+2003-11-20 Richard Henderson <rth@redhat.com>
+
+ * ssa.c, ssa-dce.c, ssa-ccp.c: Remove files.
+ * Makefile.in (OBJS-common, GTFILES): Don't reference them.
+ (gtype-desc.o, toplev.o, flow.o): Remove ssa.h.
+ (ssa.o, ssa-dce.o, ssa-ccp.o): Remove.
+ * flow.c: Don't include ssa.h.
+ (set_phi_alternative_reg): Remove.
+ (calculate_global_regs_live): Don't call it.
+ (mark_used_regs): Don't handle PHI.
+ * gengtype.c (open_base_files): Don't reference ssa.h.
+ * rtl.def (PHI): Remove.
+ * timevar.def (TV_TO_SSA, TV_SSA_CCP, TV_SSA_DCE, TV_FROM_SSA): Kill.
+ * common.opt: Remove -fssa, -fssa-ccp, -fssa-dce.
+ * opts.c (common_handle_option): Likewise.
+ * toplev.c (f_options): Likewise.
+ (DFI_ssa, DFI_ssa_ccp, DFI_ssa_dce, DFI_ussa): Remove.
+ (dump_file): Update to match.
+ (flag_ssa, flag_ssa_ccp, flag_ssa_dce): Remove.
+ (rest_of_handle_ssa): Remove.
+ (rest_of_compilation): Don't call it.
+ * toplev.h (flag_ssa, flag_ssa_dce, flag_ssa_ccp): Remove.
+ * doc/invoke.texi: Remove -fssa, -fssa-ccp, -fssa-dce.
+ * doc/passes.texi (SSA optimizations): Remove.
+
2003-11-20 Bob Wilson <bob.wilson@acm.org>
* configure.in: Add xtensa-*-* targets to test for dwarf2 debug_line.
insn-extract.o insn-opinit.o insn-output.o insn-peep.o insn-recog.o \
integrate.o intl.o jump.o langhooks.o lcm.o lists.o local-alloc.o \
loop.o optabs.o options.o opts.o params.o postreload.o predict.o \
- print-rtl.o print-tree.o value-prof.o \
+ print-rtl.o print-tree.o value-prof.o \
profile.o ra.o ra-build.o ra-colorize.o ra-debug.o ra-rewrite.o \
real.o recog.o reg-stack.o regclass.o regmove.o regrename.o \
reload.o reload1.o reorg.o resource.o rtl.o rtlanal.o rtl-error.o \
sbitmap.o sched-deps.o sched-ebb.o sched-rgn.o sched-vis.o sdbout.o \
- sibcall.o simplify-rtx.o sreal.o ssa.o ssa-ccp.o ssa-dce.o stmt.o \
- stor-layout.o stringpool.o targhooks.o timevar.o toplev.o tracer.o tree.o tree-dump.o \
- unroll.o varasm.o varray.o version.o vmsdbgout.o xcoffout.o \
- alloc-pool.o et-forest.o cfghooks.o bt-load.o pretty-print.o $(GGC) web.o
+ sibcall.o simplify-rtx.o sreal.o stmt.o stor-layout.o stringpool.o \
+ targhooks.o timevar.o toplev.o tracer.o tree.o tree-dump.o unroll.o \
+ varasm.o varray.o version.o vmsdbgout.o xcoffout.o alloc-pool.o \
+ et-forest.o cfghooks.o bt-load.o pretty-print.o $(GGC) web.o
OBJS-md = $(out_object_file)
OBJS-archive = $(EXTRA_OBJS) $(host_hook_obj) hashtable.o tree-inline.o \
gtype-desc.o: gtype-desc.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) varray.h \
$(HASHTAB_H) $(TREE_H) $(RTL_H) function.h insn-config.h $(EXPR_H) $(OPTABS_H) \
libfuncs.h debug.h $(GGC_H) bitmap.h $(BASIC_BLOCK_H) hard-reg-set.h \
- ssa.h cselib.h insn-addr.h
+ cselib.h insn-addr.h
ggc-common.o: ggc-common.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(GGC_H) \
$(HASHTAB_H) toplev.h $(PARAMS_H) hosthooks.h
debug.h insn-config.h intl.h $(RECOG_H) Makefile toplev.h \
dwarf2out.h sdbout.h dbxout.h $(EXPR_H) hard-reg-set.h $(BASIC_BLOCK_H) \
graph.h $(LOOP_H) except.h $(REGS_H) $(TIMEVAR_H) value-prof.h \
- ssa.h $(PARAMS_H) $(TM_P_H) reload.h dwarf2asm.h $(TARGET_H) \
+ $(PARAMS_H) $(TM_P_H) reload.h dwarf2asm.h $(TARGET_H) \
langhooks.h insn-flags.h cfglayout.h real.h cfgloop.h \
hosthooks.h $(LANGHOOKS_DEF_H) cgraph.h $(COVERAGE_H)
$(CC) $(ALL_CFLAGS) $(ALL_CPPFLAGS) $(INCLUDES) \
lcm.o : lcm.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) $(REGS_H) \
hard-reg-set.h flags.h real.h insn-config.h $(INSN_ATTR_H) $(RECOG_H) $(EXPR_H) \
$(BASIC_BLOCK_H) $(TM_P_H) df.h function.h
-ssa.o : ssa.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) $(REGS_H) varray.h \
- $(EXPR_H) hard-reg-set.h flags.h function.h real.h insn-config.h $(RECOG_H) \
- $(BASIC_BLOCK_H) output.h ssa.h
-ssa-dce.o : ssa-dce.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) hard-reg-set.h \
- $(BASIC_BLOCK_H) ssa.h insn-config.h $(RECOG_H) output.h
-ssa-ccp.o : ssa-ccp.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) hard-reg-set.h \
- $(BASIC_BLOCK_H) ssa.h insn-config.h $(RECOG_H) output.h \
- errors.h $(GGC_H) df.h function.h
df.o : df.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) \
insn-config.h $(RECOG_H) function.h $(REGS_H) alloc-pool.h hard-reg-set.h \
$(BASIC_BLOCK_H) df.h $(FIBHEAP_H)
alloc-pool.o : alloc-pool.c $(CONFIG_H) $(SYSTEM_H) alloc-pool.h
flow.o : flow.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) $(TREE_H) \
flags.h insn-config.h $(BASIC_BLOCK_H) $(REGS_H) hard-reg-set.h output.h toplev.h \
- $(RECOG_H) function.h except.h $(EXPR_H) ssa.h $(GGC_H) $(TM_P_H)
+ $(RECOG_H) function.h except.h $(EXPR_H) $(GGC_H) $(TM_P_H)
cfg.o : cfg.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) flags.h insn-config.h \
$(BASIC_BLOCK_H) $(REGS_H) hard-reg-set.h output.h toplev.h $(RECOG_H) \
function.h except.h $(GGC_H) $(TM_P_H) alloc-pool.h
$(host_xm_file_list) $(tm_file_list) $(HASHTAB_H) $(SPLAY_TREE_H) \
$(srcdir)/bitmap.h $(srcdir)/coverage.c $(srcdir)/function.h $(srcdir)/rtl.h \
$(srcdir)/optabs.h $(srcdir)/tree.h $(srcdir)/libfuncs.h $(srcdir)/hashtable.h \
- $(srcdir)/real.h $(srcdir)/varray.h $(srcdir)/ssa.h $(srcdir)/insn-addr.h \
+ $(srcdir)/real.h $(srcdir)/varray.h $(srcdir)/insn-addr.h \
$(srcdir)/cselib.h $(srcdir)/basic-block.h $(srcdir)/cgraph.h \
$(srcdir)/c-common.h $(srcdir)/c-tree.h \
$(srcdir)/alias.c $(srcdir)/bitmap.c $(srcdir)/cselib.c $(srcdir)/cgraph.c \
Common
Convert floating point constants to single precision constants
-fssa
-Common
-Enable static single assignment optimizations
-
-fssa-ccp
-Common
-Enable SSA conditional constant propagation
-
-fssa-dce
-Common
-Enable aggressive SSA dead code elimination
-
fstack-check
Common
Insert stack checking code into the program
-fsched-stalled-insns=@var{n} -sched-stalled-insns-dep=@var{n} @gol
-fsched2-use-superblocks @gol
-fsched2-use-traces -fsignaling-nans @gol
--fsingle-precision-constant -fssa -fssa-ccp -fssa-dce @gol
+-fsingle-precision-constant @gol
-fstrength-reduce -fstrict-aliasing -ftracer -fthread-jumps @gol
-funroll-all-loops -funroll-loops -fpeel-loops @gol
-funswitch-loops -fold-unroll-loops -fold-unroll-all-loops @gol
specify this option and you may have problems with debugging if
you specify both this option and @option{-g}.
