X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fsparc%2Fsparc_emitter.c;h=9f78eafcdfdafb58d541e0f869a6159630e66958;hb=fc81b817119d8635eaeb345c8623255fc51bdb22;hp=07eef6137e5e6d277a69580e5753c5526c247b7c;hpb=e50e95fbed5212b07773f5a118ee338c814ff368;p=libfirm

diff --git a/ir/be/sparc/sparc_emitter.c b/ir/be/sparc/sparc_emitter.c
index 07eef6137..9f78eafcd 100644
--- a/ir/be/sparc/sparc_emitter.c
+++ b/ir/be/sparc/sparc_emitter.c
@@ -20,6 +20,7 @@
 /**
  * @file
  * @brief   emit assembler for a backend graph
+ * @author  Hannes Rapp, Matthias Braun
  * @version $Id$
  */
 #include "config.h"
@@ -39,6 +40,7 @@
 #include "error.h"
 #include "raw_bitset.h"
 #include "dbginfo.h"
+#include "heights.h"
 
 #include "../besched.h"
 #include "../beblocksched.h"
@@ -46,6 +48,7 @@
 #include "../begnuas.h"
 #include "../be_dbgout.h"
 #include "../benode.h"
+#include "../bestack.h"
 
 #include "sparc_emitter.h"
 #include "gen_sparc_emitter.h"
@@ -55,6 +58,12 @@
 
 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
+static ir_heights_t  *heights;
+static const ir_node *delay_slot_filler; /**< this node has been choosen to fill
+                                              the next delay slot */
+
+static void sparc_emit_node(const ir_node *node);
+
 /**
  * Returns the register at in position pos.
  */
@@ -109,11 +118,6 @@ static const arch_register_t *get_out_reg(const ir_node *node, int pos)
 	return reg;
 }
 
-static bool is_valid_immediate(int32_t value)
-{
-	return -4096 <= value && value < 4096;
-}
-
 void sparc_emit_immediate(const ir_node *node)
 {
 	const sparc_attr_t *attr   = get_sparc_attr_const(node);
@@ -121,7 +125,7 @@ void sparc_emit_immediate(const ir_node *node)
 
 	if (entity == NULL) {
 		int32_t value = attr->immediate_value;
-		assert(is_valid_immediate(value));
+		assert(sparc_is_value_imm_encodeable(value));
 		be_emit_irprintf("%d", value);
 	} else {
 		be_emit_cstring("%lo(");
@@ -172,22 +176,15 @@ void sparc_emit_dest_register(const ir_node *node, int pos)
  */
 void sparc_emit_reg_or_imm(const ir_node *node, int pos)
 {
-	if (get_irn_arity(node) > pos) {
-		// we have reg input
-		sparc_emit_source_register(node, pos);
-	} else {
+	if (arch_irn_get_flags(node) & ((arch_irn_flags_t)sparc_arch_irn_flag_immediate_form)) {
 		// we have a imm input
 		sparc_emit_immediate(node);
+	} else {
+		// we have reg input
+		sparc_emit_source_register(node, pos);
 	}
 }
 
-static bool is_stack_pointer_relative(const ir_node *node)
-{
-	const arch_register_t *sp = &sparc_gp_regs[REG_SP];
-	return (is_sparc_St(node) && get_in_reg(node, n_sparc_St_ptr) == sp)
-	    || (is_sparc_Ld(node) && get_in_reg(node, n_sparc_Ld_ptr) == sp);
-}
-
 /**
  * emit SP offset
  */
@@ -203,12 +200,8 @@ void sparc_emit_offset(const ir_node *node, int offset_node_pos)
 		sparc_emit_source_register(node, offset_node_pos);
 	} else if (attr->is_frame_entity) {
 		int32_t offset = attr->base.immediate_value;
-		/* bad hack: the real stack stuff is behind the always-there spill
-		 * space for the register window and stack */
-		if (is_stack_pointer_relative(node))
-			offset += SPARC_MIN_STACKSIZE;
 		if (offset != 0) {
-			assert(is_valid_immediate(offset));
+			assert(sparc_is_value_imm_encodeable(offset));
 			be_emit_irprintf("%+ld", offset);
 		}
 	} else if (attr->base.immediate_value != 0
@@ -324,7 +317,7 @@ void sparc_emit_fp_mode_suffix(const ir_node *node)
 
 static ir_node *get_jump_target(const ir_node *jump)
 {
-	return get_irn_link(jump);
+	return (ir_node*)get_irn_link(jump);
 }
 
