Update jiterpreter interp_entry wrapper generator and re-enable jiterp

src/mono/mono/mini/interp/interp-internals.h

@@ -339,12 +339,15 @@ mono_jiterp_overflow_check_u4 (guint32 lhs, guint32 rhs, int opcode);
 void
 mono_jiterp_ld_delegate_method_ptr (gpointer *destination, MonoDelegate **source);
 
-int
+void
 mono_jiterp_stackval_to_data (MonoType *type, stackval *val, void *data);
 
-int
+void
 mono_jiterp_stackval_from_data (MonoType *type, stackval *result, const void *data);
 
+int
+mono_jiterp_get_arg_offset (InterpMethod *imethod, MonoMethodSignature *sig, int index);
+
 gpointer
 mono_jiterp_frame_data_allocator_alloc (FrameDataAllocator *stack, InterpFrame *frame, int size);
 

src/mono/mono/mini/interp/interp.c

@@ -8687,18 +8687,22 @@ mono_ee_interp_init (const char *opts)
 }
 
 #ifdef HOST_BROWSER
-EMSCRIPTEN_KEEPALIVE int
+EMSCRIPTEN_KEEPALIVE void
 mono_jiterp_stackval_to_data (MonoType *type, stackval *val, void *data)
 {
-	g_error ("FIXME");
-	return 0; //stackval_to_data (type, val, data, FALSE);
+	stackval_to_data (type, val, data, FALSE);
 }
 
-EMSCRIPTEN_KEEPALIVE int
+EMSCRIPTEN_KEEPALIVE void
 mono_jiterp_stackval_from_data (MonoType *type, stackval *result, const void *data)
 {
-	g_error ("FIXME");
-	return 0; //stackval_from_data (type, result, data, FALSE);
+	stackval_from_data (type, result, data, FALSE);
+}
+
+EMSCRIPTEN_KEEPALIVE int
+mono_jiterp_get_arg_offset (InterpMethod *imethod, MonoMethodSignature *sig, int index)
+{
+	return get_arg_offset_fast (imethod, sig, index);
 }
 
 EMSCRIPTEN_KEEPALIVE int
@@ -8778,7 +8782,7 @@ mono_jiterp_isinst (MonoObject* object, MonoClass* klass)
 //  in the correct place and compute the stack offset, then it passes that in to this
 //  function in order to actually enter the interpreter and process the return value
 EMSCRIPTEN_KEEPALIVE void
-mono_jiterp_interp_entry (JiterpEntryData *_data, stackval *sp_args, void *res)
+mono_jiterp_interp_entry (JiterpEntryData *_data, void *res)
 {
 	JiterpEntryDataHeader header;
 	MonoType *type;
@@ -8791,7 +8795,6 @@ mono_jiterp_interp_entry (JiterpEntryData *_data, stackval *sp_args, void *res)
 
 	g_assert(header.rmethod);
 	g_assert(header.rmethod->method);
-	g_assert(sp_args);
 
 	stackval *sp = (stackval*)header.context->stack_pointer;
 
@@ -8800,8 +8803,11 @@ mono_jiterp_interp_entry (JiterpEntryData *_data, stackval *sp_args, void *res)
 	frame.stack = sp;
 	frame.retval = sp;
 
-	header.context->stack_pointer = (guchar*)sp_args;
-	g_assert ((guchar*)sp_args < header.context->stack_end);
+	int params_size = get_arg_offset_fast (header.rmethod, NULL, header.params_count);
+	// g_printf ("jiterp_interp_entry: rmethod=%d, params_count=%d, params_size=%d\n", header.rmethod, header.params_count, params_size);
+	header.context->stack_pointer = (guchar*)ALIGN_TO ((guchar*)sp + params_size, MINT_STACK_ALIGNMENT);
+;
+	g_assert (header.context->stack_pointer < header.context->stack_end);
 
 	MONO_ENTER_GC_UNSAFE;
 	mono_interp_exec_method (&frame, header.context, NULL);

src/mono/mono/mini/interp/jiterpreter.c

@@ -1169,6 +1169,7 @@ mono_jiterp_trace_transfer (
 #define JITERP_MEMBER_BACKWARD_BRANCH_OFFSETS 10
 #define JITERP_MEMBER_BACKWARD_BRANCH_OFFSETS_COUNT 11
 #define JITERP_MEMBER_CLAUSE_DATA_OFFSETS 12
+#define JITERP_MEMBER_PARAMS_COUNT 13
 