-@item -fssa
-@opindex fssa
-Perform optimizations in static single assignment form. Each function's
-flow graph is translated into SSA form, optimizations are performed, and
-the flow graph is translated back from SSA form. Users should not
-specify this option, since it is not yet ready for production use.
-
-@item -fssa-ccp
-@opindex fssa-ccp
-Perform Sparse Conditional Constant Propagation in SSA form. Requires
-@option{-fssa}. Like @option{-fssa}, this is an experimental feature.
-
-@item -fssa-dce
-@opindex fssa-dce
-Perform aggressive dead-code elimination in SSA form. Requires @option{-fssa}.
-Like @option{-fssa}, this is an experimental feature.
-
@item -fbranch-target-load-optimize
@opindex fbranch-target-load-optimize
Perform branch target register load optimization before prologue / epilogue
Perform branch target register load optimization after prologue / epilogue
threading.
-
-
-
@item --param @var{name}=@var{value}
@opindex param
In some places, GCC uses various constants to control the amount of
@file{jump.c}. This optimization is only performed if
@option{-fthread-jumps} is enabled.
-@cindex SSA optimizations
-@cindex Single Static Assignment optimizations
-@opindex fssa
-@item
-Static Single Assignment (SSA) based optimization passes. The
-SSA conversion passes (to/from) are turned on by the @option{-fssa}
-option (it is also done automatically if you enable an SSA optimization pass).
-These passes utilize a form called Static Single Assignment. In SSA form,
-each variable (pseudo register) is only set once, giving you def-use
-and use-def chains for free, and enabling a lot more optimization
-passes to be run in linear time.
-Conversion to and from SSA form is handled by functions in
-@file{ssa.c}.
-
-@opindex de
-The option @option{-de} causes a debugging dump of the RTL code after
-this pass. This dump file's name is made by appending @samp{.ssa} to
-the input file name.
-@itemize @bullet
-@cindex SSA Conditional Constant Propagation
-@cindex Conditional Constant Propagation, SSA based
-@cindex conditional constant propagation
-@opindex fssa-ccp
-@item
-SSA Conditional Constant Propagation. Turned on by the @option{-fssa-ccp}
-option. This pass performs conditional constant propagation to simplify
-instructions including conditional branches. This pass is more aggressive
-than the constant propagation done by the CSE and GCSE passes, but operates
-in linear time.
-
-@opindex dW
-The option @option{-dW} causes a debugging dump of the RTL code after
-this pass. This dump file's name is made by appending @samp{.ssaccp} to
-the input file name.
-
-@cindex SSA DCE
-@cindex DCE, SSA based
-@cindex dead code elimination
-@opindex fssa-dce
-@item
-SSA Aggressive Dead Code Elimination. Turned on by the @option{-fssa-dce}
-option. This pass performs elimination of code considered unnecessary because
-it has no externally visible effects on the program. It operates in
-linear time.
-
-@opindex dX
-The option @option{-dX} causes a debugging dump of the RTL code after
-this pass. This dump file's name is made by appending @samp{.ssadce} to
-the input file name.
-@end itemize
-
@cindex common subexpression elimination
@cindex constant propagation
@item
#include "toplev.h"
#include "recog.h"
#include "expr.h"
-#include "ssa.h"
#include "timevar.h"
#include "obstack.h"
static void notice_stack_pointer_modification (rtx);
static void mark_reg (rtx, void *);
static void mark_regs_live_at_end (regset);
-static int set_phi_alternative_reg (rtx, int, int, void *);
static void calculate_global_regs_live (sbitmap, sbitmap, int);
static void propagate_block_delete_insn (rtx);
static rtx propagate_block_delete_libcall (rtx, rtx);
diddle_return_value (mark_reg, set);
}
-/* Callback function for for_each_successor_phi. DATA is a regset.
- Sets the SRC_REGNO, the regno of the phi alternative for phi node
- INSN, in the regset. */
-
-static int
-set_phi_alternative_reg (rtx insn ATTRIBUTE_UNUSED,
- int dest_regno ATTRIBUTE_UNUSED, int src_regno,
- void *data)
-{
- regset live = (regset) data;
- SET_REGNO_REG_SET (live, src_regno);
- return 0;
-}
-
/* Propagate global life info around the graph of basic blocks. Begin
considering blocks with their corresponding bit set in BLOCKS_IN.
If BLOCKS_IN is null, consider it the universal set.
SET_REGNO_REG_SET (new_live_at_end, PIC_OFFSET_TABLE_REGNUM);
}
- /* Regs used in phi nodes are not included in
- global_live_at_start, since they are live only along a
- particular edge. Set those regs that are live because of a
- phi node alternative corresponding to this particular block. */
- if (in_ssa_form)
- for_each_successor_phi (bb, &set_phi_alternative_reg,
- new_live_at_end);
-
if (bb == ENTRY_BLOCK_PTR)
{
COPY_REG_SET (bb->global_live_at_end, new_live_at_end);
x = COND_EXEC_CODE (x);
goto retry;
- case PHI:
- /* We _do_not_ want to scan operands of phi nodes. Operands of
- a phi function are evaluated only when control reaches this
- block along a particular edge. Therefore, regs that appear
- as arguments to phi should not be added to the global live at
- start. */
- return;
-
default:
break;
}
"config.h", "system.h", "coretypes.h", "tm.h", "varray.h",
"hashtab.h", "splay-tree.h", "bitmap.h", "tree.h", "rtl.h",
"function.h", "insn-config.h", "expr.h", "hard-reg-set.h",
- "basic-block.h", "cselib.h", "insn-addr.h", "ssa.h", "optabs.h",
+ "basic-block.h", "cselib.h", "insn-addr.h", "optabs.h",
"libfuncs.h", "debug.h", "ggc.h", "cgraph.h",
NULL
};
flag_single_precision_constant = value;
break;
- case OPT_fssa:
- flag_ssa = value;
- break;
-
- case OPT_fssa_ccp:
- flag_ssa_ccp = value;
- break;
-
- case OPT_fssa_dce:
- flag_ssa_dce = value;
- break;
-
case OPT_fstack_check:
flag_stack_check = value;
break;
/* Unsigned saturating truncate. */
DEF_RTL_EXPR(US_TRUNCATE, "us_truncate", "e", '1')
-/* The SSA phi operator.
