bescripts: Copy all common node attributes into the constructor variants.
[libfirm] / ir / be / sparc / sparc_cconv.c
index d28e8ca..564c71f 100644 (file)
  * @file
  * @brief   calling convention helpers
  * @author  Matthias Braun
- * @version $Id$
  */
 #include "config.h"
 
+#include "be_t.h"
+#include "beirg.h"
 #include "sparc_cconv.h"
 #include "irmode.h"
 #include "irgwalk.h"
 #include "util.h"
 #include "error.h"
 #include "gen_sparc_regalloc_if.h"
+#include "bitfiddle.h"
 
 static const unsigned ignore_regs[] = {
        REG_G0,
-       /* reserved for sparc ABI: */
+       /* used in case an address offset does not fit into an immediate: */
+       REG_G4,
+       /* reserved for SPARC ABI: */
        REG_G5,
        REG_G6,
        REG_G7,
@@ -49,7 +53,86 @@ static const unsigned ignore_regs[] = {
        REG_FPFLAGS,
        REG_FLAGS,
        REG_Y,
+
+       REG_F31,
+};
+
+static const arch_register_t* const param_regs[] = {
+       &sparc_registers[REG_I0],
+       &sparc_registers[REG_I1],
+       &sparc_registers[REG_I2],
+       &sparc_registers[REG_I3],
+       &sparc_registers[REG_I4],
+       &sparc_registers[REG_I5],
 };
+COMPILETIME_ASSERT(ARRAY_SIZE(param_regs) == SPARC_N_PARAM_REGS, sparcparamregs)
+
+static const arch_register_t* const float_result_regs[] = {
+       &sparc_registers[REG_F0],
+       &sparc_registers[REG_F1],
+       &sparc_registers[REG_F2],
+       &sparc_registers[REG_F3],
+       &sparc_registers[REG_F4],
+       &sparc_registers[REG_F5],
+       &sparc_registers[REG_F6],
+       &sparc_registers[REG_F7],
+};
+static arch_register_req_t float_result_reqs_double[8];
+static arch_register_req_t float_result_reqs_quad[8];
+
+static const unsigned caller_saves[] = {
+       REG_G1,
+       REG_G2,
+       REG_G3,
+       REG_O0,
+       REG_O1,
+       REG_O2,
+       REG_O3,
+       REG_O4,
+       REG_O5,
+       REG_F0,
+       REG_F1,
+       REG_F2,
+       REG_F3,
+       REG_F4,
+       REG_F5,
+       REG_F6,
+       REG_F7,
+       REG_F8,
+       REG_F9,
+       REG_F10,
+       REG_F11,
+       REG_F12,
+       REG_F13,
+       REG_F14,
+       REG_F15,
+       REG_F16,
+       REG_F17,
+       REG_F18,
+       REG_F19,
+       REG_F20,
+       REG_F21,
+       REG_F22,
+       REG_F23,
+       REG_F24,
+       REG_F25,
+       REG_F26,
+       REG_F27,
+       REG_F28,
+       REG_F29,
+       REG_F30,
+       REG_FLAGS,
+       REG_FPFLAGS,
+       REG_Y,
+};
+static unsigned default_caller_saves[BITSET_SIZE_ELEMS(N_SPARC_REGISTERS)];
+
+static const unsigned returns_twice_saved[] = {
+       REG_SP,
+       REG_FRAME_POINTER,
+       REG_I7
+};
+static unsigned default_returns_twice_saves[BITSET_SIZE_ELEMS(N_SPARC_REGISTERS)];
 
 /**
  * Maps an input register representing the i'th register input
@@ -66,8 +149,6 @@ static const arch_register_t *map_i_to_o_reg(const arch_register_t *reg)
 
 static void check_omit_fp(ir_node *node, void *env)
 {
-       bool *can_omit_fp = (bool*) env;
-
        /* omit-fp is not possible if:
         *  - we have allocations on the stack
         *  - we have calls (with the exception of tail-calls once we support them)
@@ -75,115 +156,187 @@ static void check_omit_fp(ir_node *node, void *env)
        if ((is_Alloc(node) && get_Alloc_where(node) == stack_alloc)
                        || (is_Free(node) && get_Free_where(node) == stack_alloc)
                        || is_Call(node)) {
+               bool *can_omit_fp = (bool*) env;
                *can_omit_fp = false;
        }
 }
 
