can't spill a register once it's fixed to a physical register.
This is still affecting ffmpeg... is it on someone's todo list?
Does the problem persist with -disable-physical-join?
Created attachment 5430 [details] Re-re-reduced testcase These bugpoint-reduced testcases have a tendency to get out-of-date...
And yes, -disable-physical-join fixes the issue.
Still reproducable on trunk.
This is likely to have been fixed by disabling physical register coalescing, but I can no longer make the test case fail. Please verify with trunk.