X-Git-Url: http://wien.tomnetworks.com/gitweb/?a=blobdiff_plain;ds=sidebyside;f=mono%2Fmini%2Fmini-gc.c;h=0d439acc63618b64b2c2904b25cab84699fa8ef2;hb=d0771cfe0d6f833be7bdcc4f19369f38dba89a6a;hp=b1cd75a88089257f3902a03dfec7274520be13d1;hpb=3a0faf0d72dfe90cde98a46b267a2b380a244b8e;p=mono.git

diff --git a/mono/mini/mini-gc.c b/mono/mini/mini-gc.c
index b1cd75a8808..0d439acc636 100644
--- a/mono/mini/mini-gc.c
+++ b/mono/mini/mini-gc.c
@@ -123,10 +123,8 @@ typedef struct {
  * only the collection thread is active.
  */
 typedef struct {
-	MonoLMF *lmf;
-	MonoContext ctx;
-	gboolean has_context;
-	MonoJitTlsData *jit_tls;
+	MonoThreadUnwindState unwind_state;
+	MonoThreadInfo *info;
 	/* For debugging */
 	mgreg_t tid;
 	gpointer ref_to_track;
@@ -264,6 +262,14 @@ static JITGCStats stats;
 
 static FILE *logfile;
 
+static gboolean enable_gc_maps_for_aot;
+
+void
+mini_gc_enable_gc_maps_for_aot (void)
+{
+	enable_gc_maps_for_aot = TRUE;
+}
+
 // FIXME: Move these to a shared place
 
 static inline void
@@ -373,6 +379,16 @@ encode_frame_reg (int frame_reg)
 		return 0;
 	else if (frame_reg == X86_ESP)
 		return 1;
+#elif defined(TARGET_ARM)
+	if (frame_reg == ARMREG_SP)
+		return 0;
+	else if (frame_reg == ARMREG_FP)
+		return 1;
+#elif defined(TARGET_S390X)
+	if (frame_reg == S390_SP)
+		return 0;
+	else if (frame_reg == S390_FP)
+		return 1;
 #else
 	NOT_IMPLEMENTED;
 #endif
@@ -393,6 +409,16 @@ decode_frame_reg (int encoded)
 		return X86_EBP;
 	else if (encoded == 1)
 		return X86_ESP;
+#elif defined(TARGET_ARM)
+	if (encoded == 0)
+		return ARMREG_SP;
+	else if (encoded == 1)
+		return ARMREG_FP;
+#elif defined(TARGET_S390X)
+	if (encoded == 0)
+		return S390_SP;
+	else if (encoded == 1)
+		return S390_FP;
 #else
 	NOT_IMPLEMENTED;
 #endif
@@ -408,6 +434,10 @@ static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13,
 #endif
 #elif defined(TARGET_X86)
 static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
+#elif defined(TARGET_ARM)
+static int callee_saved_regs [] = { ARMREG_V1, ARMREG_V2, ARMREG_V3, ARMREG_V4, ARMREG_V5, ARMREG_V7, ARMREG_FP };
+#elif defined(TARGET_S390X)
+static int callee_saved_regs [] = { s390_r6, s390_r7, s390_r8, s390_r9, s390_r10, s390_r11, s390_r12, s390_r13, s390_r14 };
 #endif
 
 static guint32
@@ -539,6 +569,7 @@ thread_attach_func (void)
 
 	tls = g_new0 (TlsData, 1);
 	tls->tid = GetCurrentThreadId ();
+	tls->info = mono_thread_info_current ();
 	stats.tlsdata_size += sizeof (TlsData);
 
 	return tls;
@@ -553,22 +584,46 @@ thread_detach_func (gpointer user_data)
 }
 
 static void
-thread_suspend_func (gpointer user_data, void *sigctx)
+thread_suspend_func (gpointer user_data, void *sigctx, MonoContext *ctx)
 {
 	TlsData *tls = user_data;
 
