X-Git-Url: http://wien.tomnetworks.com/gitweb/?a=blobdiff_plain;f=mono%2Fmini%2Fmini-x86.c;h=de32a0a1fea0e52eab0ddb85324bb8347b45a008;hb=ceaff7fe4d7cd2de06de70d6c9e9432c4eb062f0;hp=5a11ae2ef4a86d5e50562c077bba2707a9b09b18;hpb=ced6bbeeaf32d05cd63730532553ad101a49bcac;p=mono.git

diff --git a/mono/mini/mini-x86.c b/mono/mini/mini-x86.c
index 5a11ae2ef4a..de32a0a1fea 100644
--- a/mono/mini/mini-x86.c
+++ b/mono/mini/mini-x86.c
@@ -4,6 +4,7 @@
  * Authors:
  *   Paolo Molaro (lupus@ximian.com)
  *   Dietmar Maurer (dietmar@ximian.com)
+ *   Patrik Torstensson
  *
  * (C) 2003 Ximian, Inc.
  */
@@ -14,11 +15,23 @@
 #include <mono/metadata/appdomain.h>
 #include <mono/metadata/debug-helpers.h>
 #include <mono/metadata/profiler-private.h>
+#include <mono/utils/mono-math.h>
 
 #include "mini-x86.h"
 #include "inssel.h"
 #include "cpu-pentium.h"
 
+static gint lmf_tls_offset = -1;
+
+#ifdef PLATFORM_WIN32
+/* Under windows, the default pinvoke calling convention is stdcall */
+#define CALLCONV_IS_STDCALL(call_conv) (((call_conv) == MONO_CALL_STDCALL) || ((call_conv) == MONO_CALL_DEFAULT))
+#else
+#define CALLCONV_IS_STDCALL(call_conv) ((call_conv) == MONO_CALL_STDCALL)
+#endif
+
+static gpointer mono_arch_get_lmf_addr (void);
+
 const char*
 mono_arch_regname (int reg) {
 	switch (reg) {
@@ -26,8 +39,7 @@ mono_arch_regname (int reg) {
 	case X86_EBX: return "%ebx";
 	case X86_ECX: return "%ecx";
 	case X86_EDX: return "%edx";
-	case X86_ESP: return "%esp";
-	case X86_EBP: return "%ebp";
+	case X86_ESP: return "%esp";	case X86_EBP: return "%ebp";
 	case X86_EDI: return "%edi";
 	case X86_ESI: return "%esi";
 	}
@@ -102,13 +114,19 @@ arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgum
 static int indent_level = 0;
 
 static void indent (int diff) {
-	int v = indent_level;
+	int v;
+	if (diff < 0)
+		indent_level += diff;
+	v = indent_level;
 	while (v-- > 0) {
 		printf (". ");
 	}
-	indent_level += diff;
+	if (diff > 0)
+		indent_level += diff;
 }
 
+static gboolean enable_trace = TRUE;
+
 static void
 enter_method (MonoMethod *method, char *ebp)
 {
@@ -119,6 +137,9 @@ enter_method (MonoMethod *method, char *ebp)
 	MonoMethodSignature *sig;
 	char *fname;
 
+	if (!enable_trace)
+		return;
+
 	fname = mono_method_full_name (method, TRUE);
 	indent (1);
 	printf ("ENTER: %s(", fname);
@@ -153,7 +174,7 @@ enter_method (MonoMethod *method, char *ebp)
 				if (class == mono_defaults.string_class) {
 					printf ("this:[STRING:%p:%s], ", o, mono_string_to_utf8 ((MonoString *)o));
 				} else {
-					printf ("this:%p[%s.%s], ", o, class->name_space, class->name);
+					printf ("this:%p[%s.%s %s], ", o, class->name_space, class->name, o->vtable->domain->friendly_name);
 				}
 			} else 
 				printf ("this:NULL, ");
@@ -247,6 +268,9 @@ leave_method (MonoMethod *method, ...)
 	char *fname;
 	va_list ap;
 
+	if (!enable_trace)
+		return;
+
 	va_start(ap, method);
 
 	fname = mono_method_full_name (method, TRUE);
@@ -503,10 +527,12 @@ mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
 		    (ins->inst_vtype->type == MONO_TYPE_I2) || (ins->inst_vtype->type == MONO_TYPE_CHAR)) {
 			g_assert (MONO_VARINFO (cfg, i)->reg == -1);
 			g_assert (i == vmv->idx);
-			vars = mono_varlist_insert_sorted (cfg, vars, vmv, FALSE);
+			vars = g_list_prepend (vars, vmv);
 		}
 	}
 
+	vars = mono_varlist_sort (cfg, vars, 0);
+
 	return vars;
 }
 
@@ -538,7 +564,7 @@ mono_arch_allocate_vars (MonoCompile *m)
 	header = ((MonoMethodNormal *)m->method)->header;
 
 	sig = m->method->signature;
-	
+
 	offset = 8;
 	curinst = 0;
 	if (MONO_TYPE_ISSTRUCT (sig->ret)) {
@@ -648,23 +674,35 @@ mono_arch_allocate_vars (MonoCompile *m)
  */
 MonoCallInst*
 mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call, int is_virtual) {
-	MonoInst *arg, *in, **rev_args;
+	MonoInst *arg, *in;
 	MonoMethodSignature *sig;
 	int i, n, stack_size, type;
 	MonoType *ptype;
 
