/**
* @file
* @brief emit assembler for a backend graph
+ * @author Hannes Rapp, Matthias Braun
* @version $Id$
*/
#include "config.h"
#include <limits.h>
+#include "bitfiddle.h"
#include "xmalloc.h"
#include "tv.h"
#include "iredges.h"
#include "dbginfo.h"
#include "heights.h"
-#include "../besched.h"
-#include "../beblocksched.h"
-#include "../beirg.h"
-#include "../begnuas.h"
-#include "../be_dbgout.h"
-#include "../benode.h"
-#include "../bestack.h"
+#include "besched.h"
+#include "beblocksched.h"
+#include "beirg.h"
+#include "begnuas.h"
+#include "be_dbgout.h"
+#include "benode.h"
+#include "bestack.h"
+#include "bepeephole.h"
#include "sparc_emitter.h"
#include "gen_sparc_emitter.h"
the next delay slot */
static void sparc_emit_node(const ir_node *node);
+static bool emitting_delay_slot;
-/**
- * Returns the register at in position pos.
- */
-static const arch_register_t *get_in_reg(const ir_node *node, int pos)
+void sparc_emit_indent(void)
{
- ir_node *op;
- const arch_register_t *reg = NULL;
-
- assert(get_irn_arity(node) > pos && "Invalid IN position");
-
- /* The out register of the operator at position pos is the
- in register we need. */
- op = get_irn_n(node, pos);
-
- reg = arch_get_irn_register(op);
-
- assert(reg && "no in register found");
- return reg;
-}
-
-/**
- * Returns the register at out position pos.
- */
-static const arch_register_t *get_out_reg(const ir_node *node, int pos)
-{
- ir_node *proj;
- const arch_register_t *reg = NULL;
-
- /* 1st case: irn is not of mode_T, so it has only */
- /* one OUT register -> good */
- /* 2nd case: irn is of mode_T -> collect all Projs and ask the */
- /* Proj with the corresponding projnum for the register */
-
- if (get_irn_mode(node) != mode_T) {
- reg = arch_get_irn_register(node);
- } else if (is_sparc_irn(node)) {
- reg = arch_irn_get_register(node, pos);
- } else {
- const ir_edge_t *edge;
-
- foreach_out_edge(node, edge) {
- proj = get_edge_src_irn(edge);
- assert(is_Proj(proj) && "non-Proj from mode_T node");
- if (get_Proj_proj(proj) == pos) {
- reg = arch_get_irn_register(proj);
- break;
- }
- }
- }
-
- assert(reg && "no out register found");
- return reg;
-}
-
-static bool is_valid_immediate(int32_t value)
-{
- return -4096 <= value && value < 4096;
+ be_emit_char('\t');
+ if (emitting_delay_slot)
+ be_emit_char(' ');
}
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(");
+ if (get_entity_owner(entity) == get_tls_type()) {
+ be_emit_cstring("%tle_lox10(");
+ } else {
+ be_emit_cstring("%lo(");
+ }
be_gas_emit_entity(entity);
if (attr->immediate_value != 0) {
be_emit_irprintf("%+d", attr->immediate_value);
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);
+ be_emit_irprintf("%%hi(0x%X)", value);
} else {
+ if (get_entity_owner(entity) == get_tls_type()) {
+ be_emit_cstring("%tle_hix22(");
+ } else {
+ be_emit_cstring("%hi(");
+ }
be_gas_emit_entity(entity);
if (attr->immediate_value != 0) {
be_emit_irprintf("%+d", attr->immediate_value);
}
+ be_emit_char(')');
}
- be_emit_char(')');
}
void sparc_emit_source_register(const ir_node *node, int pos)
{
- const arch_register_t *reg = get_in_reg(node, pos);
+ const arch_register_t *reg = arch_get_irn_register_in(node, pos);
be_emit_char('%');
be_emit_string(arch_register_get_name(reg));
}
void sparc_emit_dest_register(const ir_node *node, int pos)
{
- const arch_register_t *reg = get_out_reg(node, pos);
+ const arch_register_t *reg = arch_get_irn_register_out(node, pos);
be_emit_char('%');
be_emit_string(arch_register_get_name(reg));
}
*/
