[llvm] [ARM]Adjust cost of muls in (U/S)MLAL and patterns (PR #122713)

[Nashe Mncube via llvm-commits]

================
@@ -1458,16 +1458,73 @@ InstructionCost ARMTTIImpl::getArithmeticInstrCost(
   if (LooksLikeAFreeShift())
     return 0;
 
+  // When targets have both DSP and MVE we find that the
+  // the compiler will attempt to vectorize as well as using
+  // scalar (S/U)MLAL operations. This is in cases where we have
+  // the pattern ext(mul(ext(i16), ext(i16))) we find
+  // that codegen performs better when only using (S/U)MLAL scalar
+  // ops instead of trying to mix vector ops with (S/U)MLAL ops. We therefore
+  // check if a mul instruction is used in a (U/S)MLAL pattern.
+  auto MulInDSPMLALPattern = [&](const Instruction *I, unsigned Opcode,
+                                 Type *Ty) -> bool {
+    if (!ST->hasDSP())
+      return false;
+
+    if (!I)
+      return false;
+
+    if (Opcode != Instruction::Mul)
+      return false;
+
+    if (Ty->isVectorTy())
+      return false;
+
+    auto ValueOpcodesEqual = [](const Value *LHS, const Value *RHS) -> bool {
+      return cast<Instruction>(LHS)->getOpcode() ==
+             cast<Instruction>(RHS)->getOpcode();
+    };
+    auto IsExtInst = [](const Value *V) -> bool {
+      return isa<ZExtInst>(V) || isa<SExtInst>(V);
+    };
+    auto IsExtensionFromHalf = [&, IsExtInst](const Value *V) -> bool {
----------------
nasherm wrote:

Hi Florian, yes.  I saw it triggered some buildbot failures. I opened #132018 to fix this. Is it preferable to use this merged branch?

https://github.com/llvm/llvm-project/pull/122713