/*
- * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2010 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
/**
* @file
* @brief emit assembler for a backend graph
+ * @author Hannes Rapp, Matthias Braun
* @version $Id$
*/
#include "config.h"
#include "error.h"
#include "raw_bitset.h"
#include "dbginfo.h"
+#include "heights.h"
#include "../besched.h"
#include "../beblocksched.h"
#include "../begnuas.h"
#include "../be_dbgout.h"
#include "../benode.h"
+#include "../bestack.h"
#include "sparc_emitter.h"
#include "gen_sparc_emitter.h"
#include "sparc_new_nodes.h"
#include "gen_sparc_regalloc_if.h"
-
-#define SNPRINTF_BUF_LEN 128
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
-/**
- * attribute of SAVE node which follows immediatelly after the START node
- * we need this to correct all offsets since SPARC expects
- * some reserved stack space after the stackpointer
- */
-const sparc_save_attr_t *save_attr;
+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.
return reg;
}
-/*************************************************************
- * _ _ __ _ _
- * (_) | | / _| | | | |
- * _ __ _ __ _ _ __ | |_| |_ | |__ ___| |_ __ ___ _ __
- * | '_ \| '__| | '_ \| __| _| | '_ \ / _ \ | '_ \ / _ \ '__|
- * | |_) | | | | | | | |_| | | | | | __/ | |_) | __/ |
- * | .__/|_| |_|_| |_|\__|_| |_| |_|\___|_| .__/ \___|_|
- * | | | |
- * |_| |_|
- *************************************************************/
-
void sparc_emit_immediate(const ir_node *node)
{
- const sparc_attr_t *attr = get_sparc_attr_const(node);
- assert(!(attr->immediate_value < -4096 || attr->immediate_value > 4096));
- be_emit_irprintf("%d", attr->immediate_value);
+ const sparc_attr_t *attr = get_sparc_attr_const(node);
+ ir_entity *entity = attr->immediate_value_entity;
+
+ if (entity == NULL) {
+ int32_t value = attr->immediate_value;
+ assert(sparc_is_value_imm_encodeable(value));
+ be_emit_irprintf("%d", value);
+ } else {
+ be_emit_cstring("%lo(");
+ be_gas_emit_entity(entity);
+ if (attr->immediate_value != 0) {
+ be_emit_irprintf("%+d", attr->immediate_value);
+ }
+ be_emit_char(')');
+ }
+}
+
+void sparc_emit_high_immediate(const ir_node *node)
+{
+ const sparc_attr_t *attr = get_sparc_attr_const(node);
+ ir_entity *entity = attr->immediate_value_entity;
+
+ be_emit_cstring("%hi(");
+ if (entity == NULL) {
+ uint32_t value = (uint32_t) attr->immediate_value;
+ be_emit_irprintf("0x%X", value);
+ } else {
+ be_gas_emit_entity(entity);
+ if (attr->immediate_value != 0) {
+ be_emit_irprintf("%+d", attr->immediate_value);
+ }
+ }
+ be_emit_char(')');
}
void sparc_emit_source_register(const ir_node *node, int pos)
/**
* emit SP offset
*/
-void sparc_emit_offset(const ir_node *node)
+void sparc_emit_offset(const ir_node *node, int offset_node_pos)
{
const sparc_load_store_attr_t *attr = get_sparc_load_store_attr_const(node);
- assert(attr->base.is_load_store);
- if (attr->offset > 0)
- be_emit_irprintf("+%ld", attr->offset);
+ if (attr->is_reg_reg) {
+ assert(!attr->is_frame_entity);
+ assert(attr->base.immediate_value == 0);
+ assert(attr->base.immediate_value_entity == NULL);
+ be_emit_char('+');
+ sparc_emit_source_register(node, offset_node_pos);
+ } else if (attr->is_frame_entity) {
+ int32_t offset = attr->base.immediate_value;
+ if (offset != 0) {
+ assert(sparc_is_value_imm_encodeable(offset));
+ be_emit_irprintf("%+ld", offset);
+ }
+ } else if (attr->base.immediate_value != 0
+ || attr->base.immediate_value_entity != NULL) {
+ be_emit_char('+');
+ sparc_emit_immediate(node);
+ }
}
+void sparc_emit_float_load_store_mode(const ir_node *node)
+{
+ const sparc_load_store_attr_t *attr = get_sparc_load_store_attr_const(node);