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_get_irn_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_registers[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
*/
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
}
}
+void sparc_emit_source_reg_and_offset(const ir_node *node, int regpos,
+ int offpos)
+{
+ const arch_register_t *reg = arch_get_irn_register_in(node, regpos);
+ const sparc_load_store_attr_t *attr;
+
+#ifdef DEBUG_libfirm
+ if (reg == &sparc_registers[REG_SP]) {
+ attr = get_sparc_load_store_attr_const(node);
+ if (!attr->is_reg_reg
+ && attr->base.immediate_value < SPARC_SAVE_AREA_SIZE) {
+
+ ir_fprintf(stderr, "warning: emitting stack pointer relative load/store with offset < %d\n", SPARC_SAVE_AREA_SIZE);
+ }
+ }
+#endif
+
+ sparc_emit_source_register(node, regpos);
+ sparc_emit_offset(node, offpos);
+}
+
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);
case 64: be_emit_char('d'); return;
case 128: be_emit_char('q'); return;
}
- panic("invalid flaot load/store mode %+F", mode);
+ panic("invalid float load/store mode %+F", mode);
}
/**
}
}
-/**
- * emit integer signed/unsigned prefix char
- */
-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");
-}
-
static void emit_fp_suffix(const ir_mode *mode)
{
unsigned bits = get_mode_size_bits(mode);
be_gas_emit_block_name(block);
}
+/**
+ * returns true if a sparc_call calls a register and not an immediate
+ */
+static bool is_sparc_reg_call(const ir_node *node)
+{
+ const sparc_attr_t *attr = get_sparc_attr_const(node);
+ return attr->immediate_value_entity == NULL;
+}
+
static int get_sparc_Call_dest_addr_pos(const ir_node *node)
{
+ assert(is_sparc_reg_call(node));
return get_irn_arity(node)-1;
}
{
/* 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);
+ const arch_register_t *src_reg = arch_get_irn_register_in(node, 0);
+ const arch_register_t *dest_reg = arch_get_irn_register_out(node, 0);
if (src_reg == dest_reg)
return true;
if (is_sparc_Ba(node) && ba_is_fallthrough(node))
return true;
- return be_is_Keep(node) || be_is_Barrier(node) || be_is_Start(node)
- || is_Phi(node);
+ 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;
+ if (is_sparc_Ba(node)) {
+ return !ba_is_fallthrough(node);
+ }
- 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);
+ return arch_get_irn_flags(node) & sparc_arch_irn_flag_has_delay_slot;
}
/** returns true if the emitter for this sparc node can produce more than one
if (has_delay_slot(node))
return true;
- return is_sparc_Mulh(node) || is_sparc_SDiv(node) || is_sparc_UDiv(node)
+ if (is_sparc_Call(node)) {
+ return arch_get_irn_flags(node) & sparc_arch_irn_flag_aggregate_return;
+ }
+
+ return is_sparc_SMulh(node) || is_sparc_UMulh(node)
+ || is_sparc_SDiv(node) || is_sparc_UDiv(node)
|| be_is_MemPerm(node) || be_is_Perm(node);
}
+static bool uses_reg(const ir_node *node, const arch_register_t *reg)
+{
+ int arity = get_irn_arity(node);
+ int i;
+
+ for (i = 0; i < arity; ++i) {
+ const arch_register_t *in_reg = arch_get_irn_register_in(node, i);
+ if (reg == in_reg)
+ return true;
+ }
+ return false;
+}
+
+static bool writes_reg(const ir_node *node, const arch_register_t *reg)
+{
+ unsigned n_outs = arch_get_irn_n_outs(node);
+ unsigned o;
+ for (o = 0; o < n_outs; ++o) {
+ const arch_register_t *out_reg = arch_get_irn_register_out(node, o);
+ if (out_reg == reg)
+ return true;
+ }
+ return false;
+}
+
+static bool can_move_into_delayslot(const ir_node *node, const ir_node *to)
+{
+ if (!be_can_move_before(node, to))
+ return false;
+
+ if (is_sparc_Call(to)) {
+ ir_node *check;
+ /** all deps are used after the delay slot so, we're fine */
