- Introduce cc_fixed: the calling convention must not be changed later on
[libfirm] / ir / be / ia32 / bearch_ia32.c
index 9edb3cc..e52ee6d 100644 (file)
 #include "irprintf.h"
 #include "iredges_t.h"
 #include "ircons.h"
+#include "irflag.h"
 #include "irgmod.h"
 #include "irgopt.h"
 #include "irbitset.h"
 #include "irgopt.h"
-#include "irdump_grgen.h"
 #include "pdeq.h"
 #include "pset.h"
 #include "debug.h"
@@ -76,6 +76,7 @@
 #include "ia32_new_nodes.h"
 #include "gen_ia32_regalloc_if.h"
 #include "gen_ia32_machine.h"
+#include "ia32_common_transform.h"
 #include "ia32_transform.h"
 #include "ia32_emitter.h"
 #include "ia32_map_regs.h"
@@ -502,7 +503,7 @@ static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_
 
                        /* the old SP is not needed anymore (kill the proj) */
                        assert(is_Proj(curr_sp));
-                       be_kill_node(curr_sp);
+                       kill_node(curr_sp);
 
                        /* copy ebp to esp */
                        curr_sp = be_new_Copy(&ia32_reg_classes[CLASS_ia32_gp], irg, bl, curr_bp);
@@ -926,6 +927,15 @@ static void ia32_before_abi(void *self) {
        }
 }
 
+typedef enum transformer_t {
+       TRANSFORMER_DEFAULT,
+#ifdef FIRM_GRGEN_BE
+       TRANSFORMER_PBQP
+#endif
+} transformer_t;
+
+static transformer_t be_transformer = TRANSFORMER_DEFAULT;
+
 /**
  * Transforms the standard firm graph into
  * an ia32 firm graph
@@ -933,7 +943,7 @@ static void ia32_before_abi(void *self) {
 static void ia32_prepare_graph(void *self) {
        ia32_code_gen_t *cg = self;
 
-       /* do local optimisations */
+       /* do local optimizations */
        optimize_graph_df(cg->irg);
 
        /* TODO: we often have dead code reachable through out-edges here. So for
@@ -947,18 +957,30 @@ static void ia32_prepare_graph(void *self) {
        if (cg->dump)
                be_dump(cg->irg, "-pre_transform", dump_ir_block_graph_sched);
 
+       switch (be_transformer) {
+               case TRANSFORMER_DEFAULT:
+                       /* transform remaining nodes into assembler instructions */
+                       ia32_transform_graph(cg);
+                       break;
+
 #ifdef FIRM_GRGEN_BE
-       /* transform nodes into assembler instructions by PBQP magic */
-       ia32_transform_graph_by_pbqp(cg);
-#endif
+               case TRANSFORMER_PBQP:
+                       // disable CSE, because of two-step node-construction
+                       set_opt_cse(0);
 
-       if (cg->dump)
-               be_dump(cg->irg, "-after_pbqp_transform", dump_ir_block_graph_sched);
+                       /* transform nodes into assembler instructions by PBQP magic */
+                       ia32_transform_graph_by_pbqp(cg);
+
+                       if (cg->dump)
+                               be_dump(cg->irg, "-after_pbqp_transform", dump_ir_block_graph_sched);
+                       set_opt_cse(1);
+                       break;
+#endif
 
-       /* transform remaining nodes into assembler instructions */
-       ia32_transform_graph(cg);
+               default: panic("Invalid transformer");
+       }
 
-       /* do local optimisations (mainly CSE) */
+       /* do local optimizations (mainly CSE) */
        optimize_graph_df(cg->irg);
 
        if (cg->dump)
@@ -1629,7 +1651,7 @@ static ia32_isa_t ia32_isa_template = {
                &ia32_gp_regs[REG_ESP],  /* stack pointer register */
                &ia32_gp_regs[REG_EBP],  /* base pointer register */
                -1,                      /* stack direction */
-               16,                      /* stack alignment */
+               4,                       /* power of two stack alignment, 2^4 == 16 */
                NULL,                    /* main environment */
                7,                       /* costs for a spill instruction */
                5,                       /* costs for a reload instruction */
@@ -1646,12 +1668,60 @@ static ia32_isa_t ia32_isa_template = {
 #endif
 };
 
+static void init_asm_constraints(void)
+{
+       be_init_default_asm_constraint_flags();
+
+       asm_constraint_flags['a'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['b'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['c'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['d'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['D'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['S'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['Q'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['q'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['A'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['l'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['R'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['r'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['p'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['f'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['t'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['u'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['Y'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['n'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE;
+       asm_constraint_flags['g'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE;
+
+       /* no support for autodecrement/autoincrement */
+       asm_constraint_flags['<'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       asm_constraint_flags['>'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* no float consts */
+       asm_constraint_flags['E'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       asm_constraint_flags['F'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* makes no sense on x86 */
+       asm_constraint_flags['s'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* no support for sse consts yet */
+       asm_constraint_flags['C'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* no support for x87 consts yet */
+       asm_constraint_flags['G'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* no support for mmx registers yet */
+       asm_constraint_flags['y'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* not available in 32bit mode */
+       asm_constraint_flags['Z'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       asm_constraint_flags['e'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+
+       /* no code yet to determine register class needed... */
+       asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+}
+
 /**
  * Initializes the backend ISA.
  */
 static arch_env_t *ia32_init(FILE *file_handle) {
        static int inited = 0;
        ia32_isa_t *isa;
+       int        i, n;
 
        if (inited)
                return NULL;
@@ -1690,6 +1760,14 @@ static arch_env_t *ia32_init(FILE *file_handle) {
        intrinsic_env.isa = isa;
        ia32_handle_intrinsics();
 