-	if (!tls)
+	if (!tls) {
 		/* Happens during startup */
+		tls->unwind_state.valid = FALSE;
 		return;
+	}
+
+	if (tls->tid != GetCurrentThreadId ()) {
+		/* Happens on osx because threads are not suspended using signals */
+		gboolean res;
 
-	tls->lmf = mono_get_lmf ();
-	if (sigctx) {
-		mono_arch_sigctx_to_monoctx (sigctx, &tls->ctx);
-		tls->has_context = TRUE;
+		g_assert (tls->info);
+		res = mono_thread_state_init_from_handle (&tls->unwind_state, (MonoNativeThreadId)tls->tid, tls->info->native_handle);
 	} else {
-		tls->has_context = FALSE;
+		tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF] = mono_get_lmf ();
+		if (sigctx) {
+#ifdef MONO_ARCH_HAVE_SIGCTX_TO_MONOCTX
+			mono_arch_sigctx_to_monoctx (sigctx, &tls->unwind_state.ctx);
+			tls->unwind_state.valid = TRUE;
+#else
+			tls->unwind_state.valid = FALSE;
+#endif
+		} else if (ctx) {
+			memcpy (&tls->unwind_state.ctx, ctx, sizeof (MonoContext));
+			tls->unwind_state.valid = TRUE;
+		} else {
+			tls->unwind_state.valid = FALSE;
+		}
+		tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS] = mono_native_tls_get_value (mono_jit_tls_id);
+		tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN] = mono_domain_get ();
+	}
+
+	if (!tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN]) {
+		/* Happens during startup */
+		tls->unwind_state.valid = FALSE;
+		return;
 	}
-	tls->jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
 }
 
 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
@@ -620,6 +675,16 @@ get_frame_pointer (MonoContext *ctx, int frame_reg)
 			return ctx->esp;
 		else if (frame_reg == X86_EBP)
 			return ctx->ebp;
+#elif defined(TARGET_ARM)
+		if (frame_reg == ARMREG_SP)
+			return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
+		else if (frame_reg == ARMREG_FP)
+			return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
+#elif defined(TARGET_S390X)
+		if (frame_reg == S390_SP)
+			return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
+		else if (frame_reg == S390_FP)
+			return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
 #endif
 		g_assert_not_reached ();
 		return 0;
@@ -664,7 +729,7 @@ conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
 		return;
 	}
 
-	lmf = tls->lmf;
+	lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
 	frame.domain = NULL;
 
 	/* Number of bytes scanned based on GC map data */
@@ -679,15 +744,18 @@ conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
 	/* This is one past the last address which we have scanned */
 	stack_limit = stack_start;
 
-	if (!tls->has_context)
+	if (!tls->unwind_state.valid)
 		memset (&new_ctx, 0, sizeof (ctx));
 	else
-		memcpy (&new_ctx, &tls->ctx, sizeof (MonoContext));
+		memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
 
 	memset (reg_locations, 0, sizeof (reg_locations));
 	memset (new_reg_locations, 0, sizeof (new_reg_locations));
 
 	while (TRUE) {
+		if (!tls->unwind_state.valid)
+			break;
+
 		memcpy (&ctx, &new_ctx, sizeof (ctx));
 
 		for (i = 0; i < MONO_MAX_IREGS; ++i) {
@@ -711,12 +779,14 @@ conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
 
 		g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
 
-		res = mono_find_jit_info_ext (frame.domain ? frame.domain : mono_domain_get (), tls->jit_tls, NULL, &ctx, &new_ctx, NULL, &lmf, new_reg_locations, &frame);
+		res = mono_find_jit_info_ext (frame.domain ? frame.domain : tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN], tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS], NULL, &ctx, &new_ctx, NULL, &lmf, new_reg_locations, &frame);
 		if (!res)
 			break;
 
 		ji = frame.ji;
 