+ if (!is_sparc_reg_call(to))
+ return true;
+
+ check = get_irn_n(to, get_sparc_Call_dest_addr_pos(to));
+ if (skip_Proj(check) == node)
+ return false;
+
+ /* 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. */
+ if (uses_reg(node, &sparc_registers[REG_O7]))
+ return false;
+ return true;
+ } else if (is_sparc_Return(to)) {
+ /* return uses the value of %o7, all other values are not
+ * immediately used */
+ if (writes_reg(node, &sparc_registers[REG_O7]))
+ return false;
+ return true;
+ } else {
+ /* the node must not use our computed values */
+ int arity = get_irn_arity(to);
+ int i;
+ for (i = 0; i < arity; ++i) {
+ ir_node *in = get_irn_n(to, i);
+ if (skip_Proj(in) == node)
+ return false;
+ }
+ return true;
+ }
+}
+
/**
* search for an instruction that can fill the delay slot of @p node
*/
static const ir_node *pick_delay_slot_for(const ir_node *node)
{
- 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;
+ static const unsigned PICK_DELAY_SLOT_MAX_DISTANCE = 10;
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 (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))
+ if (!can_move_into_delayslot(schedpoint, node))
continue;
/* found something */
*/
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)
+ if (offset == 0)
return;
/* SPARC stack grows downwards */
- if (offs < 0) {
- be_emit_cstring("\tsub ");
- offs = -offs;
+ sparc_emit_indent();
+ if (offset < 0) {
+ be_emit_cstring("sub ");
+ offset = -offset;
} else {
- be_emit_cstring("\tadd ");
+ be_emit_cstring("add ");
}
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);
*/
static void emit_sparc_Mulh(const ir_node *irn)
{
- be_emit_cstring("\t");
- sparc_emit_mode_sign_prefix(irn);
+ sparc_emit_indent();
+ if (is_sparc_UMulh(irn)) {
+ be_emit_char('u');
+ } else {
+ assert(is_sparc_SMulh(irn));
+ be_emit_char('s');
+ }
be_emit_cstring("mul ");
sparc_emit_source_register(irn, 0);
// our result is in the y register now
// we just copy it to the assigned target reg
- be_emit_cstring("\tmov %y, ");
+ sparc_emit_indent();
+ be_emit_cstring("mov %y, ");
sparc_emit_dest_register(irn, 0);
be_emit_finish_line_gas(irn);
}
static void fill_delay_slot(void)
{
+ emitting_delay_slot = true;
if (delay_slot_filler != NULL) {
sparc_emit_node(delay_slot_filler);
delay_slot_filler = NULL;
} else {
- be_emit_cstring("\tnop\n");
+ sparc_emit_indent();
+ be_emit_cstring("nop\n");
be_emit_write_line();
}
+ emitting_delay_slot = false;
}
static void emit_sparc_Div(const ir_node *node, bool is_signed)
unsigned wry_delay_count = 3;
unsigned i;
- be_emit_cstring("\twr ");
+ sparc_emit_indent();
+ be_emit_cstring("wr ");
sparc_emit_source_register(node, 0);
be_emit_cstring(", 0, %y");
be_emit_finish_line_gas(node);
fill_delay_slot();
}
- be_emit_irprintf("\t%s ", is_signed ? "sdiv" : "udiv");
+ sparc_emit_indent();
+ be_emit_irprintf("%s ", is_signed ? "sdiv" : "udiv");
sparc_emit_source_register(node, 1);
be_emit_cstring(", ");
sparc_emit_reg_or_imm(node, 2);
emit_sparc_Div(node, false);
}
-/**
- * Emits code for Call node
- */
static void emit_sparc_Call(const ir_node *node)
{
- const sparc_attr_t *attr = get_sparc_attr_const(node);
- ir_entity *entity = attr->immediate_value_entity;
-
- be_emit_cstring("\tcall ");
- if (entity != NULL) {
+ sparc_emit_indent();
+ be_emit_cstring("call ");
+ if (is_sparc_reg_call(node)) {
+ int dest_addr = get_sparc_Call_dest_addr_pos(node);
+ sparc_emit_source_register(node, dest_addr);
+ } else {
+ const sparc_attr_t *attr = get_sparc_attr_const(node);
+ ir_entity *entity = attr->immediate_value_entity;
