[PATCH] D95458: [PowerPC] Exploit xxsplti32dx (constant materialization) for scalars

Stefan Pintilie via Phabricator via cfe-commits cfe-commits at lists.llvm.org
Tue Feb 16 13:08:57 PST 2021


stefanp requested changes to this revision.
stefanp added inline comments.
This revision now requires changes to proceed.


================
Comment at: llvm/lib/Target/PowerPC/PPCISelLowering.h:1321
   bool convertToNonDenormSingle(APFloat &ArgAPFloat);
+  bool checkNonDenormCannotConvertToSingle(APInt &ArgAPInt);
+  bool checkNonDenormCannotConvertToSingle(APFloat &ArgAPFloat);
----------------
Is the APInt version of this function used anywere?



================
Comment at: llvm/lib/Target/PowerPC/PPCInstrInfo.td:412
+  APFloat APFloatOfN = N->getValueAPF();
+  checkNonDenormCannotConvertToSingle(APFloatOfN);
+  uint32_t Hi = (uint32_t)((APFloatOfN.bitcastToAPInt().getZExtValue() &
----------------
Why are we running this here?
We don't check the return of the function so we must assume that it returns true.
In that case the value of `APFloatOfN` won't change because `convertToNonDenormSingle` will only change the value of the parameter if it returns true. But `checkNonDenormCannotConvertToSingle` only returns true if `convertToNonDenormSingle` return false.


================
Comment at: llvm/test/CodeGen/PowerPC/constant-pool.ll:44
+; CHECK-NEXT:    xxsplti32dx vs1, 0, 56623104
+; CHECK-NEXT:    xxsplti32dx vs1, 1, -609716532
+; CHECK-NEXT:    # kill: def $f1 killed $f1 killed $vsl1
----------------
I'm looking to understand this test case.
We are trying to materialize a special PowerPC long double (double-double). It seems that we have materialized one half of it and not the other half.

Is it because the first half is a denormal?
Why are we avoiding denormals anyway? It seems like we can completely specify a 64 bit double with two `xxsplti32dx` instructions. 


================
Comment at: llvm/test/CodeGen/PowerPC/constant-pool.ll:363
+; CHECK-NEXT:    stxv vs3, 32(r1) # 16-byte Folded Spill
+; CHECK-NEXT:    xxsplti32dx vs3, 1, -343597384
+; CHECK-NEXT:    # kill: def $f3 killed $f3 killed $vsl3
----------------
What is going on here?
It almost looks like we are spilling `vs3` half way through materializing a constant.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D95458/new/

https://reviews.llvm.org/D95458



More information about the cfe-commits mailing list