[PATCH] D118635: [GlobalISel] Combine: (G_*MULO x, 0) -> 0 + no carry out

[Jay Foad via Phabricator via llvm-commits]

foad added inline comments.


================
Comment at: llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp:4589
+  MatchInfo = [=](MachineIRBuilder &B) {
+    B.buildConstant(Dst, 0);
+    B.buildConstant(Carry, 0);
----------------
Can't you copy from the RHS of the mul here? And hence there should be no need to check legality for this constant?


================
Comment at: llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp:4590
+    B.buildConstant(Dst, 0);
+    B.buildConstant(Carry, 0);
+  };
----------------
paquette wrote:
> foad wrote:
> > How does this work? Don't you need to build a single instruction that has two results (just like G_*MULO)?
> I don't think it matters; although G_*MULO instructions produce two results, those results are just regular vregs. As long as they are still defined, the uses should work out.
> 
> So, we should be able to replace this with two instructions if we know what the results are going to be. In the case of
> 
> ```
> %mul, %o = G_*MULO %x, <zero>
> ```
> 
> We know that the value of `%mul` and `%o` must both be 0. So replacing them both with appropriately-sized 0s ought to be safe.
Oh yeah, sorry. I was forgetting that (unlike sdag) it is just writing to two vregs.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D118635/new/

https://reviews.llvm.org/D118635