This currently breaks some tests because currently hex constants are treated as floating point. I'm not sure what change should be made to support both, but I think it would be greatly preferable to switch to normal hex constants to be integers and only use c99 hex syntax for FP immediates
Details
Diff Detail
Event Timeline
Hi Matt,
As far as I remember the lane mask in live-ins are only supported when subreg liveness is enabled.
Therefore, I believe we should either force the target to track subreg liveness when we see such construct or reject the entry is the target does not track subreg liveness.
What do you think?
Cheers,
-Quentin
This currently breaks some tests because currently hex constants are treated as floating point. I'm not sure what change should be made to support both, but I think it would be greatly preferable to switch to normal hex constants to be integers and only use c99 hex syntax for FP immediates
I would have wanted hex integers to fix the branch probability printing as well. However this looks tricky to solve.
I think C99 hex floats cannot represent all corner cases correctly (NaNs with payloads). However as far as I can see all instances where we want to parse hexfloats are prefixed with a keyword so we could just lex to a "hexnumber" and then interpret that as an integer or hexfloat based on context.