Fold SIMD-typed partial access to locals; delete base types from vector constants#74580
Conversation
ghost
added
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsEnabling the folding of We have positive diffs due to more CSEs, some light regressions due to more CSEs, and a couple large regressions due to more unrolling ( Closes #70052.
|
SingleAccretion
force-pushed
the
LclMorph-Simd-Use
branch
from
58307dd to
e2bbb55
Compare
|
@dotnet/jit-contrib |
SingleAccretion
force-pushed
the
LclMorph-Simd-Use
branch
from
e2bbb55 to
419de60
Compare
To be used for nodes which don't have a handle of thier own.
SingleAccretion
force-pushed
the
LclMorph-Simd-Use
branch
from
419de60 to
b402708
Compare
| CORINFO_CLASS_HANDLE* pCanonicalHnd = nullptr; | ||
| switch (size) | ||
| { | ||
| case 8: | ||
| pCanonicalHnd = &m_simdHandleCache->CanonicalSimd8Handle; | ||
| break; | ||
| case 12: | ||
| // There is no need for a canonical SIMD12 handle because it is always Vector3. | ||
| break; | ||
| case 16: | ||
| pCanonicalHnd = &m_simdHandleCache->CanonicalSimd16Handle; | ||
| break; | ||
| case 32: | ||
| pCanonicalHnd = &m_simdHandleCache->CanonicalSimd32Handle; | ||
| break; | ||
| default: | ||
| unreached(); | ||
| } | ||
|
|
||
| if ((pCanonicalHnd != nullptr) && (*pCanonicalHnd == NO_CLASS_HANDLE)) | ||
| { | ||
| *pCanonicalHnd = typeHnd; | ||
| } |
There was a problem hiding this comment.
So essentially the first SIMD handle we see in this function becomes the canonical handle for this compilation?
There was a problem hiding this comment.
Yep. It's a workaround of course, but doing this "properly" would imply breaking the assumption that (most / all that are looked at) varTypeIsStruct locals have handles. Which is work that will be done at some point, but not now.
|
/azp run runtime-coreclr superpmi-diffs |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
Failures are #76041 according to build analysis. |
…rom vector constants (dotnet#74580)" This reverts commit 78443b8.
ghost
locked as resolved and limited conversation to collaborators
3 participants
Enabling the folding of
SIMD-typed indirs in local morph, it was observed that some CSEs were getting lost. The reason was that we always usedFLOAT-derived handles forTYP_SIMDIND/LCL_FLDnodes, which would not be available in methods that only usedVector128<int>(for example). This change enhances the "canonical handle" logic to capture handles of other types to use for handle-less nodes.We have positive diffs due to more CSEs, some light regressions due to more CSEs, and a couple large regressions due to more unrolling (
IND<simd>appears cheaper togtSetEvalOrderthanOBJ<simd>).Closes #70052.