Implement new superinsns in jiterpreter

Author: kg

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

@@ -85,6 +85,16 @@ export const unopTable: { [opcode: number]: OpRec3 | undefined } = {
     [MintOpcode.MINT_CLZ_I8]:        [WasmOpcode.i64_clz,    WasmOpcode.i64_load, WasmOpcode.i64_store],
     [MintOpcode.MINT_CTZ_I8]:        [WasmOpcode.i64_ctz,    WasmOpcode.i64_load, WasmOpcode.i64_store],
     [MintOpcode.MINT_POPCNT_I8]:     [WasmOpcode.i64_popcnt, WasmOpcode.i64_load, WasmOpcode.i64_store],
+
+    [MintOpcode.MINT_ADD_I4_IMM2]:   [WasmOpcode.i32_add, WasmOpcode.i32_load, WasmOpcode.i32_store],
+    [MintOpcode.MINT_MUL_I4_IMM2]:   [WasmOpcode.i32_mul, WasmOpcode.i32_load, WasmOpcode.i32_store],
+    [MintOpcode.MINT_ADD_I8_IMM2]:   [WasmOpcode.i64_add, WasmOpcode.i64_load, WasmOpcode.i64_store],
+    [MintOpcode.MINT_MUL_I8_IMM2]:   [WasmOpcode.i64_mul, WasmOpcode.i64_load, WasmOpcode.i64_store],
+
+    [MintOpcode.MINT_AND_I4_IMM]:    [WasmOpcode.i32_and, WasmOpcode.i32_load, WasmOpcode.i32_store],
+    [MintOpcode.MINT_AND_I4_IMM2]:   [WasmOpcode.i32_and, WasmOpcode.i32_load, WasmOpcode.i32_store],
+    [MintOpcode.MINT_OR_I4_IMM]:     [WasmOpcode.i32_or,  WasmOpcode.i32_load, WasmOpcode.i32_store],
+    [MintOpcode.MINT_OR_I4_IMM2]:    [WasmOpcode.i32_or,  WasmOpcode.i32_load, WasmOpcode.i32_store],
 };
 
 // HACK: Generating correct wasm for these is non-trivial so we hand them off to C.

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

@@ -1512,7 +1512,7 @@ export function generateWasmBody(
                     } else
                         ip = abort;
                 } else if (
-                    (opcode >= MintOpcode.MINT_LDC_I4_M1) &&
+                    (opcode >= MintOpcode.MINT_LDC_I4_0) &&
                     (opcode <= MintOpcode.MINT_LDC_R8)
                 ) {
                     if (!emit_ldc(builder, ip, opcode))
@@ -2597,6 +2597,8 @@ function emit_unop(builder: WasmBuilder, ip: MintOpcodePtr, opcode: MintOpcode):
 
         case MintOpcode.MINT_ADD_I4_IMM:
         case MintOpcode.MINT_MUL_I4_IMM:
+        case MintOpcode.MINT_AND_I4_IMM:
+        case MintOpcode.MINT_OR_I4_IMM:
         case MintOpcode.MINT_SHL_I4_IMM:
         case MintOpcode.MINT_SHR_I4_IMM:
         case MintOpcode.MINT_SHR_UN_I4_IMM:
@@ -2606,6 +2608,14 @@ function emit_unop(builder: WasmBuilder, ip: MintOpcodePtr, opcode: MintOpcode):
             builder.i32_const(getArgI16(ip, 3));
             break;
 
+        case MintOpcode.MINT_ADD_I4_IMM2:
+        case MintOpcode.MINT_MUL_I4_IMM2:
+        case MintOpcode.MINT_AND_I4_IMM2:
+        case MintOpcode.MINT_OR_I4_IMM2:
+            append_ldloc(builder, getArgU16(ip, 2), loadOp);
+            builder.i32_const(getArgI32(ip, 3));
+            break;
+
         case MintOpcode.MINT_ADD_I8_IMM:
         case MintOpcode.MINT_MUL_I8_IMM:
         case MintOpcode.MINT_SHL_I8_IMM:
@@ -2617,6 +2627,12 @@ function emit_unop(builder: WasmBuilder, ip: MintOpcodePtr, opcode: MintOpcode):
             builder.i52_const(getArgI16(ip, 3));
             break;
 
+        case MintOpcode.MINT_ADD_I8_IMM2:
+        case MintOpcode.MINT_MUL_I8_IMM2:
+            append_ldloc(builder, getArgU16(ip, 2), loadOp);
+            builder.i52_const(getArgI32(ip, 3));
+            break;
+
         default:
             append_ldloc(builder, getArgU16(ip, 2), loadOp);
             break;

src/mono/mono/mini/interp/jiterpreter-opcode-values.h

@@ -35,16 +35,18 @@ OPRANGE(MINT_RET_I4_IMM, MINT_RET_I8_IMM, ABORT_OUTSIDE_BRANCH_BLOCK_NONE)
 
 // High value because interp has to do a memory load for the immediate
 //  but we can inline it into the trace
-OPRANGE(MINT_LDC_I4_M1, MINT_LDC_R8, HIGH)
+OPRANGE(MINT_LDC_I4_0, MINT_LDC_R8, HIGH)
 
 OPRANGE(MINT_MOV_I4_I1, MINT_MOV_4, NORMAL)
 // High value for large/complex moves
 OPRANGE(MINT_MOV_8, MINT_MOV_8_4, HIGH)
 
 // Binops. Assume most of them are not any faster in jiterp
 OPRANGE(MINT_ADD_I4, MINT_CLT_UN_R8, NORMAL)
-// Unops and some superinsns. Most will not be faster in jiterp.
-OPRANGE(MINT_ADD1_I4, MINT_SHR_I8_IMM, NORMAL)
+// Unops. Most will not be faster in jiterp.
+OPRANGE(MINT_ADD1_I4, MINT_CEQ0_I4, NORMAL)
+// Some superinsns that will be faster in jiterp due to inline constants
+OPRANGE(MINT_ADD_I4_IMM, MINT_ADD_MUL_I8_IMM, HIGH)
 // Math intrinsics. We implement most of these by calling libc or using wasm opcodes
 OPRANGE(MINT_ASIN, MINT_MAXF, NORMAL)
 // Field operations. Null check optimization makes these more efficient than interp