[WebAssembly] Prototype f32x4.relaxed_dot_bf16x8_add_f32

As proposed in WebAssembly/relaxed-simd#77. Only an
LLVM intrinsic and a clang builtin are implemented. Since there is no bfloat16
type, use u16 to represent the bfloats in the builtin function arguments.

Differential Revision: https://reviews.llvm.org/D133428

Author: tlively

clang/include/clang/Basic/BuiltinsWebAssembly.def

@@ -188,6 +188,7 @@ TARGET_BUILTIN(__builtin_wasm_relaxed_q15mulr_s_i16x8, "V8sV8sV8s", "nc", "relax
 
 TARGET_BUILTIN(__builtin_wasm_dot_i8x16_i7x16_s_i16x8, "V8sV16ScV16Sc", "nc", "relaxed-simd")
 TARGET_BUILTIN(__builtin_wasm_dot_i8x16_i7x16_add_s_i32x4, "V4iV16ScV16ScV4i", "nc", "relaxed-simd")
+TARGET_BUILTIN(__builtin_wasm_relaxed_dot_bf16x8_add_f32_f32x4, "V4fV8UsV8UsV4f", "nc", "relaxed-simd")
 
 #undef BUILTIN
 #undef TARGET_BUILTIN

clang/lib/CodeGen/CGBuiltin.cpp

@@ -18870,6 +18870,14 @@ Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
         CGM.getIntrinsic(Intrinsic::wasm_dot_i8x16_i7x16_add_signed);
     return Builder.CreateCall(Callee, {LHS, RHS, Acc});
   }
+  case WebAssembly::BI__builtin_wasm_relaxed_dot_bf16x8_add_f32_f32x4: {
+    Value *LHS = EmitScalarExpr(E->getArg(0));
+    Value *RHS = EmitScalarExpr(E->getArg(1));
+    Value *Acc = EmitScalarExpr(E->getArg(2));
+    Function *Callee =
+        CGM.getIntrinsic(Intrinsic::wasm_relaxed_dot_bf16x8_add_f32);
+    return Builder.CreateCall(Callee, {LHS, RHS, Acc});
+  }
   default:
     return nullptr;
   }

clang/test/CodeGen/builtins-wasm.c

@@ -794,3 +794,10 @@ i32x4 dot_i8x16_i7x16_add_s_i32x4(i8x16 a, i8x16 b, i32x4 c) {
   // WEBASSEMBLY-SAME: <16 x i8> %a, <16 x i8> %b, <4 x i32> %c)
   // WEBASSEMBLY-NEXT: ret
 }
+
+f32x4 relaxed_dot_bf16x8_add_f32_f32x4(u16x8 a, u16x8 b, f32x4 c) {
+  return __builtin_wasm_relaxed_dot_bf16x8_add_f32_f32x4(a, b, c);
+  // WEBASSEMBLY: call <4 x float> @llvm.wasm.relaxed.dot.bf16x8.add.f32
+  // WEBASSEMBLY-SAME: <8 x i16> %a, <8 x i16> %b, <4 x float> %c)
+  // WEBASSEMBLY-NEXT: ret
+}

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

@@ -285,6 +285,12 @@ def int_wasm_dot_i8x16_i7x16_add_signed:
             [llvm_v16i8_ty, llvm_v16i8_ty, llvm_v4i32_ty],
             [IntrNoMem, IntrSpeculatable]>;
 
+def int_wasm_relaxed_dot_bf16x8_add_f32:
+  Intrinsic<[llvm_v4f32_ty],
+            [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v4f32_ty],
+            [IntrNoMem, IntrSpeculatable]>;
+
+
 //===----------------------------------------------------------------------===//
 // Thread-local storage intrinsics
 //===----------------------------------------------------------------------===//

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

@@ -1466,3 +1466,15 @@ defm RELAXED_DOT_ADD :
                (v16i8 V128:$lhs), (v16i8 V128:$rhs), (v4i32 V128:$acc)))],
             "i32x4.dot_i8x16_i7x16_add_s\t$dst, $lhs, $rhs, $acc",
             "i32x4.dot_i8x16_i7x16_add_s", 0x113>;
+
+//===----------------------------------------------------------------------===//
+// Relaxed BFloat16 dot product
+//===----------------------------------------------------------------------===//
+
+defm RELAXED_DOT_BFLOAT :
+  RELAXED_I<(outs V128:$dst), (ins V128:$lhs, V128:$rhs, V128:$acc),
+            (outs), (ins),
+            [(set (v4f32 V128:$dst), (int_wasm_relaxed_dot_bf16x8_add_f32
+               (v8i16 V128:$lhs), (v8i16 V128:$rhs), (v4f32 V128:$acc)))],
+            "f32x4.relaxed_dot_bf16x8_add_f32\t$dst, $lhs, $rhs, $acc",
+            "f32x4.relaxed_dot_bf16x8_add_f32", 0x114>;

llvm/test/CodeGen/WebAssembly/simd-intrinsics.ll

@@ -786,6 +786,20 @@ define <4 x float> @relaxed_max_v4f32(<4 x float> %a, <4 x float> %b) {
   ret <4 x float> %v
 }
 
+; CHECK-LABEL: relaxed_dot_bf16x8_add_f32:
+; CHECK-NEXT: .functype relaxed_dot_bf16x8_add_f32 (v128, v128, v128) -> (v128){{$}}
+; CHECK-NEXT: f32x4.relaxed_dot_bf16x8_add_f32 $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
+; CHECK-NEXT: return $pop[[R]]{{$}}
+declare <4 x float> @llvm.wasm.relaxed.dot.bf16x8.add.f32(<8 x i16>, <8 x i16>,
+                                                          <4 x float>)
+define <4 x float> @relaxed_dot_bf16x8_add_f32(<8 x i16> %a, <8 x i16> %b,
+                                               <4 x float> %c) {
+  %v = call <4 x float> @llvm.wasm.relaxed.dot.bf16x8.add.f32(
+    <8 x i16> %a, <8 x i16> %b, <4 x float> %c
+  )
+  ret <4 x float> %v
+}
+
 ; ==============================================================================
 ; 2 x f64
 ; ==============================================================================