+	stack_size = 0;
+	/* add the vararg cookie before the non-implicit args */
+	if (call->signature->call_convention == MONO_CALL_VARARG) {
+		MonoInst *sig_arg;
+		/* FIXME: Add support for signature tokens to AOT */
+		cfg->disable_aot = TRUE;
+		MONO_INST_NEW (cfg, arg, OP_OUTARG);
+		MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
+		sig_arg->inst_p0 = call->signature;
+		arg->inst_left = sig_arg;
+		arg->type = STACK_PTR;
+		/* prepend, so they get reversed */
+		arg->next = call->out_args;
+		call->out_args = arg;
+		stack_size += sizeof (gpointer);
+	}
 	sig = call->signature;
 	n = sig->param_count + sig->hasthis;
-	rev_args = mono_mempool_alloc (cfg->mempool, sizeof (MonoInst*) * n);
 
 	if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret))
-		stack_size = 4;
-	else
-		stack_size = 0;
+		stack_size += sizeof (gpointer);
 	for (i = 0; i < n; ++i) {
 		if (is_virtual && i == 0) {
 			/* the argument will be attached to the call instrucion */
-			rev_args [n - 1] = arg = NULL;
 			in = call->args [i];
 			stack_size += 4;
 		} else {
@@ -673,7 +711,9 @@ mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call,
 			arg->cil_code = in->cil_code;
 			arg->inst_left = in;
 			arg->type = in->type;
-			rev_args [n - i - 1] = arg;
+			/* prepend, so they get reversed */
+			arg->next = call->out_args;
+			call->out_args = arg;
 			if (i >= sig->hasthis) {
 				ptype = sig->params [i - sig->hasthis];
 				if (ptype->byref)
@@ -739,6 +779,10 @@ handle_enum:
 					arg->unused = sig->pinvoke;
 					arg->inst_imm = sizeof (MonoTypedRef); 
 					break;
+				case MONO_TYPE_GENERICINST:
+					type = ptype->data.generic_inst->generic_type->type;
+					goto handle_enum;
+
 				default:
 					g_error ("unknown type 0x%02x in mono_arch_call_opcode\n", type);
 				}
@@ -748,11 +792,9 @@ handle_enum:
 			}
 		}
 	}
-	/* they need to be pushed in reverse order */
 	/* if the function returns a struct, the called method already does a ret $0x4 */
 	if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret))
 		stack_size -= 4;
-	call->args = rev_args;
 	call->stack_usage = stack_size;
 	/* 
 	 * should set more info in call, such as the stack space
@@ -1118,6 +1160,104 @@ branch_cc_table [] = {
 
 #define DEBUG(a) if (cfg->verbose_level > 1) a
 //#define DEBUG(a)
+
+/*
+ * returns the offset used by spillvar. It allocates a new
+ * spill variable if necessary. 
+ */
+static int
+mono_spillvar_offset (MonoCompile *cfg, int spillvar)
+{
+	MonoSpillInfo **si, *info;
+	int i = 0;
+
+	si = &cfg->spill_info; 
+	
+	while (i <= spillvar) {
+
+		if (!*si) {
+			*si = info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo));
+			info->next = NULL;
+			cfg->stack_offset -= sizeof (gpointer);
+			info->offset = cfg->stack_offset;
+		}
+
+		if (i == spillvar)
+			return (*si)->offset;
+
+		i++;
+		si = &(*si)->next;
+	}
+
+	g_assert_not_reached ();
+	return 0;
+}
+
+/*
+ * returns the offset used by spillvar. It allocates a new
+ * spill float variable if necessary. 
+ * (same as mono_spillvar_offset but for float)
+ */
+static int
+mono_spillvar_offset_float (MonoCompile *cfg, int spillvar)
+{
+	MonoSpillInfo **si, *info;
+	int i = 0;
+
+	si = &cfg->spill_info_float; 
+	
+	while (i <= spillvar) {
+
+		if (!*si) {
+			*si = info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo));
+			info->next = NULL;
+			cfg->stack_offset -= sizeof (double);
+			info->offset = cfg->stack_offset;
+		}
+
+		if (i == spillvar)
+			return (*si)->offset;
+
+		i++;
+		si = &(*si)->next;
+	}
+
+	g_assert_not_reached ();
+	return 0;
+}
+
+/*
+ * Creates a store for spilled floating point items
+ */
+static MonoInst*
+create_spilled_store_float (MonoCompile *cfg, int spill, int reg, MonoInst *ins)
+{
+	MonoInst *store;
+	MONO_INST_NEW (cfg, store, OP_STORER8_MEMBASE_REG);
+	store->sreg1 = reg;
+	store->inst_destbasereg = X86_EBP;
+	store->inst_offset = mono_spillvar_offset_float (cfg, spill);
+
+	DEBUG (g_print ("SPILLED FLOAT STORE (%d at 0x%08x(%%sp)) (from %d)\n", spill, store->inst_offset, reg));
+	return store;
+}
+
+/*
+ * Creates a load for spilled floating point items 
+ */
+static MonoInst*
+create_spilled_load_float (MonoCompile *cfg, int spill, int reg, MonoInst *ins)
+{
+	MonoInst *load;
+	MONO_INST_NEW (cfg, load, OP_LOADR8_SPILL_MEMBASE);
+	load->dreg = reg;
+	load->inst_basereg = X86_EBP;
+	load->inst_offset = mono_spillvar_offset_float (cfg, spill);
+
+	DEBUG (g_print ("SPILLED FLOAT LOAD (%d at 0x%08x(%%sp)) (from %d)\n", spill, load->inst_offset, reg));
+	return load;
+}
+
 #define reg_is_freeable(r) ((r) >= 0 && (r) <= 7 && X86_IS_CALLEE ((r)))
 
