2009-06-10 Zoltan Varga <vargaz@gmail.com>
[mono.git] / mono / mini / mini-amd64.c
index 8239e943c78f384adbd8c0126f68d74683d6f10b..b083dbf2f0632bb4a7e5c325ae95c227a8ddcca0 100644 (file)
@@ -1248,7 +1248,18 @@ mono_arch_allocate_vars (MonoCompile *cfg)
 
        /*
         * We use the ABI calling conventions for managed code as well.
-        * Exception: valuetypes are never passed or returned in registers.
+        * Exception: valuetypes are only sometimes passed or returned in registers.
+        */
+
+       /*
+        * The stack looks like this:
+        * <incoming arguments passed on the stack>
+        * <return value>
+        * <lmf/caller saved registers>
+        * <locals>
+        * <spill area>
+        * <localloc area>  -> grows dynamically
+        * <params area>
         */
 
        if (cfg->arch.omit_fp) {
@@ -1263,8 +1274,6 @@ mono_arch_allocate_vars (MonoCompile *cfg)
 
        if (cfg->method->save_lmf) {
                /* Reserve stack space for saving LMF */
-               /* mono_arch_find_jit_info () expects to find the LMF at a fixed offset */
-               g_assert (offset == 0);
                if (cfg->arch.omit_fp) {
                        cfg->arch.lmf_offset = offset;
                        offset += sizeof (MonoLMF);
@@ -1514,6 +1523,19 @@ mono_arch_create_vars (MonoCompile *cfg)
                        mono_print_ins (cfg->vret_addr);
                }
        }
+
+#ifdef MONO_AMD64_NO_PUSHES
+       /*
+        * When this is set, we pass arguments on the stack by moves, and by allocating 
+        * a bigger stack frame, instead of pushes.
+        * Pushes complicate exception handling because the arguments on the stack have
+        * to be popped each time a frame is unwound. They also make fp elimination
+        * impossible.
+        * FIXME: This doesn't work inside filter/finally clauses, since those execute
+        * on a new frame which doesn't include a param area.
+        */
+       cfg->arch.no_pushes = TRUE;
+#endif
 }
 
 static void
@@ -1599,9 +1621,13 @@ emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
        sig_arg->inst_p0 = tmp_sig;
        MONO_ADD_INS (cfg->cbb, sig_arg);
 
-       MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
-       arg->sreg1 = sig_arg->dreg;
-       MONO_ADD_INS (cfg->cbb, arg);
+       if (cfg->arch.no_pushes) {
+               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, cinfo->sig_cookie.offset, sig_arg->dreg);
+       } else {
+               MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
+               arg->sreg1 = sig_arg->dreg;
+               MONO_ADD_INS (cfg->cbb, arg);
+       }
 }
 
 static inline LLVMArgStorage
@@ -1729,7 +1755,39 @@ mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
        }
 
        if (cinfo->need_stack_align) {
-               MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
+               if (!cfg->arch.no_pushes)
+                       MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
+       }
+
+       /* 
+        * Emit all arguments which are passed on the stack to prevent register
+        * allocation problems.
+        */
+       if (cfg->arch.no_pushes) {
+               for (i = 0; i < n; ++i) {
+                       MonoType *t;
+                       ainfo = cinfo->args + i;
+
+                       in = call->args [i];
+
+                       if (sig->hasthis && i == 0)
+                               t = &mono_defaults.object_class->byval_arg;
+                       else
+                               t = sig->params [i - sig->hasthis];
+
+                       if (ainfo->storage == ArgOnStack && !MONO_TYPE_ISSTRUCT (t) && !call->tail_call) {
+                               if (!t->byref) {
+                                       if (t->type == MONO_TYPE_R4)
+                                               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
+                                       else if (t->type == MONO_TYPE_R8)
+                                               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
+                                       else
+                                               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
+                               } else {
+                                       MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
+                               }
+                       }
+               }
        }
 
