Optimize Vector64.Create for constants #101662
Conversation
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
PTAL @kunalspathak @tannergooding cc @dotnet/jit-contrib Empy size diffs, 283 contexts with text diffs (becuase size is the same) |
| @@ -2406,14 +2406,36 @@ void CodeGen::genSetRegToConst(regNumber targetReg, var_types targetType, GenTre | |||
| } | |||
| else | |||
| { | |||
| // Get a temp integer register to compute long address. | |||
| regNumber addrReg = tree->GetSingleTempReg(); | |||
| simd8_t val = vecCon->gtSimd8Val; | |||
There was a problem hiding this comment.
The same handling should presumably be added for TYP_SIMD16 as well, yes?
It should also be fine to expose for TYP_SIMD12, since any operations are already required to mask out the unused bit when its impactful.
There was a problem hiding this comment.
I also wonder if this is missing some cases or potentially conflicting with the IsValidConstForMovImm in LowerHWIntrinsicCreate handling
That check is supposed to recognize and catch this case, lowering it instead to Arm64.DuplicateToVectorXXX, which is itself supposed to emit the more optimized mov instruction
It might be a case where we need both since sometimes we'll have Create(...) and sometimes we'll have CNS_VEC isntead.
There was a problem hiding this comment.
Oh, good point, I'll check what we generate for other VectorX.Create
Closes #101661