+		// FIXME: For skipped frames, scan the param area of the parent frame conservatively ?
+
 		if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
 			/*
 			 * These frames are problematic for several reasons:
@@ -745,6 +815,23 @@ conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
 			}
 			DEBUG (fprintf (logfile, "\t <Last frame>\n"));
 			last = FALSE;
+			/*
+			 * new_reg_locations is not precise when a method is interrupted during its epilog, so clear it.
+			 */
+			for (i = 0; i < MONO_MAX_IREGS; ++i) {
+				if (reg_locations [i]) {
+					DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
+					mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
+					scanned_registers += SIZEOF_SLOT;
+				}
+				if (new_reg_locations [i]) {
+					DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), new_reg_locations [i]));
+					mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
+					scanned_registers += SIZEOF_SLOT;
+				}
+				reg_locations [i] = NULL;
+				new_reg_locations [i] = NULL;
+			}
 			continue;
 		}
 
@@ -948,7 +1035,7 @@ conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
 		}
 
 		/*
-		 * Clear locations of precisely stacked registers.
+		 * Clear locations of precisely tracked registers.
 		 */
 		if (precise_regmask) {
 			for (i = 0; i < NREGS; ++i) {
@@ -1016,6 +1103,9 @@ precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
 	if (!tls)
 		return;
 
+	if (!tls->unwind_state.valid)
+		return;
+
 	for (findex = 0; findex < tls->nframes; findex ++) {
 		/* Load information saved by the !precise pass */
 		fi = &tls->frames [findex];
@@ -1125,6 +1215,8 @@ thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gb
 		precise_pass (tls, stack_start, stack_end);
 }
 
+#ifndef DISABLE_JIT
+
 static void
 mini_gc_init_gc_map (MonoCompile *cfg)
 {
@@ -1134,7 +1226,10 @@ mini_gc_init_gc_map (MonoCompile *cfg)
 	if (!mono_gc_is_moving ())
 		return;
 
-	if (!cfg->compile_aot && !mono_gc_precise_stack_mark_enabled ())
+	if (cfg->compile_aot) {
+		if (!enable_gc_maps_for_aot)
+			return;
+	} else if (!mono_gc_precise_stack_mark_enabled ())
 		return;
 
 #if 1
@@ -1217,30 +1312,50 @@ slot_to_fp_offset (MonoCompile *cfg, int slot)
 	return (slot * SIZEOF_SLOT) + gcfg->min_offset;
 }
 
-static inline void
+static inline MONO_ALWAYS_INLINE void
 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
 {
 	g_assert (slot >= 0 && slot < gcfg->nslots);
 
 	if (type == SLOT_PIN) {
-		clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
-		set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+		clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
+		set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
 	} else if (type == SLOT_REF) {
-		set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
-		clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+		set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
+		clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
 	} else if (type == SLOT_NOREF) {
-		clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
-		clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
+		clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
+		clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
 	}
 }
 
 static inline void
 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
 {
+	int width, pos;
+	guint8 *ref_bitmap, *pin_bitmap;
+
+	/*
 	int cindex;
 
 	for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
 		set_slot (gcfg, slot, cindex, type);
+	*/
+	ref_bitmap = gcfg->stack_ref_bitmap;
+	pin_bitmap = gcfg->stack_pin_bitmap;
+	width = gcfg->stack_bitmap_width;
+	pos = width * slot;
+
+	if (type == SLOT_PIN) {
+		memset (ref_bitmap + pos, 0, width);
+		memset (pin_bitmap + pos, 0xff, width);
+	} else if (type == SLOT_REF) {
+		memset (ref_bitmap + pos, 0xff, width);
+		memset (pin_bitmap + pos, 0, width);
+	} else if (type == SLOT_NOREF) {
+		memset (ref_bitmap + pos, 0, width);
+		memset (pin_bitmap + pos, 0, width);
+	}
 }
 
 static inline void
@@ -1261,14 +1376,14 @@ set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type
 	g_assert (slot >= 0 && slot < gcfg->nregs);
 
 	if (type == SLOT_PIN) {
-		clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
-		set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+		clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
+		set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
 	} else if (type == SLOT_REF) {
-		set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
-		clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+		set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
+		clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
 	} else if (type == SLOT_NOREF) {
-		clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
-		clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
+		clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
+		clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
 	}
 }
 
