[llvm] [TTI] Provide conservative costs for @llvm.speculative.load. (PR #180036)

[David Sherwood via llvm-commits]

================
@@ -619,6 +619,25 @@ AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
       return InstructionCost::getInvalid();
 
   switch (ICA.getID()) {
+  case Intrinsic::speculative_load: {
+    // Speculative loads are only valid for types <= 16 bytes due to MTE
+    // (Memory Tagging Extension) using 16-byte tag granules. Loads larger
+    // than 16 bytes could cross a tag granule boundary.
+    auto LT = getTypeLegalizationCost(RetTy);
+    if (!LT.first.isValid())
+      return InstructionCost::getInvalid();
+    // For scalable vectors, check that we use a single register (which means
+    // <= 16 bytes at minimum vscale). For fixed types, compute the actual size.
+    if (isa<ScalableVectorType>(RetTy)) {
+      if (LT.first.getValue() != 1)
----------------
david-arm wrote:

Oh I see, I think I understand now. You're saying that with scalable vectors we *may* be able to enter the vector loop, but provided a separate runtime check says the vector length <= 16 bytes. In that case I think the code is fine as it is. Thanks for explaining!

https://github.com/llvm/llvm-project/pull/180036