move lea_to_add into the peephole phase
authorMatthias Braun <matze@braunis.de>
Mon, 15 Oct 2007 19:31:46 +0000 (19:31 +0000)
committerMatthias Braun <matze@braunis.de>
Mon, 15 Oct 2007 19:31:46 +0000 (19:31 +0000)
[r16212]

ir/be/ia32/bearch_ia32.c
ir/be/ia32/ia32_finish.c
ir/be/ia32/ia32_optimize.c
ir/be/ia32/ia32_optimize.h
ir/be/ia32/ia32_spec.pl

index 9d0ccda..edd26f1 100644 (file)
@@ -1469,7 +1469,7 @@ static void ia32_finish(void *self) {
        }
 
        /* do peephole optimisations */
-       ia32_peephole_optimization(irg, cg);
+       ia32_peephole_optimization(cg);
 
        /* create block schedule, this also removes empty blocks which might
         * produce critical edges */
index 81a63d1..5306d62 100644 (file)
@@ -141,214 +141,6 @@ static void ia32_transform_sub_to_neg_add(ir_node *irn, ia32_code_gen_t *cg) {
        DBG_OPT_SUB2NEGADD(irn, res);
 }
 
-static INLINE int is_noreg(ia32_code_gen_t *cg, const ir_node *node)
-{
-       return node == cg->noreg_gp;
-}
-
-static ir_node *create_immediate_from_int(ia32_code_gen_t *cg, int val)
-{
-       ir_graph *irg         = current_ir_graph;
-       ir_node  *start_block = get_irg_start_block(irg);
-       ir_node  *immediate   = new_rd_ia32_Immediate(NULL, irg, start_block, NULL,
-                                                     0, val);
-       arch_set_irn_register(cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]);
-
-       return immediate;
-}
-
-static ir_node *create_immediate_from_am(ia32_code_gen_t *cg,
-                                         const ir_node *node)
-{
-       ir_graph  *irg     = get_irn_irg(node);
-       ir_node   *block   = get_nodes_block(node);
-       int        offset  = get_ia32_am_offs_int(node);
-       int        sc_sign = is_ia32_am_sc_sign(node);
-       ir_entity *entity  = get_ia32_am_sc(node);
-       ir_node   *res;
-
-       res = new_rd_ia32_Immediate(NULL, irg, block, entity, sc_sign, offset);
-       arch_set_irn_register(cg->arch_env, res, &ia32_gp_regs[REG_GP_NOREG]);
-       return res;
-}
-
-static int is_am_one(const ir_node *node)
-{
-       int        offset  = get_ia32_am_offs_int(node);
-       ir_entity *entity  = get_ia32_am_sc(node);
-
-       return offset == 1 && entity == NULL;
-}
-
-static int is_am_minus_one(const ir_node *node)
-{
-       int        offset  = get_ia32_am_offs_int(node);
-       ir_entity *entity  = get_ia32_am_sc(node);
-
-       return offset == -1 && entity == NULL;
-}
-
-/**
- * Transforms a LEA into an Add or SHL if possible.
- * THIS FUNCTIONS MUST BE CALLED AFTER REGISTER ALLOCATION.
- */
-static void ia32_transform_lea_to_add_or_shl(ir_node *node, ia32_code_gen_t *cg)
-{
-       const arch_env_t      *arch_env = cg->arch_env;
-       ir_graph              *irg      = current_ir_graph;
-       ir_node               *base;
-       ir_node               *index;
-       const arch_register_t *base_reg;
-       const arch_register_t *index_reg;
-       const arch_register_t *out_reg;
-       int                    scale;
-       int                    has_immediates;
-       ir_node               *op1;
-       ir_node               *op2;
-       dbg_info              *dbgi;
-       ir_node               *block;
-       ir_node               *res;
-       ir_node               *noreg;
-       ir_node               *nomem;
-
-       if(!is_ia32_Lea(node))
-               return;
-
-       base  = get_irn_n(node, n_ia32_Lea_base);
-       index = get_irn_n(node, n_ia32_Lea_index);
-
-       if(is_noreg(cg, base)) {
-               base     = NULL;
-               base_reg = NULL;
-       } else {
-               base_reg = arch_get_irn_register(arch_env, base);
-       }
-       if(is_noreg(cg, index)) {
-               index     = NULL;
-               index_reg = NULL;
-       } else {
-               index_reg = arch_get_irn_register(arch_env, index);
-       }
-
-       if(base == NULL && index == NULL) {
-               /* we shouldn't construct these in the first place... */
-#ifdef DEBUG_libfirm
-               ir_fprintf(stderr, "Optimisation warning: found immediate only lea\n");
-#endif
-               return;
-       }
-
-       out_reg = arch_get_irn_register(arch_env, node);
-       scale   = get_ia32_am_scale(node);
-       assert(!is_ia32_need_stackent(node) || get_ia32_frame_ent(node) != NULL);