be_gas_emit_entity(entity);
if (attr->immediate_value != 0) {
be_emit_irprintf("%+d", attr->immediate_value);
}
be_emit_cstring(", 0");
- } else {
- 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();
+
+ if (arch_get_irn_flags(node) & sparc_arch_irn_flag_aggregate_return) {
+ sparc_emit_indent();
+ be_emit_cstring("unimp 8\n");
+ be_emit_write_line();
+ }
}
-/**
- * Emit code for Perm node
- */
static void emit_be_Perm(const ir_node *irn)
{
- be_emit_cstring("\txor ");
+ sparc_emit_indent();
+ be_emit_cstring("xor ");
sparc_emit_source_register(irn, 1);
be_emit_cstring(", ");
sparc_emit_source_register(irn, 0);
sparc_emit_source_register(irn, 0);
be_emit_finish_line_gas(NULL);
- be_emit_cstring("\txor ");
+ sparc_emit_indent();
+ be_emit_cstring("xor ");
sparc_emit_source_register(irn, 1);
be_emit_cstring(", ");
sparc_emit_source_register(irn, 0);
sparc_emit_source_register(irn, 1);
be_emit_finish_line_gas(NULL);
- be_emit_cstring("\txor ");
+ sparc_emit_indent();
+ be_emit_cstring("xor ");
sparc_emit_source_register(irn, 1);
be_emit_cstring(", ");
sparc_emit_source_register(irn, 0);
be_emit_finish_line_gas(irn);
}
-static void emit_be_MemPerm(const ir_node *node)
+/* The stack pointer must always be SPARC_STACK_ALIGNMENT bytes aligned, so get
+ * the next bigger integer that's evenly divisible by it. */
+static unsigned get_aligned_sp_change(const unsigned num_regs)
+{
+ const unsigned bytes = num_regs * SPARC_REGISTER_SIZE;
+ return round_up2(bytes, SPARC_STACK_ALIGNMENT);
+}
+
+/* Spill register l0 or both l0 and l1, depending on n_spilled and n_to_spill.*/
+static void memperm_emit_spill_registers(const ir_node *node, int n_spilled,
+ int n_to_spill)
+{
+ assert(n_spilled < n_to_spill);
+
+ if (n_spilled == 0) {
+ /* We always reserve stack space for two registers because during copy
+ * processing we don't know yet if we also need to handle a cycle which
+ * needs two registers. More complicated code in emit_MemPerm would
+ * prevent wasting SPARC_REGISTER_SIZE bytes of stack space but
+ * it is not worth the worse readability of emit_MemPerm. */
+
+ /* Keep stack pointer aligned. */
+ unsigned sp_change = get_aligned_sp_change(2);
+ sparc_emit_indent();
+ be_emit_irprintf("sub %%sp, %u, %%sp", sp_change);
+ be_emit_finish_line_gas(node);
+
+ /* Spill register l0. */
+ sparc_emit_indent();
+ be_emit_irprintf("st %%l0, [%%sp%+d]", SPARC_MIN_STACKSIZE);
+ be_emit_finish_line_gas(node);
+ }
+
+ if (n_to_spill == 2) {
+ /* Spill register l1. */
+ sparc_emit_indent();
+ be_emit_irprintf("st %%l1, [%%sp%+d]", SPARC_MIN_STACKSIZE + SPARC_REGISTER_SIZE);
+ be_emit_finish_line_gas(node);
+ }
+}
+
+/* Restore register l0 or both l0 and l1, depending on n_spilled. */
+static void memperm_emit_restore_registers(const ir_node *node, int n_spilled)
+{
+ unsigned sp_change;
+
+ if (n_spilled == 2) {
+ /* Restore register l1. */
+ sparc_emit_indent();
+ be_emit_irprintf("ld [%%sp%+d], %%l1", SPARC_MIN_STACKSIZE + SPARC_REGISTER_SIZE);
+ be_emit_finish_line_gas(node);
+ }
+
+ /* Restore register l0. */
+ sparc_emit_indent();
+ be_emit_irprintf("ld [%%sp%+d], %%l0", SPARC_MIN_STACKSIZE);
+ be_emit_finish_line_gas(node);
+
+ /* Restore stack pointer. */
+ sp_change = get_aligned_sp_change(2);
+ sparc_emit_indent();
+ be_emit_irprintf("add %%sp, %u, %%sp", sp_change);
+ be_emit_finish_line_gas(node);
+}
+
+/* Emit code to copy in_ent to out_ent. Only uses l0. */
+static void memperm_emit_copy(const ir_node *node, ir_entity *in_ent,
+ ir_entity *out_ent)
+{
+ ir_graph *irg = get_irn_irg(node);