+ ir_mode *mode = attr->load_store_mode;
+ int bits = get_mode_size_bits(mode);
+
+ assert(mode_is_float(mode));
+
+ switch (bits) {
+ case 32: return;
+ case 64: be_emit_char('d'); return;
+ case 128: be_emit_char('q'); return;
+ }
+ panic("invalid flaot load/store mode %+F", mode);
+}
/**
* Emit load mode char
void sparc_emit_load_mode(const ir_node *node)
{
const sparc_load_store_attr_t *attr = get_sparc_load_store_attr_const(node);
- ir_mode *mode = attr->load_store_mode;
- int bits = get_mode_size_bits(mode);
- bool is_signed = mode_is_signed(mode);
-
- if (bits == 16) {
- be_emit_string(is_signed ? "sh" : "uh");
- } else if (bits == 8) {
- be_emit_string(is_signed ? "sb" : "ub");
- } else if (bits == 64) {
- be_emit_string("d");
- } else {
- assert(bits == 32);
- }
+ ir_mode *mode = attr->load_store_mode;
+ int bits = get_mode_size_bits(mode);
+ bool is_signed = mode_is_signed(mode);
+
+ if (bits == 16) {
+ be_emit_string(is_signed ? "sh" : "uh");
+ } else if (bits == 8) {
+ be_emit_string(is_signed ? "sb" : "ub");
+ } else if (bits == 64) {
+ be_emit_char('d');
+ } else {
+ assert(bits == 32);
+ }
}
/**
void sparc_emit_store_mode(const ir_node *node)
{
const sparc_load_store_attr_t *attr = get_sparc_load_store_attr_const(node);
- ir_mode *mode = attr->load_store_mode;
- int bits = get_mode_size_bits(mode);
-
- if (bits == 16) {
- be_emit_string("h");
- } else if (bits == 8) {
- be_emit_string("b");
- } else if (bits == 64) {
- be_emit_string("d");
- } else {
- assert(bits == 32);
- }
+ ir_mode *mode = attr->load_store_mode;
+ int bits = get_mode_size_bits(mode);
+
+ if (bits == 16) {
+ be_emit_string("h");
+ } else if (bits == 8) {
+ be_emit_string("b");
+ } else if (bits == 64) {
+ be_emit_char('d');
+ } else {
+ assert(bits == 32);
+ }
}
/**
*/
void sparc_emit_mode_sign_prefix(const ir_node *node)
{
- ir_mode *mode = get_irn_mode(node);
- bool is_signed = mode_is_signed(mode);
- be_emit_string(is_signed ? "s" : "u");
+ ir_mode *mode = get_irn_mode(node);
+ bool is_signed = mode_is_signed(mode);
+ be_emit_string(is_signed ? "s" : "u");
}
-/**
- * emit FP load mode char
- */
-void sparc_emit_fp_load_mode(const ir_node *node)
+static void emit_fp_suffix(const ir_mode *mode)
{
- const sparc_load_store_attr_t *attr = get_sparc_load_store_attr_const(node);
- ir_mode *mode = attr->load_store_mode;
- int bits = get_mode_size_bits(mode);
-
- assert(mode_is_float(mode));
-
- if (bits == 32) {
- be_emit_string("f");
- } else if (bits == 64) {
- be_emit_string("df");
- } else {
- panic("FP load mode > 64bits not implemented yet");
- }
+ unsigned bits = get_mode_size_bits(mode);
+ assert(mode_is_float(mode));
+
+ if (bits == 32) {
+ be_emit_char('s');
+ } else if (bits == 64) {
+ be_emit_char('d');
+ } else if (bits == 128) {
+ be_emit_char('q');
+ } else {
+ panic("invalid FP mode");
+ }
}
-/**
- * emit FP store mode char
- */
-void sparc_emit_fp_store_mode(const ir_node *node)
+void sparc_emit_fp_conv_source(const ir_node *node)
{
- const sparc_load_store_attr_t *attr = get_sparc_load_store_attr_const(node);
- ir_mode *mode = attr->load_store_mode;
- int bits = get_mode_size_bits(mode);
-
- assert(mode_is_float(mode));
+ const sparc_fp_conv_attr_t *attr = get_sparc_fp_conv_attr_const(node);
+ emit_fp_suffix(attr->src_mode);
+}
- if (bits == 32) {
- be_emit_string("f");
- } else if (bits == 64) {
- be_emit_string("df");
- } else {
- panic("FP store mode > 64bits not implemented yet");
- }
+void sparc_emit_fp_conv_destination(const ir_node *node)
+{
+ const sparc_fp_conv_attr_t *attr = get_sparc_fp_conv_attr_const(node);
+ emit_fp_suffix(attr->dest_mode);
}
/**
*/
void sparc_emit_fp_mode_suffix(const ir_node *node)