        /*
@@ -1800,22 +1858,26 @@ mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
                                        MONO_ADD_INS (cfg->cbb, arg);
                                }
                        } else {
-                               MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
-                               arg->sreg1 = in->dreg;
-                               if (!sig->params [i - sig->hasthis]->byref) {
-                                       if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R4) {
-                                               MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
-                                               arg->opcode = OP_STORER4_MEMBASE_REG;
-                                               arg->inst_destbasereg = X86_ESP;
-                                               arg->inst_offset = 0;
-                                       } else if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R8) {
-                                               MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
-                                               arg->opcode = OP_STORER8_MEMBASE_REG;
-                                               arg->inst_destbasereg = X86_ESP;
-                                               arg->inst_offset = 0;
+                               if (cfg->arch.no_pushes) {
+                                       /* Already done */
+                               } else {
+                                       MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
+                                       arg->sreg1 = in->dreg;
+                                       if (!sig->params [i - sig->hasthis]->byref) {
+                                               if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R4) {
+                                                       MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
+                                                       arg->opcode = OP_STORER4_MEMBASE_REG;
+                                                       arg->inst_destbasereg = X86_ESP;
+                                                       arg->inst_offset = 0;
+                                               } else if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R8) {
+                                                       MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
+                                                       arg->opcode = OP_STORER8_MEMBASE_REG;
+                                                       arg->inst_destbasereg = X86_ESP;
+                                                       arg->inst_offset = 0;
+                                               }
                                        }
+                                       MONO_ADD_INS (cfg->cbb, arg);
                                }
-                               MONO_ADD_INS (cfg->cbb, arg);
                        }
                        break;
                default:
@@ -1928,6 +1990,8 @@ mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
                MonoInst *vtaddr, *load;
                vtaddr = mono_compile_create_var (cfg, &ins->klass->byval_arg, OP_LOCAL);
                
+               g_assert (!cfg->arch.no_pushes);
+
                MONO_INST_NEW (cfg, load, OP_LDADDR);
                load->inst_p0 = vtaddr;
                vtaddr->flags |= MONO_INST_INDIRECT;
@@ -1951,20 +2015,36 @@ mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
                }
        } else {
                if (size == 8) {
-                       /* Can't use this for < 8 since it does an 8 byte memory load */
-                       MONO_INST_NEW (cfg, arg, OP_X86_PUSH_MEMBASE);
-                       arg->inst_basereg = src->dreg;
-                       arg->inst_offset = 0;
-                       MONO_ADD_INS (cfg->cbb, arg);
+                       if (cfg->arch.no_pushes) {
+                               int dreg = mono_alloc_ireg (cfg);
+
+                               MONO_EMIT_NEW_LOAD_MEMBASE (cfg, dreg, src->dreg, 0);
+                               MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, dreg);
+                       } else {
+                               /* Can't use this for < 8 since it does an 8 byte memory load */
+                               MONO_INST_NEW (cfg, arg, OP_X86_PUSH_MEMBASE);
+                               arg->inst_basereg = src->dreg;
+                               arg->inst_offset = 0;
+                               MONO_ADD_INS (cfg->cbb, arg);
+                       }
                } else if (size <= 40) {
-                       MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, ALIGN_TO (size, 8));
-                       mini_emit_memcpy (cfg, X86_ESP, 0, src->dreg, 0, size, 4);
+                       if (cfg->arch.no_pushes) {
+                               mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
+                       } else {
+                               MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, ALIGN_TO (size, 8));
+                               mini_emit_memcpy (cfg, X86_ESP, 0, src->dreg, 0, size, 4);
+                       }
                } else {
-                       MONO_INST_NEW (cfg, arg, OP_X86_PUSH_OBJ);
-                       arg->inst_basereg = src->dreg;
-                       arg->inst_offset = 0;
-                       arg->inst_imm = size;
-                       MONO_ADD_INS (cfg->cbb, arg);
+                       if (cfg->arch.no_pushes) {
+                               // FIXME: Code growth
+                               mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
+                       } else {
+                               MONO_INST_NEW (cfg, arg, OP_X86_PUSH_OBJ);
+                               arg->inst_basereg = src->dreg;
+                               arg->inst_offset = 0;
+                               arg->inst_imm = size;
+                               MONO_ADD_INS (cfg->cbb, arg);
+                       }
                }
        }
 }
@@ -1991,28 +2071,15 @@ mono_arch_emit_setret (MonoCompile *cfg, MonoMethod *method, MonoInst *val)
 }
 
