[llvm] [TLI] Pass replace-with-veclib works with Scalable Vectors. (PR #73642)

Maciej Gabka via llvm-commits llvm-commits at lists.llvm.org
Thu Dec 14 06:11:13 PST 2023


================
@@ -38,138 +42,153 @@ STATISTIC(NumTLIFuncDeclAdded,
 STATISTIC(NumFuncUsedAdded,
           "Number of functions added to `llvm.compiler.used`");
 
-static bool replaceWithTLIFunction(CallInst &CI, const StringRef TLIName) {
-  Module *M = CI.getModule();
-
-  Function *OldFunc = CI.getCalledFunction();
-
-  // Check if the vector library function is already declared in this module,
-  // otherwise insert it.
+/// Returns a vector Function that it adds to the Module \p M. When an \p
+/// OptOldFunc is given, it copies its attributes to the newly created Function.
+Function *getTLIFunction(Module *M, FunctionType *VectorFTy,
+                         std::optional<Function *> OptOldFunc,
+                         const StringRef TLIName) {
   Function *TLIFunc = M->getFunction(TLIName);
   if (!TLIFunc) {
-    TLIFunc = Function::Create(OldFunc->getFunctionType(),
-                               Function::ExternalLinkage, TLIName, *M);
-    TLIFunc->copyAttributesFrom(OldFunc);
+    TLIFunc =
+        Function::Create(VectorFTy, Function::ExternalLinkage, TLIName, *M);
+    if (OptOldFunc)
+      TLIFunc->copyAttributesFrom(*OptOldFunc);
 
     LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Added vector library function `"
                       << TLIName << "` of type `" << *(TLIFunc->getType())
                       << "` to module.\n");
 
     ++NumTLIFuncDeclAdded;
-
-    // Add the freshly created function to llvm.compiler.used,
-    // similar to as it is done in InjectTLIMappings
+    // Add the freshly created function to llvm.compiler.used, similar to as it
+    // is done in InjectTLIMappings
     appendToCompilerUsed(*M, {TLIFunc});
-
     LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Adding `" << TLIName
                       << "` to `@llvm.compiler.used`.\n");
     ++NumFuncUsedAdded;
   }
+  return TLIFunc;
+}
 
-  // Replace the call to the vector intrinsic with a call
-  // to the corresponding function from the vector library.
+/// Replace the call to the vector intrinsic ( \p FuncToReplace ) with a call to
+/// the corresponding function from the vector library ( \p TLIFunc ).
+static void replaceWithTLIFunction(CallInst &CI, VFInfo &Info,
+                                   Function *TLIFunc, FunctionType *VecFTy) {
   IRBuilder<> IRBuilder(&CI);
   SmallVector<Value *> Args(CI.args());
+  if (auto OptMaskpos = Info.getParamIndexForOptionalMask()) {
+    if (Args.size() == VecFTy->getNumParams())
+      static_assert(true && "mask was already in place");
+
+    auto *MaskTy =
+        VectorType::get(Type::getInt1Ty(CI.getContext()), Info.Shape.VF);
+    Args.insert(Args.begin() + OptMaskpos.value(),
+                Constant::getAllOnesValue(MaskTy));
+  }
+
   // Preserve the operand bundles.
   SmallVector<OperandBundleDef, 1> OpBundles;
   CI.getOperandBundlesAsDefs(OpBundles);
   CallInst *Replacement = IRBuilder.CreateCall(TLIFunc, Args, OpBundles);
-  assert(OldFunc->getFunctionType() == TLIFunc->getFunctionType() &&
+  assert(VecFTy == TLIFunc->getFunctionType() &&
          "Expecting function types to be identical");
   CI.replaceAllUsesWith(Replacement);
-  if (isa<FPMathOperator>(Replacement)) {
-    // Preserve fast math flags for FP math.
+  // Preserve fast math flags for FP math.
+  if (isa<FPMathOperator>(Replacement))
     Replacement->copyFastMathFlags(&CI);
-  }
+}
 
-  LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Replaced call to `"
-                    << OldFunc->getName() << "` with call to `" << TLIName
-                    << "`.\n");
-  ++NumCallsReplaced;
-  return true;
+/// Utility method to get the VecDesc, depending on whether there is such a TLI
+/// mapping, prioritizing a masked version.
+static std::optional<const VecDesc *> getVecDesc(const TargetLibraryInfo &TLI,
+                                                 const StringRef &ScalarName,
+                                                 const ElementCount &VF) {
+  if (auto *VDMasked = TLI.getVectorMappingInfo(ScalarName, VF, true))
----------------
mgabka wrote:

I think the order here should be opposite, since in intrinsics won't be masked, lets firstly check for non masked version and then for masked.

https://github.com/llvm/llvm-project/pull/73642


More information about the llvm-commits mailing list