-       /* check if we have immediates values (frame entities should already be
-        * expressed in the offsets) */
-       if(get_ia32_am_offs_int(node) != 0 || get_ia32_am_sc(node) != NULL) {
-               has_immediates = 1;
-       } else {
-               has_immediates = 0;
-       }
-
-       /* we can transform leas where the out register is the same as either the
-        * base or index register back to an Add or Shl */
-       if(out_reg == base_reg) {
-               if(index == NULL) {
-#ifdef DEBUG_libfirm
-                       if(!has_immediates) {
-                               ir_fprintf(stderr, "Optimisation warning: found lea which is "
-                                          "just a copy\n");
-                       }
-#endif
-                       op1 = base;
-                       goto make_add_immediate;
-               }
-               if(scale == 0 && !has_immediates) {
-                       op1 = base;
-                       op2 = index;
-                       goto make_add;
-               }
-               /* can't create an add */
-               return;
-       } else if(out_reg == index_reg) {
-               if(base == NULL) {
-                       if(has_immediates && scale == 0) {
-                               op1 = index;
-                               goto make_add_immediate;
-                       } else if(!has_immediates && scale > 0) {
-                               op1 = index;
-                               op2 = create_immediate_from_int(cg, scale);
-                               goto make_shl;
-                       } else if(!has_immediates) {
-#ifdef DEBUG_libfirm
-                               ir_fprintf(stderr, "Optimisation warning: found lea which is "
-                                          "just a copy\n");
-#endif
-                       }
-               } else if(scale == 0 && !has_immediates) {
-                       op1 = index;
-                       op2 = base;
-                       goto make_add;
-               }
-               /* can't create an add */
-               return;
-       } else {
-               /* can't create an add */
-               return;
-       }
-
-make_add_immediate:
-       if(cg->isa->opt & IA32_OPT_INCDEC) {
-               if(is_am_one(node)) {
-                       dbgi  = get_irn_dbg_info(node);
-                       block = get_nodes_block(node);
-                       res   = new_rd_ia32_Inc(dbgi, irg, block, op1);
-                       arch_set_irn_register(arch_env, res, out_reg);
-                       goto exchange;
-               }
-               if(is_am_minus_one(node)) {
-                       dbgi  = get_irn_dbg_info(node);
-                       block = get_nodes_block(node);
-                       res   = new_rd_ia32_Dec(dbgi, irg, block, op1);
-                       arch_set_irn_register(arch_env, res, out_reg);
-                       goto exchange;
-               }
-       }
-       op2 = create_immediate_from_am(cg, node);
-
-make_add:
-       dbgi  = get_irn_dbg_info(node);
-       block = get_nodes_block(node);
-       noreg = ia32_new_NoReg_gp(cg);
-       nomem = new_NoMem();
-       res   = new_rd_ia32_Add(dbgi, irg, block, noreg, noreg, nomem, op1, op2);
-       arch_set_irn_register(arch_env, res, out_reg);
-       set_ia32_commutative(res);
-       goto exchange;
-
-make_shl:
-       dbgi  = get_irn_dbg_info(node);
-       block = get_nodes_block(node);
-       noreg = ia32_new_NoReg_gp(cg);
-       nomem = new_NoMem();
-       res   = new_rd_ia32_Shl(dbgi, irg, block, op1, op2);
-       arch_set_irn_register(arch_env, res, out_reg);
-       goto exchange;
-
-exchange:
-       SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, node));
-
-       /* add new ADD/SHL to schedule */
-       sched_add_before(node, res);
-
-       DBG_OPT_LEA2ADD(node, res);
-
-       /* remove the old LEA */
-       sched_remove(node);
-
-       /* exchange the Add and the LEA */
-       exchange(node, res);
-}
-
 static INLINE int need_constraint_copy(ir_node *irn) {
        return  ! is_ia32_Lea(irn)      &&
                ! is_ia32_Conv_I2I(irn)     &&
@@ -646,9 +438,6 @@ static void ia32_finish_irg_walker(ir_node *block, void *env) {
 