{
- ir_mode *mode = get_irn_mode(node);
- int bits = get_mode_size_bits(mode);
-
- assert(mode_is_float(mode));
+ const sparc_fp_attr_t *attr = get_sparc_fp_attr_const(node);
+ emit_fp_suffix(attr->fp_mode);
+}
- if (bits == 32) {
- be_emit_string("s");
- } else if (bits == 64) {
- be_emit_string("d");
- } else {
- panic("FP mode > 64bits not implemented yet");
- }
+static ir_node *get_jump_target(const ir_node *jump)
+{
+ return (ir_node*)get_irn_link(jump);
}
/**
*/
static void sparc_emit_cfop_target(const ir_node *node)
{
- ir_node *block = get_irn_link(node);
+ ir_node *block = get_jump_target(node);
be_gas_emit_block_name(block);
}
-/**
- * Emit single entity
- */
-static void sparc_emit_entity(ir_entity *entity)
+static int get_sparc_Call_dest_addr_pos(const ir_node *node)
{
- be_gas_emit_entity(entity);
+ return get_irn_arity(node)-1;
}
-/***********************************************************************************
- * _ __ _
- * (_) / _| | |
- * _ __ ___ __ _ _ _ __ | |_ _ __ __ _ _ __ ___ _____ _____ _ __| | __
- * | '_ ` _ \ / _` | | '_ \ | _| '__/ _` | '_ ` _ \ / _ \ \ /\ / / _ \| '__| |/ /
- * | | | | | | (_| | | | | | | | | | | (_| | | | | | | __/\ V V / (_) | | | <
- * |_| |_| |_|\__,_|_|_| |_| |_| |_| \__,_|_| |_| |_|\___| \_/\_/ \___/|_| |_|\_\
- *
- ***********************************************************************************/
-
-
-/**
- * Emits code for stack space management
- */
-static void emit_be_IncSP(const ir_node *irn)
+static bool ba_is_fallthrough(const ir_node *node)
{
- int offs = -be_get_IncSP_offset(irn);
+ ir_node *block = get_nodes_block(node);
+ ir_node *next_block = (ir_node*)get_irn_link(block);
+ return get_irn_link(node) == next_block;
+}
- if (offs == 0)
- return;
+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);
- /* SPARC stack grows downwards */
- if (offs < 0) {
- be_emit_cstring("\tsub ");
- offs = -offs;
- } else {
- be_emit_cstring("\tadd ");
+ 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;
- sparc_emit_source_register(irn, 0);
- be_emit_irprintf(", %d", offs);
- be_emit_cstring(", ");
- sparc_emit_dest_register(irn, 0);
- be_emit_finish_line_gas(irn);
+ return be_is_Keep(node) || be_is_Start(node) || is_Phi(node);
}
-/**
- * emits code for save instruction with min. required stack space
- */
-static void emit_sparc_Save(const ir_node *irn)
+static bool has_delay_slot(const ir_node *node)
{
- 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);
-}
+ if (is_sparc_Ba(node) && ba_is_fallthrough(node))
+ return false;
-/**
- * emits code to load hi 22 bit of a constant
- */
-static void emit_sparc_HiImm(const ir_node *irn)
-{
- const sparc_attr_t *attr = get_sparc_attr_const(irn);
- be_emit_cstring("\tsethi ");
- be_emit_irprintf("%%hi(%d), ", attr->immediate_value);
- sparc_emit_dest_register(irn, 0);
- be_emit_finish_line_gas(irn);
+ 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);
}
-/**
- * emits code to load lo 10bits of a constant
+/** 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 void emit_sparc_LoImm(const ir_node *irn)
+static bool emits_multiple_instructions(const ir_node *node)
{
- const sparc_attr_t *attr = get_sparc_attr_const(irn);
- be_emit_cstring("\tor ");
- sparc_emit_source_register(irn, 0);
- be_emit_irprintf(", %%lo(%d), ", attr->immediate_value);
- sparc_emit_dest_register(irn, 0);
- be_emit_finish_line_gas(irn);
+ 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 code for div with the correct sign prefix
+ * search for an instruction that can fill the delay slot of @p node
*/
-static void emit_sparc_Div(const ir_node *irn)
+static const ir_node *pick_delay_slot_for(const ir_node *node)
{