-
- The argument is a vector of 2N rtxes. Element 2N+1 is a CONST_INT
- containing the block number of the predecessor through which control
- has passed when the register at element 2N is used.
-
- Note that PHI may only appear at the beginning of a basic block.
-
- ??? There may be multiple PHI insns, but they are all evaluated
- in parallel. This probably ought to be changed to use a real
- PARALLEL, as that would be less confusing and more in the spirit
- of canonical RTL. It is, however, easier to manipulate this way. */
-DEF_RTL_EXPR(PHI, "phi", "E", 'x')
-
/*
Local variables:
+++ /dev/null
-/* Conditional constant propagation pass for the GNU compiler.
- Copyright (C) 2000, 2001, 2002, 2003 Free Software Foundation, Inc.
- Original framework by Daniel Berlin <dan@cgsoftware.com>
- Fleshed out and major cleanups by Jeff Law <law@redhat.com>
-
-This file is part of GCC.
-
-GCC is free software; you can redistribute it and/or modify it under
-the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
-version.
-
-GCC is distributed in the hope that it will be useful, but WITHOUT ANY
-WARRANTY; without even the implied warranty of MERCHANTABILITY or
-FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
-for more details.
-
-You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
-
-/* Conditional constant propagation.
-
- References:
-
- Constant propagation with conditional branches,
- Wegman and Zadeck, ACM TOPLAS 13(2):181-210.
-
- Building an Optimizing Compiler,
- Robert Morgan, Butterworth-Heinemann, 1998, Section 8.9.
-
- Advanced Compiler Design and Implementation,
- Steven Muchnick, Morgan Kaufmann, 1997, Section 12.6
-
- The overall structure is as follows:
-
- 1. Run a simple SSA based DCE pass to remove any dead code.
- 2. Run CCP to compute what registers are known constants
- and what edges are not executable. Remove unexecutable
- edges from the CFG and simplify PHI nodes.
- 3. Replace registers with constants where possible.
- 4. Remove unreachable blocks computed in step #2.
- 5. Another simple SSA DCE pass to remove dead code exposed
- by CCP.
-
- When we exit, we are still in SSA form.
-
-
- Potential further enhancements:
-
- 1. Handle SUBREGs, STRICT_LOW_PART, etc in destinations more
- gracefully.
-
- 2. Handle insns with multiple outputs more gracefully.
-
- 3. Handle CONST_DOUBLE and symbolic constants.
-
- 4. Fold expressions after performing constant substitutions. */
-
-
-#include "config.h"
-#include "system.h"
-#include "coretypes.h"
-#include "tm.h"
-
-#include "rtl.h"
-#include "hard-reg-set.h"
-#include "basic-block.h"
-#include "ssa.h"
-#include "insn-config.h"
-#include "recog.h"
-#include "output.h"
-#include "errors.h"
-#include "ggc.h"
-#include "df.h"
-#include "function.h"
-\f
-/* Possible lattice values. */
-
-typedef enum
-{
- UNDEFINED,
- CONSTANT,
- VARYING
-} latticevalue;
-
-/* Main structure for CCP.
-
- Contains the lattice value and, if it's a constant, the constant
- value. */
-typedef struct
-{
- latticevalue lattice_val;
- rtx const_value;
-} value;
-
-/* Array of values indexed by register number. */
-static value *values;
-
-/* A bitmap to keep track of executable blocks in the CFG. */
-static sbitmap executable_blocks;
-
-/* A bitmap for all executable edges in the CFG. */
-static sbitmap executable_edges;
-
-/* Array of edges on the work list. */
-static edge *edge_info;
-
-/* We need an edge list to be able to get indexes easily. */
-static struct edge_list *edges;
-
-/* For building/following use-def and def-use chains. */
-static struct df *df_analyzer;
-
-/* Current edge we are operating on, from the worklist */
-static edge flow_edges;
-
-/* Bitmap of SSA edges which will need reexamination as their definition
- has changed. */
-static sbitmap ssa_edges;
-
-/* Simple macros to simplify code */
-#define SSA_NAME(x) REGNO (SET_DEST (x))
-#define EIE(x,y) EDGE_INDEX (edges, x, y)
-
-static void visit_phi_node (rtx, basic_block);
-static void visit_expression (rtx, basic_block);
-static void defs_to_undefined (rtx);
-static void defs_to_varying (rtx);
-static void examine_flow_edges (void);
-static int mark_references (rtx *, void *);
-static void follow_def_use_chains (void);
-static void optimize_unexecutable_edges (struct edge_list *, sbitmap);
-static void ssa_ccp_substitute_constants (void);
-static void ssa_ccp_df_delete_unreachable_insns (void);
-static void ssa_fast_dce (struct df *);
-
-/* Loop through the PHI_NODE's parameters for BLOCK and compare their
- lattice values to determine PHI_NODE's lattice value. */
-static void
-visit_phi_node (rtx phi_node, basic_block block)
-{
- unsigned int i;
- rtx phi_node_expr = NULL;
- unsigned int phi_node_name = SSA_NAME (PATTERN (phi_node));
- latticevalue phi_node_lattice_val = UNDEFINED;
- rtx pat = PATTERN (phi_node);
- rtvec phi_vec = XVEC (SET_SRC (pat), 0);
- unsigned int num_elem = GET_NUM_ELEM (phi_vec);
-
- for (i = 0; i < num_elem; i += 2)
- {
- if (TEST_BIT (executable_edges,
- EIE (BASIC_BLOCK (INTVAL (RTVEC_ELT (phi_vec, i + 1))),
- block)))
- {
- unsigned int current_parm
- = REGNO (RTVEC_ELT (phi_vec, i));
-
- latticevalue current_parm_lattice_val
- = values[current_parm].lattice_val;
-
- /* If any node is VARYING, then new value of PHI_NODE
- is VARYING. */
- if (current_parm_lattice_val == VARYING)
- {
- phi_node_lattice_val = VARYING;
- phi_node_expr = NULL;
- break;
- }
-
- /* If we have more than one distinct constant, then the new
- value of PHI_NODE is VARYING. */
- if (current_parm_lattice_val == CONSTANT
- && phi_node_lattice_val == CONSTANT
- && values[current_parm].const_value != phi_node_expr)
- {
- phi_node_lattice_val = VARYING;
- phi_node_expr = NULL;
- break;
- }
-
- /* If the current value of PHI_NODE is UNDEFINED and one
- node in PHI_NODE is CONSTANT, then the new value of the
- PHI is that CONSTANT. Note this can turn into VARYING
- if we find another distinct constant later. */
- if (phi_node_lattice_val == UNDEFINED
- && phi_node_expr == NULL
- && current_parm_lattice_val == CONSTANT)
- {
- phi_node_expr = values[current_parm].const_value;
- phi_node_lattice_val = CONSTANT;
- continue;
- }
- }
- }
-
- /* If the value of PHI_NODE changed, then we will need to
- re-execute uses of the output of PHI_NODE. */
- if (phi_node_lattice_val != values[phi_node_name].lattice_val)
- {
- values[phi_node_name].lattice_val = phi_node_lattice_val;
- values[phi_node_name].const_value = phi_node_expr;
- SET_BIT (ssa_edges, phi_node_name);
- }
-}
-
-/* Sets all defs in an insn to UNDEFINED. */
-static void
-defs_to_undefined (rtx insn)
-{
- struct df_link *currdef;
- for (currdef = DF_INSN_DEFS (df_analyzer, insn); currdef;
- currdef = currdef->next)
- {
- if (values[DF_REF_REGNO (currdef->ref)].lattice_val != UNDEFINED)
- SET_BIT (ssa_edges, DF_REF_REGNO (currdef->ref));
- values[DF_REF_REGNO (currdef->ref)].lattice_val = UNDEFINED;
- }
-}
-
-/* Sets all defs in an insn to VARYING. */
-static void
-defs_to_varying (rtx insn)
-{
- struct df_link *currdef;
- for (currdef = DF_INSN_DEFS (df_analyzer, insn); currdef;
- currdef = currdef->next)
- {
- if (values[DF_REF_REGNO (currdef->ref)].lattice_val != VARYING)
- SET_BIT (ssa_edges, DF_REF_REGNO (currdef->ref));
- values[DF_REF_REGNO (currdef->ref)].lattice_val = VARYING;
- }
-}
-
-/* Go through the expression, call the appropriate evaluation routines
- to attempt cprop */
-static void
-visit_expression (rtx insn, basic_block block)
-{
- rtx src, dest, set;
-
-
- /* Ugh. CALL_INSNs may end a basic block and have multiple edges
- leading out from them.