                /* check if there is a sub which need to be transformed */
                ia32_transform_sub_to_neg_add(irn, cg);
-
-               /* transform a LEA into an Add if possible */
-               ia32_transform_lea_to_add_or_shl(irn, cg);
        }
 
        /* second: insert copies and finish irg */
index ba52e08..daecdbd 100644 (file)
@@ -38,6 +38,7 @@
 #include "irgwalk.h"
 #include "height.h"
 #include "irbitset.h"
+#include "irprintf.h"
 
 #include "../be_t.h"
 #include "../beabi.h"
@@ -46,6 +47,7 @@
 #include "../bepeephole.h"
 
 #include "ia32_new_nodes.h"
+#include "ia32_optimize.h"
 #include "bearch_ia32_t.h"
 #include "gen_ia32_regalloc_if.h"
 #include "ia32_transform.h"
@@ -88,7 +90,7 @@ static INLINE int be_is_NoReg(ia32_code_gen_t *cg, const ir_node *irn) {
 /**
  * Tries to create pushs from IncSP,Store combinations
  */
-static void ia32_create_Pushs(ir_node *irn)
+static void peephole_IncSP_Store_to_push(ir_node *irn)
 {
        int i;
        int offset;
@@ -198,34 +200,8 @@ static void ia32_create_Pushs(ir_node *irn)
        }
 
        be_set_IncSP_offset(irn, offset);
-
-       // can we remove the IncSP now?
-       if(offset == 0) {
-               const ir_edge_t *edge, *next;
-
-               foreach_out_edge_safe(irn, edge, next) {
-                       ir_node *arg = get_edge_src_irn(edge);
-                       int pos = get_edge_src_pos(edge);
-
-                       set_irn_n(arg, pos, curr_sp);
-               }
-
-               set_irn_n(irn, 0, new_Bad());
-               sched_remove(irn);
-       } else {
-               set_irn_n(irn, 0, curr_sp);
-       }
-}
-
-/**
- * Performs Peephole Optimizations for IncSP nodes.
- */
-static void ia32_peephole_optimize_node(ir_node *node, void *env)
-{
-       (void) env;
-       if (be_is_IncSP(node)) {
-               ia32_create_Pushs(node);
-       }
+       be_set_IncSP_pred(irn, curr_sp);
+       be_peephole_node_replaced(irn, irn);
 }
 
 /**
@@ -309,6 +285,9 @@ static void peephole_be_IncSP(ir_node *node)
        ir_node               *stack;
        int                    offset;
 
+       /* transform IncSP->Store combinations to Push where possible */
+       peephole_IncSP_Store_to_push(node);
+
        /* first optimize incsp->incsp combinations */
        peephole_IncSP_IncSP(node);
 
@@ -397,6 +376,217 @@ static void peephole_ia32_Const(ir_node *node)
        sched_remove(node);
 }
 
