[PATCH] D17485: [x86, InstCombine] transform x86 AVX masked stores to LLVM intrinsics

[Sanjay Patel via llvm-commits]

spatel added inline comments.

================
Comment at: lib/Transforms/InstCombine/InstCombineCalls.cpp:824
@@ +823,3 @@
+  return ConstantVector::get(BoolVec);
+}
+
----------------
RKSimon wrote:
> Not a priority but can the blendv instcombines share this helper (or something very similar)?
Yes - I'll do that as a precursor or follow-up patch.

================
Comment at: lib/Transforms/InstCombine/InstCombineCalls.cpp:1609
@@ +1608,3 @@
+  // TODO: The AVX2 integer variants can go here too.
+    return simplifyX86MaskedStore(*II, *this);
+
----------------
RKSimon wrote:
> If this returns nullptr shouldn't the case break? 
This is interesting - simplifyX86MaskedStore() *always* returns nullptr because eraseInstFromFunction() always returns nullptr. I'll change it to return a bool to indicate if a change was made.


http://reviews.llvm.org/D17485