 /**
@@ -336,12 +329,136 @@ static void sparc_emit_cfop_target(const ir_node *node)
 	be_gas_emit_block_name(block);
 }
 
+static int get_sparc_Call_dest_addr_pos(const ir_node *node)
+{
+	return get_irn_arity(node)-1;
+}
+
+static bool ba_is_fallthrough(const ir_node *node)
+{
+	ir_node *block      = get_nodes_block(node);
+	ir_node *next_block = (ir_node*)get_irn_link(block);
+	return get_irn_link(node) == next_block;
+}
+
+static bool is_no_instruction(const ir_node *node)
+{
+	/* copies are nops if src_reg == dest_reg */
+	if (be_is_Copy(node) || be_is_CopyKeep(node)) {
+		const arch_register_t *src_reg  = get_in_reg(node, 0);
+		const arch_register_t *dest_reg = get_out_reg(node, 0);
+
+		if (src_reg == dest_reg)
+			return true;
+	}
+	if (be_is_IncSP(node) && be_get_IncSP_offset(node) == 0)
+		return true;
+	/* Ba is not emitted if it is a simple fallthrough */
+	if (is_sparc_Ba(node) && ba_is_fallthrough(node))
+		return true;
+
+	return be_is_Keep(node) || be_is_Start(node) || is_Phi(node);
+}
+
+static bool has_delay_slot(const ir_node *node)
+{
+	if (is_sparc_Ba(node) && ba_is_fallthrough(node))
+		return false;
+
+	return is_sparc_Bicc(node) || is_sparc_fbfcc(node) || is_sparc_Ba(node)
+		|| is_sparc_SwitchJmp(node) || is_sparc_Call(node)
+		|| is_sparc_SDiv(node) || is_sparc_UDiv(node)
+		|| be_is_Return(node);
+}
+
+/** returns true if the emitter for this sparc node can produce more than one
+ * actual sparc instruction.
+ * Usually it is a bad sign if we have to add instructions here. We should
+ * rather try to get them lowered down. So we can actually put them into
+ * delay slots and make them more accessible to the scheduler.
+ */
+static bool emits_multiple_instructions(const ir_node *node)
+{
+	if (has_delay_slot(node))
+		return true;
+
+	return is_sparc_Mulh(node) || is_sparc_SDiv(node) || is_sparc_UDiv(node)
+		|| be_is_MemPerm(node) || be_is_Perm(node);
+}
+
 /**
- * Emit single entity
+ * search for an instruction that can fill the delay slot of @p node
  */
-static void sparc_emit_entity(ir_entity *entity)
+static const ir_node *pick_delay_slot_for(const ir_node *node)
 {
-	be_gas_emit_entity(entity);
+	const ir_node *check      = node;
+	const ir_node *schedpoint = node;
+	unsigned       tries      = 0;
+	/* currently we don't track which registers are still alive, so we can't
+	 * pick any other instructions other than the one directly preceding */
+	static const unsigned PICK_DELAY_SLOT_MAX_DISTANCE = 1;
+
+	assert(has_delay_slot(node));
+
+	if (is_sparc_Call(node)) {
+		const sparc_attr_t *attr   = get_sparc_attr_const(node);
+		ir_entity          *entity = attr->immediate_value_entity;
+		if (entity != NULL) {
+			check = NULL; /* pick any instruction, dependencies on Call
+			                 don't matter */
+		} else {
+			/* we only need to check the value for the call destination */
+			check = get_irn_n(node, get_sparc_Call_dest_addr_pos(node));
+		}
+
+		/* the Call also destroys the value of %o7, but since this is currently
+		 * marked as ignore register in the backend, it should never be used by
+		 * the instruction in the delay slot. */
+	} else if (be_is_Return(node)) {
+		/* we only have to check the jump destination value */
+		int arity = get_irn_arity(node);
+		int i;
+
+		check = NULL;
+		for (i = 0; i < arity; ++i) {
+			ir_node               *in  = get_irn_n(node, i);
+			const arch_register_t *reg = arch_get_irn_register(in);
+			if (reg == &sparc_registers[REG_O7]) {
+				check = skip_Proj(in);
+				break;
+			}
+		}
+	} else {
+		check = node;
+	}
+
+	while (sched_has_prev(schedpoint)) {
+		schedpoint = sched_prev(schedpoint);
+
+		if (has_delay_slot(schedpoint))
+			break;
+
+		/* skip things which don't really result in instructions */
+		if (is_no_instruction(schedpoint))
+			continue;
+
+		if (tries++ >= PICK_DELAY_SLOT_MAX_DISTANCE)
+			break;
+
+		if (emits_multiple_instructions(schedpoint))
+			continue;
+
+		/* allowed for delayslot: any instruction which is not necessary to
+		 * compute an input to the branch. */
+		if (check != NULL
+				&& heights_reachable_in_block(heights, check, schedpoint))
+			continue;
+
+		/* found something */
+		return schedpoint;
+	}
+
+	return NULL;
 }
 
