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

[Alex Bradbury via llvm-commits]

================
@@ -232,6 +233,59 @@ static bool canEmitLibcall(const TargetMachine *TM, Function *F,
   return TLI->getLibcallName(LC) != nullptr;
 }
 
+// Return a value appropriate for use with the memset_pattern16 libcall, if
+// possible and if we know how. (Adapted from equivalent helper in
+// LoopIdiomRecognize).
+static Constant *getMemSetPattern16Value(MemSetPatternInst *Inst,
+                                         const TargetLibraryInfo &TLI) {
+  // FIXME: This could check for UndefValue because it can be merged into any
+  // other valid pattern.
+
+  // Don't emit libcalls if a non-default address space is being used.
+  if (Inst->getRawDest()->getType()->getPointerAddressSpace() != 0)
+    return nullptr;
+
+  Value *V = Inst->getValue();
+  const DataLayout &DL = Inst->getDataLayout();
+  Module *M = Inst->getModule();
+
+  if (!isLibFuncEmittable(M, &TLI, LibFunc_memset_pattern16))
+    return nullptr;
+
+  // If the value isn't a constant, we can't promote it to being in a constant
+  // array.  We could theoretically do a store to an alloca or something, but
+  // that doesn't seem worthwhile.
+  Constant *C = dyn_cast<Constant>(V);
+  if (!C || isa<ConstantExpr>(C))
+    return nullptr;
+
+  // Only handle simple values that are a power of two bytes in size.
+  uint64_t Size = DL.getTypeSizeInBits(V->getType());
+  if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
+    return nullptr;
+
+  // Don't care enough about darwin/ppc to implement this.
+  if (DL.isBigEndian())
+    return nullptr;
+
+  // Convert to size in bytes.
+  Size /= 8;
+
+  // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
+  // if the top and bottom are the same (e.g. for vectors and large integers).
+  if (Size > 16)
+    return nullptr;
+
+  // If the constant is exactly 16 bytes, just use it.
+  if (Size == 16)
+    return C;
+
+  // Otherwise, we'll use an array of the constants.
+  unsigned ArraySize = 16 / Size;
+  ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
+  return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
----------------
asb wrote:

As Nikita rightly pointed out, I was missing an important test case that would have perhaps made this clearer. This logic (ported across from LoopIdiomRecognize) handles what icalls "splatting". e.g. if you have an i16 used in a memset in a loop, it will create a 128-bit pattern by duplicating it that is appropriate for use with memset_pattern16. You could alternatively assembly a new i128 constant, but it's not clear that would be better or simpler than the ConstantArray approach here.

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