[clang] [llvm] [AArch64] Improve bcvtn2 and remove aarch64_neon_bfcvt intrinsics (PR #120363)

[David Green via llvm-commits]

================
@@ -4064,31 +4072,59 @@ static Value *upgradeX86IntrinsicCall(StringRef Name, CallBase *CI, Function *F,
 
 static Value *upgradeAArch64IntrinsicCall(StringRef Name, CallBase *CI,
                                           Function *F, IRBuilder<> &Builder) {
-  Intrinsic::ID NewID =
-      StringSwitch<Intrinsic::ID>(Name)
-          .Case("sve.fcvt.bf16f32", Intrinsic::aarch64_sve_fcvt_bf16f32_v2)
-          .Case("sve.fcvtnt.bf16f32", Intrinsic::aarch64_sve_fcvtnt_bf16f32_v2)
-          .Default(Intrinsic::not_intrinsic);
-  if (NewID == Intrinsic::not_intrinsic)
-    llvm_unreachable("Unhandled Intrinsic!");
-
-  SmallVector<Value *, 3> Args(CI->args());
-
-  // The original intrinsics incorrectly used a predicate based on the smallest
-  // element type rather than the largest.
-  Type *BadPredTy = ScalableVectorType::get(Builder.getInt1Ty(), 8);
-  Type *GoodPredTy = ScalableVectorType::get(Builder.getInt1Ty(), 4);
-
-  if (Args[1]->getType() != BadPredTy)
-    llvm_unreachable("Unexpected predicate type!");
-
-  Args[1] = Builder.CreateIntrinsic(Intrinsic::aarch64_sve_convert_to_svbool,
-                                    BadPredTy, Args[1]);
-  Args[1] = Builder.CreateIntrinsic(Intrinsic::aarch64_sve_convert_from_svbool,
-                                    GoodPredTy, Args[1]);
-
-  return Builder.CreateIntrinsic(NewID, {}, Args, /*FMFSource=*/nullptr,
-                                 CI->getName());
+  if (Name.starts_with("neon.bfcvt")) {
+    if (Name.starts_with("neon.bfcvtn2")) {
+      SmallVector<int, 32> LoMask(4);
+      std::iota(LoMask.begin(), LoMask.end(), 0);
+      SmallVector<int, 32> ConcatMask(8);
+      std::iota(ConcatMask.begin(), ConcatMask.end(), 0);
+      Value *Inactive = Builder.CreateShuffleVector(CI->getOperand(0), LoMask);
+      Value *Trunc =
+          Builder.CreateFPTrunc(CI->getOperand(1), Inactive->getType());
+      return Builder.CreateShuffleVector(Inactive, Trunc, ConcatMask);
+    } else if (Name.starts_with("neon.bfcvtn")) {
+      SmallVector<int, 32> ConcatMask(8);
+      std::iota(ConcatMask.begin(), ConcatMask.end(), 0);
+      Type *V4BF16 =
+          FixedVectorType::get(Type::getBFloatTy(F->getContext()), 4);
+      Value *Trunc = Builder.CreateFPTrunc(CI->getOperand(0), V4BF16);
+      dbgs() << "Trunc: " << *Trunc << "\n";
+      return Builder.CreateShuffleVector(
+          Trunc, ConstantAggregateZero::get(V4BF16), ConcatMask);
+    } else {
+      return Builder.CreateFPTrunc(CI->getOperand(0),
+                                   Type::getBFloatTy(F->getContext()));
+    }
+  } else if (Name.starts_with("sve.fcvt")) {
----------------
davemgreen wrote:

The sve.fcvt is what this function (upgradeAArch64IntrinsicCall) was previously handling, so will be NFC in that regard.  It gets here because of these lines from upgradeArmOrAarch64IntrinsicFunction.
```
      // 'aarch64.sve.fcvt.bf16f32' || 'aarch64.sve.fcvtnt.bf16f32'
      if (Name == "fcvt.bf16f32" || Name == "fcvtnt.bf16f32") {
        NewFn = nullptr;
```
It's now just inside an if to be more clear. It has its own tests that are still doing OK.

https://github.com/llvm/llvm-project/pull/120363