 /**
@@ -349,39 +466,26 @@ static void sparc_emit_entity(ir_entity *entity)
  */
 static void emit_be_IncSP(const ir_node *irn)
 {
-	int offs = -be_get_IncSP_offset(irn);
+	int offset = be_get_IncSP_offset(irn);
 
-	if (offs == 0)
-			return;
+	if (offset == 0)
+		return;
 
 	/* SPARC stack grows downwards */
-	if (offs < 0) {
+	if (offset < 0) {
 		be_emit_cstring("\tsub ");
-		offs = -offs;
+		offset = -offset;
 	} else {
 		be_emit_cstring("\tadd ");
 	}
 
 	sparc_emit_source_register(irn, 0);
-	be_emit_irprintf(", %d", offs);
+	be_emit_irprintf(", %d", -offset);
 	be_emit_cstring(", ");
 	sparc_emit_dest_register(irn, 0);
 	be_emit_finish_line_gas(irn);
 }
 
-/**
- * emits code for save instruction with min. required stack space
- */
-static void emit_sparc_Save(const ir_node *irn)
-{
-	const sparc_save_attr_t *save_attr = get_sparc_save_attr_const(irn);
-	be_emit_cstring("\tsave ");
-	sparc_emit_source_register(irn, 0);
-	be_emit_irprintf(", %d, ", -save_attr->initial_stacksize);
-	sparc_emit_dest_register(irn, 0);
-	be_emit_finish_line_gas(irn);
-}
-
 /**
  * emits code for mulh
  */
@@ -407,8 +511,13 @@ static void emit_sparc_Mulh(const ir_node *irn)
 
 static void fill_delay_slot(void)
 {
-	be_emit_cstring("\tnop\n");
-	be_emit_write_line();
+	if (delay_slot_filler != NULL) {
+		sparc_emit_node(delay_slot_filler);
+		delay_slot_filler = NULL;
+	} else {
+		be_emit_cstring("\tnop\n");
+		be_emit_write_line();
+	}
 }
 
 static void emit_sparc_Div(const ir_node *node, bool is_signed)
@@ -445,18 +554,6 @@ static void emit_sparc_UDiv(const ir_node *node)
 	emit_sparc_Div(node, false);
 }
 
-/**
- * Emits code for return node
- */
-static void emit_be_Return(const ir_node *irn)
-{
-	be_emit_cstring("\tret");
-	//be_emit_cstring("\tjmp %i7+8");
-	be_emit_finish_line_gas(irn);
-	be_emit_cstring("\trestore");
-	be_emit_finish_line_gas(irn);
-}
-
 /**
  * Emits code for Call node
  */
@@ -467,14 +564,14 @@ static void emit_sparc_Call(const ir_node *node)
 
 	be_emit_cstring("\tcall ");
 	if (entity != NULL) {
-	    sparc_emit_entity(entity);
+	    be_gas_emit_entity(entity);
 	    if (attr->immediate_value != 0) {
 			be_emit_irprintf("%+d", attr->immediate_value);
 		}
 		be_emit_cstring(", 0");
 	} else {
-		int last = get_irn_arity(node);
-		sparc_emit_source_register(node, last-1);
+		int dest_addr = get_sparc_Call_dest_addr_pos(node);
+		sparc_emit_source_register(node, dest_addr);
 	}
 	be_emit_finish_line_gas(node);
 
