[WebAssembly] Implement trunc_sat and convert instructions for f16x8. (llvm#95180)

These instructions can be generated using regular LL intrinsics.

Specified at:

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

Author: brendandahl

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

@@ -1320,16 +1320,23 @@ def : Pat<(v8f16 (int_wasm_pmax (v8f16 V128:$lhs), (v8f16 V128:$rhs))),
 //===----------------------------------------------------------------------===//
 
 multiclass SIMDConvert<Vec vec, Vec arg, SDPatternOperator op, string name,
-                       bits<32> simdop> {
+                       bits<32> simdop, list<Predicate> reqs = []> {
   defm op#_#vec :
     SIMD_I<(outs V128:$dst), (ins V128:$vec), (outs), (ins),
            [(set (vec.vt V128:$dst), (vec.vt (op (arg.vt V128:$vec))))],
-           vec.prefix#"."#name#"\t$dst, $vec", vec.prefix#"."#name, simdop>;
+           vec.prefix#"."#name#"\t$dst, $vec", vec.prefix#"."#name, simdop, reqs>;
+}
+
+multiclass HalfPrecisionConvert<Vec vec, Vec arg, SDPatternOperator op,
+                                string name, bits<32> simdop> {
+  defm "" : SIMDConvert<vec, arg, op, name, simdop, [HasHalfPrecision]>;
 }
 
 // Floating point to integer with saturation: trunc_sat
 defm "" : SIMDConvert<I32x4, F32x4, fp_to_sint, "trunc_sat_f32x4_s", 248>;
 defm "" : SIMDConvert<I32x4, F32x4, fp_to_uint, "trunc_sat_f32x4_u", 249>;
+defm "" : HalfPrecisionConvert<I16x8, F16x8, fp_to_sint, "trunc_sat_f16x8_s", 0x148>;
+defm "" : HalfPrecisionConvert<I16x8, F16x8, fp_to_uint, "trunc_sat_f16x8_u", 0x149>;
 
 // Support the saturating variety as well.
 def trunc_s_sat32 : PatFrag<(ops node:$x), (fp_to_sint_sat $x, i32)>;
@@ -1355,6 +1362,8 @@ defm "" : SIMDConvert<F32x4, I32x4, sint_to_fp, "convert_i32x4_s", 250>;
 defm "" : SIMDConvert<F32x4, I32x4, uint_to_fp, "convert_i32x4_u", 251>;
 defm "" : SIMDConvert<F64x2, I32x4, convert_low_s, "convert_low_i32x4_s", 0xfe>;
 defm "" : SIMDConvert<F64x2, I32x4, convert_low_u, "convert_low_i32x4_u", 0xff>;
+defm "" : HalfPrecisionConvert<F16x8, I16x8, sint_to_fp, "convert_i16x8_s", 0x14a>;
+defm "" : HalfPrecisionConvert<F16x8, I16x8, uint_to_fp, "convert_i16x8_u", 0x14b>;
 
 // Extending operations
 // TODO: refactor this to be uniform for i64x2 if the numbering is not changed.

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

@@ -246,3 +246,39 @@ define <8 x half> @nearest_v8f16_via_roundeven(<8 x half> %a) {
   %v = call <8 x half> @llvm.roundeven.v8f16(<8 x half> %a)
   ret <8 x half> %v
 }
+
+define <8 x half> @convert_s_v8f16(<8 x i16> %x) {
+; CHECK-LABEL: convert_s_v8f16:
+; CHECK:         .functype convert_s_v8f16 (v128) -> (v128)
+; CHECK-NEXT:    f16x8.convert_i16x8_s $push0=, $0
+; CHECK-NEXT:    return $pop[[R]]{{$}}
+  %a = sitofp <8 x i16> %x to <8 x half>
+  ret <8 x half> %a
+}
+
+define <8 x half> @convert_u_v8f16(<8 x i16> %x) {
+; CHECK-LABEL: convert_u_v8f16:
+; CHECK:         .functype convert_u_v8f16 (v128) -> (v128)
+; CHECK-NEXT:    f16x8.convert_i16x8_u $push0=, $0
+; CHECK-NEXT:    return $pop[[R]]{{$}}
+  %a = uitofp <8 x i16> %x to <8 x half>
+  ret <8 x half> %a
+}
+
+define <8 x i16> @trunc_sat_s_v8i16(<8 x half> %x) {
+; CHECK-LABEL: trunc_sat_s_v8i16:
+; CHECK:         .functype trunc_sat_s_v8i16 (v128) -> (v128)
+; CHECK-NEXT:    i16x8.trunc_sat_f16x8_s $push0=, $0
+; CHECK-NEXT:    return $pop[[R]]{{$}}
+  %a = fptosi <8 x half> %x to <8 x i16>
+  ret <8 x i16> %a
+}
+
+define <8 x i16> @trunc_sat_u_v8i16(<8 x half> %x) {
+; CHECK-LABEL: trunc_sat_u_v8i16:
+; CHECK:         .functype trunc_sat_u_v8i16 (v128) -> (v128)
+; CHECK-NEXT:    i16x8.trunc_sat_f16x8_u $push0=, $0
+; CHECK-NEXT:    return $pop[[R]]{{$}}
+  %a = fptoui <8 x half> %x to <8 x i16>
+  ret <8 x i16> %a
+}

llvm/test/MC/WebAssembly/simd-encodings.s

@@ -920,4 +920,16 @@ main:
     # CHECK: f16x8.relaxed_nmadd # encoding: [0xfd,0xc7,0x02]
     f16x8.relaxed_nmadd
 
+    # CHECK: i16x8.trunc_sat_f16x8_s # encoding: [0xfd,0xc8,0x02]
+    i16x8.trunc_sat_f16x8_s
+
+    # CHECK: i16x8.trunc_sat_f16x8_u # encoding: [0xfd,0xc9,0x02]
+    i16x8.trunc_sat_f16x8_u
+
+    # CHECK: f16x8.convert_i16x8_s # encoding: [0xfd,0xca,0x02]
+    f16x8.convert_i16x8_s
+
+    # CHECK: f16x8.convert_i16x8_u # encoding: [0xfd,0xcb,0x02]
+    f16x8.convert_i16x8_u
+
     end_function