 typedef struct {
@@ -1125,6 +1265,7 @@ typedef struct {
 	int killed_in;
 	int last_use;
 	int prev_use;
+	int flags;		/* used to track fp spill/load */
 } RegTrack;
 
 static const char*const * ins_spec = pentium_desc;
@@ -1342,6 +1483,7 @@ insert_before_ins (MonoInst *ins, InstList *item, MonoInst* to_insert)
 	item->data = to_insert; 
 }
 
+
 #if  0
 static int
 alloc_int_reg (MonoCompile *cfg, InstList *curinst, MonoInst *ins, int sym_reg, guint32 allow_mask)
@@ -1366,7 +1508,12 @@ alloc_int_reg (MonoCompile *cfg, InstList *curinst, MonoInst *ins, int sym_reg,
 }
 #endif
 
-#include "cprop.c"
+/* flags used in reginfo->flags */
+#define MONO_X86_FP_NEEDS_LOAD_SPILL	1
+#define MONO_X86_FP_NEEDS_SPILL			2
+#define MONO_X86_FP_NEEDS_LOAD			4
+
+/*#include "cprop.c"*/
 
 /*
  * Local register allocation.
@@ -1386,31 +1533,43 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 	InstList *tmp, *reversed = NULL;
 	const char *spec;
 	guint32 src1_mask, src2_mask, dest_mask;
+	GList *fspill_list = NULL;
+	int fspill = 0;
 
 	if (!bb->code)
 		return;
 	rs->next_vireg = bb->max_ireg;
 	rs->next_vfreg = bb->max_freg;
 	mono_regstate_assign (rs);
-	reginfo = mono_mempool_alloc0 (cfg->mempool, sizeof (RegTrack) * rs->next_vireg);
-	reginfof = mono_mempool_alloc0 (cfg->mempool, sizeof (RegTrack) * rs->next_vfreg);
+	reginfo = g_malloc0 (sizeof (RegTrack) * rs->next_vireg);
+	reginfof = g_malloc0 (sizeof (RegTrack) * rs->next_vfreg);
 	rs->ifree_mask = X86_CALLEE_REGS;
 
 	ins = bb->code;
 
-	if (cfg->opt & MONO_OPT_COPYPROP)
-		local_copy_prop (cfg, ins);
-	
+	/*if (cfg->opt & MONO_OPT_COPYPROP)
+		local_copy_prop (cfg, ins);*/
+
 	i = 1;
-	fpcount = 0; /* FIXME: track fp stack utilization */
+	fpcount = 0;
 	DEBUG (g_print ("LOCAL regalloc: basic block: %d\n", bb->block_num));
 	/* forward pass on the instructions to collect register liveness info */
 	while (ins) {
 		spec = ins_spec [ins->opcode];
 		DEBUG (print_ins (i, ins));
+
 		if (spec [MONO_INST_SRC1]) {
-			if (spec [MONO_INST_SRC1] == 'f')
+			if (spec [MONO_INST_SRC1] == 'f') {
+				GList *spill;
 				reginfo1 = reginfof;
+
+				spill = g_list_first (fspill_list);
+				if (spill && fpcount < MONO_MAX_FREGS) {
+					reginfo1 [ins->sreg1].flags |= MONO_X86_FP_NEEDS_LOAD;
+					fspill_list = g_list_remove (fspill_list, spill->data);
+				} else
+					fpcount--;
+			}
 			else
 				reginfo1 = reginfo;
 			reginfo1 [ins->sreg1].prev_use = reginfo1 [ins->sreg1].last_use;
@@ -1419,8 +1578,21 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 			ins->sreg1 = -1;
 		}
 		if (spec [MONO_INST_SRC2]) {
-			if (spec [MONO_INST_SRC2] == 'f')
+			if (spec [MONO_INST_SRC2] == 'f') {
+				GList *spill;
 				reginfo2 = reginfof;
+				spill = g_list_first (fspill_list);
+				if (spill) {
+					reginfo2 [ins->sreg2].flags |= MONO_X86_FP_NEEDS_LOAD;
+					fspill_list = g_list_remove (fspill_list, spill->data);
+					if (fpcount >= MONO_MAX_FREGS) {
+						fspill++;
+						fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
+						reginfo2 [ins->sreg2].flags |= MONO_X86_FP_NEEDS_LOAD_SPILL;
+					}
+				} else
+					fpcount--;
+			}
 			else
 				reginfo2 = reginfo;
 			reginfo2 [ins->sreg2].prev_use = reginfo2 [ins->sreg2].last_use;
@@ -1429,8 +1601,16 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 			ins->sreg2 = -1;
 		}
 		if (spec [MONO_INST_DEST]) {
-			if (spec [MONO_INST_DEST] == 'f')
+			if (spec [MONO_INST_DEST] == 'f') {
 				reginfod = reginfof;
+				if (fpcount >= MONO_MAX_FREGS) {
+					reginfod [ins->dreg].flags |= MONO_X86_FP_NEEDS_SPILL;
+					fspill++;
+					fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
+					fpcount--;
+				}
+				fpcount++;
+			}
 			else
 				reginfod = reginfo;
 			if (spec [MONO_INST_DEST] != 'b') /* it's not just a base register */
@@ -1445,7 +1625,7 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 				reginfod [ins->dreg + 1].last_use = i;
 				if (reginfod [ins->dreg + 1].born_in == 0 || reginfod [ins->dreg + 1].born_in > i)
 					reginfod [ins->dreg + 1].born_in = i;
-			}
+			} 
 		} else {
 			ins->dreg = -1;
 		}
@@ -1454,6 +1634,9 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 		ins = ins->next;
 	}
 