@@ -1424,10 +1539,39 @@ process_other_slots (MonoCompile *cfg)
 		if (cfg->verbose_level > 1) {
 			if (type == SLOT_REF)
 				printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
+			else if (type == SLOT_NOREF)
+				printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
 		}
 	}
 }
 
+static gsize*
+get_vtype_bitmap (MonoType *t, int *numbits)
+{
+	MonoClass *klass = mono_class_from_mono_type (t);
+
+	if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
+		/* FIXME: Generic sharing */
+		return NULL;
+	} else {
+		mono_class_compute_gc_descriptor (klass);
+
+		return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
+	}
+}
+
+static inline const char*
+get_offset_sign (int offset)
+{
+	return offset < 0 ? "-" : "+";
+}
+
+static inline int
+get_offset_val (int offset)
+{
+	return offset < 0 ? (- offset) : offset;
+}
+
 static void
 process_variables (MonoCompile *cfg)
 {
@@ -1475,8 +1619,10 @@ process_variables (MonoCompile *cfg)
 		gboolean byref, is_this = FALSE;
 		gboolean is_arg = i < cfg->locals_start;
 
-		if (ins == cfg->ret)
-			continue;
+		if (ins == cfg->ret) {
+			if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
+				continue;
+		}
 
 		vmv = MONO_VARINFO (cfg, i);
 
@@ -1530,10 +1676,6 @@ process_variables (MonoCompile *cfg)
 		if (ins->inst_offset % SIZEOF_SLOT != 0)
 			continue;
 
-		if (is_arg && ins->inst_offset >= gcfg->max_offset)
-			/* In parent frame */
-			continue;
-
 		pos = fp_offset_to_slot (cfg, ins->inst_offset);
 
 		if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
@@ -1541,7 +1683,7 @@ process_variables (MonoCompile *cfg)
 			set_slot_everywhere (gcfg, pos, SLOT_NOREF);
 
 			if (cfg->verbose_level > 1) {
-				printf ("\tdead arg at fp%s0x%x (slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "+", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, pos, mono_type_full_name (ins->inst_vtype));
+				printf ("\tdead arg at fp%s0x%x (slot = %d): %s\n", get_offset_sign (ins->inst_offset), get_offset_val (ins->inst_offset), pos, mono_type_full_name (ins->inst_vtype));
 			}
 			continue;
 		}
@@ -1559,6 +1701,9 @@ process_variables (MonoCompile *cfg)
 				size = mono_class_value_size (ins->klass, NULL);
 			size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
 
+			if (cfg->verbose_level > 1)
+				printf ("\tvtype R%d at %s0x%x(fp)-%s0x%x(fp) (slot %d-%d): %s\n", vmv->vreg, get_offset_sign (ins->inst_offset), get_offset_val (ins->inst_offset), get_offset_sign (ins->inst_offset), get_offset_val (ins->inst_offset + (size_in_slots * SIZEOF_SLOT)), pos, pos + size_in_slots, mono_type_full_name (ins->inst_vtype));
+
 			if (!ins->klass->has_references) {
 				if (is_arg) {
 					for (j = 0; j < size_in_slots; ++j)
@@ -1567,32 +1712,24 @@ process_variables (MonoCompile *cfg)
 				continue;
 			}
 
-			if (ins->klass->generic_container || mono_class_is_open_constructed_type (t)) {
-				/* FIXME: Generic sharing */
+			bitmap = get_vtype_bitmap (t, &numbits);
+			if (!bitmap)
 				pin = TRUE;
-			} else {
-				mono_class_compute_gc_descriptor (ins->klass);
 
-				bitmap = mono_gc_get_bitmap_for_descr (ins->klass->gc_descr, &numbits);
-				if (!bitmap)
+			/*
+			 * Most vtypes are marked volatile because of the LDADDR instructions,
+			 * and they have no liveness information since they are decomposed
+			 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
+			 * them during VZERO decomposition.
+			 */
+			if (!is_arg) {
+				if (!pc_offsets [vmv->vreg])
 					pin = TRUE;
 