@@ -511,9 +608,6 @@ static void emit_be_Perm(const ir_node *irn)
 	be_emit_finish_line_gas(irn);
 }
 
-/**
- * TODO: not really tested but seems to work with memperm_arity == 1
- */
 static void emit_be_MemPerm(const ir_node *node)
 {
 	int i;
@@ -570,24 +664,41 @@ static void emit_be_MemPerm(const ir_node *node)
 	assert(sp_change == 0);
 }
 
-/**
- * Emits code for FrameAddr fix
- */
+static void emit_be_Return(const ir_node *node)
+{
+	const char *destreg = "%o7";
+
+	/* hack: we don't explicitely model register changes because of the
+	 * restore node. So we have to do it manually here */
+	if (delay_slot_filler != NULL &&
+			(is_sparc_Restore(delay_slot_filler)
+			 || is_sparc_RestoreZero(delay_slot_filler))) {
+		destreg = "%i7";
+	}
+	be_emit_cstring("\tjmp ");
+	be_emit_string(destreg);
+	be_emit_cstring("+8");
+	be_emit_finish_line_gas(node);
+	fill_delay_slot();
+}
+
 static void emit_sparc_FrameAddr(const ir_node *node)
 {
-	const sparc_attr_t *attr = get_sparc_attr_const(node);
+	const sparc_attr_t *attr   = get_sparc_attr_const(node);
+	int32_t             offset = attr->immediate_value;
 
-	// no need to fix offset as we are adressing via the framepointer
-	if (attr->immediate_value >= 0) {
+	if (offset < 0) {
 		be_emit_cstring("\tadd ");
 		sparc_emit_source_register(node, 0);
 		be_emit_cstring(", ");
-		be_emit_irprintf("%ld", attr->immediate_value);
+		assert(sparc_is_value_imm_encodeable(offset));
+		be_emit_irprintf("%ld", offset);
 	} else {
 		be_emit_cstring("\tsub ");
 		sparc_emit_source_register(node, 0);
 		be_emit_cstring(", ");
-		be_emit_irprintf("%ld", -attr->immediate_value);
+		assert(sparc_is_value_imm_encodeable(-offset));
+		be_emit_irprintf("%ld", -offset);
 	}
 
 	be_emit_cstring(", ");
@@ -595,70 +706,65 @@ static void emit_sparc_FrameAddr(const ir_node *node)
 	be_emit_finish_line_gas(node);
 }
 
-static const char *get_icc_unsigned(pn_Cmp pnc)
+static const char *get_icc_unsigned(ir_relation relation)
 {
-	switch (pnc) {
-	case pn_Cmp_False: return "bn";
-	case pn_Cmp_Eq:    return "be";
-	case pn_Cmp_Lt:    return "blu";
-	case pn_Cmp_Le:    return "bleu";
-	case pn_Cmp_Gt:    return "bgu";
-	case pn_Cmp_Ge:    return "bgeu";
-	case pn_Cmp_Lg:    return "bne";
-	case pn_Cmp_Leg:   return "ba";
-	default: panic("Cmp has unsupported pnc");
+	switch (relation & (ir_relation_less_equal_greater)) {
+	case ir_relation_false:              return "bn";
+	case ir_relation_equal:              return "be";
+	case ir_relation_less:               return "blu";
+	case ir_relation_less_equal:         return "bleu";
+	case ir_relation_greater:            return "bgu";
+	case ir_relation_greater_equal:      return "bgeu";
+	case ir_relation_less_greater:       return "bne";
+	case ir_relation_less_equal_greater: return "ba";
+	default: panic("Cmp has unsupported relation");
 	}
 }
 
