[PATCH] D118617: [AArch64][SVE] Remove false register dependency for unary FP convert operations

[Sander de Smalen via Phabricator via llvm-commits]

sdesmalen added inline comments.


================
Comment at: llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td:1676
 
-  def : Pat<(nxv2f32 (AArch64fcvte_mt (nxv2i1 PPR:$Pg), (nxv2f16 ZPR:$Zs), (nxv2f32 ZPR:$Zd))),
-            (FCVT_ZPmZ_HtoS ZPR:$Zd, PPR:$Pg, ZPR:$Zs)>;
+  def : Pat<(nxv2f32 (AArch64fcvte_mt (nxv2i1 (SVEAllActive):$Pg), (nxv2f16 ZPR:$Zs), (nxv2f32 ZPR:$Zd))),
+            (FCVT_ZPmZ_HtoS_UNDEF ZPR:$Zd, PPR:$Pg, ZPR:$Zs)>;
----------------
Shouldn't we retain the original pattern in case $Pg is not all active?


================
Comment at: llvm/test/CodeGen/AArch64/sve-fpext-load.ll:50
 ; CHECK-NEXT:    uunpkhi z4.d, z0.s
+; CHECK-NEXT:    movprfx z0, z2
 ; CHECK-NEXT:    fcvt z0.d, p0/m, z2.h
----------------
I see how this can be an improvement, but it's also a bit of a hammer. I would at least have expected a check somewhere to ask if a MOVPRFX is free/cheap, is it worth adding that?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D118617/new/

https://reviews.llvm.org/D118617