 #define EMIT_COND_BRANCH(ins,cond,sign) \
-if (ins->flags & MONO_INST_BRLABEL) { \
-        if (ins->inst_i0->inst_c0) { \
-               x86_branch (code, cond, cfg->native_code + ins->inst_i0->inst_c0, sign); \
-        } else { \
-               mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_LABEL, ins->inst_i0); \
-               if ((cfg->opt & MONO_OPT_BRANCH) && \
-                    x86_is_imm8 (ins->inst_i0->inst_c1 - cpos)) \
-                       x86_branch8 (code, cond, 0, sign); \
-                else \
-                       x86_branch32 (code, cond, 0, sign); \
-        } \
-} else { \
         if (ins->inst_true_bb->native_offset) { \
                x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
         } else { \
                mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
                if ((cfg->opt & MONO_OPT_BRANCH) && \
-                    x86_is_imm8 (ins->inst_true_bb->max_offset - cpos)) \
+            x86_is_imm8 (ins->inst_true_bb->max_offset - offset)) \
                        x86_branch8 (code, cond, 0, sign); \
                 else \
                        x86_branch32 (code, cond, 0, sign); \
-        } \
 }
 
 /* emit an exception if condition is fail */
@@ -2457,7 +2524,7 @@ emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int sreg, int size,
 }
 
 static unsigned char*
-mono_emit_stack_alloc (guchar *code, MonoInst* tree)
+mono_emit_stack_alloc (MonoCompile *cfg, guchar *code, MonoInst* tree)
 {
        int sreg = tree->sreg1;
        int need_touch = FALSE;
@@ -2529,6 +2596,8 @@ mono_emit_stack_alloc (guchar *code, MonoInst* tree)
                amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
                                
                amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, offset);
+               if (cfg->param_area && cfg->arch.no_pushes)
+                       amd64_alu_reg_imm (code, X86_ADD, AMD64_RDI, cfg->param_area);
                amd64_cld (code);
                amd64_prefix (code, X86_REP_PREFIX);
                amd64_stosl (code);
@@ -2671,7 +2740,17 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
        guint8 *code = cfg->native_code + cfg->code_len;
        MonoInst *last_ins = NULL;
        guint last_offset = 0;
-       int max_len, cpos;
+       int max_len;
+
+       /* Fix max_offset estimate for each successor bb */
+       if (cfg->opt & MONO_OPT_BRANCH) {
+               int current_offset = cfg->code_len;
+               MonoBasicBlock *current_bb;
+               for (current_bb = bb; current_bb != NULL; current_bb = current_bb->next_bb) {
+                       current_bb->max_offset = current_offset;
+                       current_offset += current_bb->max_length;
+               }
+       }
 
        if (cfg->opt & MONO_OPT_LOOP) {
                int pad, align = LOOP_ALIGNMENT;
@@ -2688,12 +2767,9 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
        if (cfg->verbose_level > 2)
                g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
 
-       cpos = bb->max_offset;
-
        if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
                MonoProfileCoverageInfo *cov = cfg->coverage_info;
                g_assert (!cfg->compile_aot);
-               cpos += 6;
 
                cov->data [bb->dfn].cil_code = bb->cil_code;
                amd64_mov_reg_imm (code, AMD64_R11, (guint64)&cov->data [bb->dfn].count);
@@ -3471,7 +3547,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 
                        /* FIXME: no tracing support... */
                        if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
-                               code = mono_arch_instrument_epilog (cfg, mono_profiler_method_leave, code, FALSE);
+                               code = mono_arch_instrument_epilog_full (cfg, mono_profiler_method_leave, code, FALSE, FALSE);
 
                        g_assert (!cfg->method->save_lmf);
 
@@ -3553,7 +3629,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method, FALSE);
                        else
                                code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr, FALSE);
-                       if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
+                       if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
                                amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
                        code = emit_move_return_value (cfg, ins, code);
                        break;
@@ -3601,7 +3677,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        }
 