-static const char *get_icc_signed(pn_Cmp pnc)
+static const char *get_icc_signed(ir_relation relation)
 {
-	switch (pnc) {
-	case pn_Cmp_False: return "bn";
-	case pn_Cmp_Eq:    return "be";
-	case pn_Cmp_Lt:    return "bl";
-	case pn_Cmp_Le:    return "ble";
-	case pn_Cmp_Gt:    return "bg";
-	case pn_Cmp_Ge:    return "bge";
-	case pn_Cmp_Lg:    return "bne";
-	case pn_Cmp_Leg:   return "ba";
-	default: panic("Cmp has unsupported pnc");
+	switch (relation & (ir_relation_less_equal_greater)) {
+	case ir_relation_false:              return "bn";
+	case ir_relation_equal:              return "be";
+	case ir_relation_less:               return "bl";
+	case ir_relation_less_equal:         return "ble";
+	case ir_relation_greater:            return "bg";
+	case ir_relation_greater_equal:      return "bge";
+	case ir_relation_less_greater:       return "bne";
+	case ir_relation_less_equal_greater: return "ba";
+	default: panic("Cmp has unsupported relation");
 	}
 }
 
-static const char *get_fcc(pn_Cmp pnc)
+static const char *get_fcc(ir_relation relation)
 {
-	switch (pnc) {
-	case pn_Cmp_False: return "fbn";
-	case pn_Cmp_Eq:    return "fbe";
-	case pn_Cmp_Lt:    return "fbl";
-	case pn_Cmp_Le:    return "fble";
-	case pn_Cmp_Gt:    return "fbg";
-	case pn_Cmp_Ge:    return "fbge";
-	case pn_Cmp_Lg:    return "fblg";
-	case pn_Cmp_Leg:   return "fbo";
-	case pn_Cmp_Uo:    return "fbu";
-	case pn_Cmp_Ue:    return "fbue";
-	case pn_Cmp_Ul:    return "fbul";
-	case pn_Cmp_Ule:   return "fbule";
-	case pn_Cmp_Ug:    return "fbug";
-	case pn_Cmp_Uge:   return "fbuge";
-	case pn_Cmp_Ne:    return "fbne";
-	case pn_Cmp_True:  return "fba";
-	case pn_Cmp_max:
-		break;
+	switch (relation) {
+	case ir_relation_false:                   return "fbn";
+	case ir_relation_equal:                   return "fbe";
+	case ir_relation_less:                    return "fbl";
+	case ir_relation_less_equal:              return "fble";
+	case ir_relation_greater:                 return "fbg";
+	case ir_relation_greater_equal:           return "fbge";
+	case ir_relation_less_greater:            return "fblg";
+	case ir_relation_less_equal_greater:      return "fbo";
+	case ir_relation_unordered:               return "fbu";
+	case ir_relation_unordered_equal:         return "fbue";
+	case ir_relation_unordered_less:          return "fbul";
+	case ir_relation_unordered_less_equal:    return "fbule";
+	case ir_relation_unordered_greater:       return "fbug";
+	case ir_relation_unordered_greater_equal: return "fbuge";
+	case ir_relation_unordered_less_greater:  return "fbne";
+	case ir_relation_true:                    return "fba";
 	}
-	panic("invalid pnc");
+	panic("invalid relation");
 }
 
-typedef const char* (*get_cc_func)(pn_Cmp pnc);
+typedef const char* (*get_cc_func)(ir_relation relation);
 
-/**
- * Emits code for Branch
- */
 static void emit_sparc_branch(const ir_node *node, get_cc_func get_cc)
 {
 	const sparc_jmp_cond_attr_t *attr = get_sparc_jmp_cond_attr_const(node);
-	pn_Cmp           pnc         = attr->pnc;
+	ir_relation      relation    = attr->relation;
 	const ir_node   *proj_true   = NULL;
 	const ir_node   *proj_false  = NULL;
 	const ir_edge_t *edge;
@@ -679,7 +785,7 @@ static void emit_sparc_branch(const ir_node *node, get_cc_func get_cc)
 	block = get_nodes_block(node);
 
