[WebAssembly] Implement prototype f16x8.splat instruction. (llvm#93228)

Adds a builtin and intrinsic for the f16x8.splat instruction.

Specified at:

https://github.com/WebAssembly/half-precision/blob/29a9b9462c9285d4ccc1a5dc39214ddfd1892658/proposals/half-precision/Overview.md

Note: the current spec has f16x8.splat as opcode 0x123, but this is
incorrect and will be changed to 0x120 soon.

Author: brendandahl

clang/include/clang/Basic/BuiltinsWebAssembly.def

@@ -193,6 +193,7 @@ TARGET_BUILTIN(__builtin_wasm_relaxed_dot_bf16x8_add_f32_f32x4, "V4fV8UsV8UsV4f"
 // Half-Precision (fp16)
 TARGET_BUILTIN(__builtin_wasm_loadf16_f32, "fh*", "nU", "half-precision")
 TARGET_BUILTIN(__builtin_wasm_storef16_f32, "vfh*", "n", "half-precision")
+TARGET_BUILTIN(__builtin_wasm_splat_f16x8, "V8hf", "nc", "half-precision")
 
 // Reference Types builtins
 // Some builtins are custom type-checked - see 't' as part of the third argument,

clang/lib/Basic/Targets/WebAssembly.h

@@ -90,6 +90,9 @@ class LLVM_LIBRARY_VISIBILITY WebAssemblyTargetInfo : public TargetInfo {
 
   StringRef getABI() const override;
   bool setABI(const std::string &Name) override;
+  bool useFP16ConversionIntrinsics() const override {
+    return !HasHalfPrecision;
+  }
 
 protected:
   void getTargetDefines(const LangOptions &Opts,

clang/lib/CodeGen/CGBuiltin.cpp

@@ -21230,6 +21230,11 @@ Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_storef16_f32);
     return Builder.CreateCall(Callee, {Val, Addr});
   }
+  case WebAssembly::BI__builtin_wasm_splat_f16x8: {
+    Value *Val = EmitScalarExpr(E->getArg(0));
+    Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_splat_f16x8);
+    return Builder.CreateCall(Callee, {Val});
+  }
   case WebAssembly::BI__builtin_wasm_table_get: {
     assert(E->getArg(0)->getType()->isArrayType());
     Value *Table = EmitArrayToPointerDecay(E->getArg(0)).emitRawPointer(*this);

clang/test/CodeGen/builtins-wasm.c

@@ -11,6 +11,7 @@ typedef unsigned char u8x16 __attribute((vector_size(16)));
 typedef unsigned short u16x8 __attribute((vector_size(16)));
 typedef unsigned int u32x4 __attribute((vector_size(16)));
 typedef unsigned long long u64x2 __attribute((vector_size(16)));
+typedef __fp16 f16x8 __attribute((vector_size(16)));
 typedef float f32x4 __attribute((vector_size(16)));
 typedef double f64x2 __attribute((vector_size(16)));
 
@@ -813,6 +814,11 @@ void store_f16_f32(float val, __fp16 *addr) {
   // WEBASSEMBLY-NEXT: ret
 }
 
+f16x8 splat_f16x8(float a) {
+  // WEBASSEMBLY: %0 = tail call <8 x half> @llvm.wasm.splat.f16x8(float %a)
+  // WEBASSEMBLY-NEXT: ret <8 x half> %0
+  return __builtin_wasm_splat_f16x8(a);
+}
 __externref_t externref_null() {
   return __builtin_wasm_ref_null_extern();
   // WEBASSEMBLY: tail call ptr addrspace(10) @llvm.wasm.ref.null.extern()

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

@@ -337,6 +337,10 @@ def int_wasm_storef16_f32:
             [llvm_float_ty, llvm_ptr_ty],
             [IntrWriteMem, IntrArgMemOnly],
              "", [SDNPMemOperand]>;
+def int_wasm_splat_f16x8:
+  DefaultAttrsIntrinsic<[llvm_v8f16_ty],
+                        [llvm_float_ty],
+                        [IntrNoMem, IntrSpeculatable]>;
 
 
 //===----------------------------------------------------------------------===//

llvm/lib/Target/WebAssembly/Utils/WebAssemblyTypeUtilities.cpp

@@ -50,6 +50,7 @@ wasm::ValType WebAssembly::toValType(MVT Type) {
   case MVT::v8i16:
   case MVT::v4i32:
   case MVT::v2i64:
+  case MVT::v8f16:
   case MVT::v4f32:
   case MVT::v2f64:
     return wasm::ValType::V128;