- be_emit_cstring("\t");
- sparc_emit_mode_sign_prefix(irn);
- be_emit_cstring("div ");
+ 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));
+ }
- sparc_emit_source_register(irn, 0);
- be_emit_cstring(", ");
- sparc_emit_reg_or_imm(irn, 1);
- be_emit_cstring(", ");
- sparc_emit_dest_register(irn, 0);
- be_emit_finish_line_gas(irn);
+ /* 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;
}
/**
- * emit code for mul with the correct sign prefix
+ * Emits code for stack space management
*/
-static void emit_sparc_Mul(const ir_node *irn)
+static void emit_be_IncSP(const ir_node *irn)
{
- be_emit_cstring("\t");
- sparc_emit_mode_sign_prefix(irn);
- be_emit_cstring("mul ");
+ int offset = be_get_IncSP_offset(irn);
+
+ if (offset == 0)
+ return;
+
+ /* SPARC stack grows downwards */
+ if (offset < 0) {
+ be_emit_cstring("\tsub ");
+ offset = -offset;
+ } else {
+ be_emit_cstring("\tadd ");
+ }
sparc_emit_source_register(irn, 0);
- be_emit_cstring(", ");
- sparc_emit_reg_or_imm(irn, 1);
+ be_emit_irprintf(", %d", -offset);
be_emit_cstring(", ");
sparc_emit_dest_register(irn, 0);
be_emit_finish_line_gas(irn);
// our result is in the y register now
// we just copy it to the assigned target reg
- be_emit_cstring("\tmov ");
- be_emit_char('%');
- be_emit_string(arch_register_get_name(&sparc_flags_regs[REG_Y]));
- be_emit_cstring(", ");
+ be_emit_cstring("\tmov %y, ");
sparc_emit_dest_register(irn, 0);
be_emit_finish_line_gas(irn);
}
-/**
- * Emits code for return node
- */
-static void emit_be_Return(const ir_node *irn)
+static void fill_delay_slot(void)
{
- 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);
+ 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)
+{
+ /* can we get the delay count of the wr instruction somewhere? */
+ unsigned wry_delay_count = 3;
+ unsigned i;
+
+ be_emit_cstring("\twr ");
+ sparc_emit_source_register(node, 0);
+ be_emit_cstring(", 0, %y");
+ be_emit_finish_line_gas(node);
+
+ for (i = 0; i < wry_delay_count; ++i) {
+ fill_delay_slot();
+ }
+
+ be_emit_irprintf("\t%s ", is_signed ? "sdiv" : "udiv");
+ sparc_emit_source_register(node, 1);
+ be_emit_cstring(", ");
+ sparc_emit_reg_or_imm(node, 2);
+ be_emit_cstring(", ");
+ sparc_emit_dest_register(node, 0);
+ be_emit_finish_line_gas(node);
+}
+
+static void emit_sparc_SDiv(const ir_node *node)
+{
+ emit_sparc_Div(node, true);
+}
+
+static void emit_sparc_UDiv(const ir_node *node)
+{
+ emit_sparc_Div(node, false);
}
/**
* Emits code for Call node
*/
-static void emit_be_Call(const ir_node *irn)
+static void emit_sparc_Call(const ir_node *node)
{
- ir_entity *entity = be_Call_get_entity(irn);
+ const sparc_attr_t *attr = get_sparc_attr_const(node);
+ ir_entity *entity = attr->immediate_value_entity;
+ be_emit_cstring("\tcall ");
if (entity != NULL) {
- be_emit_cstring("\tcall ");
- 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");
- be_emit_finish_line_gas(irn);
- be_emit_cstring("\tnop");
- be_emit_pad_comment();
- be_emit_cstring("/* TODO: use delay slot */\n");
} else {
- be_emit_cstring("\tnop\n");
- be_emit_pad_comment();
- be_emit_cstring("/* TODO: Entity == NULL */\n");
- be_emit_finish_line_gas(irn);
+ int dest_addr = get_sparc_Call_dest_addr_pos(node);
+ sparc_emit_source_register(node, dest_addr);
}
+ be_emit_finish_line_gas(node);
+
+ fill_delay_slot();
}
/**
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;
int memperm_arity;
int sp_change = 0;
+ ir_graph *irg = get_irn_irg(node);
+ be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
+
+ /* this implementation only works with frame pointers currently */
+ assert(layout->sp_relative == false);
/* TODO: this implementation is slower than necessary.