 	/* we have a block schedule */
-	next_block = get_irn_link(block);
+	next_block = (ir_node*)get_irn_link(block);
 
 	if (get_irn_link(proj_true) == next_block) {
 		/* exchange both proj's so the second one can be omitted */
@@ -687,16 +793,12 @@ static void emit_sparc_branch(const ir_node *node, get_cc_func get_cc)
 
 		proj_true  = proj_false;
 		proj_false = t;
-		if (is_sparc_fbfcc(node)) {
-			pnc = get_negated_pnc(pnc, mode_F);
-		} else {
-			pnc = get_negated_pnc(pnc, mode_Iu);
-		}
+		relation   = get_negated_relation(relation);
 	}
 
 	/* emit the true proj */
 	be_emit_cstring("\t");
-	be_emit_string(get_cc(pnc));
+	be_emit_string(get_cc(relation));
 	be_emit_char(' ');
 	sparc_emit_cfop_target(proj_true);
 	be_emit_finish_line_gas(proj_true);
@@ -728,122 +830,53 @@ static void emit_sparc_fbfcc(const ir_node *node)
 	emit_sparc_branch(node, get_fcc);
 }
 
-/**
- * emit Jmp (which actually is a branch always (ba) instruction)
- */
 static void emit_sparc_Ba(const ir_node *node)
 {
-	ir_node *block, *next_block;
-
-	/* for now, the code works for scheduled and non-schedules blocks */
-	block = get_nodes_block(node);
-
-	/* we have a block schedule */
-	next_block = get_irn_link(block);
-	if (get_irn_link(node) != next_block) {
+	if (ba_is_fallthrough(node)) {
+		be_emit_cstring("\t/* fallthrough to ");
+		sparc_emit_cfop_target(node);
+		be_emit_cstring(" */");
+	} else {
 		be_emit_cstring("\tba ");
 		sparc_emit_cfop_target(node);
 		be_emit_finish_line_gas(node);
 		fill_delay_slot();
-	} else {
-		be_emit_cstring("\t/* fallthrough to ");
-		sparc_emit_cfop_target(node);
-		be_emit_cstring(" */");
 	}
 	be_emit_finish_line_gas(node);
 }
 
-static void emit_jump_table(const ir_node *node)
+static void emit_sparc_SwitchJmp(const ir_node *node)
 {
 	const sparc_switch_jmp_attr_t *attr = get_sparc_switch_jmp_attr_const(node);
-	long             switch_min    = LONG_MAX;
-	long             switch_max    = LONG_MIN;
-	long             default_pn    = attr->default_proj_num;
-	ir_entity       *entity        = attr->jump_table;
-	ir_node         *default_block = NULL;
-	unsigned         length;
-	const ir_edge_t *edge;
-	unsigned         i;
-	ir_node        **table;
-
-	/* go over all proj's and collect them */
-	foreach_out_edge(node, edge) {
-		ir_node *proj = get_edge_src_irn(edge);
-		long     pn   = get_Proj_proj(proj);
-
-		/* check for default proj */
-		if (pn == default_pn) {
-			assert(default_block == NULL); /* more than 1 default_pn? */
-			default_block = get_jump_target(proj);
-		} else {
-			switch_min = pn < switch_min ? pn : switch_min;
-			switch_max = pn > switch_max ? pn : switch_max;
-		}
-	}
-	length = (unsigned long) (switch_max - switch_min) + 1;
-	assert(switch_min < LONG_MAX || switch_max > LONG_MIN);
-
-	table = XMALLOCNZ(ir_node*, length);
-	foreach_out_edge(node, edge) {
-		ir_node *proj = get_edge_src_irn(edge);
-		long     pn   = get_Proj_proj(proj);
-		if (pn == default_pn)
-			continue;
 
-		table[pn - switch_min] = get_jump_target(proj);
-	}
-
-	/* emit table */
-	be_gas_emit_switch_section(GAS_SECTION_RODATA);
-	be_emit_cstring("\t.align 4\n");
-	be_gas_emit_entity(entity);
-	be_emit_cstring(":\n");
-	for (i = 0; i < length; ++i) {
-		ir_node *block = table[i];
-		if (block == NULL)
-			block = default_block;
-		be_emit_cstring("\t.long ");
-		be_gas_emit_block_name(block);
-		be_emit_char('\n');
-		be_emit_write_line();
-	}
-	be_gas_emit_switch_section(GAS_SECTION_TEXT);
-
-	xfree(table);
-}
-
-static void emit_sparc_SwitchJmp(const ir_node *node)
-{
 	be_emit_cstring("\tjmp ");
 	sparc_emit_source_register(node, 0);
 	be_emit_finish_line_gas(node);
 	fill_delay_slot();
 