                        amd64_call_reg (code, ins->sreg1);
-                       if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
+                       if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
                                amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
                        code = emit_move_return_value (cfg, ins, code);
                        break;
@@ -3630,7 +3706,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        amd64_nop (code);
 
                        amd64_call_membase (code, ins->sreg1, ins->inst_offset);
-                       if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
+                       if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
                                amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
                        code = emit_move_return_value (cfg, ins, code);
                        break;
@@ -3638,17 +3714,23 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        amd64_mov_membase_reg (code, cfg->frame_reg, cfg->arch.lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
                        break;
                case OP_X86_PUSH:
+                       g_assert (!cfg->arch.no_pushes);
                        amd64_push_reg (code, ins->sreg1);
                        break;
                case OP_X86_PUSH_IMM:
+                       g_assert (!cfg->arch.no_pushes);
                        g_assert (amd64_is_imm32 (ins->inst_imm));
                        amd64_push_imm (code, ins->inst_imm);
                        break;
                case OP_X86_PUSH_MEMBASE:
+                       g_assert (!cfg->arch.no_pushes);
                        amd64_push_membase (code, ins->inst_basereg, ins->inst_offset);
                        break;
                case OP_X86_PUSH_OBJ: {
                        int size = ALIGN_TO (ins->inst_imm, 8);
+
+                       g_assert (!cfg->arch.no_pushes);
+
                        amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
                        amd64_push_reg (code, AMD64_RDI);
                        amd64_push_reg (code, AMD64_RSI);
@@ -3680,8 +3762,10 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        /* keep alignment */
                        amd64_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
                        amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
-                       code = mono_emit_stack_alloc (code, ins);
+                       code = mono_emit_stack_alloc (cfg, code, ins);
                        amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
+                       if (cfg->param_area && cfg->arch.no_pushes)
+                               amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
                        break;
                case OP_LOCALLOC_IMM: {
                        guint32 size = ins->inst_imm;
@@ -3701,13 +3785,15 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                                        amd64_mov_reg_imm (code, ins->dreg, size);
                                        ins->sreg1 = ins->dreg;
 
-                                       code = mono_emit_stack_alloc (code, ins);
+                                       code = mono_emit_stack_alloc (cfg, code, ins);
                                        amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
                                }
                        } else {
                                amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
                                amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
                        }
+                       if (cfg->param_area && cfg->arch.no_pushes)
+                               amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
                        break;
                }
                case OP_THROW: {
@@ -3733,6 +3819,12 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                case OP_START_HANDLER: {
                        MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
                        amd64_mov_membase_reg (code, spvar->inst_basereg, spvar->inst_offset, AMD64_RSP, 8);
+
+                       if ((MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY) ||
+                                MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY)) &&
+                               cfg->param_area && cfg->arch.no_pushes) {
+                               amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ALIGN_TO (cfg->param_area, MONO_ARCH_FRAME_ALIGNMENT));
+                       }
                        break;
                }
                case OP_ENDFINALLY: {
@@ -3756,28 +3848,15 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
                        //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
                        //break;
-                       if (ins->flags & MONO_INST_BRLABEL) {
-                               if (ins->inst_i0->inst_c0) {
-                                       amd64_jump_code (code, cfg->native_code + ins->inst_i0->inst_c0);
-                               } else {
-                                       mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_LABEL, ins->inst_i0);
-                                       if ((cfg->opt & MONO_OPT_BRANCH) &&
-                                           x86_is_imm8 (ins->inst_i0->inst_c1 - cpos))
-                                               x86_jump8 (code, 0);
-                                       else 
-                                               x86_jump32 (code, 0);
-                               }
-                       } else {
                                if (ins->inst_target_bb->native_offset) {
                                        amd64_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset); 
                                } else {
                                        mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
                                        if ((cfg->opt & MONO_OPT_BRANCH) &&
-                                           x86_is_imm8 (ins->inst_target_bb->max_offset - cpos))
+                                           x86_is_imm8 (ins->inst_target_bb->max_offset - offset))
                                                x86_jump8 (code, 0);
                                        else 
                                                x86_jump32 (code, 0);
-                               } 
                        }
                        break;
                case OP_BR_REG:
@@ -4395,8 +4474,6 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
                        g_assert_not_reached ();
                }
               