The longterm goal is however to avoid the memperm node completely */
if (memperm_arity > 8)
panic("memperm with more than 8 inputs not supported yet");
+ be_emit_irprintf("\tsub %%sp, %d, %%sp", memperm_arity*4);
+ be_emit_finish_line_gas(node);
+
for (i = 0; i < memperm_arity; ++i) {
- int offset;
ir_entity *entity = be_get_MemPerm_in_entity(node, i);
+ int offset = be_get_stack_entity_offset(layout, entity, 0);
/* spill register */
- sp_change += 4;
- be_emit_irprintf("\tst %%l%d, [%%sp-%d]", i, sp_change);
+ be_emit_irprintf("\tst %%l%d, [%%sp%+d]", i, sp_change + SPARC_MIN_STACKSIZE);
be_emit_finish_line_gas(node);
/* load from entity */
- offset = get_entity_offset(entity) + sp_change;
- be_emit_irprintf("\tld [%%sp+%d], %%l%d", offset, i);
+ be_emit_irprintf("\tld [%%fp%+d], %%l%d", offset, i);
be_emit_finish_line_gas(node);
+ sp_change += 4;
}
for (i = memperm_arity-1; i >= 0; --i) {
- int offset;
ir_entity *entity = be_get_MemPerm_out_entity(node, i);
+ int offset = be_get_stack_entity_offset(layout, entity, 0);
+
+ sp_change -= 4;
/* store to new entity */
- offset = get_entity_offset(entity) + sp_change;
- be_emit_irprintf("\tst %%l%d, [%%sp+%d]", i, offset);
+ be_emit_irprintf("\tst %%l%d, [%%fp%+d]", i, offset);
be_emit_finish_line_gas(node);
/* restore register */
- be_emit_irprintf("\tld [%%sp-%d], %%l%d", sp_change, i);
- sp_change -= 4;
+ be_emit_irprintf("\tld [%%sp%+d], %%l%d", sp_change + SPARC_MIN_STACKSIZE, i);
be_emit_finish_line_gas(node);
}
+
+ be_emit_irprintf("\tadd %%sp, %d, %%sp", memperm_arity*4);
+ be_emit_finish_line_gas(node);
+
assert(sp_change == 0);
}
-/**
- * Emit a SymConst.
- */
-static void emit_sparc_SymConst(const ir_node *irn)
+static void emit_be_Return(const ir_node *node)
{
- const sparc_symconst_attr_t *attr = get_sparc_symconst_attr_const(irn);
-
- //sethi %hi(const32),%reg
- //or %reg,%lo(const32),%reg
-
- be_emit_cstring("\tsethi %hi(");
- be_gas_emit_entity(attr->entity);
- be_emit_cstring("), ");
- sparc_emit_dest_register(irn, 0);
- be_emit_cstring("\n ");
-
- // TODO: could be combined with the following load/store instruction
- be_emit_cstring("\tor ");
- sparc_emit_dest_register(irn, 0);
- be_emit_cstring(", %lo(");
- be_gas_emit_entity(attr->entity);
- be_emit_cstring("), ");
- sparc_emit_dest_register(irn, 0);
- be_emit_finish_line_gas(irn);
+ 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();
}
-/**
- * Emits code for FrameAddr fix
- */
-static void emit_sparc_FrameAddr(const ir_node *irn)
+static void emit_sparc_FrameAddr(const ir_node *node)
{
- const sparc_symconst_attr_t *attr = get_irn_generic_attr_const(irn);
+ 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->fp_offset >= 0) {
+ if (offset < 0) {
be_emit_cstring("\tadd ");
- sparc_emit_source_register(irn, 0);
+ sparc_emit_source_register(node, 0);
be_emit_cstring(", ");
- be_emit_irprintf("%ld", attr->fp_offset + save_attr->initial_stacksize);
+ assert(sparc_is_value_imm_encodeable(offset));
+ be_emit_irprintf("%ld", offset);
} else {
be_emit_cstring("\tsub ");
- sparc_emit_source_register(irn, 0);
+ sparc_emit_source_register(node, 0);
be_emit_cstring(", ");
- be_emit_irprintf("%ld", -attr->fp_offset);
+ assert(sparc_is_value_imm_encodeable(-offset));
+ be_emit_irprintf("%ld", -offset);
}
be_emit_cstring(", ");
