[PATCH] D87479: [InstCombine] Don't sink the fdiv from (fmul (fdiv 1.0, %x), %y) if the fdiv isn't in the same basic block as the fmul

[Craig Topper via Phabricator via llvm-commits]

craig.topper added inline comments.


================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp:531
 
-    Value *Z;
-    if (match(&I, m_c_FMul(m_OneUse(m_FDiv(m_Value(X), m_Value(Y))),
-                           m_Value(Z)))) {
-      // Sink division: (X / Y) * Z --> (X * Z) / Y
-      Value *NewFMul = Builder.CreateFMulFMF(X, Z, &I);
+    // Sink division: (X / Y) * Z --> (X * Z) / Y
+    // Don't sink if the fdiv is a reciprocal in a different basic block as this
----------------
spatel wrote:
> craig.topper wrote:
> > I found another variation of this
> > 
> > (X/Y) outside loop
> > Mul with Z inside loop.
> > 
> > ->
> > 
> > X*Z ends up inside loop
> > / Y ends up in loop
> > 
> > ->
> > 
> > recip of Y outside loop
> > 
> > X*Z in loop
> > mul with reciprocal of Y inside loop.
> > 
> > 
> > 
> > 
> > So we went from an fdiv outside loop and one fmul inside the loop to 1 div outside the loop and 2 fmuls inside the loop.
> Does this model it:
> https://godbolt.org/z/W4Ta9jzEn ?
> ...although MachineLICM is managing to pull the fmuls out of the loop there, so there must be some other complication?
MachineLICM doesn't fix it on pure scalar.

https://godbolt.org/z/zGvxxqaTP


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D87479/new/

https://reviews.llvm.org/D87479