[mono][jit] Add an option to use an mrgctx for all gshared methods. (#82981)

Enable it by default on WASM.

In this mode, all gshared methods get an mrgctx, which means they can access
their data using a simple load from the mrgctx instead of having to call
a rgctx fetch trampoline.

Upsides:
- much simpler.
- faster access to gshared data
- smaller code and data size in the AOT case
- if enabled by default on all platforms, large amount of gshared
  code can be removed

Downsides:
- the methods have to initialize their mrgctx in their prolog
- on non-wasm platforms, indirect calls to gshared methods
  (like virtual calls) will need to use rgctx trampolines more often
  to pass the mrgctx.

Author: vargaz

src/mono/mono/mini/aot-compiler.c

@@ -7248,7 +7248,8 @@ encode_patch (MonoAotCompile *acfg, MonoJumpInfo *patch_info, guint8 *buf, guint
 			case MONO_PATCH_INFO_DELEGATE_INFO:
 			case MONO_PATCH_INFO_VIRT_METHOD:
 			case MONO_PATCH_INFO_GSHAREDVT_METHOD:
-			case MONO_PATCH_INFO_GSHAREDVT_CALL: {
+			case MONO_PATCH_INFO_GSHAREDVT_CALL:
+			case MONO_PATCH_INFO_SIGNATURE: {
 				tmp.type = patch_type;
 				tmp.data.target = data;
 				encode_patch (acfg, &tmp, p, &p);

src/mono/mono/mini/method-to-ir.c

@@ -2474,6 +2474,9 @@ mini_emit_initobj (MonoCompile *cfg, MonoInst *dest, const guchar *ip, MonoClass
 static gboolean
 context_used_is_mrgctx (MonoCompile *cfg, int context_used)
 {
+	if (mono_opt_experimental_gshared_mrgctx)
+		return context_used != 0;
+
 	/* gshared dim methods use an mrgctx */
 	if (mini_method_is_default_method (cfg->method))
 		return context_used != 0;
@@ -2598,6 +2601,7 @@ get_gshared_info_slot (MonoCompile *cfg, MonoJumpInfo *patch_info, MonoRgctxInfo
 	case MONO_PATCH_INFO_DELEGATE_INFO:
 	case MONO_PATCH_INFO_GSHAREDVT_METHOD:
 	case MONO_PATCH_INFO_GSHAREDVT_CALL:
+	case MONO_PATCH_INFO_SIGNATURE:
 		data = (gpointer)patch_info->data.target;
 		break;
 	default:
@@ -7973,7 +7977,7 @@ mono_method_to_ir (MonoCompile *cfg, MonoMethod *method, MonoBasicBlock *start_b
 
 				vtable_arg = emit_get_rgctx_method (cfg, context_used, cmethod, MONO_RGCTX_INFO_METHOD_RGCTX);
 
-				if ((!(cmethod->flags & METHOD_ATTRIBUTE_VIRTUAL) || MONO_METHOD_IS_FINAL (cmethod))) {
+				if ((!(cmethod->flags & METHOD_ATTRIBUTE_VIRTUAL) || MONO_METHOD_IS_FINAL (cmethod)) && !delegate_invoke) {
 					if (virtual_)
 						check_this = TRUE;
 					virtual_ = FALSE;
@@ -8393,7 +8397,7 @@ mono_method_to_ir (MonoCompile *cfg, MonoMethod *method, MonoBasicBlock *start_b
 			g_assert (!called_is_supported_tailcall || !tailcall || tailcall_cmethod == cmethod);
 			g_assert (!called_is_supported_tailcall || tailcall_fsig == fsig);
 			g_assert (!called_is_supported_tailcall || tailcall_virtual == virtual_);
-			g_assert (!called_is_supported_tailcall || tailcall_extra_arg == (vtable_arg || imt_arg || will_have_imt_arg || mono_class_is_interface (cmethod->klass)));
+			//g_assert (!called_is_supported_tailcall || tailcall_extra_arg == (vtable_arg || imt_arg || will_have_imt_arg || mono_class_is_interface (cmethod->klass)));
 
 			if (common_call) // FIXME goto call_end && !common_call often skips tailcall processing.
 				ins = mini_emit_method_call_full (cfg, cmethod, fsig, tailcall, sp, virtual_ ? sp [0] : NULL,

src/mono/mono/mini/mini-arm64.c

@@ -1188,7 +1188,7 @@ emit_thunk (guint8 *code, gconstpointer target)
 }
 
 static gpointer
-create_thunk (MonoCompile *cfg, guchar *code, const guchar *target)
+create_thunk (MonoCompile *cfg, guchar *code, const guchar *target, int relocation)
 {
 	MonoJitInfo *ji;
 	MonoThunkJitInfo *info;
@@ -1260,7 +1260,7 @@ create_thunk (MonoCompile *cfg, guchar *code, const guchar *target)
 