- sparc_emit_dest_register(irn, 0);
- be_emit_finish_line_gas(irn);
+ sparc_emit_dest_register(node, 0);
+ be_emit_finish_line_gas(node);
}
+static const char *get_icc_unsigned(ir_relation relation)
+{
+ 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");
+ }
+}
-/**
- * Emits code for Branch
- */
-static void emit_sparc_Branch(const ir_node *irn)
+static const char *get_icc_signed(ir_relation relation)
{
+ 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(ir_relation relation)
+{
+ 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 relation");
+}
+
+typedef const char* (*get_cc_func)(ir_relation relation);
+
+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);
+ ir_relation relation = attr->relation;
+ const ir_node *proj_true = NULL;
+ const ir_node *proj_false = NULL;
const ir_edge_t *edge;
- const ir_node *proj_true = NULL;
- const ir_node *proj_false = NULL;
- const ir_node *block;
- const ir_node *next_block;
- ir_node *op1 = get_irn_n(irn, 0);
- const char *suffix;
- int proj_num = get_sparc_jmp_cond_proj_num(irn);
- const sparc_cmp_attr_t *cmp_attr = get_irn_generic_attr_const(op1);
- // bool is_signed = !cmp_attr->is_unsigned;
-
- assert(is_sparc_Cmp(op1) || is_sparc_Tst(op1));
-
- foreach_out_edge(irn, edge) {
+ const ir_node *block;
+ const ir_node *next_block;
+
+ foreach_out_edge(node, edge) {
ir_node *proj = get_edge_src_irn(edge);
long nr = get_Proj_proj(proj);
if (nr == pn_Cond_true) {
}
}
- if (cmp_attr->ins_permuted) {
- proj_num = get_mirrored_pnc(proj_num);
- }
-
/* for now, the code works for scheduled and non-schedules blocks */
- block = get_nodes_block(irn);
+ block = get_nodes_block(node);
/* we have a block schedule */
- next_block = get_irn_link(block);
-
- assert(proj_num != pn_Cmp_False);
- assert(proj_num != pn_Cmp_True);
+ 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 */
proj_true = proj_false;
proj_false = t;
- proj_num = get_negated_pnc(proj_num, mode_Iu);
- }
-
-
- switch (proj_num) {
- case pn_Cmp_Eq: suffix = "e"; break;
- case pn_Cmp_Lt: suffix = "l"; break;
- case pn_Cmp_Le: suffix = "le"; break;
- case pn_Cmp_Gt: suffix = "g"; break;
- case pn_Cmp_Ge: suffix = "ge"; break;
- case pn_Cmp_Lg: suffix = "ne"; break;
- case pn_Cmp_Leg: suffix = "a"; break;
- default: panic("Cmp has unsupported pnc");
+ relation = get_negated_relation(relation);
}
/* emit the true proj */
- be_emit_irprintf("\tb%s ", suffix);
+ be_emit_cstring("\t");
+ be_emit_string(get_cc(relation));
+ be_emit_char(' ');
sparc_emit_cfop_target(proj_true);
be_emit_finish_line_gas(proj_true);
- be_emit_cstring("\tnop");
- be_emit_pad_comment();
- be_emit_cstring("/* TODO: use delay slot */\n");
+ fill_delay_slot();
if (get_irn_link(proj_false) == next_block) {
- be_emit_cstring("\t/* false-fallthrough to ");
+ be_emit_cstring("\t/* fallthrough to ");
sparc_emit_cfop_target(proj_false);
be_emit_cstring(" */");
be_emit_finish_line_gas(proj_false);
be_emit_cstring("\tba ");
sparc_emit_cfop_target(proj_false);
be_emit_finish_line_gas(proj_false);
- be_emit_cstring("\tnop\t\t/* TODO: use delay slot */\n");
- be_emit_finish_line_gas(proj_false);
+ fill_delay_slot();
}
}
-/**
- * emit Jmp (which actually is a branch always (ba) instruction)
- */
-static void emit_sparc_Jmp(const ir_node *node)
+static void emit_sparc_Bicc(const ir_node *node)
{