+	// todo: check if we have anything left on fp stack, in verify mode?
+	fspill = 0;
+
 	DEBUG (print_regtrack (reginfo, rs->next_vireg));
 	DEBUG (print_regtrack (reginfof, rs->next_vfreg));
 	tmp = reversed;
@@ -1463,6 +1646,7 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 		--i;
 		ins = tmp->data;
 		spec = ins_spec [ins->opcode];
+		prev_dreg = -1;
 		DEBUG (g_print ("processing:"));
 		DEBUG (print_ins (i, ins));
 		if (spec [MONO_INST_CLOB] == 's') {
@@ -1566,6 +1750,7 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 					}
 				} else {
 					DEBUG (g_print ("\tshortcut assignment of R%d to %s\n", ins->dreg, mono_arch_regname (dest_reg)));
+					prev_dreg = ins->dreg;
 					rs->iassign [ins->dreg] = dest_reg;
 					rs->isymbolic [dest_reg] = ins->dreg;
 					ins->dreg = dest_reg;
@@ -1597,8 +1782,22 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 				mono_regstate_free_int (rs, X86_EDX);
 			}
 		}
-		/* update for use with FP regs... */
-		if (spec [MONO_INST_DEST] != 'f' && ins->dreg >= MONO_MAX_IREGS) {
+
+		/* Track dreg */
+		if (spec [MONO_INST_DEST] == 'f') {
+			if (reginfof [ins->dreg].flags & MONO_X86_FP_NEEDS_SPILL) {
+				GList *spill_node;
+				MonoInst *store;
+				spill_node = g_list_first (fspill_list);
+				g_assert (spill_node);
+
+				store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->dreg, ins);
+				insert_before_ins (ins, tmp, store);
+				fspill_list = g_list_remove (fspill_list, spill_node->data);
+				fspill--;
+			}
+		} 
+		else if (ins->dreg >= MONO_MAX_IREGS) {
 			val = rs->iassign [ins->dreg];
 			prev_dreg = ins->dreg;
 			if (val < 0) {
@@ -1667,8 +1866,6 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 			} else if (spec [MONO_INST_DEST] == 'd' && ins->dreg != X86_EDX && spec [MONO_INST_CLOB] != 'd') {
 				create_copy_ins (cfg, ins->dreg, X86_EDX, ins);
 			}
-		} else {
-			prev_dreg = -1;
 		}
 		if (spec [MONO_INST_DEST] != 'f' && reg_is_freeable (ins->dreg) && prev_dreg >= 0 && reginfo [prev_dreg].born_in >= i) {
 			DEBUG (g_print ("\tfreeable %s (R%d) (born in %d)\n", mono_arch_regname (ins->dreg), prev_dreg, reginfo [prev_dreg].born_in));
@@ -1687,7 +1884,31 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 			ins->sreg1 = X86_EAX;
 			rs->ifree_mask &= ~ (1 << X86_EAX);
 		}
-		if (spec [MONO_INST_SRC1] != 'f' && ins->sreg1 >= MONO_MAX_IREGS) {
+
+		/* Track sreg1 */
+		if (spec [MONO_INST_SRC1] == 'f') {
+			if (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD) {
+				MonoInst *load;
+				MonoInst *store = NULL;
+
+				if (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD_SPILL) {
+					GList *spill_node;
+					spill_node = g_list_first (fspill_list);
+					g_assert (spill_node);
+
+					store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->sreg1, ins);		
+					fspill_list = g_list_remove (fspill_list, spill_node->data);
+				}
+
+				fspill++;
+				fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
+				load = create_spilled_load_float (cfg, fspill, ins->sreg1, ins);
+				insert_before_ins (ins, tmp, load);
+				if (store) 
+					insert_before_ins (load, tmp, store);
+			}
+		} 
+		else if (ins->sreg1 >= MONO_MAX_IREGS) {
 			val = rs->iassign [ins->sreg1];
 			prev_sreg1 = ins->sreg1;
 			if (val < 0) {
@@ -1742,7 +1963,33 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 				ins->next = copy;
 			}
 		}
-		if (spec [MONO_INST_SRC2] != 'f' && ins->sreg2 >= MONO_MAX_IREGS) {
+		/* track sreg2 */
+		if (spec [MONO_INST_SRC2] == 'f') {
+			if (reginfof [ins->sreg2].flags & MONO_X86_FP_NEEDS_LOAD) {
+				MonoInst *load;
+				MonoInst *store = NULL;
+
+				if (reginfof [ins->sreg2].flags & MONO_X86_FP_NEEDS_LOAD_SPILL) {
+					GList *spill_node;
+
+					spill_node = g_list_first (fspill_list);
+					g_assert (spill_node);
+					if (spec [MONO_INST_SRC1] == 'f' && (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD_SPILL))
+						spill_node = g_list_next (spill_node);
+	
+					store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->sreg2, ins);
+					fspill_list = g_list_remove (fspill_list, spill_node->data);
+				} 
+				
+				fspill++;
+				fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
+				load = create_spilled_load_float (cfg, fspill, ins->sreg2, ins);
+				insert_before_ins (ins, tmp, load);
+				if (store) 
+					insert_before_ins (load, tmp, store);
+			}
+		} 
+		else if (ins->sreg2 >= MONO_MAX_IREGS) {
 			val = rs->iassign [ins->sreg2];
 			prev_sreg2 = ins->sreg2;
 			if (val < 0) {
@@ -1786,13 +2033,17 @@ mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
 			DEBUG (g_print ("freeable %s\n", mono_arch_regname (ins->sreg2)));
 			mono_regstate_free_int (rs, ins->sreg2);
 		}*/
-		
+	
 		//DEBUG (print_ins (i, ins));
 		/* this may result from a insert_before call */
 		if (!tmp->next)
 			bb->code = tmp->data;
 		tmp = tmp->next;
 	}
+
+	g_free (reginfo);
+	g_free (reginfof);
+	g_list_free (fspill_list);
 }
 