-
- Mark all the outgoing edges as executable, then fall into the
- normal processing below. */
- if (GET_CODE (insn) == CALL_INSN && block->end == insn)
- {
- edge curredge;
-
- for (curredge = block->succ; curredge;
- curredge = curredge->succ_next)
- {
- int index = EIE (curredge->src, curredge->dest);
-
- if (TEST_BIT (executable_edges, index))
- continue;
-
- SET_BIT (executable_edges, index);
- edge_info[index] = flow_edges;
- flow_edges = curredge;
- }
- }
-
- set = single_set (insn);
- if (! set)
- {
- defs_to_varying (insn);
- return;
- }
-
- src = SET_SRC (set);
- dest = SET_DEST (set);
-
- /* We may want to refine this some day. */
- if (GET_CODE (dest) != REG && dest != pc_rtx)
- {
- defs_to_varying (insn);
- return;
- }
-
- /* Hard registers are not put in SSA form and thus we must consider
- them varying. All the more reason to avoid hard registers in
- RTL until as late as possible in the compilation. */
- if (GET_CODE (dest) == REG && REGNO (dest) < FIRST_PSEUDO_REGISTER)
- {
- defs_to_varying (insn);
- return;
- }
-
- /* If this is assigning DEST to a constant, record that fact. */
- if (GET_CODE (src) == CONST_INT && GET_CODE (insn) == INSN)
- {
- unsigned int resultreg = REGNO (dest);
-
- values[resultreg].lattice_val = CONSTANT;
- values[resultreg].const_value = SET_SRC (PATTERN (insn));
- SET_BIT (ssa_edges, resultreg);
- }
-
- /* If this is a copy operation, then we can copy the lattice values. */
- else if (GET_CODE (src) == REG && GET_CODE (dest) == REG)
- {
- unsigned int old_value = REGNO (src);
- latticevalue old_lattice_value = values[old_value].lattice_val;
- unsigned int new_value = REGNO (dest);
-
- /* Unless the lattice value is going to change, don't bother
- adding the "new value" into the worklist. */
- if (values[new_value].lattice_val != old_lattice_value
- || values[new_value].const_value != values[old_value].const_value)
- SET_BIT (ssa_edges, new_value);
-
- /* Copy the old lattice node info into the new value lattice node. */
- values[new_value].lattice_val = old_lattice_value;
- values[new_value].const_value = values[old_value].const_value;
- }
-
- /* Handle jumps. */
- else if (GET_CODE (insn) == JUMP_INSN)
- {
- rtx x = pc_set (insn);
- if (GET_CODE (src) != IF_THEN_ELSE)
- {
- edge curredge;
-
- /* This is a computed jump, table jump, or an unconditional
- jump. For all these cases we want to mark all successor
- blocks as executable if they have not already been
- marked.
-
- One day we may try do better with switch tables and
- other computed jumps. */
- for (curredge = block->succ; curredge;
- curredge = curredge->succ_next)
- {
- int index = EIE (curredge->src, curredge->dest);
-
- if (TEST_BIT (executable_edges, index))
- continue;
-
- SET_BIT (executable_edges, index);
- edge_info[index] = flow_edges;
- flow_edges = curredge;
- }
- }
- else
- {
- edge curredge;
- enum rtx_code comparison_code;
- rtx comparison_src0;
- rtx comparison_src1;
-
- comparison_code = GET_CODE (XEXP (src, 0));
- comparison_src0 = XEXP (XEXP (src, 0), 0);
- comparison_src1 = XEXP (XEXP (src, 0), 1);
-
- /* If either operand is undefined, then there is nothing to
- do right now. If/when operands are later defined we will
- revaluate the condition and take the appropriate action. */
- if ((GET_CODE (comparison_src0) == REG
- && values[REGNO (comparison_src0)].lattice_val == UNDEFINED)
- || (GET_CODE (comparison_src1) == REG
- && values[REGNO (comparison_src1)].lattice_val == UNDEFINED))
- return;
-
- /* If either operand is varying, then we must consider all
- paths as executable. */
- if ((GET_CODE (comparison_src0) == REG
- && values[REGNO (comparison_src0)].lattice_val == VARYING)
- || (GET_CODE (comparison_src1) == REG
- && values[REGNO (comparison_src1)].lattice_val == VARYING))
- {
- for (curredge = block->succ; curredge;
- curredge = curredge->succ_next)
- {
- int index = EIE (curredge->src, curredge->dest);
-
- if (TEST_BIT (executable_edges, index))
- continue;
-
- SET_BIT (executable_edges, index);
- edge_info[index] = flow_edges;
- flow_edges = curredge;
- }
- return;
- }
-
- /* Try to simplify the comparison. */
- if (GET_CODE (comparison_src0) == REG
- && values[REGNO (comparison_src0)].lattice_val == CONSTANT)
- comparison_src0 = values[REGNO (comparison_src0)].const_value;
-
- if (GET_CODE (comparison_src1) == REG
- && values[REGNO (comparison_src1)].lattice_val == CONSTANT)
- comparison_src1 = values[REGNO (comparison_src1)].const_value;
-
- x = simplify_ternary_operation (IF_THEN_ELSE,
- VOIDmode,
- GET_MODE (XEXP (src, 0)),
- gen_rtx (comparison_code,
- GET_MODE (XEXP (src, 0)),
- comparison_src0,
- comparison_src1),
- XEXP (src, 1),
- XEXP (src, 2));
-
- /* Walk through all the outgoing edges from this block and see
- which (if any) we should mark as executable. */
- for (curredge = block->succ; curredge;
- curredge = curredge->succ_next)
- {
- int index = EIE (curredge->src, curredge->dest);
-
- if (TEST_BIT (executable_edges, index))
- continue;
-
- /* If we were unable to simplify the expression at this
- point, it's highly unlikely we'll be able to simplify
- it later. So consider all edges as executable if the
- expression did not simplify.