+ be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
+ int off_in = be_get_stack_entity_offset(layout, in_ent, 0);
+ int off_out = be_get_stack_entity_offset(layout, out_ent, 0);
+
+ /* Load from input entity. */
+ sparc_emit_indent();
+ be_emit_irprintf("ld [%%fp%+d], %%l0", off_in);
+ be_emit_finish_line_gas(node);
+
+ /* Store to output entity. */
+ sparc_emit_indent();
+ be_emit_irprintf("st %%l0, [%%fp%+d]", off_out);
+ be_emit_finish_line_gas(node);
+}
+
+/* Emit code to swap ent1 and ent2. Uses l0 and l1. */
+static void memperm_emit_swap(const ir_node *node, ir_entity *ent1,
+ ir_entity *ent2)
+{
+ ir_graph *irg = get_irn_irg(node);
+ be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
+ int off1 = be_get_stack_entity_offset(layout, ent1, 0);
+ int off2 = be_get_stack_entity_offset(layout, ent2, 0);
+
+ /* Load from first input entity. */
+ sparc_emit_indent();
+ be_emit_irprintf("ld [%%fp%+d], %%l0", off1);
+ be_emit_finish_line_gas(node);
+
+ /* Load from second input entity. */
+ sparc_emit_indent();
+ be_emit_irprintf("ld [%%fp%+d], %%l1", off2);
+ be_emit_finish_line_gas(node);
+
+ /* Store first value to second output entity. */
+ sparc_emit_indent();
+ be_emit_irprintf("st %%l0, [%%fp%+d]", off2);
+ be_emit_finish_line_gas(node);
+
+ /* Store second value to first output entity. */
+ sparc_emit_indent();
+ be_emit_irprintf("st %%l1, [%%fp%+d]", off1);
+ be_emit_finish_line_gas(node);
+}
+
+/* Find the index of ent in ents or return -1 if not found. */
+static int get_index(ir_entity **ents, int n, ir_entity *ent)
{
int i;
- int memperm_arity;
- int sp_change = 0;
+
+ for (i = 0; i < n; ++i)
+ if (ents[i] == ent)
+ return i;
+
+ return -1;
+}
+
+/*
+ * Emit code for a MemPerm node.
+ *
+ * Analyze MemPerm for copy chains and cyclic swaps and resolve them using
+ * loads and stores.
+ * This function is conceptually very similar to permute_values in
+ * beprefalloc.c.
+ */
+static void emit_be_MemPerm(const ir_node *node)
+{
+ int memperm_arity = be_get_MemPerm_entity_arity(node);
+ /* Upper limit for the number of participating entities is twice the
+ * arity, e.g., for a simple copying MemPerm node with one input/output. */
+ int max_size = 2 * memperm_arity;
+ ir_entity **entities = ALLOCANZ(ir_entity *, max_size);
+ /* sourceof contains the input entity for each entity. If an entity is
+ * never used as an output, its entry in sourceof is a fix point. */
+ int *sourceof = ALLOCANZ(int, max_size);
+ /* n_users counts how many output entities use this entity as their input.*/
+ int *n_users = ALLOCANZ(int, max_size);
+ /* n_spilled records the number of spilled registers, either 1 or 2. */
+ int n_spilled = 0;
+ int i, n, oidx;
+
+ /* This implementation currently only works with frame pointers. */
ir_graph *irg = get_irn_irg(node);
be_stack_layout_t *layout = be_get_irg_stack_layout(irg);
+ assert(!layout->sp_relative && "MemPerms currently do not work without frame pointers");
- /* this implementation only works with frame pointers currently */
- assert(layout->sp_relative == false);
+ for (i = 0; i < max_size; ++i) {
+ sourceof[i] = i;
+ }
- /* TODO: this implementation is slower than necessary.
- The longterm goal is however to avoid the memperm node completely */
+ for (i = n = 0; i < memperm_arity; ++i) {
+ ir_entity *out = be_get_MemPerm_out_entity(node, i);
+ ir_entity *in = be_get_MemPerm_in_entity(node, i);
+ int oidx; /* Out index */
+ int iidx; /* In index */
- memperm_arity = be_get_MemPerm_entity_arity(node);
- // we use our local registers - so this is limited to 8 inputs !