 static unsigned char*
@@ -2315,18 +2566,13 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 		case OP_VOIDCALL:
 		case CEE_CALL:
 			call = (MonoCallInst*)ins;
-			if (call->signature->call_convention == MONO_CALL_VARARG) {
-				/* FIXME: hack */
-				x86_push_imm (code, call->signature);
-				offset = code - cfg->native_code;
-			}
 			if (ins->flags & MONO_INST_HAS_METHOD)
 				mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_METHOD, call->method);
 			else {
 				mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_ABS, call->fptr);
 			}
 			x86_call_code (code, 0);
-			if (call->stack_usage && (call->signature->call_convention != MONO_CALL_STDCALL))
+			if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
 				x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
 			break;
 		case OP_FCALL_REG:
@@ -2336,7 +2582,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 		case OP_CALL_REG:
 			call = (MonoCallInst*)ins;
 			x86_call_reg (code, ins->sreg1);
-			if (call->stack_usage && (call->signature->call_convention != MONO_CALL_STDCALL))
+			if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
 				x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
 			break;
 		case OP_FCALL_MEMBASE:
@@ -2346,7 +2592,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 		case OP_CALL_MEMBASE:
 			call = (MonoCallInst*)ins;
 			x86_call_membase (code, ins->sreg1, ins->inst_offset);
-			if (call->stack_usage && (call->signature->call_convention != MONO_CALL_STDCALL))
+			if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
 				x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
 			break;
 		case OP_OUTARG:
@@ -2491,7 +2737,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 		case OP_R8CONST: {
 			double d = *(double *)ins->inst_p0;
 
-			if ((d == 0.0) && (signbit (d) == 0)) {
+			if ((d == 0.0) && (mono_signbit (d) == 0)) {
 				x86_fldz (code);
 			} else if (d == 1.0) {
 				x86_fld1 (code);
@@ -2504,7 +2750,7 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 		case OP_R4CONST: {
 			float f = *(float *)ins->inst_p0;
 
-			if ((f == 0.0) && (signbit (f) == 0)) {
+			if ((f == 0.0) && (mono_signbit (f) == 0)) {
 				x86_fldz (code);
 			} else if (f == 1.0) {
 				x86_fld1 (code);
@@ -2517,6 +2763,10 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 		case OP_STORER8_MEMBASE_REG:
 			x86_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, TRUE, TRUE);
 			break;
+		case OP_LOADR8_SPILL_MEMBASE:
+			x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
+			x86_fxch (code, 1);
+			break;
 		case OP_LOADR8_MEMBASE:
 			x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
 			break;
@@ -2966,12 +3216,13 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 void
 mono_arch_register_lowlevel_calls (void)
 {
+	mono_register_jit_icall (mono_arch_get_lmf_addr, "mono_arch_get_lmf_addr", NULL, TRUE);
 	mono_register_jit_icall (enter_method, "mono_enter_method", NULL, TRUE);
 	mono_register_jit_icall (leave_method, "mono_leave_method", NULL, TRUE);
 }
 
 void
-mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji)
+mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gboolean run_cctors)
 {
 	MonoJumpInfo *patch_info;
 
