[PATCH] D105525: [GISel] Add fpext/fptrunc combines

[Jay Foad via Phabricator via llvm-commits]

foad added a comment.

> Removed the combine applying when unsafe math is enabled.

Thanks. Please remove it from the commit message too.



================
Comment at: llvm/include/llvm/Target/GlobalISel/Combine.td:476
+
+// Fold (fptrunc ([su]itofp x)) -> ([su]itofp x)
+def fptrunc_int_to_float: GICombineRule<
----------------
Is this safe? E.g. does converting i32->f32->f16 always give the same result as i32->f16, even though the former does a kind of double rounding (from 32 bits to 24 bit mantissa to 11 bit mantissa)? Does SelectionDAG do the equivalent optimization?


================
Comment at: llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp:2467
+  unsigned SrcOpc = SrcMI->getOpcode();
+  if (SrcOpc == TargetOpcode::G_SITOFP || SrcOpc == TargetOpcode::G_UITOFP) {
+    Register SrcInputReg = SrcMI->getOperand(1).getReg();
----------------
Could return early here.


================
Comment at: llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp:2490
+  LLT SrcInputTy = MRI.getType(SrcInputReg);
+  if (SrcInputTy.getSizeInBits() == 8) {
+    MatchInfo = [=](MachineIRBuilder &B) {
----------------
Could return early here.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D105525/new/

https://reviews.llvm.org/D105525