-				/*
-				 * Most vtypes are marked volatile because of the LDADDR instructions,
-				 * and they have no liveness information since they are decomposed
-				 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
-				 * them during VZERO decomposition.
-				 */
-				if (!pc_offsets [vmv->vreg])
+				if (ins->backend.is_pinvoke)
 					pin = TRUE;
 			}
 
-			if (ins->backend.is_pinvoke)
-				pin = TRUE;
-
-			if (cfg->verbose_level > 1)
-				printf ("\tvtype R%d at fp+0x%x-0x%x: %s\n", vmv->vreg, (int)ins->inst_offset, (int)(ins->inst_offset + (size / SIZEOF_SLOT)), mono_type_full_name (ins->inst_vtype));
-
 			if (bitmap) {
 				for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
 					if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
@@ -1661,7 +1798,12 @@ process_variables (MonoCompile *cfg)
 		if (!mini_type_is_reference (cfg, t)) {
 			set_slot_everywhere (gcfg, pos, SLOT_NOREF);
 			if (cfg->verbose_level > 1)
-				printf ("\tnoref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
+				printf ("\tnoref%s at %s0x%x(fp) (R%d, slot = %d): %s\n", (is_arg ? " arg" : ""), ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
+			if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
+				set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
+				if (cfg->verbose_level > 1)
+					printf ("\tnoref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)(ins->inst_offset + 4) : (int)ins->inst_offset + 4, vmv->vreg, pos + 1, mono_type_full_name (ins->inst_vtype));
+			}
 			continue;
 		}
 
@@ -1694,7 +1836,7 @@ process_variables (MonoCompile *cfg)
 		}
 
 		if (cfg->verbose_level > 1) {
-			printf ("\tref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
+			printf ("\tref%s at %s0x%x(fp) (R%d, slot = %d): %s\n", (is_arg ? " arg" : ""), ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
 		}
 	}
 
@@ -1722,28 +1864,6 @@ sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
 #endif
 }
 
-static GCSlotType
-type_to_gc_slot_type (MonoCompile *cfg, MonoType *t)
-{
-	if (t->byref)
-		return SLOT_PIN;
-	t = mini_type_get_underlying_type (NULL, t);
-	if (mini_type_is_reference (cfg, t))
-		return SLOT_REF;
-	else {
-		if (MONO_TYPE_ISSTRUCT (t)) {
-			MonoClass *klass = mono_class_from_mono_type (t);
-			if (!klass->has_references) {
-				return SLOT_NOREF;
-			} else {
-				// FIXME:
-				return SLOT_PIN;
-			}
-		}
-		return SLOT_NOREF;
-	}
-}
-
 static void
 process_param_area_slots (MonoCompile *cfg)
 {
@@ -1793,39 +1913,10 @@ process_param_area_slots (MonoCompile *cfg)
 			set_slot_everywhere (gcfg, i, SLOT_NOREF);
 	}
 
