[llvm] [SVE] Don't require lookup when demangling vector function mappings (PR #72260)

Graham Hunter via llvm-commits llvm-commits at lists.llvm.org
Wed Nov 22 03:22:47 PST 2023


================
@@ -273,49 +275,88 @@ ParseRet tryParseAlign(StringRef &ParseString, Align &Alignment) {
   return ParseRet::None;
 }
 
-#ifndef NDEBUG
-// Verify the assumtion that all vectors in the signature of a vector
-// function have the same number of elements.
-bool verifyAllVectorsHaveSameWidth(FunctionType *Signature) {
-  SmallVector<VectorType *, 2> VecTys;
-  if (auto *RetTy = dyn_cast<VectorType>(Signature->getReturnType()))
-    VecTys.push_back(RetTy);
-  for (auto *Ty : Signature->params())
-    if (auto *VTy = dyn_cast<VectorType>(Ty))
-      VecTys.push_back(VTy);
-
-  if (VecTys.size() <= 1)
-    return true;
-
-  assert(VecTys.size() > 1 && "Invalid number of elements.");
-  const ElementCount EC = VecTys[0]->getElementCount();
-  return llvm::all_of(llvm::drop_begin(VecTys), [&EC](VectorType *VTy) {
-    return (EC == VTy->getElementCount());
-  });
+// Given a type, return the size in bits if it is a supported element type
+// for vectorized function calls, or nullopt if not.
+std::optional<unsigned> getSizeFromScalarType(Type *Ty) {
+  // The scalar function should only take scalar arguments.
+  if (!Ty->isIntegerTy() && !Ty->isFloatingPointTy() && !Ty->isPointerTy())
+    return std::nullopt;
+
+  unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
+  switch (SizeInBits) {
+  // Legal power-of-two scalars are supported.
+  case 64:
+  case 32:
+  case 16:
+  case 8:
+    return SizeInBits;
+  case 0:
+    // We're assuming a 64b pointer size here for SVE; if another non-64b
+    // target adds support for scalable vectors, we may need DataLayout to
+    // determine the size.
+    if (Ty->isPointerTy())
+      return 64;
+    break;
+  default:
+    break;
+  }
+
+  return std::nullopt;
 }
-#endif // NDEBUG
-
-// Extract the VectorizationFactor from a given function signature,
-// under the assumtion that all vectors have the same number of
-// elements, i.e. same ElementCount.Min.
-ElementCount getECFromSignature(FunctionType *Signature) {
-  assert(verifyAllVectorsHaveSameWidth(Signature) &&
-         "Invalid vector signature.");
-
-  if (auto *RetTy = dyn_cast<VectorType>(Signature->getReturnType()))
-    return RetTy->getElementCount();
-  for (auto *Ty : Signature->params())
-    if (auto *VTy = dyn_cast<VectorType>(Ty))
-      return VTy->getElementCount();
-
-  return ElementCount::getFixed(/*Min=*/1);
+
+// Extract the VectorizationFactor from a given function signature, based
+// on the widest scalar element types that will become vector parameters.
+std::optional<ElementCount>
+getScalableECFromSignature(FunctionType *Signature, const VFISAKind ISA,
+                           const SmallVectorImpl<VFParameter> &Params) {
+  // Look up the minimum known register size in order to calculate minimum VF.
+  // Only AArch64 SVE is supported at present.
+  unsigned MinRegSizeInBits;
+  switch (ISA) {
+  case VFISAKind::SVE:
+    MinRegSizeInBits = 128;
+    break;
+  default:
+    return std::nullopt;
+  }
+
+  unsigned WidestTypeInBits = 0;
+  for (auto &Param : Params) {
+    // Check any parameters that will be widened to vectors. Uniform or linear
+    // parameters may be misleading for determining the VF of a given function.
+    if (Param.ParamKind == VFParamKind::Vector) {
+      // If the scalar function doesn't actually have a corresponding argument,
+      // reject the mapping.
+      if (Param.ParamPos + 1 > Signature->getNumParams())
----------------
huntergr-arm wrote:

I tried your suggestion to move the check to the parameter parsing code, and unfortunately discovered that many of the unit tests are broken -- VectorFunctionABITest.cpp has a method called invokeParser, which takes a mangled name, a scalar function name, and a scalar function type. However, it's often invoked with the mangled name + the vector variant name and no type, using the default 'void()'.

So I'll avoid doing that in this PR and raise another later to fix this.

https://github.com/llvm/llvm-project/pull/72260


More information about the llvm-commits mailing list