[llvm] [RISCV] Sink vp.splat operands of VP intrinsic. (PR #133245)

[Luke Lau via llvm-commits]

================
@@ -2837,39 +2837,43 @@ bool RISCVTTIImpl::isProfitableToSinkOperands(
   if (!ST->sinkSplatOperands())
     return false;
 
-  for (auto OpIdx : enumerate(I->operands())) {
-    if (!canSplatOperand(I, OpIdx.index()))
-      continue;
-
-    Instruction *Op = dyn_cast<Instruction>(OpIdx.value().get());
+  for (auto &U : I->operands()) {
+    auto *Op = dyn_cast<Instruction>(U.get());
     // Make sure we are not already sinking this operand
     if (!Op || any_of(Ops, [&](Use *U) { return U->get() == Op; }))
       continue;
 
-    // We are looking for a splat that can be sunk.
-    if (!match(Op, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()),
+    // We are looking for a splat/vp.splat that can be sunk.
+    // FIXME: Should we care about the poison value of inactive elements in
+    // vp.splat?
----------------
lukel97 wrote:

I just did a quick check and I think we still match even on all-disabled VP intrinsics, e.g.:

```llvm
define <vscale x 8 x i32> @f(<vscale x 8 x i32> %a, i32 zeroext %b, i32 zeroext %evl) {
  %splat = tail call <vscale x 8 x i32> @llvm.experimental.vp.splat.nxv8i32(i32 %b, <vscale x 8 x i1> splat(i1 false), i32 0)
  %add = call <vscale x 8 x i32> @llvm.vp.add(<vscale x 8 x i32> %a, <vscale x 8 x i32> %splat, <vscale x 8 x i1> splat (i1 true), i32 %evl)
  ret <vscale x 8 x i32> %add
}
```

Still matches to 

```asm
f:                                      # @f
	.cfi_startproc
# %bb.0:
	vsetvli	zero, a1, e32, m4, ta, ma
	vadd.vx	v8, v8, a0
	ret
```

https://github.com/llvm/llvm-project/pull/133245