- ir_node *block, *next_block;
+ const sparc_jmp_cond_attr_t *attr = get_sparc_jmp_cond_attr_const(node);
+ bool is_unsigned = attr->is_unsigned;
+ emit_sparc_branch(node, is_unsigned ? get_icc_unsigned : get_icc_signed);
+}
- /* for now, the code works for scheduled and non-schedules blocks */
- block = get_nodes_block(node);
+static void emit_sparc_fbfcc(const ir_node *node)
+{
+ emit_sparc_branch(node, get_fcc);
+}
- /* we have a block schedule */
- next_block = get_irn_link(block);
- if (get_irn_link(node) != next_block) {
- be_emit_cstring("\tba ");
- sparc_emit_cfop_target(node);
- be_emit_finish_line_gas(node);
- be_emit_cstring("\tnop\t\t/* TODO: use delay slot */\n");
- } else {
+static void emit_sparc_Ba(const ir_node *node)
+{
+ 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();
}
be_emit_finish_line_gas(node);
}
-/**
- * emit copy node
- */
-static void emit_be_Copy(const ir_node *irn)
+static void emit_jump_table(const ir_node *node)
{
- ir_mode *mode = get_irn_mode(irn);
+ const sparc_switch_jmp_attr_t *attr = get_sparc_switch_jmp_attr_const(node);
+ long switch_max = LONG_MIN;
+ long default_pn = attr->default_proj_num;
+ ir_entity *entity = attr->jump_table;
+ ir_node *default_block = NULL;
+ unsigned long length;
+ const ir_edge_t *edge;
+ unsigned i;
+ ir_node **table;
- if (get_in_reg(irn, 0) == get_out_reg(irn, 0)) {
- /* omitted Copy */
- return;
+ /* 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_max = pn > switch_max ? pn : switch_max;
+ }
+ }
+ assert(switch_max > LONG_MIN);
+
+ length = (unsigned long) switch_max + 1;
+ /* the 16000 isn't a real limit of the architecture. But should protect us
+ * from seamingly endless compiler runs */
+ if (length > 16000) {
+ /* switch lowerer should have broken this monster to pieces... */
+ panic("too large switch encountered");
}
+ 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] = 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);
+}
+
+static void emit_fmov(const ir_node *node, const arch_register_t *src_reg,
+ const arch_register_t *dst_reg)
+{
+ be_emit_cstring("\tfmovs %");
+ be_emit_string(arch_register_get_name(src_reg));
+ 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->global_index;
+ assert(reg == &sparc_registers[index]);
+ index++;
+ assert(index - REG_F0 < N_sparc_fp_REGS);
+ return &sparc_registers[index];
+}
+
+static void emit_be_Copy(const ir_node *node)
+{
+ ir_mode *mode = get_irn_mode(node);
+ const arch_register_t *src_reg = get_in_reg(node, 0);
+ const arch_register_t *dst_reg = get_out_reg(node, 0);
+
+ if (src_reg == dst_reg)
+ return;
+
if (mode_is_float(mode)) {
- panic("emit_be_Copy: move not supported for FP");
+ unsigned bits = get_mode_size_bits(mode);
+ int n = bits > 32 ? bits > 64 ? 3 : 1 : 0;
+ int i;
+ emit_fmov(node, src_reg, dst_reg);
+ for (i = 0; i < n; ++i) {
+ src_reg = get_next_fp_reg(src_reg);
+ dst_reg = get_next_fp_reg(dst_reg);
+ emit_fmov(node, src_reg, dst_reg);
+ }
} else if (mode_is_data(mode)) {
be_emit_cstring("\tmov ");
- sparc_emit_source_register(irn, 0);
+ sparc_emit_source_register(node, 0);
be_emit_cstring(", ");
- sparc_emit_dest_register(irn, 0);
- be_emit_finish_line_gas(irn);
+ sparc_emit_dest_register(node, 0);
+ be_emit_finish_line_gas(node);
} else {
- panic("emit_be_Copy: move not supported for this mode");