+static INLINE int is_noreg(ia32_code_gen_t *cg, const ir_node *node)
+{
+       return node == cg->noreg_gp;
+}
+
+static ir_node *create_immediate_from_int(ia32_code_gen_t *cg, int val)
+{
+       ir_graph *irg         = current_ir_graph;
+       ir_node  *start_block = get_irg_start_block(irg);
+       ir_node  *immediate   = new_rd_ia32_Immediate(NULL, irg, start_block, NULL,
+                                                     0, val);
+       arch_set_irn_register(cg->arch_env, immediate, &ia32_gp_regs[REG_GP_NOREG]);
+
+       return immediate;
+}
+
+static ir_node *create_immediate_from_am(ia32_code_gen_t *cg,
+                                         const ir_node *node)
+{
+       ir_graph  *irg     = get_irn_irg(node);
+       ir_node   *block   = get_nodes_block(node);
+       int        offset  = get_ia32_am_offs_int(node);
+       int        sc_sign = is_ia32_am_sc_sign(node);
+       ir_entity *entity  = get_ia32_am_sc(node);
+       ir_node   *res;
+
+       res = new_rd_ia32_Immediate(NULL, irg, block, entity, sc_sign, offset);
+       arch_set_irn_register(cg->arch_env, res, &ia32_gp_regs[REG_GP_NOREG]);
+       return res;
+}
+
+static int is_am_one(const ir_node *node)
+{
+       int        offset  = get_ia32_am_offs_int(node);
+       ir_entity *entity  = get_ia32_am_sc(node);
+
+       return offset == 1 && entity == NULL;
+}
+
+static int is_am_minus_one(const ir_node *node)
+{
+       int        offset  = get_ia32_am_offs_int(node);
+       ir_entity *entity  = get_ia32_am_sc(node);
+
+       return offset == -1 && entity == NULL;
+}
+
+/**
+ * Transforms a LEA into an Add or SHL if possible.
+ */
+static void peephole_ia32_Lea(ir_node *node)
+{
+       const arch_env_t      *arch_env = cg->arch_env;
+       ir_graph              *irg      = current_ir_graph;
+       ir_node               *base;
+       ir_node               *index;
+       const arch_register_t *base_reg;
+       const arch_register_t *index_reg;
+       const arch_register_t *out_reg;
+       int                    scale;
+       int                    has_immediates;
+       ir_node               *op1;
+       ir_node               *op2;
+       dbg_info              *dbgi;
+       ir_node               *block;
+       ir_node               *res;
+       ir_node               *noreg;
+       ir_node               *nomem;
+
+       assert(is_ia32_Lea(node));
+
+       /* we can only do this if are allowed to globber the flags */
+       if(be_peephole_get_value(CLASS_ia32_flags, REG_EFLAGS) != NULL)
+               return;
+
+       base  = get_irn_n(node, n_ia32_Lea_base);
+       index = get_irn_n(node, n_ia32_Lea_index);
+
+       if(is_noreg(cg, base)) {
+               base     = NULL;
+               base_reg = NULL;
+       } else {
+               base_reg = arch_get_irn_register(arch_env, base);
+       }
+       if(is_noreg(cg, index)) {
+               index     = NULL;
+               index_reg = NULL;
+       } else {
+               index_reg = arch_get_irn_register(arch_env, index);
+       }
+
+       if(base == NULL && index == NULL) {
+               /* we shouldn't construct these in the first place... */
+#ifdef DEBUG_libfirm
+               ir_fprintf(stderr, "Optimisation warning: found immediate only lea\n");
+#endif
+               return;
+       }
+
+       out_reg = arch_get_irn_register(arch_env, node);
+       scale   = get_ia32_am_scale(node);
+       assert(!is_ia32_need_stackent(node) || get_ia32_frame_ent(node) != NULL);
+       /* check if we have immediates values (frame entities should already be
+        * expressed in the offsets) */
+       if(get_ia32_am_offs_int(node) != 0 || get_ia32_am_sc(node) != NULL) {
+               has_immediates = 1;
+       } else {
+               has_immediates = 0;
+       }
+
+       /* we can transform leas where the out register is the same as either the
+        * base or index register back to an Add or Shl */
+       if(out_reg == base_reg) {
+               if(index == NULL) {
+#ifdef DEBUG_libfirm
+                       if(!has_immediates) {
+                               ir_fprintf(stderr, "Optimisation warning: found lea which is "
+                                          "just a copy\n");
+                       }
+#endif
+                       op1 = base;
+                       goto make_add_immediate;
+               }
+               if(scale == 0 && !has_immediates) {
+                       op1 = base;
+                       op2 = index;
+                       goto make_add;
+               }
+               /* can't create an add */
+               return;
+       } else if(out_reg == index_reg) {
+               if(base == NULL) {
+                       if(has_immediates && scale == 0) {
+                               op1 = index;
+                               goto make_add_immediate;
+                       } else if(!has_immediates && scale > 0) {
+                               op1 = index;
+                               op2 = create_immediate_from_int(cg, scale);
+                               goto make_shl;
+                       } else if(!has_immediates) {
+#ifdef DEBUG_libfirm
+                               ir_fprintf(stderr, "Optimisation warning: found lea which is "
+                                          "just a copy\n");
+#endif
+                       }
+               } else if(scale == 0 && !has_immediates) {
+                       op1 = index;
+                       op2 = base;
+                       goto make_add;
+               }
+               /* can't create an add */
+               return;
+       } else {
+               /* can't create an add */
+               return;
+       }
+
+make_add_immediate:
+       if(cg->isa->opt & IA32_OPT_INCDEC) {
+               if(is_am_one(node)) {
+                       dbgi  = get_irn_dbg_info(node);
+                       block = get_nodes_block(node);
+                       res   = new_rd_ia32_Inc(dbgi, irg, block, op1);
+                       arch_set_irn_register(arch_env, res, out_reg);
+                       goto exchange;
+               }
+               if(is_am_minus_one(node)) {
+                       dbgi  = get_irn_dbg_info(node);
+                       block = get_nodes_block(node);
+                       res   = new_rd_ia32_Dec(dbgi, irg, block, op1);
+                       arch_set_irn_register(arch_env, res, out_reg);
+                       goto exchange;
+               }
+       }
+       op2 = create_immediate_from_am(cg, node);
+
+make_add:
+       dbgi  = get_irn_dbg_info(node);
+       block = get_nodes_block(node);
+       noreg = ia32_new_NoReg_gp(cg);
+       nomem = new_NoMem();
+       res   = new_rd_ia32_Add(dbgi, irg, block, noreg, noreg, nomem, op1, op2);
+       arch_set_irn_register(arch_env, res, out_reg);
+       set_ia32_commutative(res);
+       goto exchange;
+
+make_shl:
+       dbgi  = get_irn_dbg_info(node);
+       block = get_nodes_block(node);
+       noreg = ia32_new_NoReg_gp(cg);
+       nomem = new_NoMem();
+       res   = new_rd_ia32_Shl(dbgi, irg, block, op1, op2);
+       arch_set_irn_register(arch_env, res, out_reg);
+       goto exchange;
+
+exchange:
+       SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(cg, node));
+
+       /* add new ADD/SHL to schedule */
+       sched_add_before(node, res);
+
+       DBG_OPT_LEA2ADD(node, res);
+
+       /* remove the old LEA */
+       sched_remove(node);
+
+       /* exchange the Add and the LEA */
+       be_peephole_node_replaced(node, res);
+       exchange(node, res);
+}
+
 /**
  * Register a peephole optimisation function.
  */
@@ -406,7 +596,7 @@ static void register_peephole_optimisation(ir_op *op, peephole_opt_func func) {
 }
 