-	emit_jump_table(node);
+	emit_jump_table(node, attr->default_proj_num, attr->jump_table,
+	                get_jump_target);
 }
 
 static void emit_fmov(const ir_node *node, const arch_register_t *src_reg,
                       const arch_register_t *dst_reg)
 {
-	be_emit_cstring("\tfmov ");
+	be_emit_cstring("\tfmovs %");
 	be_emit_string(arch_register_get_name(src_reg));
-	be_emit_cstring(", ");
+	be_emit_cstring(", %");
 	be_emit_string(arch_register_get_name(dst_reg));
 	be_emit_finish_line_gas(node);
 }
 
 static const arch_register_t *get_next_fp_reg(const arch_register_t *reg)
 {
-	unsigned index = reg->index;
-	assert(reg == &sparc_fp_regs[index]);
-	index++;
-	assert(index < N_sparc_fp_REGS);
-	return &sparc_fp_regs[index];
+	unsigned idx = reg->global_index;
+	assert(reg == &sparc_registers[idx]);
+	idx++;
+	assert(idx - REG_F0 < N_sparc_fp_REGS);
+	return &sparc_registers[idx];
 }
 
-/**
- * emit copy node
- */
 static void emit_be_Copy(const ir_node *node)
 {
 	ir_mode               *mode    = get_irn_mode(node);
@@ -874,23 +907,13 @@ static void emit_be_Copy(const ir_node *node)
 	}
 }
 
-
-/**
- * dummy emitter for ignored nodes
- */
 static void emit_nothing(const ir_node *irn)
 {
 	(void) irn;
 }
 
-/**
- * type of emitter function
- */
 typedef void (*emit_func) (const ir_node *);
 
-/**
- * Set a node emitter. Make it a bit more type safe.
- */
 static inline void set_emitter(ir_op *op, emit_func sparc_emit_node)
 {
 	op->ops.generic = (op_func)sparc_emit_node;
@@ -920,13 +943,11 @@ static void sparc_register_emitters(void)
 	set_emitter(op_sparc_fbfcc,     emit_sparc_fbfcc);
 	set_emitter(op_sparc_FrameAddr, emit_sparc_FrameAddr);
 	set_emitter(op_sparc_Mulh,      emit_sparc_Mulh);
-	set_emitter(op_sparc_Save,      emit_sparc_Save);
 	set_emitter(op_sparc_SDiv,      emit_sparc_SDiv);
 	set_emitter(op_sparc_SwitchJmp, emit_sparc_SwitchJmp);
 	set_emitter(op_sparc_UDiv,      emit_sparc_UDiv);
 
 	/* no need to emit anything for the following nodes */
-	set_emitter(op_be_Barrier, emit_nothing);
 	set_emitter(op_be_Keep,    emit_nothing);
 	set_emitter(op_be_Start,   emit_nothing);
 	set_emitter(op_Phi,        emit_nothing);
@@ -944,33 +965,57 @@ static void sparc_emit_node(const ir_node *node)
 		be_dbg_set_dbg_info(get_irn_dbg_info(node));
 		(*func) (node);
 	} else {
-		panic("No emit handler for node %+F (graph %+F)\n",	node,
+		panic("No emit handler for node %+F (graph %+F)\n", node,
 		      current_ir_graph);
 	}
 }
 
+static ir_node *find_next_delay_slot(ir_node *from)
+{
+	ir_node *schedpoint = from;
+	while (!has_delay_slot(schedpoint)) {
+		if (!sched_has_next(schedpoint))
+			return NULL;
+		schedpoint = sched_next(schedpoint);
+	}
+	return schedpoint;
+}
+
 /**
  * Walks over the nodes in a block connected by scheduling edges
  * and emits code for each node.
  */
-static void sparc_gen_block(ir_node *block, void *data)
+static void sparc_emit_block(ir_node *block)
 {
 	ir_node *node;
-	(void) data;
+	ir_node *next_delay_slot;
 