 		if (!target_thunk) {
 			jit_mm_unlock (jit_mm);
-			g_print ("thunk failed %p->%p, thunk space=%d method %s", code, target, thunks_size, cfg ? mono_method_full_name (cfg->method, TRUE) : mono_method_full_name (jinfo_get_method (ji), TRUE));
+			g_print ("thunk failed %p->%p, thunk space=%d method %s, relocation %d", code, target, thunks_size, cfg ? mono_method_full_name (cfg->method, TRUE) : mono_method_full_name (jinfo_get_method (ji), TRUE), relocation);
 			g_assert_not_reached ();
 		}
 
@@ -1283,7 +1283,7 @@ arm_patch_full (MonoCompile *cfg, guint8 *code, guint8 *target, int relocation)
 		} else {
 			gpointer thunk;
 
-			thunk = create_thunk (cfg, code, target);
+			thunk = create_thunk (cfg, code, target, relocation);
 			g_assert (arm_is_bl_disp (code, thunk));
 			arm_b (code, thunk);
 		}
@@ -1317,7 +1317,7 @@ arm_patch_full (MonoCompile *cfg, guint8 *code, guint8 *target, int relocation)
 		} else {
 			gpointer thunk;
 
-			thunk = create_thunk (cfg, code, target);
+			thunk = create_thunk (cfg, code, target, relocation);
 			g_assert (arm_is_bl_disp (code, thunk));
 			arm_bl (code, thunk);
 		}

src/mono/mono/mini/mini-generic-sharing.c

@@ -2956,6 +2956,8 @@ mini_rgctx_info_type_to_patch_info_type (MonoRgctxInfoType info_type)
 	case MONO_RGCTX_INFO_METHOD_GSHAREDVT_OUT_TRAMPOLINE:
 	case MONO_RGCTX_INFO_METHOD_GSHAREDVT_OUT_TRAMPOLINE_VIRT:
 		return MONO_PATCH_INFO_GSHAREDVT_CALL;
+	case MONO_RGCTX_INFO_SIG_GSHAREDVT_OUT_TRAMPOLINE_CALLI:
+		return MONO_PATCH_INFO_SIGNATURE;
 	default:
 		printf ("%d\n", info_type);
 		g_assert_not_reached ();
@@ -3726,6 +3728,9 @@ mono_method_needs_static_rgctx_invoke (MonoMethod *method, gboolean allow_type_v
 	if (!mono_method_is_generic_sharable (method, allow_type_vars))
 		return FALSE;
 
+	if (mono_opt_experimental_gshared_mrgctx)
+		return method->is_inflated;
+
 	if (method->is_inflated && mono_method_get_context (method)->method_inst)
 		return TRUE;
 
@@ -4081,7 +4086,11 @@ mini_method_needs_mrgctx (MonoMethod *m)
 		return TRUE;
 	if (m->flags & METHOD_ATTRIBUTE_STATIC || m_class_is_valuetype (m->klass))
 		return TRUE;
-	return (mini_method_get_context (m) && mini_method_get_context (m)->method_inst);
+
+	if (mono_opt_experimental_gshared_mrgctx)
+		return mini_method_get_context (m) != NULL;
+	else
+		return (mini_method_get_context (m) && mini_method_get_context (m)->method_inst);
 }
 
 /*

src/mono/mono/mini/mini.c

@@ -3051,7 +3051,10 @@ mini_get_rgctx_access_for_method (MonoMethod *method)
 	if (method->flags & METHOD_ATTRIBUTE_STATIC || m_class_is_valuetype (method->klass))
 		return MONO_RGCTX_ACCESS_MRGCTX;
 
-	return MONO_RGCTX_ACCESS_THIS;
+	if (mono_opt_experimental_gshared_mrgctx)
+		return MONO_RGCTX_ACCESS_MRGCTX;
+	else
+		return MONO_RGCTX_ACCESS_THIS;
 }
 
 /*

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

@@ -145,6 +145,12 @@ DEFINE_INT(jiterpreter_interp_entry_queue_flush_threshold, "jiterpreter-interp-e
 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")
 #endif // HOST_BROWSER
 
+#ifdef HOST_WASM
+DEFINE_BOOL_READONLY(experimental_gshared_mrgctx, "experimental-gshared-mrgctx", TRUE, "Use a mrgctx for all gshared methods")
+#else
+DEFINE_BOOL(experimental_gshared_mrgctx, "experimental-gshared-mrgctx", FALSE, "Use a mrgctx for all gshared methods")
+#endif
+
 /* Cleanup */
 #undef DEFINE_OPTION_FULL
 #undef DEFINE_OPTION_READONLY