Track bool-returning and scalar-T-returning HW intrinsics via flags#128848
Track bool-returning and scalar-T-returning HW intrinsics via flags#128848Copilot wants to merge 5 commits into
Conversation
Co-authored-by: tannergooding <10487869+tannergooding@users.noreply.github.com>
|
CC. @EgorBo, the downside of this is that we were out of flag space and so it extends it up to I think there's likely cleanup possible and we can free up space, but I expect that's more involved and requires careful thought. |
There was a problem hiding this comment.
Pull request overview
This PR centralizes knowledge about certain hardware intrinsics that return either a boolean (conceptually TYP_INT in range [0, 1]) or a scalar of the SIMD base type, by encoding that information as flags in the HW intrinsic table and switching consumers over to flag queries.
Changes:
- Widen
HWIntrinsicFlagtouint64_tand add common flags forReturnsBooleanandReturnsScalarT, exposed viaHWIntrinsicInfo. - Tag relevant xarch/arm64 intrinsic table entries with the new return-shape flags.
- Refactor
assertionprop.cpp,rangecheck.cpp, andvaluenum.cppto queryHWIntrinsicInfo::{ReturnsBoolean,ReturnsScalarT}instead of maintaining long intrinsic-ID switches.
Reviewed changes
Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.
Show a summary per file
| File | Description |
|---|---|
| src/coreclr/jit/valuenum.cpp | Uses HW intrinsic return-shape flags to simplify IsVNNeverNegative handling for HWI VNs. |
| src/coreclr/jit/rangecheck.cpp | Uses HW intrinsic return-shape flags to derive ranges for HWI VNs without ID enumeration. |
| src/coreclr/jit/assertionprop.cpp | Uses HW intrinsic return-shape flags to derive symbolic integer ranges for GT_HWINTRINSIC. |
| src/coreclr/jit/hwintrinsiclistxarch.h | Tags xarch intrinsics (equality/test/compare, GetElement/ToScalar/Extract) with new return-shape flags. |
| src/coreclr/jit/hwintrinsiclistarm64.h | Tags arm64 intrinsics (equality, GetElement/ToScalar/Extract) with new return-shape flags. |
| src/coreclr/jit/hwintrinsic.h | Widens flag type and adds ReturnsBoolean / ReturnsScalarT APIs on HWIntrinsicInfo. |
Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>
…heck Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>
| } | ||
|
|
||
| #if defined(TARGET_XARCH) | ||
| case VNF_HWI_AVX2_LeadingZeroCount: |
There was a problem hiding this comment.
@EgorBo do you think it's worth handling these other VNF_HWI_* cases under the IsVNHWIntrinsicFunc path, seeing as they can only trigger for that anyways?
Probably doesn't make a big difference here, but at least keeps the hwintrinsic code grouped and the main switch case smaller?
No strong opinion, just something that is possible now that we have such an explicit check for the other hwintrinsic handling.
…unc blocks Co-authored-by: tannergooding <10487869+tannergooding@users.noreply.github.com>
Several hardware intrinsics are known to return
boolor a scalarT, but consumers (assertionprop, rangecheck, valuenum) tracked them via long ID switches that had to be kept in sync. Replace those with flags on the intrinsic table queryable throughHWIntrinsicInfo.hwintrinsic.hHWIntrinsicFlagtouint64_t(arm64 was out of bits in the 32-bit space).HW_Flag_ReturnsBooleanandHW_Flag_ReturnsScalarTas common flags.HWIntrinsicInfo::ReturnsBoolean/ReturnsScalarTstatic APIs, grouped afterReturnsPerElementMask.hwintrinsiclist*.hTag the entries previously enumerated in the consumer switches (skipping
HW_Flag_InvalidNodeIdentries per the issue):ReturnsBoolean(Vector{128,256,512}_op_{Equality,Inequality},X86Base.CompareScalar{Ordered,Unordered}*,X86Base/AVXTest{C,Z,NotZAndNotC}) and 8ReturnsScalarT(Vector{128,256,512}.{GetElement,ToScalar},X86Base[_X64].Extract).ReturnsBoolean(Vector{64,128}_op_{Equality,Inequality}) and 5ReturnsScalarT(Vector{64,128}.{GetElement,ToScalar},AdvSimd.Extract).Consumer refactors
assertionprop.cpp,rangecheck.cpp, andvaluenum.cppno longer enumerate intrinsic IDs in theirGT_HWINTRINSIC/VNF_HWI_*switches. A flag check runs before the remaining switch and preserves prior semantics, e.g.:For
IsVNNeverNegativethe scalar-T path keeps its base-type guard (onlyTYP_UBYTE/TYP_USHORTcontinue), matching the previous case-by-case behavior. The HW intrinsic checks inIsVNNeverNegativeandrangecheck.cppnow useIsVNHWIntrinsicFunc, eliminating the manualVNF_HWI_FIRST/VNF_HWI_LASTrange check, the explicitNamedIntrinsiccast, the separateGetVNHWIntrinsicSizeAndBaseTypecall. BothIsVNHWIntrinsicFuncblocks are correctly wrapped in#if defined(FEATURE_HW_INTRINSICS)/#endifguards.SVE bool/scalar-T intrinsics are intentionally not tagged here — they were never in the original switches and remain covered by the existing
TODO-SVEnotes.