[llvm] [PreISelIntrinsicLowering] Produce a memset_pattern16 libcall for llvm.experimental.memset.pattern when available (PR #120420)

[Alex Bradbury via llvm-commits]

================
@@ -322,7 +376,41 @@ bool PreISelIntrinsicLowering::expandMemIntrinsicUses(Function &F) const {
     }
     case Intrinsic::experimental_memset_pattern: {
       auto *Memset = cast<MemSetPatternInst>(Inst);
-      expandMemSetPatternAsLoop(Memset);
+      const TargetLibraryInfo &TLI = LookupTLI(*Memset->getFunction());
+      if (Constant *PatternValue = getMemSetPattern16Value(Memset, TLI)) {
+        // FIXME: There is currently no profitability calculation for emitting
+        // the libcall vs expanding the memset.pattern directly.
+        IRBuilder<> Builder(Inst);
+        Module *M = Memset->getModule();
+        const DataLayout &DL = Memset->getDataLayout();
+
+        StringRef FuncName = "memset_pattern16";
+        FunctionCallee MSP = getOrInsertLibFunc(
+            M, TLI, LibFunc_memset_pattern16, Builder.getVoidTy(),
+            Memset->getRawDest()->getType(), Builder.getPtrTy(),
+            Memset->getLength()->getType());
+        inferNonMandatoryLibFuncAttrs(M, FuncName, TLI);
+
+        // Otherwise we should form a memset_pattern16.  PatternValue is known
+        // to be an constant array of 16-bytes. Put the value into a mergable
+        // global.
+        GlobalVariable *GV = new GlobalVariable(
+            *M, PatternValue->getType(), true, GlobalValue::PrivateLinkage,
+            PatternValue, ".memset_pattern");
+        GV->setUnnamedAddr(
+            GlobalValue::UnnamedAddr::Global); // Ok to merge these.
+        GV->setAlignment(Align(16));
+        Value *PatternPtr = GV;
+        Value *NumBytes = Builder.CreateMul(
+            Builder.getInt64(
+                DL.getTypeSizeInBits(Memset->getValue()->getType()) / 8),
+            Memset->getLength());
+        CallInst *MemsetPattern16Call = Builder.CreateCall(
+            MSP, {Memset->getRawDest(), PatternPtr, NumBytes});
----------------
asb wrote:

Were there particular attributes you imagine we'd lose out by not preserving? I'd be concerned about doing something semantically incorrect by blindly copying them, though it seems mergeAttributesAndFlags in SimplifyLibCalls would be one eample to copy. As I think failing to preserve attributes is conservatively correct I've put a TODO with a reference to this for the time being.

https://github.com/llvm/llvm-project/pull/120420