 // we use these helpers at JIT time to figure out where to do memory loads and stores
 EMSCRIPTEN_KEEPALIVE size_t
@@ -1196,6 +1197,8 @@ mono_jiterp_get_member_offset (int member) {
 			return offsetof (InterpMethod, clause_data_offsets);
 		case JITERP_MEMBER_RMETHOD:
 			return offsetof (JiterpEntryDataHeader, rmethod);
+		case JITERP_MEMBER_PARAMS_COUNT:
+			return offsetof (JiterpEntryDataHeader, params_count);
 		case JITERP_MEMBER_SPAN_LENGTH:
 			return offsetof (MonoSpanOfVoid, _length);
 		case JITERP_MEMBER_SPAN_DATA:

src/mono/mono/mini/interp/jiterpreter.h

@@ -106,6 +106,7 @@ typedef struct {
 	ThreadContext *context;
 	gpointer orig_domain;
 	gpointer attach_cookie;
+	int params_count;
 } JiterpEntryDataHeader;
 
 // we optimize delegate calls by attempting to cache the delegate invoke
@@ -136,7 +137,7 @@ void
 mono_jiterp_do_safepoint (InterpFrame *frame, guint16 *ip);
 
 void
-mono_jiterp_interp_entry (JiterpEntryData *_data, stackval *sp_args, void *res);
+mono_jiterp_interp_entry (JiterpEntryData *_data, void *res);
 
 gpointer
 mono_jiterp_imethod_to_ftnptr (InterpMethod *imethod);

src/mono/mono/utils/options-def.h

@@ -73,11 +73,11 @@ DEFINE_BOOL(interp_simd_packedsimd, "interp-simd-packedsimd", FALSE, "Enable int
 //  and wasm modules between threads. before these can be enabled we need to implement all that
 #ifdef DISABLE_THREADS
 // traces_enabled controls whether the jiterpreter will JIT individual interpreter opcode traces
-DEFINE_BOOL(jiterpreter_traces_enabled, "jiterpreter-traces-enabled", FALSE, "JIT interpreter opcode traces into WASM")
+DEFINE_BOOL(jiterpreter_traces_enabled, "jiterpreter-traces-enabled", TRUE, "JIT interpreter opcode traces into WASM")
 // interp_entry_enabled controls whether specialized interp_entry wrappers will be jitted
-DEFINE_BOOL(jiterpreter_interp_entry_enabled, "jiterpreter-interp-entry-enabled", FALSE, "JIT specialized WASM interp_entry wrappers")
+DEFINE_BOOL(jiterpreter_interp_entry_enabled, "jiterpreter-interp-entry-enabled", TRUE, "JIT specialized WASM interp_entry wrappers")
 // jit_call_enabled controls whether do_jit_call will use specialized trampolines for hot call sites
-DEFINE_BOOL(jiterpreter_jit_call_enabled, "jiterpreter-jit-call-enabled", FALSE, "JIT specialized WASM do_jit_call trampolines")
+DEFINE_BOOL(jiterpreter_jit_call_enabled, "jiterpreter-jit-call-enabled", TRUE, "JIT specialized WASM do_jit_call trampolines")
 #else
 // traces_enabled controls whether the jiterpreter will JIT individual interpreter opcode traces
 DEFINE_BOOL_READONLY(jiterpreter_traces_enabled, "jiterpreter-traces-enabled", FALSE, "JIT interpreter opcode traces into WASM")
@@ -101,7 +101,7 @@ DEFINE_BOOL(jiterpreter_call_resume_enabled, "jiterpreter-call-resume-enabled",
 //  stats for options like estimateHeat, but raises overhead.
 DEFINE_BOOL(jiterpreter_disable_heuristic, "jiterpreter-disable-heuristic", FALSE, "Always insert trace entry points for more accurate statistics")
 // Automatically prints stats at app exit or when jiterpreter_dump_stats is called
-DEFINE_BOOL(jiterpreter_stats_enabled, "jiterpreter-stats-enabled", FALSE, "Automatically print jiterpreter statistics")
+DEFINE_BOOL(jiterpreter_stats_enabled, "jiterpreter-stats-enabled", TRUE, "Automatically print jiterpreter statistics")
 // Continue counting hits for traces that fail to compile and use it to estimate
 //  the relative importance of the opcode that caused them to abort
 DEFINE_BOOL(jiterpreter_estimate_heat, "jiterpreter-estimate-heat", FALSE, "Maintain accurate hit count for all trace entry points")
@@ -145,12 +145,12 @@ DEFINE_INT(jiterpreter_jit_call_trampoline_hit_count, "jiterpreter-jit-call-hit-
 // After a do_jit_call call site is hit this many times without being jitted, we will flush the JIT queue
 DEFINE_INT(jiterpreter_jit_call_queue_flush_threshold, "jiterpreter-jit-call-queue-flush-threshold", 5000, "Flush the do_jit_call JIT queue after an unJITted call site has this many hits")
 // After a generic interp_entry wrapper is hit this many times, we will queue it to be jitted
-DEFINE_INT(jiterpreter_interp_entry_trampoline_hit_count, "jiterpreter-interp-entry-hit-count", 1000, "Queue specialized interp_entry wrapper for JIT after this many hits")
+DEFINE_INT(jiterpreter_interp_entry_trampoline_hit_count, "jiterpreter-interp-entry-hit-count", 1, "Queue specialized interp_entry wrapper for JIT after this many hits")
 // After a generic interp_entry wrapper is hit this many times without being jitted, we will flush the JIT queue
-DEFINE_INT(jiterpreter_interp_entry_queue_flush_threshold, "jiterpreter-interp-entry-queue-flush-threshold", 3000, "Flush the interp_entry JIT queue after an unJITted call site has this many hits")
+DEFINE_INT(jiterpreter_interp_entry_queue_flush_threshold, "jiterpreter-interp-entry-queue-flush-threshold", 300, "Flush the interp_entry JIT queue after an unJITted call site has this many hits")
 // In degenerate cases the jiterpreter could end up generating lots of WASM, so shut off jitting once it reaches this limit
 // Each wasm byte likely maps to multiple bytes of native code, so it's important for this limit not to be too high
-DEFINE_INT(jiterpreter_wasm_bytes_limit, "jiterpreter-wasm-bytes-limit", 6 * 1024 * 1024, "Disable jiterpreter code generation once this many bytes of WASM have been generated")
+DEFINE_INT(jiterpreter_wasm_bytes_limit, "jiterpreter-wasm-bytes-limit", 16 * 1024 * 1024, "Disable jiterpreter code generation once this many bytes of WASM have been generated")
 #endif // HOST_BROWSER
 
