[llvm] [llvm][NFC] Refactor AutoUpgrader arm/aarch64 (PR #74145)

[Nathan Sidwell via llvm-commits]

================
@@ -621,6 +621,318 @@ static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
   return false;
 }
 
+// Upgrade ARM (IsArm) or Aarch64 (!IsArm) intrinsic fns. Return true iff so.
+// IsArm: 'arm.*', !IsArm: 'aarch64.*'.
+static bool UpgradeArmOrAarch64IntrinsicFunction(bool IsArm, Function *F,
+                                                 StringRef Name,
+                                                 Function *&NewFn) {
+  if (Name.starts_with("rbit")) {
+    // '(arm|aarch64).rbit'.
+    NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::bitreverse,
+                                      F->arg_begin()->getType());
+    return true;
+  }
+
+  if (Name == "thread.pointer") {
+    // '(arm|aarch64).thread.pointer'.
+    NewFn =
+        Intrinsic::getDeclaration(F->getParent(), Intrinsic::thread_pointer);
+    return true;
+  }
+
+  bool Neon = Name.consume_front("neon.");
+  if (Neon) {
+    // '(arm|aarch64).neon.*'.
+    // Changed in 12.0: bfdot accept v4bf16 and v8bf16 instead of v8i8 and
+    // v16i8 respectively.
+    if (Name.consume_front("bfdot.")) {
+      // (arm|aarch64).neon.bfdot.*'.
+      Intrinsic::ID ID = StringSwitch<Intrinsic::ID>(Name)
+                             .Cases("v2f32.v8i8", "v4f32.v16i8",
+                                    IsArm ? Intrinsic::arm_neon_bfdot
+                                          : Intrinsic::aarch64_neon_bfdot)
+                             .Default(Intrinsic::not_intrinsic);
+      if (ID != Intrinsic::not_intrinsic) {
+        size_t OperandWidth = F->getReturnType()->getPrimitiveSizeInBits();
+        assert((OperandWidth == 64 || OperandWidth == 128) &&
+               "Unexpected operand width");
+        LLVMContext &Ctx = F->getParent()->getContext();
+        std::array<Type *, 2> Tys{
+            {F->getReturnType(),
+             FixedVectorType::get(Type::getBFloatTy(Ctx), OperandWidth / 16)}};
+        NewFn = Intrinsic::getDeclaration(F->getParent(), ID, Tys);
+        return true;
+      }
+      return false; // No other '(arm|aarch64).neon.bfdot.*'.
+    }
+
+    // Changed in 12.0: bfmmla, bfmlalb and bfmlalt are not polymorphic
+    // anymore and accept v8bf16 instead of v16i8.
+    if (Name.consume_front("bfm")) {
+      // (arm|aarch64).neon.bfm*'.
+      if (Name.consume_back(".v4f32.v16i8")) {
+        // (arm|aarch64).neon.bfm*.v4f32.v16i8'.
+        Intrinsic::ID ID =
+            StringSwitch<Intrinsic::ID>(Name)
+                .Case("mla", IsArm ? Intrinsic::arm_neon_bfmmla
+                                   : Intrinsic::aarch64_neon_bfmmla)
+                .Case("lalb", IsArm ? Intrinsic::arm_neon_bfmlalb
+                                    : Intrinsic::aarch64_neon_bfmlalb)
+                .Case("lalt", IsArm ? Intrinsic::arm_neon_bfmlalt
+                                    : Intrinsic::aarch64_neon_bfmlalt)
+                .Default(Intrinsic::not_intrinsic);
+        if (ID != Intrinsic::not_intrinsic) {
+          std::array<Type *, 0> Tys;
+          NewFn = Intrinsic::getDeclaration(F->getParent(), ID, Tys);
+          return true;
+        }
+        return false; // No other '(arm|aarch64).neon.bfm*.v16i8'.
+      }
+      return false; // No other '(arm|aarch64).neon.bfm*.
+    }
+    // Continue on to Aarch64 Neon or Arm Neon.
+  }
+  // Continue on to Arm or Aarch64.
+
+  if (IsArm) {
+    // 'arm.*'.
+    if (Neon) {
+      // 'arm.neon.*'.
+      if (Name.consume_front("vclz.")) {
+        // 'arm.neon.vclz.*'.
+        Type *args[2] = {F->arg_begin()->getType(),
+                         Type::getInt1Ty(F->getContext())};
+        // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