-
- If the expression simplified to (pc), then we know we
- will take the fall-thru edge, so mark it. Similarly,
- if the expression simplified to (label_ref ...), then
- we know the branch will be taken and we mark that
- edge as taken. */
- if (!x
- || (x == pc_rtx
- && (curredge->flags & EDGE_FALLTHRU))
- || (GET_CODE (x) == LABEL_REF
- && ! (curredge->flags & EDGE_FALLTHRU)))
- {
- SET_BIT (executable_edges, index);
- edge_info[index] = flow_edges;
- flow_edges = curredge;
- }
- }
- }
- }
- else if (!PHI_NODE_P (insn))
- {
- rtx simplified = NULL;
-
- /* We've got some kind of INSN. If it's simple, try to evaluate
- it and record the results.
-
- We already know this insn is a single_set and that it sets
- a pseudo register. So we just need to extract the source
- arguments, simplify them to constants if possible, then
- simplify the expression as a whole if possible. */
- switch (GET_RTX_CLASS (GET_CODE (src)))
- {
- case '<':
- {
- rtx src0 = XEXP (src, 0);
- rtx src1 = XEXP (src, 1);
- enum machine_mode mode;
-
- /* If either is undefined, then the result is undefined. */
- if ((GET_CODE (src0) == REG
- && values[REGNO (src0)].lattice_val == UNDEFINED)
- || (GET_CODE (src1) == REG
- && values[REGNO (src1)].lattice_val == UNDEFINED))
- {
- defs_to_undefined (insn);
- break;
- }
-
- /* Determine the mode for the operation before we simplify
- our arguments to constants. */
- mode = GET_MODE (src0);
- if (mode == VOIDmode)
- mode = GET_MODE (src1);
-
- /* Simplify source operands to whatever known values they
- may have. */
- if (GET_CODE (src0) == REG
- && values[REGNO (src0)].lattice_val == CONSTANT)
- src0 = values[REGNO (src0)].const_value;
-
- if (GET_CODE (src1) == REG
- && values[REGNO (src1)].lattice_val == CONSTANT)
- src1 = values[REGNO (src1)].const_value;
-
- /* See if the simplifier can determine if this operation
- computes a constant value. */
- simplified = simplify_relational_operation (GET_CODE (src),
- mode, src0, src1);
- break;
-
- }
-
- case '1':
- {
- rtx src0 = XEXP (src, 0);
- enum machine_mode mode0 = GET_MODE (src0);
-
- /* If the operand is undefined, then the result is undefined. */
- if (GET_CODE (src0) == REG
- && values[REGNO (src0)].lattice_val == UNDEFINED)
- {
- defs_to_undefined (insn);
- break;
- }
-
- /* Simplify source operands to whatever known values they
- may have. */
- if (GET_CODE (src0) == REG
- && values[REGNO (src0)].lattice_val == CONSTANT)
- src0 = values[REGNO (src0)].const_value;
-
- /* See if the simplifier can determine if this operation
- computes a constant value. */
- simplified = simplify_unary_operation (GET_CODE (src),
- GET_MODE (src),
- src0,
- mode0);
- break;
- }
-
- case '2':
- case 'c':
- {
- rtx src0 = XEXP (src, 0);
- rtx src1 = XEXP (src, 1);
-
- /* If either is undefined, then the result is undefined. */
- if ((GET_CODE (src0) == REG
- && values[REGNO (src0)].lattice_val == UNDEFINED)
- || (GET_CODE (src1) == REG
- && values[REGNO (src1)].lattice_val == UNDEFINED))
- {
- defs_to_undefined (insn);
- break;
- }
-
- /* Simplify source operands to whatever known values they
- may have. */
- if (GET_CODE (src0) == REG
- && values[REGNO (src0)].lattice_val == CONSTANT)
- src0 = values[REGNO (src0)].const_value;
-
- if (GET_CODE (src1) == REG
- && values[REGNO (src1)].lattice_val == CONSTANT)
- src1 = values[REGNO (src1)].const_value;
-
- /* See if the simplifier can determine if this operation
- computes a constant value. */
- simplified = simplify_binary_operation (GET_CODE (src),
- GET_MODE (src),
- src0, src1);
- break;
- }
-
- case '3':
- case 'b':
- {
- rtx src0 = XEXP (src, 0);
- rtx src1 = XEXP (src, 1);
- rtx src2 = XEXP (src, 2);
-
- /* If either is undefined, then the result is undefined. */
- if ((GET_CODE (src0) == REG
- && values[REGNO (src0)].lattice_val == UNDEFINED)
- || (GET_CODE (src1) == REG
- && values[REGNO (src1)].lattice_val == UNDEFINED)
- || (GET_CODE (src2) == REG
- && values[REGNO (src2)].lattice_val == UNDEFINED))
- {
- defs_to_undefined (insn);
- break;
- }
-
- /* Simplify source operands to whatever known values they
- may have. */
- if (GET_CODE (src0) == REG
- && values[REGNO (src0)].lattice_val == CONSTANT)
- src0 = values[REGNO (src0)].const_value;
-
- if (GET_CODE (src1) == REG
- && values[REGNO (src1)].lattice_val == CONSTANT)
- src1 = values[REGNO (src1)].const_value;
-
- if (GET_CODE (src2) == REG
- && values[REGNO (src2)].lattice_val == CONSTANT)
- src2 = values[REGNO (src2)].const_value;
-
- /* See if the simplifier can determine if this operation
- computes a constant value. */
- simplified = simplify_ternary_operation (GET_CODE (src),
- GET_MODE (src),
- GET_MODE (src),
- src0, src1, src2);
- break;
- }
-
- default:
- defs_to_varying (insn);
- }
-
- if (simplified && GET_CODE (simplified) == CONST_INT)
- {
- if (values[REGNO (dest)].lattice_val != CONSTANT
- || values[REGNO (dest)].const_value != simplified)
- SET_BIT (ssa_edges, REGNO (dest));
-
- values[REGNO (dest)].lattice_val = CONSTANT;
- values[REGNO (dest)].const_value = simplified;
- }
- else
- defs_to_varying (insn);
- }
-}
-
-/* Iterate over the FLOW_EDGES work list. Simulate the target block
- for each edge. */
-static void
-examine_flow_edges (void)
-{
- while (flow_edges != NULL)
- {
- basic_block succ_block;
- rtx curr_phi_node;
-
- /* Pull the next block to simulate off the worklist. */
- succ_block = flow_edges->dest;
- flow_edges = edge_info[EIE (flow_edges->src, flow_edges->dest)];
-
- /* There is nothing to do for the exit block. */
- if (succ_block == EXIT_BLOCK_PTR)
- continue;
-
- /* Always simulate PHI nodes, even if we have simulated this block
- before. Note that all PHI nodes are consecutive within a block. */
- for (curr_phi_node = first_insn_after_basic_block_note (succ_block);
- PHI_NODE_P (curr_phi_node);
- curr_phi_node = NEXT_INSN (curr_phi_node))
- visit_phi_node (curr_phi_node, succ_block);
-
- /* If this is the first time we've simulated this block, then we
- must simulate each of its insns. */
- if (!TEST_BIT (executable_blocks, succ_block->index))
- {
- rtx currinsn;
- edge succ_edge = succ_block->succ;
-
- /* Note that we have simulated this block. */
- SET_BIT (executable_blocks, succ_block->index);