+       /* emit asm includes */
+       n = get_irp_n_asms();
+       for (i = 0; i < n; ++i) {
+               be_emit_cstring("#APP\n");
+               be_emit_ident(get_irp_asm(i));
+               be_emit_cstring("\n#NO_APP\n");
+       }
+
        /* needed for the debug support */
        be_gas_emit_switch_section(GAS_SECTION_TEXT);
        be_emit_cstring(".Ltext0:\n");
@@ -1786,7 +1864,9 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
        ir_mode  *mode;
        unsigned  cc;
        int       n, i, regnum;
+       int                 pop_amount = 0;
        be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
+
        (void) self;
 
        /* set abi flags for calls */
@@ -1804,8 +1884,9 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
                cc = cc_cdecl_set;
        } else {
                cc = get_method_calling_convention(method_type);
-               if (get_method_additional_properties(method_type) & mtp_property_private
-                               && (ia32_cg_config.optimize_cc)) {
+               if (!(cc & cc_fixed)                                                     &&
+                   get_method_additional_properties(method_type) & mtp_property_private &&
+                   ia32_cg_config.optimize_cc) {
                        /* set the calling conventions to register parameter */
                        cc = (cc & ~cc_bits) | cc_reg_param;
                }
@@ -1814,7 +1895,7 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
        /* we have to pop the shadow parameter ourself for compound calls */
        if( (get_method_calling_convention(method_type) & cc_compound_ret)
                        && !(cc & cc_reg_param)) {
-               be_abi_call_set_pop(abi, get_mode_size_bytes(mode_P_data));
+               pop_amount += get_mode_size_bytes(mode_P_data);
        }
 
        n = get_method_n_params(method_type);
@@ -1833,12 +1914,21 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
                } else {
                        /* Micro optimisation: if the mode is shorter than 4 bytes, load 4 bytes.
                         * movl has a shorter opcode than mov[sz][bw]l */
-                       ir_mode *load_mode = mode;
-                       if (mode != NULL && get_mode_size_bytes(mode) < 4) load_mode = mode_Iu;
+                       ir_mode  *load_mode = mode;
+                       unsigned  size      = get_mode_size_bytes(mode);
+
+                       if (cc & cc_callee_clear_stk) {
+                               pop_amount += (size + 3U) & ~3U;
+                       }
+
+                       if (mode != NULL && size < 4) load_mode = mode_Iu;
+
                        be_abi_call_param_stack(abi, i, load_mode, 4, 0, 0);
                }
        }
 
+       be_abi_call_set_pop(abi, pop_amount);
+
        /* set return registers */
        n = get_method_n_ress(method_type);
 
@@ -2152,6 +2242,23 @@ static int ia32_is_psi_allowed(ir_node *sel, ir_node *phi_list, int i, int j)
        return 0;
 }
 
+static asm_constraint_flags_t ia32_parse_asm_constraint(const void *self, const char **c)
+{
+       (void) self;
+       (void) c;
+
+       /* we already added all our simple flags to the flags modifier list in
+        * init, so this flag we don't know. */
+       return ASM_CONSTRAINT_FLAG_INVALID;
+}
+
+static int ia32_is_valid_clobber(const void *self, const char *clobber)
+{
+       (void) self;
+
+       return ia32_get_clobber_register(clobber) != NULL;
+}
+
 /**
  * Returns the libFirm configuration parameter for this backend.
  */
@@ -2173,15 +2280,21 @@ static const backend_params *ia32_get_libfirm_params(void) {
        static backend_params p = {
                1,     /* need dword lowering */
                1,     /* support inline assembly */
+               0,     /* no immediate floating point mode. */
                NULL,  /* no additional opcodes */
                NULL,  /* will be set later */
                ia32_create_intrinsic_fkt,
                &intrinsic_env,  /* context for ia32_create_intrinsic_fkt */
                NULL,  /* will be set below */
+               NULL   /* will be set below */
        };
 
        ia32_setup_cg_config();
 
+       /* doesn't really belong here, but this is the earliest place the backend
+        * is called... */
+       init_asm_constraints();
+
        p.dep_param    = &ad;
        p.if_conv_info = &ifconv;
        return &p;
@@ -2199,9 +2312,22 @@ static lc_opt_enum_int_var_t gas_var = {
        (int*) &be_gas_flavour, gas_items
 };
 
+static const lc_opt_enum_int_items_t transformer_items[] = {
+       { "default", TRANSFORMER_DEFAULT },
+#ifdef FIRM_GRGEN_BE
+       { "pbqp",    TRANSFORMER_PBQP    },
+#endif
+       { NULL,      0                   }
+};
+
+static lc_opt_enum_int_var_t transformer_var = {
+       (int*)&be_transformer, transformer_items
+};
+
 static const lc_opt_table_entry_t ia32_options[] = {
        LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
-       LC_OPT_ENT_INT("stackalign", "set stack alignment for calls",
+       LC_OPT_ENT_ENUM_INT("transformer", "the transformer used for code selection", &transformer_var),
+       LC_OPT_ENT_INT("stackalign", "set power of two stack alignment for calls",
                       &ia32_isa_template.arch_env.stack_alignment),
        LC_OPT_LAST
 };
@@ -2221,6 +2347,8 @@ const arch_isa_if_t ia32_isa_if = {
        ia32_get_allowed_execution_units,
        ia32_get_machine,
        ia32_get_irg_list,
+       ia32_parse_asm_constraint,
+       ia32_is_valid_clobber
 };
 
 void ia32_init_emitter(void);