 /* Perform peephole-optimizations. */
-void ia32_peephole_optimization(ir_graph *irg, ia32_code_gen_t *new_cg)
+void ia32_peephole_optimization(ia32_code_gen_t *new_cg)
 {
        cg       = new_cg;
        arch_env = cg->arch_env;
@@ -415,9 +605,9 @@ void ia32_peephole_optimization(ir_graph *irg, ia32_code_gen_t *new_cg)
        clear_irp_opcodes_generic_func();
        register_peephole_optimisation(op_ia32_Const, peephole_ia32_Const);
        register_peephole_optimisation(op_be_IncSP, peephole_be_IncSP);
+       register_peephole_optimisation(op_ia32_Lea, peephole_ia32_Lea);
 
        be_peephole_opt(cg->birg);
-       irg_walk_graph(irg, ia32_peephole_optimize_node, NULL, NULL);
 }
 
 /**
index 37a43c9..2b22071 100644 (file)
@@ -48,6 +48,6 @@ void ia32_optimize_graph(ia32_code_gen_t *cg);
  * @param irg   the graph
  * @param cg    the code generator object
  */
-void ia32_peephole_optimization(ir_graph *irg, ia32_code_gen_t *cg);
+void ia32_peephole_optimization(ia32_code_gen_t *cg);
 
 #endif /* FIRM_BE_IA32_IA32_OPTIMIZE_H */
index 983d4ad..eec6649 100644 (file)
@@ -1237,8 +1237,8 @@ Lea => {
        latency   => 2,
        units     => [ "GP" ],
        mode      => $mode_gp,
-# well this isn't true for Lea, but we often transform Lea back to Add, Inc
-# or Dec, so we set the flag
+# lea doesn't modify the flags, but setting this seems advantageous since it
+# increases chances that the Lea is transformed back to an Add
        modified_flags => 1,
 },