-	for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
-		GCCallSite *callsite = gcfg->callsites [cindex];
-		GSList *l;
-
-		for (l = callsite->param_slots; l; l = l->next) {
-			MonoInst *def = l->data;
-			MonoType *t = def->inst_vtype;
-			int sp_offset = def->inst_offset;
-			int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
-			int slot = fp_offset_to_slot (cfg, fp_offset);
-			GCSlotType type = type_to_gc_slot_type (cfg, t);
-
-			if (MONO_TYPE_ISSTRUCT (t)) {
-				guint32 align;
-				guint32 size;
-				int size_in_slots;
-
-				size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
-				size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
-				// FIXME: slot type
-				for (i = 0; i < size_in_slots; ++i) {
-					set_slot_in_range (gcfg, slot + i, def->backend.pc_offset, callsite->pc_offset + 1, type);
-				}
-				if (cfg->verbose_level > 1)
-					printf ("\t%s param area slots at %s0x%x(fp)=0x%x(sp) (slot = %d-%d) [0x%x-0x%x]\n", slot_type_to_string (type), fp_offset >= 0 ? "+" : "-", ABS (fp_offset), sp_offset, slot, slot + (size / (int)sizeof (mgreg_t)), def->backend.pc_offset, callsite->pc_offset + 1);
-			} else {
-				/* The slot is live between the def instruction and the call */
-				set_slot_in_range (gcfg, slot, def->backend.pc_offset, callsite->pc_offset + 1, type);
-				if (cfg->verbose_level > 1)
-					printf ("\t%s param area slot at %s0x%x(fp)=0x%x(sp) (slot = %d) [0x%x-0x%x]\n", slot_type_to_string (type), fp_offset >= 0 ? "+" : "-", ABS (fp_offset), sp_offset, slot, def->backend.pc_offset, callsite->pc_offset + 1);
-			}
-		}
-	}
+	/*
+	 * We treat param area slots as being part of the callee's frame, to be able to handle tail calls which overwrite
+	 * the argument area of the caller.
+	 */
 }
 
 static void
@@ -1898,7 +1989,7 @@ compute_frame_size (MonoCompile *cfg)
 	/* Compute min/max offsets from the fp */
 
 	/* Locals */
-#if defined(TARGET_AMD64) || defined(TARGET_X86)
+#if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM) || defined(TARGET_S390X)
 	locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
 	locals_max_offset = cfg->locals_max_stack_offset;
 #else
@@ -1916,8 +2007,14 @@ compute_frame_size (MonoCompile *cfg)
 	for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
 		MonoInst *ins = cfg->args [i];
 
-		if (ins->opcode == OP_REGOFFSET)
+		if (ins->opcode == OP_REGOFFSET) {
+			int size, size_in_slots;
+			size = mini_type_stack_size_full (cfg->generic_sharing_context, ins->inst_vtype, NULL, ins->backend.is_pinvoke);
+			size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
+
 			min_offset = MIN (min_offset, ins->inst_offset);
+			max_offset = MAX ((int)max_offset, (int)(ins->inst_offset + (size_in_slots * SIZEOF_SLOT)));
+		}
 	}
 
 	/* Cfa slots */
@@ -1948,6 +2045,12 @@ compute_frame_size (MonoCompile *cfg)
 	min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
 #elif defined(TARGET_X86)
 	min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
+#elif defined(TARGET_ARM)
+	// FIXME:
+#elif defined(TARGET_s390X)
+	// FIXME:
+#else
+	NOT_IMPLEMENTED;
 #endif
 
 	gcfg->min_offset = min_offset;
@@ -2006,15 +2109,15 @@ init_gcfg (MonoCompile *cfg)
 	gcfg->nregs = nregs;
 	gcfg->callsites = callsites;
 	gcfg->ncallsites = ncallsites;
-	gcfg->stack_bitmap_width = ALIGN_TO (nslots, 8) / 8;
-	gcfg->reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
-	gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
-	gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
-	gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
-	gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
+	gcfg->stack_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
+	gcfg->reg_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
+	gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
+	gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
+	gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
+	gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
 
 	/* All slots start out as PIN */
-	memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * ncallsites);
+	memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * nregs);
 	for (i = 0; i < nregs; ++i) {
 		/*
 		 * By default, registers are NOREF.
@@ -2026,6 +2129,19 @@ init_gcfg (MonoCompile *cfg)
 	}
 }
 
+static inline gboolean
+has_bit_set (guint8 *bitmap, int width, int slot)
+{
+	int i;
+	int pos = width * slot;
+
+	for (i = 0; i < width; ++i) {
+		if (bitmap [pos + i])
+			break;
+	}
+	return i < width;
+}
+
 static void
 create_map (MonoCompile *cfg)
 {
@@ -2060,12 +2176,10 @@ create_map (MonoCompile *cfg)
 		gboolean has_ref = FALSE;
 		gboolean has_pin = FALSE;
 