-
- /* Simulate each insn within the block. */
- currinsn = succ_block->head;
- while (currinsn != succ_block->end)
- {
- if (INSN_P (currinsn))
- visit_expression (currinsn, succ_block);
-
- currinsn = NEXT_INSN (currinsn);
- }
-
- /* Don't forget the last insn in the block. */
- if (INSN_P (currinsn))
- visit_expression (currinsn, succ_block);
-
- /* If we haven't looked at the next block, and it has a
- single successor, add it onto the worklist. This is because
- if we only have one successor, we know it gets executed,
- so we don't have to wait for cprop to tell us. */
- if (succ_edge != NULL
- && succ_edge->succ_next == NULL
- && !TEST_BIT (executable_edges,
- EIE (succ_edge->src, succ_edge->dest)))
- {
- SET_BIT (executable_edges,
- EIE (succ_edge->src, succ_edge->dest));
- edge_info[EIE (succ_edge->src, succ_edge->dest)] = flow_edges;
- flow_edges = succ_edge;
- }
- }
- }
-}
-
-/* Follow the def-use chains for each definition on the worklist and
- simulate the uses of the definition. */
-
-static void
-follow_def_use_chains (void)
-{
- /* Iterate over all the entries on the SSA_EDGES worklist. */
- while (sbitmap_first_set_bit (ssa_edges) >= 0)
- {
- int member;
- struct df_link *curruse;
-
- /* Pick an entry off the worklist (it does not matter which
- entry we pick). */
- member = sbitmap_first_set_bit (ssa_edges);
- RESET_BIT (ssa_edges, member);
-
- /* Iterate through all the uses of this entry. */
- for (curruse = df_analyzer->regs[member].uses; curruse;
- curruse = curruse->next)
- {
- rtx useinsn;
-
- useinsn = DF_REF_INSN (curruse->ref);
- if (PHI_NODE_P (useinsn))
- {
- if (TEST_BIT (executable_blocks, BLOCK_NUM (useinsn)))
- visit_phi_node (useinsn, BLOCK_FOR_INSN (useinsn));
- }
- else
- {
- if (TEST_BIT (executable_blocks, BLOCK_NUM (useinsn)))
- visit_expression (useinsn, BLOCK_FOR_INSN (useinsn));
- }
- }
- }
-}
-
-/* Examine each edge to see if we were able to prove any were
- not executable.
-
- If an edge is not executable, then we can remove its alternative
- in PHI nodes as the destination of the edge, we can simplify the
- conditional branch at the source of the edge, and we can remove
- the edge from the CFG. Note we do not delete unreachable blocks
- yet as the DF analyzer can not deal with that yet. */
-static void
-optimize_unexecutable_edges (struct edge_list *edges,
- sbitmap executable_edges)
-{
- int i;
- basic_block bb;
-
- for (i = 0; i < NUM_EDGES (edges); i++)
- {
- if (!TEST_BIT (executable_edges, i))
- {
- edge edge = INDEX_EDGE (edges, i);
-
- if (edge->flags & EDGE_ABNORMAL)
- continue;
-
- /* We found an edge that is not executable. First simplify
- the PHI nodes in the target block. */
- if (edge->dest != EXIT_BLOCK_PTR)
- {
- rtx insn = first_insn_after_basic_block_note (edge->dest);
-
- while (PHI_NODE_P (insn))
- {
- remove_phi_alternative (PATTERN (insn), edge->src);
- if (rtl_dump_file)
- fprintf (rtl_dump_file,
- "Removing alternative for bb %d of phi %d\n",
- edge->src->index, SSA_NAME (PATTERN (insn)));
- insn = NEXT_INSN (insn);
- }
- }
- if (rtl_dump_file)
- fprintf (rtl_dump_file,
- "Removing unexecutable edge from %d to %d\n",
- edge->src->index, edge->dest->index);
- /* Since the edge was not executable, remove it from the CFG. */
- remove_edge (edge);
- }
- }
-
- /* We have removed all the unexecutable edges from the CFG. Fix up
- the conditional jumps at the end of any affected block.
-
- We have three cases to deal with:
-
- a. Both outgoing edges are not executable. This happens if the
- source block is not reachable. We will deal with this by
- deleting all the insns in the block later.
-
- b. The fall-thru edge is not executable. In this case we
- change the conditional jump into an unconditional jump and
- add a BARRIER after the unconditional jump. Note that since
- we are working on generic RTL we can change the jump in-place
- instead of dealing with the headache of reemitting the jump.
-
- c. The branch taken edge is not executable. In this case
- we turn the jump into (set (pc) (pc)) which is a nop-jump
- and we will remove the unrecognizable insn later.
-
- In cases B & C we are removing uses of registers, so make sure
- to note those changes for the DF analyzer. */
-
- FOR_EACH_BB (bb)
- {
- rtx insn = bb->end;
- edge edge = bb->succ;
-
- /* If we have no predecessors, then this block is unreachable and
- will be cleaned up when we remove unreachable blocks. */
- if (bb->pred == NULL || GET_CODE (insn) != JUMP_INSN)
- continue;
-
- /* If this block ends in a conditional jump, but only has one
- successor, then the jump needs adjustment. */
- if (condjump_p (insn) && ! simplejump_p (insn)
- && bb->succ && bb->succ->succ_next == NULL)
- {
- /* If the fallthru edge is the executable edge, then turn
- this jump into a nop jump, otherwise make it an unconditional
- jump to its target. */
- if (edge->flags & EDGE_FALLTHRU)
- {
- PUT_CODE (insn, NOTE);
- NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
- }
- else
- {
- SET_SRC (PATTERN (insn)) = gen_rtx_LABEL_REF (Pmode,
- JUMP_LABEL (insn));
- emit_barrier_after (insn);
- INSN_CODE (insn) = -1;
- }
-
- /* Inform the DF analyzer that this insn changed. */
- df_insn_modify (df_analyzer, BLOCK_FOR_INSN (insn), insn);
- }
- }
-}
-
-/* Perform substitution of known values for pseudo registers.
-
- ??? Note we do not do simplifications or constant folding here, it
- is unlikely that any significant simplifications can be done here
- anyway. Consider that if the simplification would result in an
- expression that produces a constant value that the value would
- have been discovered and recorded already.
-
- We perform two transformations. First, we initialize pseudos to their
- known constant values at their definition point. Second, we try to
- replace uses with the known constant value. */
-
-static void
-ssa_ccp_substitute_constants (void)
-{
- unsigned int i;
-
- for (i = FIRST_PSEUDO_REGISTER; i < VARRAY_SIZE (ssa_definition); i++)
- {
- if (values[i].lattice_val == CONSTANT)
- {
- rtx def = VARRAY_RTX (ssa_definition, i);
- rtx set = single_set (def);
- struct df_link *curruse;
-
- if (! set)
- continue;
-
- /* Do not try to simplify PHI nodes down to a constant load.