+ panic("emit_be_Copy: invalid mode");
}
}
-
-/**
- * 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;
*/
static void sparc_register_emitters(void)
{
-
/* first clear the generic function pointer for all ops */
clear_irp_opcodes_generic_func();
-
/* register all emitter functions defined in spec */
sparc_register_spec_emitters();
/* custom emitter */
- set_emitter(op_be_IncSP, emit_be_IncSP);
- set_emitter(op_be_Return, emit_be_Return);
- set_emitter(op_be_Call, emit_be_Call);
- set_emitter(op_sparc_FrameAddr, emit_sparc_FrameAddr);
- set_emitter(op_sparc_Branch, emit_sparc_Branch);
- set_emitter(op_sparc_SymConst, emit_sparc_SymConst);
- set_emitter(op_sparc_Jmp, emit_sparc_Jmp);
- set_emitter(op_sparc_Save, emit_sparc_Save);
-
- set_emitter(op_sparc_HiImm, emit_sparc_HiImm);
- set_emitter(op_sparc_LoImm, emit_sparc_LoImm);
- set_emitter(op_sparc_Div, emit_sparc_Div);
- set_emitter(op_sparc_Mul, emit_sparc_Mul);
- set_emitter(op_sparc_Mulh, emit_sparc_Mulh);
-
- set_emitter(op_be_Copy, emit_be_Copy);
- set_emitter(op_be_CopyKeep, emit_be_Copy);
-
- set_emitter(op_be_Perm, emit_be_Perm);
- set_emitter(op_be_MemPerm, emit_be_MemPerm);
-
-/*
- set_emitter(op_arm_B, emit_arm_B);
- set_emitter(op_arm_CopyB, emit_arm_CopyB);
- set_emitter(op_arm_fpaConst, emit_arm_fpaConst);
- set_emitter(op_arm_fpaDbl2GP, emit_arm_fpaDbl2GP);
- set_emitter(op_arm_LdTls, emit_arm_LdTls);
- set_emitter(op_arm_SwitchJmp, emit_arm_SwitchJmp);
-
-*/
- /* no need to emit anything for the following nodes */
- set_emitter(op_Phi, emit_nothing);
- set_emitter(op_be_Keep, emit_nothing);
- set_emitter(op_be_Start, emit_nothing);
- set_emitter(op_be_Barrier, emit_nothing);
-
+ set_emitter(op_be_Copy, emit_be_Copy);
+ set_emitter(op_be_CopyKeep, emit_be_Copy);
+ set_emitter(op_be_IncSP, emit_be_IncSP);
+ set_emitter(op_be_MemPerm, emit_be_MemPerm);
+ set_emitter(op_be_Perm, emit_be_Perm);
+ set_emitter(op_be_Return, emit_be_Return);
+ set_emitter(op_sparc_Ba, emit_sparc_Ba);
+ set_emitter(op_sparc_Bicc, emit_sparc_Bicc);
+ set_emitter(op_sparc_Call, emit_sparc_Call);
+ 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_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_Keep, emit_nothing);
+ set_emitter(op_be_Start, emit_nothing);
+ set_emitter(op_Phi, emit_nothing);
}
/**
*/
static void sparc_emit_node(const ir_node *node)
{
- ir_op *op = get_irn_op(node);
+ ir_op *op = get_irn_op(node);
if (op->ops.generic) {
emit_func func = (emit_func) op->ops.generic;
be_dbg_set_dbg_info(get_irn_dbg_info(node));
(*func) (node);
} else {
- panic("Error: No emit handler for node %+F (graph %+F)\n",
- node, current_ir_graph);
+ 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.
*/
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;
}
}
-
-/**
- * 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;
-
- /* 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;
+ /* 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);
}
+ for (i = 0; i < n; ++i) {
+ ir_node *block = block_schedule[i];
+ if (block == get_irg_end_block(irg))
+ continue;
+ sparc_emit_block(block);
+ }
- //irg_walk_blkwise_graph(irg, NULL, sparc_gen_block, NULL);
-
- // emit function epilog
+ /* emit function epilog */
sparc_emit_func_epilog(irg);
+
+ heights_free(heights);
}
void sparc_init_emitter(void)