[mono][llvm] Add a fastpath to calling mini_init_method_rgctx (). (#8…

…4226)

* [mono][llvm] Add a fastpath to calling mini_init_method_rgctx ().

* Add arm64 JIT support.

src/mono/mono/mini/cpu-arm64.mdesc

@@ -523,5 +523,6 @@ expand_r8: dest:x src1:f len:4
 
 generic_class_init: src1:a len:44 clob:c
 gc_safe_point: src1:i len:12 clob:c
+init_mrgctx: src1:a src2:i len:44 clob:c
 
 fill_prof_call_ctx: src1:i len:128

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

@@ -6750,7 +6750,15 @@ mono_method_to_ir (MonoCompile *cfg, MonoMethod *method, MonoBasicBlock *start_b
 			EMIT_NEW_PCONST (cfg, args [1], info);
 
 		cfg->init_method_rgctx_ins_arg = args [1];
-		cfg->init_method_rgctx_ins = mono_emit_jit_icall (cfg, mini_init_method_rgctx, args);
+		if (COMPILE_LLVM (cfg) || cfg->backend->have_init_mrgctx) {
+			MONO_INST_NEW (cfg, ins, OP_INIT_MRGCTX);
+			ins->sreg1 = args [0]->dreg;
+			ins->sreg2 = args [1]->dreg;
+			MONO_ADD_INS (cfg->cbb, ins);
+			cfg->init_method_rgctx_ins = ins;
+		} else {
+			cfg->init_method_rgctx_ins = mono_emit_jit_icall (cfg, mini_init_method_rgctx, args);
+		}
 	}
 
 	if (cfg->gsharedvt && cfg->method == method) {

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

@@ -5023,6 +5023,27 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 			mono_arm_patch (jump, code, MONO_R_ARM64_CBZ);
 			break;
 		}
+		case OP_INIT_MRGCTX: {
+			int field_offset;
+			guint8 *jump;
+
+			field_offset = MONO_STRUCT_OFFSET (MonoMethodRuntimeGenericContext, entries);
+
+			/* Load mrgxtx->entries */
+			arm_ldrx (code, ARMREG_IP0, sreg1, field_offset);
+			jump = code;
+			arm_cbnzx (code, ARMREG_IP0, 0);
+
+			/* Slowpath */
+			g_assert (sreg1 == ARMREG_R0);
+			if (sreg2 != ARMREG_R1)
+				arm_movx (code, ARMREG_R1, sreg2);
+			code = emit_call (cfg, code, MONO_PATCH_INFO_JIT_ICALL_ID,
+							  GUINT_TO_POINTER (MONO_JIT_ICALL_mini_init_method_rgctx));
+
+			mono_arm_patch (jump, code, MONO_R_ARM64_CBZ);
+			break;
+		}
 
 		case OP_CHECK_THIS:
 			arm_ldrb (code, ARMREG_LR, sreg1, 0);

src/mono/mono/mini/mini-arm64.h

@@ -179,6 +179,7 @@ typedef struct {
 #define MONO_ARCH_HAVE_DECOMPOSE_LONG_OPTS 1
 #define MONO_ARCH_FLOAT32_SUPPORTED 1
 #define MONO_ARCH_HAVE_INTERP_PINVOKE_TRAMP 1
+#define MONO_ARCH_HAVE_INIT_MRGCTX 1
 #define MONO_ARCH_LLVM_TARGET_LAYOUT "e-i64:64-i128:128-n32:64-S128"
 #ifdef TARGET_OSX
 #define MONO_ARCH_FORCE_FLOAT32 1