- That will be done later as we translate out of SSA. Also,
- doing that here could violate the rule that all PHI nodes
- are consecutive at the start of the basic block.
-
- Don't do anything to nodes that were already sets to
- constants. */
- if (! PHI_NODE_P (def)
- && ! ((GET_CODE (def) == INSN
- && GET_CODE (SET_SRC (set)) == CONST_INT)))
- {
- if (rtl_dump_file)
- fprintf (rtl_dump_file,
- "Register %d is now set to a constant\n",
- SSA_NAME (PATTERN (def)));
- SET_SRC (set) = values[i].const_value;
- INSN_CODE (def) = -1;
- df_insn_modify (df_analyzer, BLOCK_FOR_INSN (def), def);
- }
-
- /* Iterate through all the uses of this entry and try replacements
- there too. Note it is not particularly profitable to try
- and fold/simplify expressions here as most of the common
- cases were handled above. */
- for (curruse = df_analyzer->regs[i].uses;
- curruse;
- curruse = curruse->next)
- {
- rtx useinsn;
-
- useinsn = DF_REF_INSN (curruse->ref);
-
- if (!INSN_DELETED_P (useinsn)
- && ! (GET_CODE (useinsn) == NOTE
- && NOTE_LINE_NUMBER (useinsn) == NOTE_INSN_DELETED)
- && (GET_CODE (useinsn) == INSN
- || GET_CODE (useinsn) == JUMP_INSN))
- {
-
- if (validate_replace_src (regno_reg_rtx [i],
- values[i].const_value,
- useinsn))
- {
- if (rtl_dump_file)
- fprintf (rtl_dump_file,
- "Register %d in insn %d replaced with constant\n",
- i, INSN_UID (useinsn));
- INSN_CODE (useinsn) = -1;
- df_insn_modify (df_analyzer,
- BLOCK_FOR_INSN (useinsn),
- useinsn);
- }
-
- }
- }
- }
- }
-}
-
-/* Now find all unreachable basic blocks. All the insns in those
- blocks are unreachable, so delete them and mark any necessary
- updates for the DF analyzer. */
-
-static void
-ssa_ccp_df_delete_unreachable_insns (void)
-{
- basic_block b;
-
- /* Use the CFG to find all the reachable blocks. */
- find_unreachable_blocks ();
-
- /* Now we know what blocks are not reachable. Mark all the insns
- in those blocks as deleted for the DF analyzer. We'll let the
- normal flow code actually remove the unreachable blocks. */
- FOR_EACH_BB_REVERSE (b)
- {
- if (!(b->flags & BB_REACHABLE))
- {
- rtx start = b->head;
- rtx end = b->end;
- rtx tmp;
-
- /* Include any jump table following the basic block. */
- end = b->end;
- if (tablejump_p (end, NULL, &tmp))
- end = tmp;
-
- while (1)
- {
- rtx next = NEXT_INSN (start);
-
- if (GET_CODE (start) == INSN
- || GET_CODE (start) == CALL_INSN
- || GET_CODE (start) == JUMP_INSN)
- df_insn_delete (df_analyzer, BLOCK_FOR_INSN (start), start);
-
- if (start == end)
- break;
- start = next;
- }
- }
- }
-}
-
-
-/* Main entry point for SSA Conditional Constant Propagation.
-
- Long term it should accept as input the specific flow graph to
- operate on so that it can be called for sub-graphs. */
-
-void
-ssa_const_prop (void)
-{
- unsigned int i;
- edge curredge;
-
- /* We need alias analysis (for what?) */
- init_alias_analysis ();
-
- df_analyzer = df_init ();
- df_analyse (df_analyzer, 0,
- DF_RD_CHAIN | DF_RU_CHAIN | DF_REG_INFO | DF_HARD_REGS);
-
- /* Perform a quick and dirty dead code elimination pass. This is not
- as aggressive as it could be, but it's good enough to clean up a
- lot of unwanted junk and it is fast. */
- ssa_fast_dce (df_analyzer);
-
- /* Build an edge list from the CFG. */
- edges = create_edge_list ();
-
- /* Initialize the values array with everything as undefined. */
- values = xmalloc (VARRAY_SIZE (ssa_definition) * sizeof (value));
- for (i = 0; i < VARRAY_SIZE (ssa_definition); i++)
- {
- if (i < FIRST_PSEUDO_REGISTER)
- values[i].lattice_val = VARYING;
- else
- values[i].lattice_val = UNDEFINED;
- values[i].const_value = NULL;
- }
-
- ssa_edges = sbitmap_alloc (VARRAY_SIZE (ssa_definition));
- sbitmap_zero (ssa_edges);
-
- executable_blocks = sbitmap_alloc (last_basic_block);
- sbitmap_zero (executable_blocks);
-
- executable_edges = sbitmap_alloc (NUM_EDGES (edges));
- sbitmap_zero (executable_edges);
-
- edge_info = xmalloc (NUM_EDGES (edges) * sizeof (edge));
- flow_edges = ENTRY_BLOCK_PTR->succ;
-
- /* Add the successors of the entry block to the edge worklist. That
- is enough of a seed to get SSA-CCP started. */
- for (curredge = ENTRY_BLOCK_PTR->succ; curredge;
- curredge = curredge->succ_next)
- {
- int index = EIE (curredge->src, curredge->dest);
- SET_BIT (executable_edges, index);
- edge_info[index] = curredge->succ_next;
- }
-
- /* Iterate until until the worklists are empty. */
- do
- {
- examine_flow_edges ();
- follow_def_use_chains ();
- }
- while (flow_edges != NULL);
-
- /* Now perform substitutions based on the known constant values. */
- ssa_ccp_substitute_constants ();
-
- /* Remove unexecutable edges from the CFG and make appropriate
- adjustments to PHI nodes. */
- optimize_unexecutable_edges (edges, executable_edges);
-
- /* Now remove all unreachable insns and update the DF information.
- as appropriate. */
- ssa_ccp_df_delete_unreachable_insns ();
-
-#if 0
- /* The DF analyzer expects the number of blocks to remain constant,
- so we can't remove unreachable blocks.
-
- Code the DF analyzer calls expects there to be no unreachable
- blocks in the CFG. So we can't leave unreachable blocks in the
- CFG.
