[PATCH] D82458: [ARM] Adjust default fp extend and trunc costs

[Sjoerd Meijer via Phabricator via llvm-commits]

SjoerdMeijer added inline comments.


================
Comment at: llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp:253
+  // Single to/from double precision conversions.
+  if (Src->isVectorTy() && ST->hasNEON() &&
+      ((ISD == ISD::FP_ROUND && SrcTy.getScalarType() == MVT::f64 &&
----------------
Quickly looking at this, I was wondering 2 things here:
- how about MVE?
- and how about f16?


================
Comment at: llvm/test/Analysis/CostModel/ARM/cast.ll:134
 ; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %r84df = fptrunc <16 x double> undef to <16 x float>
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %r80dh = fptrunc double undef to half
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %r81dh = fptrunc <2 x double> undef to <2 x half>
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %r82dh = fptrunc <4 x double> undef to <4 x half>
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %r83dh = fptrunc <8 x double> undef to <8 x half>
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %r84dh = fptrunc <16 x double> undef to <16 x half>
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %r80fh = fptrunc float undef to half
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %r81fh = fptrunc <2 x float> undef to <2 x half>
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %r82fh = fptrunc <4 x float> undef to <4 x half>
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 14 for instruction: %r83fh = fptrunc <8 x float> undef to <8 x half>
-; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 28 for instruction: %r84fh = fptrunc <16 x float> undef to <16 x half>
+; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %r80dh = fptrunc double undef to half
+; CHECK-NEON-RECIP-NEXT:  Cost Model: Found an estimated cost of 20 for instruction: %r81dh = fptrunc <2 x double> undef to <2 x half>
----------------
I can't remember if Sam asked the same question on a different patch, but the cost looks a lot higher than the unchanged instructions here. Is that correct?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D82458/new/

https://reviews.llvm.org/D82458