-	if (! is_Block(block))
-		return;
+	assert(is_Block(block));
 
 	be_gas_emit_block_name(block);
 	be_emit_cstring(":\n");
 	be_emit_write_line();
 
+	next_delay_slot = find_next_delay_slot(sched_first(block));
+	if (next_delay_slot != NULL)
+		delay_slot_filler = pick_delay_slot_for(next_delay_slot);
+
 	sched_foreach(block, node) {
+		if (node == delay_slot_filler) {
+			continue;
+		}
+
 		sparc_emit_node(node);
+
+		if (node == next_delay_slot) {
+			assert(delay_slot_filler == NULL);
+			next_delay_slot = find_next_delay_slot(sched_next(node));
+			if (next_delay_slot != NULL)
+				delay_slot_filler = pick_delay_slot_for(next_delay_slot);
+		}
 	}
 }
 
-
 /**
  * Emits code for function start.
  */
@@ -996,11 +1041,6 @@ static void sparc_emit_func_epilog(ir_graph *irg)
 	be_emit_write_line();
 }
 
-/**
- * Block-walker:
- * TODO: Sets labels for control flow nodes (jump target).
- * Links control predecessors to there destination blocks.
- */
 static void sparc_gen_labels(ir_node *block, void *env)
 {
 	ir_node *pred;
@@ -1013,51 +1053,47 @@ static void sparc_gen_labels(ir_node *block, void *env)
 	}
 }
 
-
-/**
- * Main driver
- */
-void sparc_gen_routine(const sparc_code_gen_t *cg, ir_graph *irg)
+void sparc_emit_routine(ir_graph *irg)
 {
-	ir_node **blk_sched;
-	ir_node *last_block = NULL;
-	ir_entity *entity     = get_irg_entity(irg);
-	int i, n;
-	(void) cg;
+	ir_entity  *entity = get_irg_entity(irg);
+	ir_node   **block_schedule;
+	size_t      i;
+	size_t      n;
 
-	be_gas_elf_type_char = '#';
+	be_gas_elf_type_char      = '#';
 	be_gas_object_file_format = OBJECT_FILE_FORMAT_ELF_SPARC;
 
+	heights = heights_new(irg);
+
 	/* register all emitter functions */
 	sparc_register_emitters();
 	be_dbg_method_begin(entity);
 
 	/* create the block schedule. For now, we don't need it earlier. */
-	blk_sched = be_create_block_schedule(irg);
+	block_schedule = be_create_block_schedule(irg);
 
-	// emit function prolog
 	sparc_emit_func_prolog(irg);
-
-	// generate BLOCK labels
 	irg_block_walk_graph(irg, sparc_gen_labels, NULL, NULL);
 
-	// inject block scheduling links & emit code of each block
-	n = ARR_LEN(blk_sched);
-	for (i = 0; i < n;) {
-		ir_node *block, *next_bl;
-
-		block = blk_sched[i];
-		++i;
-		next_bl = i < n ? blk_sched[i] : NULL;
+	/* inject block scheduling links & emit code of each block */
+	n = ARR_LEN(block_schedule);
+	for (i = 0; i < n; ++i) {
+		ir_node *block      = block_schedule[i];
+		ir_node *next_block = i+1 < n ? block_schedule[i+1] : NULL;
+		set_irn_link(block, next_block);
+	}
 
-		/* set here the link. the emitter expects to find the next block here */
-		set_irn_link(block, next_bl);
-		sparc_gen_block(block, last_block);
-		last_block = block;
+	for (i = 0; i < n; ++i) {
+		ir_node *block = block_schedule[i];
+		if (block == get_irg_end_block(irg))
+			continue;
+		sparc_emit_block(block);
 	}
 
-	// emit function epilog
+	/* emit function epilog */
 	sparc_emit_func_epilog(irg);
+
+	heights_free(heights);
 }
 
 void sparc_init_emitter(void)