@@ -2992,6 +3243,9 @@ mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, Mono
 		case MONO_PATCH_INFO_IP:
 			*((gpointer *)(ip)) = ip;
 			continue;
+		case MONO_PATCH_INFO_METHOD_REL:
+			*((gpointer *)(ip)) = code + patch_info->data.offset;
+			continue;
 		case MONO_PATCH_INFO_INTERNAL_METHOD: {
 			MonoJitICallInfo *mi = mono_find_jit_icall_by_name (patch_info->data.name);
 			if (!mi) {
@@ -3001,37 +3255,35 @@ mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, Mono
 			target = mono_icall_get_wrapper (mi);
 			break;
 		}
-		case MONO_PATCH_INFO_METHOD_JUMP:
+		case MONO_PATCH_INFO_METHOD_JUMP: {
+			GSList *list;
+
 			/* get the trampoline to the method from the domain */
-			if (!(target = g_hash_table_lookup (domain->jit_code_hash, patch_info->data.method))) {
-				GSList *list;
-				target = mono_arch_create_jump_trampoline (patch_info->data.method);
-				if (!domain->jump_target_hash)
-					domain->jump_target_hash = g_hash_table_new (NULL, NULL);
-				list = g_hash_table_lookup (domain->jump_target_hash, patch_info->data.method);
-				list = g_slist_prepend (list, ip);
-				g_hash_table_insert (domain->jump_target_hash, patch_info->data.method, list);
-			}
+			target = mono_create_jump_trampoline (domain, patch_info->data.method, TRUE);
+			if (!domain->jump_target_hash)
+				domain->jump_target_hash = g_hash_table_new (NULL, NULL);
+			list = g_hash_table_lookup (domain->jump_target_hash, patch_info->data.method);
+			list = g_slist_prepend (list, ip);
+			g_hash_table_insert (domain->jump_target_hash, patch_info->data.method, list);
 			break;
+		}
 		case MONO_PATCH_INFO_METHOD:
 			if (patch_info->data.method == method) {
 				target = code;
-			} else {
+			} else
 				/* get the trampoline to the method from the domain */
-				if (!(target = g_hash_table_lookup (domain->jit_code_hash, patch_info->data.method)))
-					target = mono_arch_create_jit_trampoline (patch_info->data.method);
-			}
+				target = mono_arch_create_jit_trampoline (patch_info->data.method);
 			break;
 		case MONO_PATCH_INFO_SWITCH: {
-			gpointer *table = (gpointer *)patch_info->data.target;
+			gpointer *jump_table = mono_mempool_alloc (domain->code_mp, sizeof (gpointer) * patch_info->table_size);
 			int i;
 
-			*((gconstpointer *)(ip + 2)) = patch_info->data.target;
+			*((gconstpointer *)(ip + 2)) = jump_table;
 