src/mono/mono/mini/mini-llvm.c

@@ -7415,6 +7415,52 @@ MONO_RESTORE_WARNING
 			break;
 		}
 
+		case OP_INIT_MRGCTX: {
+			LLVMValueRef val, cmp, callee;
+			LLVMBasicBlockRef poll_bb, cont_bb;
+			LLVMValueRef args [2];
+			static LLVMTypeRef icall_sig;
+
+			g_assert (cfg->compile_aot);
+
+			/*
+			 * if (!((MonoMethodRuntimeGenericContext)->lhs)->entries)
+			 *   mini_init_method_rgctx (lhs, rhs);
+			 */
+
+			LLVMValueRef offset = const_int32 (MONO_STRUCT_OFFSET (MonoMethodRuntimeGenericContext, entries));
+			LLVMValueRef ptr = LLVMBuildAdd (builder, convert (ctx, lhs, IntPtrType ()), convert (ctx, offset, IntPtrType ()), "");
+
+			val = emit_load (builder, IntPtrType (), convert (ctx, ptr, pointer_type (IntPtrType ())), "", TRUE);
+			cmp = LLVMBuildICmp (builder, LLVMIntEQ, val, LLVMConstNull (LLVMTypeOf (val)), "");
+			poll_bb = gen_bb (ctx, "POLL_BB");
+			cont_bb = gen_bb (ctx, "CONT_BB");
+
+			args [0] = cmp;
+			args [1] = LLVMConstInt (LLVMInt1Type (), 0, FALSE);
+			cmp = call_intrins (ctx, INTRINS_EXPECT_I1, args, "");
+
+			mono_llvm_build_weighted_branch (builder, cmp, poll_bb, cont_bb, 1, 1000);
+
+			ctx->builder = builder = create_builder (ctx);
+			LLVMPositionBuilderAtEnd (builder, poll_bb);
+
+			if (!icall_sig)
+				icall_sig = LLVMFunctionType2 (LLVMVoidType (), IntPtrType (), IntPtrType (), FALSE);
+
+			args [0] = convert (ctx, lhs, IntPtrType ());
+			args [1] = convert (ctx, rhs, IntPtrType ());
+
+			callee = get_callee (ctx, icall_sig, MONO_PATCH_INFO_JIT_ICALL_ID, GUINT_TO_POINTER (MONO_JIT_ICALL_mini_init_method_rgctx));
+			LLVMBuildCall2 (builder, icall_sig, callee, args, 2, "");
+			LLVMBuildBr (builder, cont_bb);
+
+			ctx->builder = builder = create_builder (ctx);
+			LLVMPositionBuilderAtEnd (builder, cont_bb);
+			ctx->bblocks [bb->block_num].end_bblock = cont_bb;
+			break;
+		}
+
 			/*
 			 * Overflow opcodes.
 			 */

src/mono/mono/mini/mini-ops.h

@@ -1307,6 +1307,13 @@ MINI_OP(OP_GC_SAFE_POINT, "gc_safe_point", NONE, IREG, NONE)
  */
 MINI_OP(OP_GENERIC_CLASS_INIT, "generic_class_init", NONE, IREG, NONE)
 
+/*
+ * Call mini_init_method_rgctx () if needed.
+ * sreg1 is a MonoMethodRuntimeGenericContext.
+ * sreg2 is a MonoGSharedMethodInfo.
+ */
+MINI_OP(OP_INIT_MRGCTX, "init_mrgctx", NONE, IREG, IREG)
+
 /* Arch specific opcodes */
 #if defined(TARGET_X86) || defined(TARGET_AMD64)
 MINI_OP(OP_X86_TEST_NULL,          "x86_test_null", NONE, IREG, NONE)

src/mono/mono/mini/mini.c

@@ -3018,6 +3018,9 @@ init_backend (MonoBackend *backend)
 #ifdef MONO_ARCH_FORCE_FLOAT32
 	backend->force_float32 = 1;
 #endif
+#ifdef MONO_ARCH_HAVE_INIT_MRGCTX
+	backend->have_init_mrgctx = 1;
+#endif
 }
 
 static gboolean

src/mono/mono/mini/mini.h

@@ -1246,6 +1246,7 @@ typedef struct {
 	gboolean         have_op_tailcall_membase : 1;
 	gboolean         have_op_tailcall_reg : 1;
 	gboolean         have_volatile_non_param_register : 1;
+	guint            have_init_mrgctx : 1;
 	guint            gshared_supported : 1;
 	guint            ilp32 : 1;
 	guint            need_got_var : 1;