+static unsigned determine_n_float_regs(ir_mode *mode)
+{
+       unsigned bits = get_mode_size_bits(mode);
+       switch (bits) {
+       case 32:
+               return 1;
+       case 64:
+               return 2;
+       case 128:
+               return 4;
+       default:
+               panic("Unexpected floatingpoint mode %+F", mode);
+       }
+}
+
 calling_convention_t *sparc_decide_calling_convention(ir_type *function_type,
                                                       ir_graph *irg)
 {
-       int                   stack_offset        = 0;
-       int                   n_param_regs        = ARRAY_SIZE(param_regs);
-       int                   n_float_result_regs = ARRAY_SIZE(float_result_regs);
-       bool                  omit_fp             = false;
-       reg_or_stackslot_t   *params;
-       reg_or_stackslot_t   *results;
-       int                   n_params;
-       int                   n_results;
-       int                   i;
-       int                   regnum;
-       int                   float_regnum;
-       calling_convention_t *cconv;
-
+       bool omit_fp = false;
        if (irg != NULL) {
-               const be_options_t *options = be_get_irg_options(irg);
-               omit_fp = options->omit_fp;
-               irg_walk_graph(irg, check_omit_fp, NULL, &omit_fp);
+               omit_fp = be_options.omit_fp;
+               /* our current vaarg handling needs the standard space to store the
+                * args 0-5 in it */
+               if (get_method_variadicity(function_type) == variadicity_variadic)
+                       omit_fp = false;
+               if (omit_fp == true) {
+                       irg_walk_graph(irg, check_omit_fp, NULL, &omit_fp);
+               }
+       }
+
+       mtp_additional_properties mtp
+               = get_method_additional_properties(function_type);
+       unsigned *caller_saves = rbitset_malloc(N_SPARC_REGISTERS);
+       if (mtp & mtp_property_returns_twice) {
+               rbitset_copy(caller_saves, default_returns_twice_saves,
+                            N_SPARC_REGISTERS);
+       } else {
+               rbitset_copy(caller_saves, default_caller_saves, N_SPARC_REGISTERS);
        }
 
        /* determine how parameters are passed */
-       n_params = get_method_n_params(function_type);
-       regnum   = 0;
-       params   = XMALLOCNZ(reg_or_stackslot_t, n_params);
+       int                 n_params = get_method_n_params(function_type);
+       int                 regnum   = 0;
+       reg_or_stackslot_t *params   = XMALLOCNZ(reg_or_stackslot_t, n_params);
 
-       for (i = 0; i < n_params; ++i) {
+       int      n_param_regs = ARRAY_SIZE(param_regs);
+       unsigned stack_offset = 0;
+       for (int i = 0; i < n_params; ++i) {
                ir_type            *param_type = get_method_param_type(function_type,i);
-               ir_mode            *mode       = get_type_mode(param_type);
-               int                 bits       = get_mode_size_bits(mode);
-               reg_or_stackslot_t *param      = &params[i];
+               ir_mode            *mode;
+               int                 bits;
+               reg_or_stackslot_t *param;
+
+               if (is_compound_type(param_type))
+                       panic("compound arguments not supported yet");
+
+               mode  = get_type_mode(param_type);
+               bits  = get_mode_size_bits(mode);
+               param = &params[i];
+
+               if (i == 0 &&
+                   (get_method_calling_convention(function_type) & cc_compound_ret)) {
+                       assert(mode_is_reference(mode) && bits == 32);
+                       /* special case, we have reserved space for this on the between
+                        * type */
+                       param->type   = param_type;
+                       param->offset = -SPARC_MIN_STACKSIZE+SPARC_AGGREGATE_RETURN_OFFSET;
+                       continue;
+               }
 
