================
@@ -708,7 +708,7 @@ AArch64TargetInfo::getVScaleRange(const LangOptions
&LangOpts) const {
return std::pair<unsigned, unsigned>(
LangOpts.VScaleMin ? LangOpts.VScaleMin : 1, LangOpts.VScaleMax);
- if (hasFeature("sve"))
+ if (hasFeature("sve") || hasFeature("sme"))
----------------
sdesmalen-arm wrote:
If the function only has `+sme`, then `vscale_range` should only apply to
(locally) streaming functions, because non-streaming or streaming-compatible
functions cannot use scalable vectors.
I guess the the thing to watch out for is that we don't propagate assumptions
from using `-msve-vector-bits` to streaming[-compatible] functions, because
that information may be incorrect.
The SME attributes tell which mode the statements of a function must be
executed in, so `vscale_range` should apply to all LLVM IR operations in the
body of that function. For targets that have both SVE and SME, this may no
longer be true during CodeGen because the DAG introduces streaming-mode changes
for calls, and in the prologue/epilogue of locally-streaming functions. If it
assumes a certain vector width for the body, the DAG might end up with
instructions that assume this width for operations executed in a different
mode. I've not really thought this through, but it does need further
investigation if we ever want to enable that. Anyway, that doesn't really apply
to this patch.
https://github.com/llvm/llvm-project/pull/124466
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