- if (memperm_arity > 8)
- panic("memperm with more than 8 inputs not supported yet");
+ /* Insert into entities to be able to operate on unique indices. */
+ if (get_index(entities, n, out) == -1)
+ entities[n++] = out;
+ if (get_index(entities, n, in) == -1)
+ entities[n++] = in;
- be_emit_irprintf("\tsub %%sp, %d, %%sp", memperm_arity*4);
- be_emit_finish_line_gas(node);
+ oidx = get_index(entities, n, out);
+ iidx = get_index(entities, n, in);
- for (i = 0; i < memperm_arity; ++i) {
- ir_entity *entity = be_get_MemPerm_in_entity(node, i);
- int offset = be_get_stack_entity_offset(layout, entity, 0);
+ sourceof[oidx] = iidx; /* Remember the source. */
+ ++n_users[iidx]; /* Increment number of users of this entity. */
+ }
- /* spill register */
- be_emit_irprintf("\tst %%l%d, [%%sp%+d]", i, sp_change + SPARC_MIN_STACKSIZE);
- be_emit_finish_line_gas(node);
+ /* First do all the copies. */
+ for (oidx = 0; oidx < n; /* empty */) {
+ int iidx = sourceof[oidx];
- /* load from entity */
- be_emit_irprintf("\tld [%%fp%+d], %%l%d", offset, i);
- be_emit_finish_line_gas(node);
- sp_change += 4;
+ /* Nothing to do for fix points.
+ * Also, if entities[oidx] is used as an input by another copy, we
+ * can't overwrite entities[oidx] yet.*/
+ if (iidx == oidx || n_users[oidx] > 0) {
+ ++oidx;
+ continue;
+ }
+
+ /* We found the end of a 'chain', so do the copy. */
+ if (n_spilled == 0) {
+ memperm_emit_spill_registers(node, n_spilled, /*n_to_spill=*/1);
+ n_spilled = 1;
+ }
+ memperm_emit_copy(node, entities[iidx], entities[oidx]);
+
+ /* Mark as done. */
+ sourceof[oidx] = oidx;
+
+ assert(n_users[iidx] > 0);
+ /* Decrementing the number of users might enable us to do another
+ * copy. */
+ --n_users[iidx];
+
+ if (iidx < oidx && n_users[iidx] == 0) {
+ oidx = iidx;
+ } else {
+ ++oidx;
+ }
}
- for (i = memperm_arity-1; i >= 0; --i) {
- ir_entity *entity = be_get_MemPerm_out_entity(node, i);
- int offset = be_get_stack_entity_offset(layout, entity, 0);
+ /* The rest are cycles. */
+ for (oidx = 0; oidx < n; /* empty */) {
+ int iidx = sourceof[oidx];
+ int tidx;
+
+ /* Nothing to do for fix points. */
+ if (iidx == oidx) {
+ ++oidx;
+ continue;
+ }
+
+ assert(n_users[iidx] == 1);
- sp_change -= 4;
+ /* Swap the two values to resolve the cycle. */
+ if (n_spilled < 2) {
+ memperm_emit_spill_registers(node, n_spilled, /*n_to_spill=*/2);
+ n_spilled = 2;
+ }
+ memperm_emit_swap(node, entities[iidx], entities[oidx]);
- /* store to new entity */
- 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 + SPARC_MIN_STACKSIZE, i);
- be_emit_finish_line_gas(node);
+ tidx = sourceof[iidx];
+ /* Mark as done. */
+ sourceof[iidx] = iidx;
+
+ /* The source of oidx is now the old source of iidx, because we swapped
+ * the two entities. */
+ sourceof[oidx] = tidx;
}
- be_emit_irprintf("\tadd %%sp, %d, %%sp", memperm_arity*4);
- be_emit_finish_line_gas(node);
+#ifdef DEBUG_libfirm
+ /* Only fix points should remain. */
+ for (i = 0; i < max_size; ++i) {
+ assert(sourceof[i] == i);
+ }
+#endif
- assert(sp_change == 0);
+ assert(n_spilled > 0 && "Useless MemPerm node");
+
+ memperm_emit_restore_registers(node, n_spilled);
}
-static void emit_be_Return(const ir_node *node)
+static void emit_sparc_Return(const ir_node *node)
{
+ ir_graph *irg = get_irn_irg(node);
+ ir_entity *entity = get_irg_entity(irg);
+ ir_type *type = get_entity_type(entity);
+
const char *destreg = "%o7";
/* hack: we don't explicitely model register changes because of the
|| is_sparc_RestoreZero(delay_slot_filler))) {
destreg = "%i7";
}
- be_emit_cstring("\tjmp ");
+ sparc_emit_indent();
+ be_emit_cstring("jmp ");
be_emit_string(destreg);
- be_emit_cstring("+8");
+ if (get_method_calling_convention(type) & cc_compound_ret) {
+ be_emit_cstring("+12");
