[wasm][jiterp] RelaxedSimd jiterpreter intrinsics (#115444)

Allow the Jiterpreter to use the potentially faster relaxed_swizzle in the case where we know the swizzle indices are valid and can detect at runtime the opcode is supported.

Author: lewing

src/mono/browser/runtime/jiterpreter-opcodes.ts

@@ -289,6 +289,7 @@ export const enum WasmSimdOpcode {
     v128_const                    = 0x0c,
     i8x16_shuffle                 = 0x0d,
     i8x16_swizzle                 = 0x0e,
+    i8x16_relaxed_swizzle         = 0x100,
     i8x16_splat                   = 0x0f,
     i16x8_splat                   = 0x10,
     i32x4_splat                   = 0x11,
@@ -466,6 +467,8 @@ export const enum WasmSimdOpcode {
     f64x2_pmax                    = 0xf7,
     i32x4_trunc_sat_f32x4_s       = 0xf8,
     i32x4_trunc_sat_f32x4_u       = 0xf9,
+    i32x4_relaxed_trunc_f32x4_s   = 0x101,
+    i32x4_relaxed_trunc_f32x4_u   = 0x102,
     f32x4_convert_i32x4_s         = 0xfa,
     f32x4_convert_i32x4_u         = 0xfb,
     v128_load32_zero              = 0x5c,
@@ -483,6 +486,7 @@ export const enum WasmSimdOpcode {
     i64x2_extmul_low_i32x4_u      = 0xde,
     i64x2_extmul_high_i32x4_u     = 0xdf,
     i16x8_q15mulr_sat_s           = 0x82,
+    i16x8_relaxed_q15mulr_sat_s   = 0x111,
     v128_any_true                 = 0x53,
     v128_load8_lane               = 0x54,
     v128_load16_lane              = 0x55,
@@ -503,13 +507,30 @@ export const enum WasmSimdOpcode {
     f64x2_convert_low_i32x4_u     = 0xff,
     i32x4_trunc_sat_f64x2_s_zero  = 0xfc,
     i32x4_trunc_sat_f64x2_u_zero  = 0xfd,
+    i32x4_relaxed_trunc_f64_s_zero = 0x103,
+    i32x4_relaxed_trunc_f64_u_zero = 0x104,
     f32x4_demote_f64x2_zero       = 0x5e,
     f64x2_promote_low_f32x4       = 0x5f,
     i8x16_popcnt                  = 0x62,
     i16x8_extadd_pairwise_i8x16_s = 0x7c,
     i16x8_extadd_pairwise_i8x16_u = 0x7d,
     i32x4_extadd_pairwise_i16x8_s = 0x7e,
     i32x4_extadd_pairwise_i16x8_u = 0x7f,
+    f32x4_relaxed_madd            = 0x105,
+    f32x4_relaxed_mnadd           = 0x106,
+    f64x2_relaxed_madd            = 0x107,
+    f64x2_relaxed_mnadd           = 0x108,
+    i8x16_relaxed_lane_select     = 0x109,
+    i16x8_relaxed_lane_select     = 0x10a,
+    i32x4_relaxed_lane_select     = 0x10b,
+    i64x2_relaxed_lane_select     = 0x10c,
+    f32x4_relaxed_min             = 0x10d,
+    f32x4_relaxed_max             = 0x10e,
+    f64x2_relaxed_min             = 0x10f,
+    f64x2_relaxed_max             = 0x110,
+    i16x8_relaxed_dot_i8x16_i7x16_s = 0x112,
+    i16x8_relaxed_dot_i8x16_i7x16_u = 0x113,
+    i32x4_relaxed_dot_i8x16_i7x16_s = 0x114,
 }
 
 export const enum WasmAtomicOpcode {

src/mono/browser/runtime/jiterpreter-trace-generator.ts

@@ -3919,7 +3919,11 @@ function emit_shuffle (builder: WasmBuilder, ip: MintOpcodePtr, elementCount: nu
             for (let j = 0; j < elementSize; j++)
                 builder.appendU8(i);
         }
-        builder.appendSimd(WasmSimdOpcode.i8x16_swizzle);
+        if (runtimeHelpers.featureWasmRelaxedSimd) {
+            builder.appendSimd(WasmSimdOpcode.i8x16_relaxed_swizzle);
+        } else {
+            builder.appendSimd(WasmSimdOpcode.i8x16_swizzle);
+        }
         // multiply indices by 2 or 4 to scale from elt indices to byte indices
         builder.i32_const(elementCount === 4 ? 2 : 1);
         builder.appendSimd(WasmSimdOpcode.i8x16_shl);

src/mono/browser/runtime/loader/globals.ts

@@ -4,7 +4,7 @@
 /* eslint-disable @typescript-eslint/triple-slash-reference */
 /// <reference path="../types/sidecar.d.ts" />
 
-import { exceptions, simd } from "wasm-feature-detect";
+import { exceptions, simd, relaxedSimd } from "wasm-feature-detect";
 
 import gitHash from "consts:gitHash";
 
@@ -133,6 +133,7 @@ export function setLoaderGlobals (
         // from wasm-feature-detect npm package
         exceptions,
         simd,
+        relaxedSimd
     };
     Object.assign(runtimeHelpers, rh);
     Object.assign(loaderHelpers, lh);

src/mono/browser/runtime/startup.ts

@@ -519,8 +519,12 @@ async function instantiate_wasm_module (
 }
 
 async function ensureUsedWasmFeatures () {
-    runtimeHelpers.featureWasmSimd = await loaderHelpers.simd();
-    runtimeHelpers.featureWasmEh = await loaderHelpers.exceptions();
+    const simd = loaderHelpers.simd();
+    const relaxedSimd = loaderHelpers.relaxedSimd();
+    const exceptions = loaderHelpers.exceptions();
+    runtimeHelpers.featureWasmSimd = await simd;
+    runtimeHelpers.featureWasmRelaxedSimd = await relaxedSimd;
+    runtimeHelpers.featureWasmEh = await exceptions;
     if (runtimeHelpers.emscriptenBuildOptions.wasmEnableSIMD) {
         mono_assert(runtimeHelpers.featureWasmSimd, "This browser/engine doesn't support WASM SIMD. Please use a modern version. See also https://aka.ms/dotnet-wasm-features");
     }

src/mono/browser/runtime/types/internal.ts

@@ -179,6 +179,7 @@ export type LoaderHelpers = {
     // from wasm-feature-detect npm package
     exceptions: () => Promise<boolean>,
     simd: () => Promise<boolean>,
+    relaxedSimd: () => Promise<boolean>,
 }
 export type RuntimeHelpers = {
     emscriptenBuildOptions: EmscriptenBuildOptions,
@@ -232,6 +233,7 @@ export type RuntimeHelpers = {
 
     featureWasmEh: boolean,
     featureWasmSimd: boolean,
+    featureWasmRelaxedSimd: boolean,
 
     //core
     stringify_as_error_with_stack?: (error: any) => string,