                if (regnum < n_param_regs) {
-                       const arch_register_t *reg = param_regs[regnum++];
+                       const arch_register_t *reg = param_regs[regnum];
                        if (irg == NULL || omit_fp)
                                reg = map_i_to_o_reg(reg);
-                       param->reg0 = reg;
+                       param->reg0       = reg;
+                       param->req0       = reg->single_req;
+                       param->reg_offset = regnum;
+                       ++regnum;
                } else {
                        param->type   = param_type;
                        param->offset = stack_offset;
-                       /* increase offset 4 bytes so everything is aligned */
-                       stack_offset += bits > 32 ? bits/8 : 4;
+                       /* increase offset by at least SPARC_REGISTER_SIZE bytes so everything is aligned */
+                       stack_offset += bits > 8 * SPARC_REGISTER_SIZE ? bits / 8 : SPARC_REGISTER_SIZE;
                        continue;
                }
 
                /* we might need a 2nd 32bit component (for 64bit or double values) */
                if (bits > 32) {
                        if (bits > 64)
-                               panic("only 32 and 64bit modes supported in sparc backend");
+                               panic("only 32 and 64bit modes supported");
 
                        if (regnum < n_param_regs) {
-                               const arch_register_t *reg = param_regs[regnum++];
+                               const arch_register_t *reg = param_regs[regnum];
                                if (irg == NULL || omit_fp)
                                        reg = map_i_to_o_reg(reg);
-                               param->reg1 = reg;
+                               param->reg1       = reg;
+                               param->req1       = reg->single_req;
+                               ++regnum;
                        } else {
                                ir_mode *regmode = param_regs[0]->reg_class->mode;
                                ir_type *type    = get_type_for_mode(regmode);
                                param->type      = type;
                                param->offset    = stack_offset;
                                assert(get_mode_size_bits(regmode) == 32);
-                               stack_offset += 4;
+                               stack_offset += SPARC_REGISTER_SIZE;
                        }
                }
        }
+       unsigned n_param_regs_used = regnum;
 
        /* determine how results are passed */
-       n_results    = get_method_n_ress(function_type);
-       regnum       = 0;
-       float_regnum = 0;
-       results      = XMALLOCNZ(reg_or_stackslot_t, n_results);
-       for (i = 0; i < n_results; ++i) {
+       int                 n_results           = get_method_n_ress(function_type);
+       unsigned            float_regnum        = 0;
+       unsigned            n_reg_results       = 0;
+       unsigned            n_float_result_regs = ARRAY_SIZE(float_result_regs);
+       reg_or_stackslot_t *results = XMALLOCNZ(reg_or_stackslot_t, n_results);
+       regnum        = 0;
+       for (int i = 0; i < n_results; ++i) {
                ir_type            *result_type = get_method_res_type(function_type, i);
                ir_mode            *result_mode = get_type_mode(result_type);
                reg_or_stackslot_t *result      = &results[i];
 
                if (mode_is_float(result_mode)) {
-                       if (float_regnum >= n_float_result_regs) {
-                               panic("Too many float results for sparc backend");
+                       unsigned n_regs   = determine_n_float_regs(result_mode);
+                       unsigned next_reg = round_up2(float_regnum, n_regs);
+
+                       if (next_reg >= n_float_result_regs) {
+                               panic("Too many float results");
                        } else {
-                               const arch_register_t *reg = float_result_regs[float_regnum++];
-                               result->reg0 = reg;
+                               const arch_register_t *reg = float_result_regs[next_reg];
+                               rbitset_clear(caller_saves, reg->global_index);
+                               result->reg_offset = i;
+                               if (n_regs == 1) {
+                                       result->req0 = reg->single_req;
+                               } else if (n_regs == 2) {
+                                       result->req0 = &float_result_reqs_double[next_reg];
+                                       rbitset_clear(caller_saves, reg->global_index+1);
+                               } else if (n_regs == 4) {
+                                       result->req0 = &float_result_reqs_quad[next_reg];
+                                       rbitset_clear(caller_saves, reg->global_index+1);
+                                       rbitset_clear(caller_saves, reg->global_index+2);
+                                       rbitset_clear(caller_saves, reg->global_index+3);
+                               } else {
+                                       panic("invalid number of registers in result");
+                               }
+                               float_regnum = next_reg + n_regs;
+
+                               ++n_reg_results;
                        }
                } else {
                        if (get_mode_size_bits(result_mode) > 32) {
-                               panic("Results with more than 32bits not supported by sparc backend yet");
+                               panic("Results with more than 32bits not supported yet");
                        }
 
