JIT: Build and consume FMA and Permute node operands in standard order #103100
Conversation
The building and consumption of these operands can happen in op1, op2, op3 order regardless of whether the codegen uses the registers in a different order. Fix dotnet#102773
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
| { | ||
| srcCount += op->isContained() ? BuildOperandUses(op) : BuildDelayFreeUses(op, emitOp1); | ||
| } | ||
| } |
There was a problem hiding this comment.
This could probably be cleaned up to just build the uses directly without first figuring out emitOp1, emitOp2, emitOp3, but I didn't really understand all the constraints around copiesUpperBits, the preferred target and the contained operand, so I gave up on trying to do it.
|
cc @dotnet/jit-contrib PTAL @kunalspathak @tannergooding This should hopefully fix the failures for good this time. |
|
@jakobbotsch there's another case that is doing something similar with operand switching for |
|
We also, more generally have various commutative nodes that also dynamically pick the |
Ah, good catch, same thing should be applied there. Pushed it to this PR.
Yeah, that looks fine. Seems likely we could try to clean up/unify some of these cases. |
jakobbotsch
changed the title
The building and consumption of these operands can happen in op1, op2, op3 order regardless of whether the codegen uses the registers in a different order.
Fix #102773