-               cpos += max_len;
-
                last_ins = ins;
                last_offset = offset;
        }
@@ -4516,7 +4593,7 @@ mono_arch_emit_prolog (MonoCompile *cfg)
        MonoBasicBlock *bb;
        MonoMethodSignature *sig;
        MonoInst *ins;
-       int alloc_size, pos, max_offset, i, cfa_offset, quad, max_epilog_size;
+       int alloc_size, pos, i, cfa_offset, quad, max_epilog_size;
        guint8 *code;
        CallInfo *cinfo;
        gint32 lmf_offset = cfg->arch.lmf_offset;
@@ -4590,6 +4667,15 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                        }
        }
 
+       /* The param area is always at offset 0 from sp */
+       /* This needs to be allocated here, since it has to come after the spill area */
+       if (cfg->arch.no_pushes && cfg->param_area) {
+               if (cfg->arch.omit_fp)
+                       // FIXME:
+                       g_assert_not_reached ();
+               cfg->stack_offset += ALIGN_TO (cfg->param_area, sizeof (gpointer));
+       }
+
        if (cfg->arch.omit_fp) {
                /* 
                 * On enter, the stack is misaligned by the the pushing of the return
@@ -4719,30 +4805,28 @@ mono_arch_emit_prolog (MonoCompile *cfg)
                amd64_mov_membase_reg (code, cfg->rgctx_var->inst_basereg, cfg->rgctx_var->inst_offset, MONO_ARCH_RGCTX_REG, 8);
        }
 
-       /* compute max_offset in order to use short forward jumps */
-       max_offset = 0;
+       /* compute max_length in order to use short forward jumps */
        max_epilog_size = get_max_epilog_size (cfg);
        if (cfg->opt & MONO_OPT_BRANCH) {
                for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
                        MonoInst *ins;
-                       bb->max_offset = max_offset;
+                       int max_length = 0;
 
                        if (cfg->prof_options & MONO_PROFILE_COVERAGE)
-                               max_offset += 6;
+                               max_length += 6;
                        /* max alignment for loops */
                        if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
-                               max_offset += LOOP_ALIGNMENT;
+                               max_length += LOOP_ALIGNMENT;
 
                        MONO_BB_FOR_EACH_INS (bb, ins) {
-                               if (ins->opcode == OP_LABEL)
-                                       ins->inst_c1 = max_offset;
-                               
-                               max_offset += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
+                               max_length += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
                        }
 
-                       if (mono_jit_trace_calls && bb == cfg->bb_exit)
-                               /* The tracing code can be quite large */
-                               max_offset += max_epilog_size;
+                       /* Take prolog and epilog instrumentation into account */
+                       if (bb == cfg->bb_entry || bb == cfg->bb_exit)
+                               max_length += max_epilog_size;
+                       
+                       bb->max_length = max_length;
                }
        }
 
@@ -5410,7 +5494,7 @@ enum {
 };
 
 void*
-mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
+mono_arch_instrument_epilog_full (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments, gboolean preserve_argument_registers)
 {
        guchar *code = p;
        int save_mode = SAVE_NONE;
@@ -5483,10 +5567,20 @@ mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean ena
        else
                amd64_mov_reg_imm (code, AMD64_RAX, 0);
 
+       if (preserve_argument_registers) {
+               amd64_push_reg (code, MONO_AMD64_ARG_REG1);
+               amd64_push_reg (code, MONO_AMD64_ARG_REG2);
+       }
+
        mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
        amd64_set_reg_template (code, AMD64_ARG_REG1);
        code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
 
+       if (preserve_argument_registers) {
+               amd64_pop_reg (code, MONO_AMD64_ARG_REG2);
+               amd64_pop_reg (code, MONO_AMD64_ARG_REG1);
+       }
+
        /* Restore result */
        switch (save_mode) {
        case SAVE_EAX: