static section_t curr_sec = NO_SECTION;
static const char *text[ASM_MAX][SECTION_MAX] = {
{
- ".section\t.text", ".section\t.data", ".section\t.rodata", ".section\t.text"
+ ".section\t.text",
+ ".section\t.data",
+ ".section\t.rodata",
+ ".section\t.text",
+ ".section\t.tbss,\"awT\",@nobits",
+ ".section\t.ctors,\"aw\",@progbits"
},
{
- ".section\t.text", ".section\t.data", ".section .rdata,\"dr\"", ".section\t.text"
+ ".section\t.text",
+ ".section\t.data",
+ ".section .rdata,\"dr\"",
+ ".section\t.text",
+ ".section\t.tbss,\"awT\",@nobits",
+ ".section\t.ctors,\"aw\",@progbits"
}
};
case SECTION_DATA:
case SECTION_RODATA:
case SECTION_COMMON:
+ case SECTION_TLS:
+ case SECTION_CTOR:
fprintf(F, "\t%s\n", text[asm_flavour][sec]);
break;
* Emits the jump sequence for a conditional jump (cmp + jmp_true + jmp_false)
*/
static void finish_CondJmp(FILE *F, const ir_node *irn, ir_mode *mode) {
- const ir_node *proj1, *proj2 = NULL;
- const ir_node *block, *next_bl = NULL;
+ const ir_node *proj_true;
+ const ir_node *proj_false;
+ const ir_node *block;
+ const ir_node *next_block;
char buf[SNPRINTF_BUF_LEN];
char cmd_buf[SNPRINTF_BUF_LEN];
char cmnt_buf[SNPRINTF_BUF_LEN];
int is_unsigned;
+ int pnc;
+ int flipped = 0;
/* get both Proj's */
- proj1 = get_proj(irn, pn_Cond_true);
- assert(proj1 && "CondJmp without true Proj");
+ proj_true = get_proj(irn, pn_Cond_true);
+ assert(proj_true && "CondJmp without true Proj");
- proj2 = get_proj(irn, pn_Cond_false);
- assert(proj2 && "CondJmp without false Proj");
+ proj_false = get_proj(irn, pn_Cond_false);
+ assert(proj_false && "CondJmp without false Proj");
+
+ pnc = get_ia32_pncode(irn);
/* for now, the code works for scheduled and non-schedules blocks */
block = get_nodes_block(irn);
/* we have a block schedule */
- next_bl = next_blk_sched(block);
+ next_block = next_blk_sched(block);
- if (get_cfop_target_block(proj1) == next_bl) {
+ if (get_cfop_target_block(proj_true) == next_block) {
/* exchange both proj's so the second one can be omitted */
- const ir_node *t = proj1;
- proj1 = proj2;
- proj2 = t;
+ const ir_node *t = proj_true;
+ proj_true = proj_false;
+ proj_false = t;
+
+ flipped = 1;
+ pnc = get_negated_pnc(pnc, mode);
}
/* the first Proj must always be created */
is_unsigned = mode_is_float(mode) || ! mode_is_signed(mode);
- if (get_Proj_proj(proj1) == pn_Cond_true) {
- snprintf(cmd_buf, SNPRINTF_BUF_LEN, "j%s %s",
- get_cmp_suffix(get_ia32_pncode(irn), is_unsigned),
- get_cfop_target(proj1, buf));
- snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* cmp(a, b) == TRUE */");
- }
- else {
- snprintf(cmd_buf, SNPRINTF_BUF_LEN, "j%s %s",
- get_cmp_suffix(get_negated_pnc(get_ia32_pncode(irn), mode), is_unsigned),
- get_cfop_target(proj1, buf));
- snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* cmp(a, b) == FALSE */");
- }
+ snprintf(cmd_buf, SNPRINTF_BUF_LEN, "j%s %s",
+ get_cmp_suffix(pnc, is_unsigned),
+ get_cfop_target(proj_true, buf));
+ snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* %s(a, b) %s*/",
+ get_pnc_string(pnc), flipped ? "(was flipped)" : "");
IA32_DO_EMIT(irn);
/* the second Proj might be a fallthrough */
- if (get_cfop_target_block(proj2) != next_bl) {
- snprintf(cmd_buf, SNPRINTF_BUF_LEN, "jmp %s", get_cfop_target(proj2, buf));
+ if (get_cfop_target_block(proj_false) != next_block) {
+ snprintf(cmd_buf, SNPRINTF_BUF_LEN, "jmp %s", get_cfop_target(proj_false, buf));
snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* otherwise */");
}
else {
cmd_buf[0] = '\0';
- snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* fallthrough %s */", get_cfop_target(proj2, buf));
+ snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* fallthrough %s */", get_cfop_target(proj_false, buf));
}
IA32_DO_EMIT(irn);
}
const char *instr = "fcom";
int reverse = 0;
- switch (get_ia32_pncode(irn)) {
+ switch (get_ia32_irn_opcode(irn)) {
case iro_ia32_fcomrJmp:
reverse = 1;
case iro_ia32_fcomJmp:
}
if (reverse)
- set_ia32_pncode(irn, (long)get_negated_pnc(get_ia32_pncode(irn), mode_Is));
+ set_ia32_pncode(irn, (long)get_inversed_pnc(get_ia32_pncode(irn)));
- snprintf(cmd_buf, SNPRINTF_BUF_LEN, "%s %%%s", instr, reg);
+ snprintf(cmd_buf, SNPRINTF_BUF_LEN, "%s %s%s", instr, reg[0] == '\0' ? "" : "%", reg);
lc_esnprintf(ia32_get_arg_env(), cmnt_buf, SNPRINTF_BUF_LEN, "/* %+F */", irn);
IA32_DO_EMIT(irn);
lc_esnprintf(ia32_get_arg_env(), cmd_buf, SNPRINTF_BUF_LEN, "fnstsw %%ax", irn);
snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* Store ah into flags */");
IA32_DO_EMIT(irn);
- finish_CondJmp(F, irn, mode_Is);
+ /* the compare flags must be evaluated using carry , ie unsigned */
+ finish_CondJmp(F, irn, mode_Iu);
}
static void CMov_emitter(ir_node *irn, ia32_emit_env_t *env) {
int is_unsigned = mode_is_float(mode) || ! mode_is_signed(mode);
const char *cmp_suffix = get_cmp_suffix(get_ia32_pncode(irn), is_unsigned);
int is_PsiCondCMov = is_ia32_PsiCondCMov(irn);
+ int idx_left = 2 - is_PsiCondCMov;
+ int idx_right = 3 - is_PsiCondCMov;
char cmd_buf[SNPRINTF_BUF_LEN];
char cmnt_buf[SNPRINTF_BUF_LEN];
const arch_register_t *in1, *in2, *out;
out = arch_get_irn_register(env->arch_env, irn);
- in1 = arch_get_irn_register(env->arch_env, get_irn_n(irn, 2 - is_PsiCondCMov));
- in2 = arch_get_irn_register(env->arch_env, get_irn_n(irn, 3 - is_PsiCondCMov));
+ in1 = arch_get_irn_register(env->arch_env, get_irn_n(irn, idx_left));
+ in2 = arch_get_irn_register(env->arch_env, get_irn_n(irn, idx_right));
/* we have to emit the cmp first, because the destination register */
/* could be one of the compare registers */
}
else if (is_PsiCondCMov) {
/* omit compare because flags are already set by And/Or */
- snprintf(cmd_buf, SNPRINTF_BUF_LEN, " ");
+ lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "test %1S, %1S", irn, irn);
}
else {
assert(0 && "unsupported CMov");
}
else if (REGS_ARE_EQUAL(out, in1)) {
/* true in == out -> need complement compare and exchange true and default in */
- ir_node *t = get_irn_n(irn, 2);
- set_irn_n(irn, 2, get_irn_n(irn, 3));
- set_irn_n(irn, 3, t);
+ ir_node *t = get_irn_n(irn, idx_left);
+ set_irn_n(irn, idx_left, get_irn_n(irn, idx_right));
+ set_irn_n(irn, idx_right, t);
- cmp_suffix = get_cmp_suffix(get_inversed_pnc(get_ia32_pncode(irn)), is_unsigned);
+ cmp_suffix = get_cmp_suffix(get_negated_pnc(get_ia32_pncode(irn), get_irn_mode(irn)), is_unsigned);
}
else {
/* out is different from in: need copy default -> out */
- lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "mov %1D, %4S", irn, irn);
+ if (is_PsiCondCMov)
+ lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "mov %1D, %3S", irn, irn);
+ else
+ lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "mov %1D, %4S", irn, irn);
+
lc_esnprintf(arg_env, cmnt_buf, SNPRINTF_BUF_LEN, "/* copy default -> out */" );
IA32_DO_EMIT(irn);
}
- lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "cmov%s %1D, %3S", cmp_suffix, irn, irn);
+ if (is_PsiCondCMov)
+ lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "cmov%s %1D, %2S", cmp_suffix, irn, irn);
+ else
+ lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "cmov%s %1D, %3S", cmp_suffix, irn, irn);
+
lc_esnprintf(arg_env, cmnt_buf, SNPRINTF_BUF_LEN, "/* condition is true case */" );
IA32_DO_EMIT(irn);
}
FILE *F = env->out;
const lc_arg_env_t *arg_env = ia32_get_arg_env();
int sse_pnc = -1;
+ long pnc = get_ia32_pncode(irn);
+ long unord = pnc & pn_Cmp_Uo;
char cmd_buf[SNPRINTF_BUF_LEN];
char cmnt_buf[SNPRINTF_BUF_LEN];
- switch (get_ia32_pncode(irn)) {
+ switch (pnc) {
case pn_Cmp_Leg: /* odered */
sse_pnc = 7;
break;
case pn_Cmp_Uo: /* unordered */
sse_pnc = 3;
break;
- case pn_Cmp_Ue: /* == */
+ case pn_Cmp_Ue:
+ case pn_Cmp_Eq: /* == */
sse_pnc = 0;
break;
- case pn_Cmp_Ul: /* < */
+ case pn_Cmp_Ul:
+ case pn_Cmp_Lt: /* < */
sse_pnc = 1;
break;
- case pn_Cmp_Ule: /* <= */
+ case pn_Cmp_Ule:
+ case pn_Cmp_Le: /* <= */
sse_pnc = 2;
break;
- case pn_Cmp_Ug: /* > */
+ case pn_Cmp_Ug:
+ case pn_Cmp_Gt: /* > */
sse_pnc = 6;
break;
- case pn_Cmp_Uge: /* >= */
+ case pn_Cmp_Uge:
+ case pn_Cmp_Ge: /* >= */
sse_pnc = 5;
break;
- case pn_Cmp_Ne: /* != */
+ case pn_Cmp_Ne:
+ case pn_Cmp_Lg: /* != */
sse_pnc = 4;
break;
}
- assert(sse_pnc >= 0 && "unsupported floating point compare");
-
- lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "cmps%M %s, %d", irn, ia32_emit_binop(irn, env), sse_pnc);
- lc_esnprintf(arg_env, cmnt_buf, SNPRINTF_BUF_LEN, "/* SSE compare with result in %1D */", irn);
+ assert(sse_pnc >= 0 && "unsupported compare");
+
+ if (unord && sse_pnc != 3) {
+ /*
+ We need a separate compare against unordered.
+ Quick and Dirty solution:
+ - get some memory on stack
+ - compare
+ - store result
+ - compare
+ - and result and stored result
+ - cleanup stack
+ */
+ snprintf(cmd_buf, SNPRINTF_BUF_LEN, "sub %%esp, 8");
+ snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* reserve some space for unordered compare result */");
+ IA32_DO_EMIT(NULL);
+ snprintf(cmd_buf, SNPRINTF_BUF_LEN, "cmpsd %s, 3", ia32_emit_binop(irn, env));
+ snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* SSE compare: unordered */");
+ IA32_DO_EMIT(NULL);
+ lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "movsd [%%esp], %1D", irn);
+ snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* store compare result */");
+ IA32_DO_EMIT(NULL);
+ }
+
+ snprintf(cmd_buf, SNPRINTF_BUF_LEN, "cmpsd %s, %d", ia32_emit_binop(irn, env), sse_pnc);
+ lc_esnprintf(arg_env, cmnt_buf, SNPRINTF_BUF_LEN, "/* SSE compare (%+F) with result in %1D */", irn, irn);
IA32_DO_EMIT(irn);
+
+ if (unord && sse_pnc != 3) {
+ lc_esnprintf(arg_env, cmd_buf, SNPRINTF_BUF_LEN, "andpd %1D, [%%esp]", irn);
+ snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* build the final result */");
+ IA32_DO_EMIT(NULL);
+ snprintf(cmd_buf, SNPRINTF_BUF_LEN, "add %%esp, 8");
+ snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* free allocated space */");
+ IA32_DO_EMIT(NULL);
+ }
}
/*********************************************************
conv_cmd = "cbw";
else if (n == 16 || m == 16)
conv_cmd = "cwde";
- else
+ else {
+ printf("%d -> %d unsupported\n", n, m);
assert(0 && "unsupported Conv_I2I");
+ }
}
switch(get_ia32_op_type(irn)) {
*/
static void emit_be_IncSP(const ir_node *irn, ia32_emit_env_t *emit_env) {
FILE *F = emit_env->out;
- unsigned offs = be_get_IncSP_offset(irn);
- be_stack_dir_t dir = be_get_IncSP_direction(irn);
+ int offs = be_get_IncSP_offset(irn);
char cmd_buf[SNPRINTF_BUF_LEN], cmnt_buf[SNPRINTF_BUF_LEN];
if (offs) {
- if (dir == be_stack_dir_expand)
+ if (offs > 0)
lc_esnprintf(ia32_get_arg_env(), cmd_buf, SNPRINTF_BUF_LEN, "sub %1S, %u", irn, offs);
else
- lc_esnprintf(ia32_get_arg_env(), cmd_buf, SNPRINTF_BUF_LEN, "add %1S, %u", irn, offs);
+ lc_esnprintf(ia32_get_arg_env(), cmd_buf, SNPRINTF_BUF_LEN, "add %1S, %u", irn, -offs);
lc_esnprintf(ia32_get_arg_env(), cmnt_buf, SNPRINTF_BUF_LEN, "/* %+F (IncSP) */", irn);
}
else {
"pxor %1S, %2S\n\tpxor %2S, %1S\n\tpxor %1S, %2S", irn, irn, irn, irn, irn, irn);
}
else if (cls1 == &ia32_reg_classes[CLASS_ia32_vfp]) {
- assert(0 && "Perm with vfp should not happen");
+ /* is a NOP */
+ cmd_buf[0] = '\0';
}
else if (cls1 == &ia32_reg_classes[CLASS_ia32_st]) {
- assert(0 && "Perm with st(X) should not happen");
+ /* is a NOP */
+ cmd_buf[0] = '\0';
}
lc_esnprintf(ia32_get_arg_env(), cmnt_buf, SNPRINTF_BUF_LEN, "/* %+F(%1A, %2A) */", irn, irn, irn);
* Emits code to increase stack pointer.
