[PATCH] D109067: [Analysis][SVE] Make the costs for gathers/scatters/ordered reductions less pessimistic

[Sander de Smalen via Phabricator via llvm-commits]

sdesmalen added inline comments.


================
Comment at: llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h:142
 
-    unsigned MaxNumVScale = 16;
+    unsigned MaxNumVScale = DefaultMaxVscale;
     if (F && F->hasFnAttribute(Attribute::VScaleRange)) {
----------------
In practice the vscale_range attribute is added by Clang to all functions now, so this change is artificial in that it only changes the cost for the unit-tests (which don't specify vscale_range).

I think what you're actually after is a function that returns the median value between min and max, e.g. for a vscale_range(0, 16) it chooses 8, and and for vscale_range(8, 8) it also chooses 8. It would be better if we don't start tuning for specific bit-widths based on implementations that are available today.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109067/new/

https://reviews.llvm.org/D109067