+ } else {
+ be_emit_cstring("+8");
+ }
be_emit_finish_line_gas(node);
fill_delay_slot();
}
+static const arch_register_t *map_i_to_o_reg(const arch_register_t *reg)
+{
+ unsigned idx = reg->global_index;
+ if (idx < REG_I0 || idx > REG_I7)
+ return reg;
+ idx += REG_O0 - REG_I0;
+ assert(REG_O0 <= idx && idx <= REG_O7);
+ return &sparc_registers[idx];
+}
+
+static void emit_sparc_Restore(const ir_node *node)
+{
+ const arch_register_t *destreg
+ = arch_get_irn_register_out(node, pn_sparc_Restore_res);
+ sparc_emit_indent();
+ be_emit_cstring("restore ");
+ sparc_emit_source_register(node, 1);
+ be_emit_cstring(", ");
+ sparc_emit_reg_or_imm(node, 2);
+ be_emit_cstring(", ");
+ destreg = map_i_to_o_reg(destreg);
+ be_emit_char('%');
+ be_emit_string(arch_register_get_name(destreg));
+ be_emit_finish_line_gas(node);
+}
+
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) {
- be_emit_cstring("\tadd ");
+ sparc_emit_indent();
+ if (offset < 0) {
+ be_emit_cstring("add ");
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 ");
+ be_emit_cstring("sub ");
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(", ");
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)
-{
- 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;
+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 pnc");
+ panic("invalid relation");
}
-typedef const char* (*get_cc_func)(pn_Cmp pnc);
+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);
- 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;
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));
+ sparc_emit_indent();
+ be_emit_string(get_cc(relation));
be_emit_char(' ');
sparc_emit_cfop_target(proj_true);
be_emit_finish_line_gas(proj_true);
fill_delay_slot();
+ sparc_emit_indent();
if (get_irn_link(proj_false) == next_block) {
- be_emit_cstring("\t/* fallthrough to ");
+ be_emit_cstring("/* fallthrough to ");
sparc_emit_cfop_target(proj_false);
be_emit_cstring(" */");
be_emit_finish_line_gas(proj_false);
} else {
- be_emit_cstring("\tba ");
+ be_emit_cstring("ba ");
sparc_emit_cfop_target(proj_false);
be_emit_finish_line_gas(proj_false);
fill_delay_slot();
static void emit_sparc_fbfcc(const ir_node *node)
{
+ /* if the flags producing node was immediately in front of us, emit
+ * a nop */
+ ir_node *flags = get_irn_n(node, n_sparc_fbfcc_flags);
+ ir_node *prev = sched_prev(node);
+ if (is_Block(prev)) {
+ /* TODO: when the flags come from another block, then we have to do
+ * more complicated tests to see wether the flag producing node is
+ * potentially in front of us (could happen for fallthroughs) */
+ panic("TODO: fbfcc flags come from other block");
+ }
+ if (skip_Proj(flags) == prev) {
+ sparc_emit_indent();
+ be_emit_cstring("nop\n");
+ }
emit_sparc_branch(node, get_fcc);
}
static void emit_sparc_Ba(const ir_node *node)
{
+ sparc_emit_indent();
if (ba_is_fallthrough(node)) {
- be_emit_cstring("\t/* fallthrough to ");
+ be_emit_cstring("/* fallthrough to ");
sparc_emit_cfop_target(node);
be_emit_cstring(" */");
+ be_emit_finish_line_gas(node);
} else {
- be_emit_cstring("\tba ");
+ be_emit_cstring("ba ");
sparc_emit_cfop_target(node);
be_emit_finish_line_gas(node);
fill_delay_slot();
}
- 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_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;
-
- /* 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_indent();
+ be_emit_cstring("jmp ");
sparc_emit_source_register(node, 0);
be_emit_finish_line_gas(node);
fill_delay_slot();
- emit_jump_table(node);
+ be_emit_jump_table(node, attr->table, attr->table_entity, 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("\tfmovs %");
+ sparc_emit_indent();
+ be_emit_cstring("fmovs %");
be_emit_string(arch_register_get_name(src_reg));
be_emit_cstring(", %");