 #ifdef TARGET_WASM

src/mono/wasm/runtime/cwraps.ts

@@ -127,6 +127,7 @@ const fn_signatures: SigLine[] = [
     [true, "mono_jiterp_get_opcode_value_table_entry", "number", ["number"]],
     [true, "mono_jiterp_get_simd_intrinsic", "number", ["number", "number"]],
     [true, "mono_jiterp_get_simd_opcode", "number", ["number", "number"]],
+    [true, "mono_jiterp_get_arg_offset", "number", ["number", "number", "number"]],
     ...legacy_interop_cwraps
 ];
 
@@ -250,6 +251,7 @@ export interface t_Cwraps {
     mono_jiterp_get_opcode_value_table_entry(opcode: number): number;
     mono_jiterp_get_simd_intrinsic(arity: number, index: number): VoidPtr;
     mono_jiterp_get_simd_opcode(arity: number, index: number): number;
+    mono_jiterp_get_arg_offset (imethod: number, sig: number, index: number): number;
 }
 
 const wrapped_c_functions: t_Cwraps = <any>{};

src/mono/wasm/runtime/jiterpreter-interp-entry.ts

@@ -5,7 +5,7 @@ import { mono_assert, MonoMethod, MonoType } from "./types";
 import { NativePointer } from "./types/emscripten";
 import { Module } from "./globals";
 import {
-    getU32_unaligned, _zero_region
+    setI32, getU32_unaligned, _zero_region
 } from "./memory";
 import { WasmOpcode } from "./jiterpreter-opcodes";
 import cwraps from "./cwraps";
@@ -36,6 +36,7 @@ typedef struct {
     ThreadContext *context; // 4
     gpointer orig_domain; // 8
     gpointer attach_cookie; // 12
+    int params_count; // 16
 } JiterpEntryDataHeader;
 */
 