-		for (j = 0; j < ncallsites; ++j) {
-			if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
-				has_pin = TRUE;
-			if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
-				has_ref = TRUE;
-		}
+		if (has_bit_set (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i))
+			has_pin = TRUE;
+		if (has_bit_set (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i))
+			has_ref = TRUE;
 
 		if (has_ref)
 			has_ref_slots = TRUE;
@@ -2105,18 +2219,10 @@ create_map (MonoCompile *cfg)
 		if (!(cfg->used_int_regs & (1 << i)))
 			continue;
 
-		for (j = 0; j < ncallsites; ++j) {
-			if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i)) {
-				has_ref = TRUE;
-				break;
-			}
-		}
-		for (j = 0; j < ncallsites; ++j) {
-			if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i)) {
-				has_pin = TRUE;
-				break;
-			}
-		}
+		if (has_bit_set (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i))
+			has_pin = TRUE;
+		if (has_bit_set (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i))
+			has_ref = TRUE;
 
 		if (has_ref) {
 			reg_ref_mask |= (1 << i);
@@ -2135,6 +2241,7 @@ create_map (MonoCompile *cfg)
 
 	/* Create the GC Map */
 
+	/* The work bitmaps have one row for each slot, since this is how we access them during construction */
 	stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
 	stack_bitmap_size = stack_bitmap_width * ncallsites;
 	reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
@@ -2171,7 +2278,7 @@ create_map (MonoCompile *cfg)
 		bitmap = &bitmaps [map->stack_ref_bitmap_offset];
 		for (i = 0; i < nslots; ++i) {
 			for (j = 0; j < ncallsites; ++j) {
-				if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
+				if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i, j))
 					set_bit (bitmap, stack_bitmap_width, j, i - start);
 			}
 		}
@@ -2183,7 +2290,7 @@ create_map (MonoCompile *cfg)
 		bitmap = &bitmaps [map->stack_pin_bitmap_offset];
 		for (i = 0; i < nslots; ++i) {
 			for (j = 0; j < ncallsites; ++j) {
-				if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
+				if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i, j))
 					set_bit (bitmap, stack_bitmap_width, j, i - start);
 			}
 		}
@@ -2197,7 +2304,7 @@ create_map (MonoCompile *cfg)
 		for (i = 0; i < nregs; ++i) {
 			if (reg_ref_mask & (1 << i)) {
 				for (j = 0; j < ncallsites; ++j) {
-					if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i))
+					if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i, j))
 						set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
 				}
 				bindex ++;
@@ -2213,7 +2320,7 @@ create_map (MonoCompile *cfg)
 		for (i = 0; i < nregs; ++i) {
 			if (reg_pin_mask & (1 << i)) {
 				for (j = 0; j < ncallsites; ++j) {
-					if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i))
+					if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i, j))
 						set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
 				}
 				bindex ++;
@@ -2327,6 +2434,8 @@ mini_gc_create_gc_map (MonoCompile *cfg)
 	create_map (cfg);
 }
 
+#endif /* DISABLE_JIT */
+
 static void
 parse_debug_options (void)
 {
@@ -2408,11 +2517,18 @@ mini_gc_init (void)
 
 #else
 
+void
+mini_gc_enable_gc_maps_for_aot (void)
+{
+}
+
 void
 mini_gc_init (void)
 {
 }
 
+#ifndef DISABLE_JIT
+
 static void
 mini_gc_init_gc_map (MonoCompile *cfg)
 {
@@ -2433,8 +2549,12 @@ mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType ty
 {
 }
 
+#endif /* DISABLE_JIT */
+
 #endif
 
+#ifndef DISABLE_JIT
+
 /*
  * mini_gc_init_cfg:
  *
@@ -2451,6 +2571,8 @@ mini_gc_init_cfg (MonoCompile *cfg)
 	mini_gc_init_gc_map (cfg);
 }
 
+#endif /* DISABLE_JIT */
+
 /*
  * Problems with the current code:
  * - the stack walk is slow