vectorization regression in Clang 19

[eyelash]

When compiling the following code on x86-64 with -O2

void foo(float* result, int size, float y, float delta) {
    for (int t = 0; t < size; ++t) {
        result[t] = y + delta * t;
    }
}

Clang generates a lot of code that seems suboptimal. After asking on Stack Overflow I was told that this is a regression in Clang 19 that I should report here. The comments and this answer on Stack Overflow shed more light into what's going on.

Here's a link to this example on Compiler Explorer with Clang trunk that shows the suboptimal assembly.

[nikic]

@goldsteinn This looks like fallout from sitofp to uitofp canonicalization. Presumably we don't have undo code for x86 yet?

[goldsteinn]

@goldsteinn This looks like fallout from sitofp to uitofp canonicalization. Presumably we don't have undo code for x86 yet?

We do: 6c40d46

I think the issue is we are dropping the nneg flag from uitofp when vectorizing in middle-end: https://godbolt.org/z/q97xvvsM6

[goldsteinn]

@goldsteinn This looks like fallout from sitofp to uitofp canonicalization. Presumably we don't have undo code for x86 yet?

We do: 6c40d46

I think the issue is we are dropping the nneg flag from uitofp when vectorizing in middle-end: https://godbolt.org/z/q97xvvsM6

Issue here seems to be with loop unroll.

[goldsteinn]

Seems like generally VPlan doesn't propagate ir flags.

[goldsteinn]

I think the fix is something along the lines of:

@@ -1484,6 +1484,8 @@ void VPWidenCastRecipe::execute(VPTransformState &State) {
   Value *Cast = Builder.CreateCast(Instruction::CastOps(Opcode), A, DestTy);
   State.set(this, Cast);
   State.addMetadata(Cast, cast_or_null<Instruction>(getUnderlyingValue()));
+  if (auto *CastOp = dyn_cast<Instruction>(Cast))
+    setFlags(CastOp);
 }
 
 InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,

Testing it, will have PR up soon if all goes well.