be_emit_string(arch_register_get_name(dst_reg));
static const arch_register_t *get_next_fp_reg(const arch_register_t *reg)
{
- unsigned index = reg->index;
- assert(reg == &sparc_registers[index]);
- index++;
- assert(index - REG_F0 < N_sparc_fp_REGS);
- return &sparc_registers[index];
+ unsigned idx = reg->global_index;
+ assert(reg == &sparc_registers[idx]);
+ idx++;
+ assert(idx - REG_F0 < N_sparc_fp_REGS);
+ return &sparc_registers[idx];
}
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);
+ const arch_register_t *src_reg = arch_get_irn_register_in(node, 0);
+ const arch_register_t *dst_reg = arch_get_irn_register_out(node, 0);
if (src_reg == dst_reg)
return;
emit_fmov(node, src_reg, dst_reg);
}
} else if (mode_is_data(mode)) {
- be_emit_cstring("\tmov ");
+ sparc_emit_indent();
+ be_emit_cstring("mov ");
sparc_emit_source_register(node, 0);
be_emit_cstring(", ");
sparc_emit_dest_register(node, 0);
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_SMulh, emit_sparc_Mulh);
+ set_emitter(op_sparc_UMulh, emit_sparc_Mulh);
+ set_emitter(op_sparc_Restore, emit_sparc_Restore);
+ set_emitter(op_sparc_Return, emit_sparc_Return);
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);
+ set_emitter(op_be_Keep, emit_nothing);
+ set_emitter(op_sparc_Start, emit_nothing);
+ set_emitter(op_Phi, emit_nothing);
}
/**
return schedpoint;
}
+static bool block_needs_label(const ir_node *block, const ir_node *sched_prev)
+{
+ int n_cfgpreds;
+
+ if (has_Block_entity(block))
+ return true;
+
+ n_cfgpreds = get_Block_n_cfgpreds(block);
+ if (n_cfgpreds == 0) {
+ return false;
+ } else if (n_cfgpreds > 1) {
+ return true;
+ } else {
+ ir_node *cfgpred = get_Block_cfgpred(block, 0);
+ ir_node *cfgpred_block = get_nodes_block(cfgpred);
+ if (is_Proj(cfgpred) && is_sparc_SwitchJmp(get_Proj_pred(cfgpred)))
+ return true;
+ return sched_prev != cfgpred_block || get_irn_link(cfgpred) != block;
+ }
+}
+
/**
* Walks over the nodes in a block connected by scheduling edges
* and emits code for each node.
*/
-static void sparc_emit_block(ir_node *block)
+static void sparc_emit_block(ir_node *block, ir_node *prev)
{
ir_node *node;
ir_node *next_delay_slot;
assert(is_Block(block));
- be_gas_emit_block_name(block);
- be_emit_cstring(":\n");
- be_emit_write_line();
+ if (block_needs_label(block, prev)) {
+ 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)
*/
static void sparc_emit_func_prolog(ir_graph *irg)
{
- ir_entity *ent = get_irg_entity(irg);
- be_gas_emit_function_prolog(ent, 4);
- be_emit_write_line();
+ ir_entity *entity = get_irg_entity(irg);
+ be_gas_emit_function_prolog(entity, 4);
}
/**
*/
static void sparc_emit_func_epilog(ir_graph *irg)
{
- ir_entity *ent = get_irg_entity(irg);
- const char *irg_name = get_entity_ld_name(ent);
- be_emit_write_line();
- be_emit_irprintf("\t.size %s, .-%s\n", irg_name, irg_name);
- be_emit_cstring("# -- End ");
- be_emit_string(irg_name);
- be_emit_cstring("\n");
- be_emit_write_line();
+ ir_entity *entity = get_irg_entity(irg);
+ be_gas_emit_function_epilog(entity);
}
static void sparc_gen_labels(ir_node *block, void *env)
void sparc_emit_routine(ir_graph *irg)
{
- ir_entity *entity = get_irg_entity(irg);
- ir_node **block_schedule;
- int i;
- int n;
-
- be_gas_elf_type_char = '#';
- be_gas_object_file_format = OBJECT_FILE_FORMAT_ELF_SPARC;
+ ir_node **block_schedule;
+ size_t i;
+ size_t n;
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. */
block_schedule = be_create_block_schedule(irg);
for (i = 0; i < n; ++i) {
ir_node *block = block_schedule[i];
+ ir_node *prev = i>=1 ? block_schedule[i-1] : NULL;
if (block == get_irg_end_block(irg))
continue;
- sparc_emit_block(block);
+ sparc_emit_block(block, prev);
}
/* emit function epilog */
projects / libfirm / blobdiff