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

@@ -70,6 +70,9 @@ WebAssemblyTargetLowering::WebAssemblyTargetLowering(
     addRegisterClass(MVT::v2i64, &WebAssembly::V128RegClass);
     addRegisterClass(MVT::v2f64, &WebAssembly::V128RegClass);
   }
+  if (Subtarget->hasHalfPrecision()) {
+    addRegisterClass(MVT::v8f16, &WebAssembly::V128RegClass);
+  }
   if (Subtarget->hasReferenceTypes()) {
     addRegisterClass(MVT::externref, &WebAssembly::EXTERNREFRegClass);
     addRegisterClass(MVT::funcref, &WebAssembly::FUNCREFRegClass);

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

@@ -38,6 +38,13 @@ multiclass RELAXED_I<dag oops_r, dag iops_r, dag oops_s, dag iops_s,
                             asmstr_s, simdop, HasRelaxedSIMD>;
 }
 
+multiclass HALF_PRECISION_I<dag oops_r, dag iops_r, dag oops_s, dag iops_s,
+                            list<dag> pattern_r, string asmstr_r = "",
+                            string asmstr_s = "", bits<32> simdop = -1> {
+  defm "" : ABSTRACT_SIMD_I<oops_r, iops_r, oops_s, iops_s, pattern_r, asmstr_r,
+                            asmstr_s, simdop, HasHalfPrecision>;
+}
+
 
 defm "" : ARGUMENT<V128, v16i8>;
 defm "" : ARGUMENT<V128, v8i16>;
@@ -591,6 +598,14 @@ defm "" : Splat<I64x2, 18>;
 defm "" : Splat<F32x4, 19>;
 defm "" : Splat<F64x2, 20>;
 
+// Half values are not fully supported so an intrinsic is used instead of a
+// regular Splat pattern as above.
+defm SPLAT_F16x8 :
+  HALF_PRECISION_I<(outs V128:$dst), (ins F32:$x),
+                   (outs), (ins),
+                   [(set (v8f16 V128:$dst), (int_wasm_splat_f16x8 F32:$x))],
+                   "f16x8.splat\t$dst, $x", "f16x8.splat", 0x120>;
+
 // scalar_to_vector leaves high lanes undefined, so can be a splat
 foreach vec = AllVecs in
 def : Pat<(vec.vt (scalar_to_vector (vec.lane_vt vec.lane_rc:$x))),

llvm/lib/Target/WebAssembly/WebAssemblyRegisterInfo.td

@@ -63,7 +63,8 @@ def I32 : WebAssemblyRegClass<[i32], 32, (add FP32, SP32, I32_0)>;
 def I64 : WebAssemblyRegClass<[i64], 64, (add FP64, SP64, I64_0)>;
 def F32 : WebAssemblyRegClass<[f32], 32, (add F32_0)>;
 def F64 : WebAssemblyRegClass<[f64], 64, (add F64_0)>;
-def V128 : WebAssemblyRegClass<[v4f32, v2f64, v2i64, v4i32, v16i8, v8i16], 128,
-                               (add V128_0)>;
+def V128 : WebAssemblyRegClass<[v8f16, v4f32, v2f64, v2i64, v4i32, v16i8,
+                                v8i16],
+                               128, (add V128_0)>;
 def FUNCREF : WebAssemblyRegClass<[funcref], 0, (add FUNCREF_0)>;
 def EXTERNREF : WebAssemblyRegClass<[externref], 0, (add EXTERNREF_0)>;

llvm/test/CodeGen/WebAssembly/half-precision.ll

@@ -1,5 +1,5 @@
-; RUN: llc < %s --mtriple=wasm32-unknown-unknown -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+half-precision | FileCheck %s
-; RUN: llc < %s --mtriple=wasm64-unknown-unknown -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+half-precision | FileCheck %s
+; RUN: llc < %s --mtriple=wasm32-unknown-unknown -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+half-precision,+simd128 | FileCheck %s
+; RUN: llc < %s --mtriple=wasm64-unknown-unknown -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+half-precision,+simd128 | FileCheck %s
 
 declare float @llvm.wasm.loadf32.f16(ptr)
 declare void @llvm.wasm.storef16.f32(float, ptr)
@@ -19,3 +19,11 @@ define void @stf16_32(float %v, ptr %p) {
   tail call void @llvm.wasm.storef16.f32(float %v, ptr %p)
   ret void
 }
+
+; CHECK-LABEL: splat_v8f16:
+; CHECK:       f16x8.splat $push0=, $0
+; CHECK-NEXT:  return $pop0
+define <8 x half> @splat_v8f16(float %x) {
+  %v = call <8 x half> @llvm.wasm.splat.f16x8(float %x)
+  ret <8 x half> %v
+}