*/
static void emit_ia32_AddSP(const ir_node *irn, ia32_emit_env_t *emit_env) {
- FILE *F = emit_env->out;
+ FILE *F = emit_env->out;
char cmd_buf[SNPRINTF_BUF_LEN], cmnt_buf[SNPRINTF_BUF_LEN];
if (is_ia32_ImmConst(irn)) {
IA32_DO_EMIT(irn);
}
+/**
+ * Emits code to load the TLS base
+ */
+static void emit_ia32_LdTls(const ir_node *irn, ia32_emit_env_t *emit_env) {
+ FILE *F = emit_env->out;
+ char cmd_buf[SNPRINTF_BUF_LEN], cmnt_buf[SNPRINTF_BUF_LEN];
+
+ switch (asm_flavour) {
+ case ASM_LINUX_GAS:
+ lc_esnprintf(ia32_get_arg_env(), cmd_buf, SNPRINTF_BUF_LEN, "mov %1D, DWORD PTR %%gs:0", irn);
+ break;
+ case ASM_MINGW_GAS:
+ lc_esnprintf(ia32_get_arg_env(), cmd_buf, SNPRINTF_BUF_LEN, "mov %1D, DWORD PTR %%gs:0", irn);
+ break;
+ default:
+ assert(0 && "unsupported TLS");
+ break;
+ }
+ snprintf(cmnt_buf, SNPRINTF_BUF_LEN, "/* get thread local storage base */");
+
+ IA32_DO_EMIT(irn);
+}
+
static void emit_be_Return(const ir_node *n, ia32_emit_env_t *env) {
FILE *F = env->out;
const lc_arg_env_t *arg_env = ia32_get_arg_env();
IA32_EMIT(Conv_I2I8Bit);
IA32_EMIT(Const);
IA32_EMIT(AddSP);
+ IA32_EMIT(LdTls);
IA32_EMIT(xCmp);
IA32_EMIT(xCmpSet);
IA32_EMIT(xCmpCMov);
static void ia32_emit_align_func(FILE *F, cpu_support cpu) {
unsigned align; unsigned maximum_skip;
- /* gcc doesn't emit alignment for p4 ?*/
- if (cpu == arch_pentium_4)
- return;
-
switch (cpu) {
case arch_i386:
- align = 2; maximum_skip = 3;
+ align = 2;
break;
case arch_i486:
- align = 4; maximum_skip = 15;
+ align = 4;
break;
case arch_k6:
- align = 5; maximum_skip = 31;
+ align = 5;
break;
default:
- align = 4; maximum_skip = 15;
+ align = 4;
}
+ maximum_skip = (1 << align) - 1;
ia32_emit_alignment(F, align, maximum_skip);
}
static void ia32_emit_align_label(FILE *F, cpu_support cpu) {
unsigned align; unsigned maximum_skip;
- /* gcc doesn't emit alignment for p4 ?*/
- if (cpu == arch_pentium_4)
- return;
-
switch (cpu) {
case arch_i386:
- align = 2; maximum_skip = 3;
+ align = 2;
break;
case arch_i486:
- align = 4; maximum_skip = 15;
+ align = 4;
break;
case arch_k6:
- align = 5; maximum_skip = 7;
+ align = 5;
break;
default:
- align = 4; maximum_skip = 7;
+ align = 4;
}
+ maximum_skip = (1 << align) - 1;
ia32_emit_alignment(F, align, maximum_skip);
}
ia32_emit_env_t *emit_env = env;
const ir_node *irn;
int need_label = block != get_irg_start_block(get_irn_irg(block));
+ FILE *F = emit_env->out;
if (! is_Block(block))
return;
}
if (need_label) {
+ char cmd_buf[SNPRINTF_BUF_LEN];
+ int i, arity;
+
ia32_emit_align_label(emit_env->out, emit_env->isa->opt_arch);
- fprintf(emit_env->out, BLOCK_PREFIX("%ld:\n"), get_irn_node_nr(block));
+
+ ir_snprintf(cmd_buf, sizeof(cmd_buf), BLOCK_PREFIX("%d:"),
+ get_irn_node_nr(block));
+ fprintf(F, "%-43s ", cmd_buf);
+
+ /* emit list of pred blocks in comment */
+ fprintf(F, "/* preds:");
+
+ arity = get_irn_arity(block);
+ for(i = 0; i < arity; ++i) {
+ ir_node *predblock = get_Block_cfgpred_block(block, i);
+ fprintf(F, " %ld", get_irn_node_nr(predblock));
+ }
+ fprintf(F, " */\n");
}
+ /* emit the contents of the block */
sched_foreach(block, irn) {
ia32_emit_node(irn, env);
}
projects / libfirm / blobdiff