 			for (i = 0; i < patch_info->table_size; i++) {
-				table [i] = (int)patch_info->data.table [i] + code;
+				jump_table [i] = code + (int)patch_info->data.table [i];
 			}
-			/* we put into the table the absolute address, no need fo x86_patch in this case */
+			/* we put into the table the absolute address, no need for x86_patch in this case */
 			continue;
 		}
 		case MONO_PATCH_INFO_METHODCONST:
@@ -3040,10 +3292,69 @@ mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, Mono
 		case MONO_PATCH_INFO_FIELD:
 			*((gconstpointer *)(ip + 1)) = patch_info->data.target;
 			continue;
+		case MONO_PATCH_INFO_IID:
+			mono_class_init (patch_info->data.klass);
+			*((guint32 *)(ip + 1)) = patch_info->data.klass->interface_id;
+			continue;			
+		case MONO_PATCH_INFO_VTABLE:
+			*((gconstpointer *)(ip + 1)) = mono_class_vtable (domain, patch_info->data.klass);
+			continue;
+		case MONO_PATCH_INFO_CLASS_INIT: {
+			guint8 *code = ip;
+			/* Might already been changed to a nop */
+			x86_call_imm (code, 0);
+			target = mono_create_class_init_trampoline (mono_class_vtable (domain, patch_info->data.klass));
+			break;
+		}
+		case MONO_PATCH_INFO_SFLDA: {
+			MonoVTable *vtable = mono_class_vtable (domain, patch_info->data.field->parent);
+			if (!vtable->initialized && !(vtable->klass->flags & TYPE_ATTRIBUTE_BEFORE_FIELD_INIT) && mono_class_needs_cctor_run (vtable->klass, method))
+				/* Done by the generated code */
+				;
+			else {
+				if (run_cctors)
+					mono_runtime_class_init (vtable);
+			}
+			*((gconstpointer *)(ip + 1)) = 
+				(char*)vtable->data + patch_info->data.field->offset;
+			continue;
+		}
 		case MONO_PATCH_INFO_R4:
 		case MONO_PATCH_INFO_R8:
 			*((gconstpointer *)(ip + 2)) = patch_info->data.target;
 			continue;
+		case MONO_PATCH_INFO_EXC_NAME:
+			*((gconstpointer *)(ip + 1)) = patch_info->data.name;
+			continue;
+		case MONO_PATCH_INFO_LDSTR:
+			*((gconstpointer *)(ip + 1)) = 
+				mono_ldstr (domain, patch_info->data.token->image, 
+							mono_metadata_token_index (patch_info->data.token->token));
+			continue;
+		case MONO_PATCH_INFO_TYPE_FROM_HANDLE: {
+			gpointer handle;
+			MonoClass *handle_class;
+
+			handle = mono_ldtoken (patch_info->data.token->image, 
+								   patch_info->data.token->token, &handle_class);
+			mono_class_init (handle_class);
+			mono_class_init (mono_class_from_mono_type (handle));
+
+			*((gconstpointer *)(ip + 1)) = 
+				mono_type_get_object (domain, handle);
+			continue;
+		}
+		case MONO_PATCH_INFO_LDTOKEN: {
+			gpointer handle;
+			MonoClass *handle_class;
+
+			handle = mono_ldtoken (patch_info->data.token->image,
+								   patch_info->data.token->token, &handle_class);
+			mono_class_init (handle_class);
+
+			*((gconstpointer *)(ip + 1)) = handle;
+			continue;
+		}
 		default:
 			g_assert_not_reached ();
 		}
@@ -3060,7 +3371,7 @@ mono_arch_max_epilog_size (MonoCompile *cfg)
 	if (cfg->method->save_lmf)
 		max_epilog_size += 128;
 	
-	if (mono_jit_trace_calls)
+	if (mono_jit_trace_calls != NULL)
 		max_epilog_size += 50;
 
 	if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
@@ -3116,11 +3427,28 @@ mono_arch_emit_prolog (MonoCompile *cfg)
 
 		/* save method info */
 		x86_push_imm (code, method);
-	
+
 		/* get the address of lmf for the current thread */
-		mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_INTERNAL_METHOD, 
-				     (gpointer)"mono_get_lmf_addr");
-		x86_call_code (code, 0);
+		/* 
+		 * This is performance critical so we try to use some tricks to make
+		 * it fast.
+		 */
+		if (lmf_tls_offset != -1) {
+			/* Load lmf quicky using the GS register */
+			x86_prefix (code, X86_GS_PREFIX);
+			x86_mov_reg_mem (code, X86_EAX, 0, 4);
+			x86_mov_reg_membase (code, X86_EAX, X86_EAX, lmf_tls_offset, 4);
+		}
+		else {
+#ifdef HAVE_KW_THREAD
+			mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_INTERNAL_METHOD, 
+								 (gpointer)"mono_arch_get_lmf_addr");
+#else
+			mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_INTERNAL_METHOD, 
+								 (gpointer)"mono_get_lmf_addr");
+#endif
+			x86_call_code (code, 0);
+		}
 
 		/* push lmf */
 		x86_push_reg (code, X86_EAX); 
@@ -3148,8 +3476,21 @@ mono_arch_emit_prolog (MonoCompile *cfg)
 
 	alloc_size -= pos;
 
-	if (alloc_size)
+	if (alloc_size) {
+		/* See mono_emit_stack_alloc */
+#ifdef PLATFORM_WIN32
+		guint32 remaining_size = alloc_size;
+		while (remaining_size >= 0x1000) {
+			x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x1000);
+			x86_test_membase_reg (code, X86_ESP, 0, X86_ESP);
+			remaining_size -= 0x1000;
+		}
+		if (remaining_size)
+			x86_alu_reg_imm (code, X86_SUB, X86_ESP, remaining_size);
+#else
 		x86_alu_reg_imm (code, X86_SUB, X86_ESP, alloc_size);
+#endif
+	}
 
         /* compute max_offset in order to use short forward jumps */
 	max_offset = 0;
@@ -3168,7 +3509,7 @@ mono_arch_emit_prolog (MonoCompile *cfg)
 		}
 	}
 
-	if (mono_jit_trace_calls)
+	if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
 		code = mono_arch_instrument_prolog (cfg, enter_method, code, TRUE);
 
 	/* load arguments allocated to register from the stack */
@@ -3195,12 +3536,14 @@ mono_arch_emit_epilog (MonoCompile *cfg)
 {
 	MonoJumpInfo *patch_info;
 	MonoMethod *method = cfg->method;
+	MonoMethodSignature *sig = method->signature;
 	int pos;
+	guint32 stack_to_pop;
 	guint8 *code;
 
 	code = cfg->native_code + cfg->code_len;
 
-	if (mono_jit_trace_calls)
+	if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
 		code = mono_arch_instrument_epilog (cfg, leave_method, code, TRUE);
 