                        if (regnum >= n_param_regs) {
-                               panic("Too many results for sparc backend");
+                               panic("Too many results");
                        } else {
                                const arch_register_t *reg = param_regs[regnum++];
                                if (irg == NULL || omit_fp)
                                        reg = map_i_to_o_reg(reg);
-                               result->reg0 = reg;
+                               result->req0       = reg->single_req;
+                               result->reg_offset = i;
+                               rbitset_clear(caller_saves, reg->global_index);
+                               ++n_reg_results;
                        }
                }
        }
 
-       cconv                   = XMALLOCZ(calling_convention_t);
+       calling_convention_t *cconv = XMALLOCZ(calling_convention_t);
        cconv->parameters       = params;
        cconv->param_stack_size = stack_offset;
+       cconv->n_param_regs     = n_param_regs_used;
        cconv->results          = results;
        cconv->omit_fp          = omit_fp;
+       cconv->caller_saves     = caller_saves;
+       cconv->n_reg_results    = n_reg_results;
 
        /* setup ignore register array */
        if (irg != NULL) {
@@ -192,7 +345,6 @@ calling_convention_t *sparc_decide_calling_convention(ir_type *function_type,
                struct obstack *obst      = &birg->obst;
                size_t          r;
 
-               assert(birg->allocatable_regs == NULL);
                birg->allocatable_regs = rbitset_obstack_alloc(obst, N_SPARC_REGISTERS);
                rbitset_set_all(birg->allocatable_regs, N_SPARC_REGISTERS);
                for (r = 0; r < n_ignores; ++r) {
@@ -207,5 +359,34 @@ void sparc_free_calling_convention(calling_convention_t *cconv)
 {
        free(cconv->parameters);
        free(cconv->results);
+       free(cconv->caller_saves);
        free(cconv);
 }
+
+void sparc_cconv_init(void)
+{
+       for (size_t i = 0; i < ARRAY_SIZE(caller_saves); ++i) {
+               rbitset_set(default_caller_saves, caller_saves[i]);
+       }
+
+       rbitset_set_all(default_returns_twice_saves, N_SPARC_REGISTERS);
+       for (size_t i = 0; i < ARRAY_SIZE(returns_twice_saved); ++i) {
+               rbitset_clear(default_returns_twice_saves, returns_twice_saved[i]);
+       }
+       for (size_t i = 0; i < ARRAY_SIZE(ignore_regs); ++i) {
+               rbitset_clear(default_returns_twice_saves, ignore_regs[i]);
+       }
+
+       for (size_t i = 0; i < ARRAY_SIZE(float_result_reqs_double); i += 2) {
+               arch_register_req_t *req = &float_result_reqs_double[i];
+               *req = *float_result_regs[i]->single_req;
+               req->type |= arch_register_req_type_aligned;
+               req->width = 2;
+       }
+       for (size_t i = 0; i < ARRAY_SIZE(float_result_reqs_quad); i += 4) {
+               arch_register_req_t *req = &float_result_reqs_quad[i];
+               *req = *float_result_regs[i]->single_req;
+               req->type |= arch_register_req_type_aligned;
+               req->width = 4;
+       }
+}