@@ -231,7 +232,6 @@ function flush_wasm_entry_trampoline_jit_queue() {
             "interp_entry",
             {
                 "pData": WasmValtype.i32,
-                "sp_args": WasmValtype.i32,
                 "res": WasmValtype.i32,
             },
             WasmValtype.void, true
@@ -243,7 +243,7 @@ function flush_wasm_entry_trampoline_jit_queue() {
                 "result": WasmValtype.i32,
                 "value": WasmValtype.i32
             },
-            WasmValtype.i32, true
+            WasmValtype.void, true
         );
     } else
         builder.clear(constantSlots);
@@ -411,10 +411,10 @@ function flush_wasm_entry_trampoline_jit_queue() {
 }
 
 function append_stackval_from_data(
-    builder: WasmBuilder, type: MonoType, valueName: string
+    builder: WasmBuilder, imethod: number, type: MonoType, valueName: string, argIndex: number
 ) {
-    const stackvalSize = cwraps.mono_jiterp_get_size_of_stackval();
     const rawSize = cwraps.mono_jiterp_type_get_raw_value_size(type);
+    const offset = cwraps.mono_jiterp_get_arg_offset(imethod, 0, argIndex);
 
     switch (rawSize) {
         case 256: {
@@ -423,10 +423,7 @@ function append_stackval_from_data(
             builder.local(valueName);
 
             builder.appendU8(WasmOpcode.i32_store);
-            builder.appendMemarg(0, 2);
-
-            // Fixed stackval size
-            builder.i32_const(stackvalSize);
+            builder.appendMemarg(offset, 2);
             break;
         }
 
@@ -465,30 +462,25 @@ function append_stackval_from_data(
             }
 
             builder.appendU8(WasmOpcode.i32_store);
-            builder.appendMemarg(0, 2);
-
-            // Fixed stackval size
-            builder.i32_const(stackvalSize);
+            builder.appendMemarg(offset, 2);
             break;
         }
 
         default: {
-            // Call stackval_from_data to copy the value and get its size
+            // Call stackval_from_data to copy the value
             builder.ptr_const(type);
             // result
             builder.local("sp_args");
+            // apply offset
+            builder.i32_const(offset);
+            builder.appendU8(WasmOpcode.i32_add);
             // value
             builder.local(valueName);
 
             builder.callImport("stackval_from_data");
             break;
         }
     }
-
-    // Value size is on the stack, add it to sp_args and update it
-    builder.local("sp_args");
-    builder.appendU8(WasmOpcode.i32_add);
-    builder.local("sp_args", WasmOpcode.set_local);
 }
 
 function generate_wasm_body(
@@ -505,6 +497,13 @@ function generate_wasm_body(
     const scratchBuffer = <any>Module._malloc(sizeOfJiterpEntryData);
     _zero_region(scratchBuffer, sizeOfJiterpEntryData);
 
+    // Initialize the parameter count in the data blob. This is used to calculate the new value of sp
+    //  before entering the interpreter
+    setI32(
+        scratchBuffer + getMemberOffset(JiterpMember.ParamsCount),
+        info.paramTypes.length + (info.hasThisReference ? 1 : 0)
+    );
+
     // the this-reference may be a boxed struct that needs to be unboxed, for example calling
     //  methods like object.ToString on structs will end up with the unbox flag set
     // instead of passing an extra 'unbox' argument to every wrapper, though, the flag is hidden
@@ -559,7 +558,7 @@ function generate_wasm_body(
 
     if (info.hasThisReference) {
         // null type for raw ptr copy
-        append_stackval_from_data(builder, <any>0, "this_arg");
+        append_stackval_from_data(builder, info.imethod, <any>0, "this_arg", 0);
     }
 
     /*
@@ -576,11 +575,10 @@ function generate_wasm_body(
 
     for (let i = 0; i < info.paramTypes.length; i++) {
         const type = <any>info.paramTypes[i];
-        append_stackval_from_data(builder, type, `arg${i}`);
+        append_stackval_from_data(builder, info.imethod, type, `arg${i}`, i + (info.hasThisReference ? 1 : 0));
     }
 
     builder.local("scratchBuffer");
-    builder.local("sp_args");
     if (info.hasReturnValue)
         builder.local("res");
     else

src/mono/wasm/runtime/jiterpreter-support.ts

@@ -1666,6 +1666,7 @@ export const enum JiterpMember {
     BackwardBranchOffsets = 10,
     BackwardBranchOffsetsCount = 11,
     ClauseDataOffsets = 12,
+    ParamsCount = 13,
 }
 
 const memberOffsets: { [index: number]: number } = {};