+        // the end of the name. Change name from 'llvm.arm.neon.vclz.*' to
+        //  'llvm.ctlz.*'.
+        FunctionType *fType =
+            FunctionType::get(F->getReturnType(), args, false);
+        NewFn = Function::Create(fType, F->getLinkage(), F->getAddressSpace(),
+                                 "llvm.ctlz." + Name, F->getParent());
+        return true;
+      }
+      if (Name.starts_with("vcnt")) {
+        // 'arm.neon.vcnt*'.
+        NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
+                                          F->arg_begin()->getType());
+        return true;
+      }
+
+      if (Name.consume_front("vst")) {
+        // 'arm.neon.vst*'.
+        static const Regex vstRegex("^([1234]|[234]lane)\\.v[a-z0-9]*$");
+        SmallVector<StringRef, 2> Groups;
+        if (vstRegex.match(Name, &Groups)) {
+          static const Intrinsic::ID StoreInts[] = {
+              Intrinsic::arm_neon_vst1, Intrinsic::arm_neon_vst2,
+              Intrinsic::arm_neon_vst3, Intrinsic::arm_neon_vst4};
+
+          static const Intrinsic::ID StoreLaneInts[] = {
+              Intrinsic::arm_neon_vst2lane, Intrinsic::arm_neon_vst3lane,
+              Intrinsic::arm_neon_vst4lane};
+
+          auto fArgs = F->getFunctionType()->params();
+          Type *Tys[] = {fArgs[0], fArgs[1]};
+          if (Groups[1].size() == 1)
+            NewFn = Intrinsic::getDeclaration(F->getParent(),
+                                              StoreInts[fArgs.size() - 3], Tys);
+          else
+            NewFn = Intrinsic::getDeclaration(
+                F->getParent(), StoreLaneInts[fArgs.size() - 5], Tys);
+          return true;
+        }
+        return false; // No other 'arm.neon.vst*'.
+      }
+
+      if (Name.consume_front("vq")) {
+        // 'arm.neon.vq*'.
+        Intrinsic::ID ID = StringSwitch<Intrinsic::ID>(Name)
+                               .StartsWith("adds.", Intrinsic::sadd_sat)
+                               .StartsWith("addu.", Intrinsic::uadd_sat)
+                               .StartsWith("subs.", Intrinsic::ssub_sat)
+                               .StartsWith("subu.", Intrinsic::usub_sat)
+                               .Default(Intrinsic::not_intrinsic);
+        if (ID != Intrinsic::not_intrinsic) {
+          NewFn = Intrinsic::getDeclaration(F->getParent(), ID,
+                                            F->arg_begin()->getType());
+          return true;
+        }
+        return false; // No other 'arm.neon.vq*'.
+      }
+      return false; // No other 'arm.neon.*'.
+    }
+
+    if (Name.consume_front("mve.")) {
+      // 'arm.mve.*'.
+      if (Name == "vctp64") {
+        if (cast<FixedVectorType>(F->getReturnType())->getNumElements() == 4) {
+          // A vctp64 returning a v4i1 is converted to return a v2i1. Rename
+          // the function and deal with it below in UpgradeIntrinsicCall.
+          rename(F);
+          return true;
+        }
+        return false; // Not 'arm.mve.vctp64'.
+      }
+
+      // These too are changed to accept a v2i1 instead of the old v4i1.
+      if (Name.consume_back(".v4i1")) {
+        // 'arm.mve.*.v4i1'.
+        if (Name.consume_back(".predicated.v2i64.v4i32")) {
+          // 'arm.mve.*.predicated.v2i64.v4i32.v4i1'
+          if (Name == "mull.int" || Name == "vqdmull")
+            return true;
+          return false; // No other 'arm.mve.*.predicated.v2i64.v4i32.v4i1'.
----------------
urnathan wrote:

I'm not so sure -- doing that breaks the idiom of explicit `return BOOLCONST`:
```
if (some condition) {
  maybe do a thing
  return true;
}
```
and make it more awkward for adding another upgradable intrinsic (no idea how likely)

It's all going to compile to the same code -- if the compiler's any good :)

https://github.com/llvm/llvm-project/pull/74145