[PATCH] D79357: [SveEmitter] Add builtins for svdup and svindex
Sander de Smalen via Phabricator via cfe-commits
cfe-commits at lists.llvm.org
Tue May 5 13:31:12 PDT 2020
sdesmalen marked 2 inline comments as done.
sdesmalen added inline comments.
================
Comment at: clang/include/clang/Basic/arm_sve.td:1050
def SVCOMPACT : SInst<"svcompact[_{d}]", "dPd", "ilUiUlfd", MergeNone, "aarch64_sve_compact">;
-// SVDUP_LANE (to land in D78750)
+def SVDUP_LANE : SInst<"svdup_lane[_{d}]", "ddL", "csilUcUsUiUlhfd", MergeNone, "aarch64_sve_tbl">; // Implemented using tbl
def SVDUPQ_LANE : SInst<"svdupq_lane[_{d}]", "ddn", "csilUcUsUiUlhfd", MergeNone, "aarch64_sve_dupq_lane">;
----------------
efriedma wrote:
> "Implemented using tbl" isn't really a helpful comment.
>
> I guess you're doing it this way so the lane doesn't have to be a constant? (And then I guess you can optimize it in the backend.)
Yes that's right. I can probably clarify the comment a bit more.
================
Comment at: clang/lib/CodeGen/CGBuiltin.cpp:8047
+ Value *PFalse = Constant::getNullValue(PTrue->getType());
+ Value *Sel = Builder.CreateSelect(CmpNE, PTrue, PFalse);
+ return EmitSVEPredicateCast(Sel, cast<llvm::ScalableVectorType>(Ty));
----------------
efriedma wrote:
> The select might not be the best approach. I mean, it's semantically correct, but the backend currently doesn't know how to lower a select like that. And even if it did, it's probably better to avoid generating it for now.
>
> Maybe this should use whilelo directly, since that's what the backend would eventually use anyway.
> but the backend currently doesn't know how to lower a select like that
Yes I noticed, that was the reason for D79356 (added as parent revision)
> And even if it did, it's probably better to avoid generating it for now.
Is that because there are common optimisations on the select? I thought it would be safe to use the 'full vector' form of `select` where the predicate is a scalar (and both operands are splats)
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D79357/new/
https://reviews.llvm.org/D79357
More information about the cfe-commits
mailing list