 	/* the code restoring the registers must be kept in sync with CEE_JMP */
@@ -3254,9 +3597,20 @@ mono_arch_emit_epilog (MonoCompile *cfg)
 	}
 
 	x86_leave (code);
-	/* FIXME: add another check to support stdcall convention here */
+
+	if (CALLCONV_IS_STDCALL (sig->call_convention)) {
+	  MonoJitArgumentInfo *arg_info = alloca (sizeof (MonoJitArgumentInfo) * (sig->param_count + 1));
+
+	  stack_to_pop = arch_get_argument_info (sig, sig->param_count, arg_info);
+	}
+	else
 	if (MONO_TYPE_ISSTRUCT (cfg->method->signature->ret))
-		x86_ret_imm (code, 4);
+	  stack_to_pop = 4;
+	else
+	  stack_to_pop = 0;
+
+	if (stack_to_pop)
+		x86_ret_imm (code, stack_to_pop);
 	else
 		x86_ret (code);
 
@@ -3265,7 +3619,9 @@ mono_arch_emit_epilog (MonoCompile *cfg)
 		switch (patch_info->type) {
 		case MONO_PATCH_INFO_EXC:
 			x86_patch (patch_info->ip.i + cfg->native_code, code);
+			mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_EXC_NAME, patch_info->data.target);
 			x86_push_imm (code, patch_info->data.target);
+			mono_add_patch_info (cfg, code + 1 - cfg->native_code, MONO_PATCH_INFO_METHOD_REL, (gpointer)patch_info->ip.i);
 			x86_push_imm (code, patch_info->ip.i + cfg->native_code);
 			patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
 			patch_info->data.name = "mono_arch_throw_exception_by_name";
@@ -3290,3 +3646,108 @@ mono_arch_flush_icache (guint8 *code, gint size)
 	/* not needed */
 }
 
+/*
+ * Support for fast access to the thread-local lmf structure using the GS
+ * segment register on NPTL + kernel 2.6.x.
+ */
+
+static gboolean tls_offset_inited = FALSE;
+
+#ifdef HAVE_KW_THREAD
+static __thread gpointer mono_lmf_addr;
+#endif
+
+static gpointer
+mono_arch_get_lmf_addr (void)
+{
+#ifdef HAVE_KW_THREAD
+	return mono_lmf_addr;
+#else
+	g_assert_not_reached ();
+	return NULL;
+#endif
+}
+
+void
+mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
+{
+	if (!tls_offset_inited) {
+		guint8 *code;
+
+		tls_offset_inited = TRUE;
+
+		if (getenv ("MONO_NPTL")) {
+			/* 
+			 * Determine the offset of mono_lfm_addr inside the TLS structures
+			 * by disassembling the function above.
+			 */
+			code = (guint8*)&mono_arch_get_lmf_addr;
+
+			/* This is generated by gcc 3.3.2 */
+			if ((code [0] == 0x55) && (code [1] == 0x89) && (code [2] == 0xe5) &&
+				(code [3] == 0x65) && (code [4] == 0xa1) && (code [5] == 0x00) &&
+				(code [6] == 0x00) && (code [7] == 0x00) && (code [8] == 0x00) &&
+				(code [9] == 0x8b) && (code [10] == 0x80)) {
+				lmf_tls_offset = *(int*)&(code [11]);
+			}
+		}
+	}		
+
+#ifdef HAVE_KW_THREAD
+	mono_lmf_addr = &tls->lmf;
+#endif
+}
+
+void
+mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_reg, int this_type, int vt_reg)
+{
+
+	/* add the this argument */
+	if (this_reg != -1) {
+		MonoInst *this;
+		MONO_INST_NEW (cfg, this, OP_OUTARG);
+		this->type = this_type;
+		this->sreg1 = this_reg;
+		mono_bblock_add_inst (cfg->cbb, this);
+	}
+
+	if (vt_reg != -1) {
+		MonoInst *vtarg;
+		MONO_INST_NEW (cfg, vtarg, OP_OUTARG);
+		vtarg->type = STACK_MP;
+		vtarg->sreg1 = vt_reg;
+		mono_bblock_add_inst (cfg->cbb, vtarg);
+	}
+}
+
+
+gint
+mono_arch_get_opcode_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
+{
+	if (cmethod->klass == mono_defaults.math_class) {
+		if (strcmp (cmethod->name, "Sin") == 0)
+			return OP_SIN;
+		else if (strcmp (cmethod->name, "Cos") == 0)
+			return OP_COS;
+		else if (strcmp (cmethod->name, "Tan") == 0)
+			return OP_TAN;
+		else if (strcmp (cmethod->name, "Atan") == 0)
+			return OP_ATAN;
+		else if (strcmp (cmethod->name, "Sqrt") == 0)
+			return OP_SQRT;
+		else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8)
+			return OP_ABS;
+#if 0
+		/* OP_FREM is not IEEE compatible */
+		else if (strcmp (cmethod->name, "IEEERemainder") == 0)
+			return OP_FREM;
+#endif
+		else
+			return -1;
+	} else {
+		return -1;
+	}
+	return -1;
+}
+
+