-
- So, there is no way to do an incremental update of the DF data
- at this point. */
- df_analyse (df_analyzer, 0,
- DF_RD_CHAIN | DF_RU_CHAIN | DF_REG_INFO | DF_HARD_REGS);
-#endif
-
- /* Clean up any dead code exposed by SSA-CCP, do this after updating
- the dataflow information! */
- ssa_fast_dce (df_analyzer);
-
- free (values);
- values = NULL;
-
- free (edge_info);
- edge_info = NULL;
-
- sbitmap_free (executable_blocks);
- executable_blocks = NULL;
-
- sbitmap_free (ssa_edges);
- ssa_edges = NULL;
-
- free_edge_list (edges);
- edges = NULL;
-
- sbitmap_free (executable_edges);
- executable_edges = NULL;
-
- df_finish (df_analyzer);
- end_alias_analysis ();
-}
-
-static int
-mark_references (rtx *current_rtx, void *data)
-{
- rtx x = *current_rtx;
- sbitmap worklist = (sbitmap) data;
-
- if (x == NULL_RTX)
- return 0;
-
- if (GET_CODE (x) == SET)
- {
- rtx dest = SET_DEST (x);
-
- if (GET_CODE (dest) == STRICT_LOW_PART
- || GET_CODE (dest) == SUBREG
- || GET_CODE (dest) == SIGN_EXTRACT
- || GET_CODE (dest) == ZERO_EXTRACT)
- {
- rtx reg;
-
- reg = dest;
-
- while (GET_CODE (reg) == STRICT_LOW_PART
- || GET_CODE (reg) == SUBREG
- || GET_CODE (reg) == SIGN_EXTRACT
- || GET_CODE (reg) == ZERO_EXTRACT)
- reg = XEXP (reg, 0);
-
- if (GET_CODE (reg) == REG)
- SET_BIT (worklist, REGNO (reg));
- }
-
- if (GET_CODE (dest) == REG)
- {
- for_each_rtx (&SET_SRC (x), mark_references, data);
- return -1;
- }
-
- return 0;
- }
- else if (GET_CODE (x) == REG)
- {
- SET_BIT (worklist, REGNO (x));
- return -1;
- }
- else if (GET_CODE (x) == CLOBBER)
- return -1;
- else
- return 0;
-}
-
-static void
-ssa_fast_dce (struct df *df)
-{
- sbitmap worklist = sbitmap_alloc (VARRAY_SIZE (ssa_definition));
- sbitmap_ones (worklist);
-
- /* Iterate on the worklist until there's no definitions left to
- examine. */
- while (sbitmap_first_set_bit (worklist) >= 0)
- {
- struct df_link *curruse;
- int reg, found_use;
-
- /* Remove an item from the worklist. */
- reg = sbitmap_first_set_bit (worklist);
- RESET_BIT (worklist, reg);
-
- /* We never consider deleting assignments to hard regs or things
- which do not have SSA definitions, or things we have already
- deleted, or things with unusual side effects. */
- if (reg < FIRST_PSEUDO_REGISTER
- || ! VARRAY_RTX (ssa_definition, reg)
- || INSN_DELETED_P (VARRAY_RTX (ssa_definition, reg))
- || (GET_CODE (VARRAY_RTX (ssa_definition, reg)) == NOTE
- && (NOTE_LINE_NUMBER (VARRAY_RTX (ssa_definition, reg))
- == NOTE_INSN_DELETED))
- || side_effects_p (PATTERN (VARRAY_RTX (ssa_definition, reg))))
- continue;
-
- /* Iterate over the uses of this register. If we can not find
- any uses that have not been deleted, then the definition of
- this register is dead. */
- found_use = 0;
- for (curruse = df->regs[reg].uses; curruse; curruse = curruse->next)
- {
- if (curruse->ref
- && DF_REF_INSN (curruse->ref)
- && ! INSN_DELETED_P (DF_REF_INSN (curruse->ref))
- && ! (GET_CODE (DF_REF_INSN (curruse->ref)) == NOTE
- && (NOTE_LINE_NUMBER (DF_REF_INSN (curruse->ref))
- == NOTE_INSN_DELETED))
- && DF_REF_INSN (curruse->ref) != VARRAY_RTX (ssa_definition, reg))
- {
- found_use = 1;
- break;
- }
- }
-
- /* If we did not find a use of this register, then the definition
- of this register is dead. */
-
- if (! found_use)
- {
- rtx def = VARRAY_RTX (ssa_definition, reg);
-
- /* Add all registers referenced by INSN to the work
- list. */
- for_each_rtx (&PATTERN (def), mark_references, worklist);
-
- /* Inform the analyzer that this insn is going to be
- deleted. */
- df_insn_delete (df, BLOCK_FOR_INSN (def), def);
-
- VARRAY_RTX (ssa_definition, reg) = NULL;
- }
- }
-
- sbitmap_free (worklist);
-
- /* Update the use-def chains in the df_analyzer as needed. */
- df_analyse (df_analyzer, 0,
- DF_RD_CHAIN | DF_RU_CHAIN | DF_REG_INFO | DF_HARD_REGS);
-}
+++ /dev/null
-/* Dead-code elimination pass for the GNU compiler.
- Copyright (C) 2000, 2001, 2002, 2003 Free Software Foundation, Inc.
- Written by Jeffrey D. Oldham <oldham@codesourcery.com>.
-
-This file is part of GCC.
-
-GCC is free software; you can redistribute it and/or modify it under
-the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
-version.
-
-GCC is distributed in the hope that it will be useful, but WITHOUT ANY
-WARRANTY; without even the implied warranty of MERCHANTABILITY or
-FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
-for more details.
-
-You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
-
-/* Dead-code elimination is the removal of instructions which have no
- impact on the program's output. "Dead instructions" have no impact
- on the program's output, while "necessary instructions" may have
- impact on the output.
-
- The algorithm consists of three phases:
- 1) marking as necessary all instructions known to be necessary,
- e.g., writing a value to memory,
- 2) propagating necessary instructions, e.g., the instructions
- giving values to operands in necessary instructions, and
- 3) removing dead instructions (except replacing dead conditionals
- with unconditional jumps).
-
- Side Effects:
- The last step can require adding labels, deleting insns, and
- modifying basic block structures. Some conditional jumps may be
- converted to unconditional jumps so the control-flow graph may be
- out-of-date.
-
- Edges from some infinite loops to the exit block can be added to
- the control-flow graph, but will be removed after this pass is
- complete.
-
- It Does Not Perform:
- We decided to not simultaneously perform jump optimization and dead
- loop removal during dead-code elimination. Thus, all jump
- instructions originally present remain after dead-code elimination
- but 1) unnecessary conditional jump instructions are changed to
- unconditional jump instructions and 2) all unconditional jump
- instructions remain.
-
- Assumptions:
- 1) SSA has been performed.
- 2) The basic block and control-flow graph structures are accurate.
- 3) The flow graph permits constructing an edge_list.
- 4) note rtxes should be saved.
-
- Unfinished:
- When replacing unnecessary conditional jumps with unconditional
- jumps, the control-flow graph is not updated. It should be.
-
- References:
- Building an Optimizing Compiler
- Robert Morgan
- Butterworth-Heinemann, 1998
- Section 8.9
-*/
-
-#include "config.h"
-#include "system.h"
-#include "coretypes.h"
-#include "tm.h"
-
-#include "rtl.h"
-#include "hard-reg-set.h"
-#include "basic-block.h"
-#include "ssa.h"
-#include "insn-config.h"
-#include "recog.h"
-#include "output.h"
-
-\f
-/* A map from blocks to the edges on which they are control dependent. */
-typedef struct {
- /* A dynamically allocated array. The Nth element corresponds to
- the block with index N + 2. The Ith bit in the bitmap is set if
- that block is dependent on the Ith edge. */
- bitmap *data;
- /* The number of elements in the array. */
- int length;
-} control_dependent_block_to_edge_map_s, *control_dependent_block_to_edge_map;
-
-/* Local function prototypes. */
-static control_dependent_block_to_edge_map control_dependent_block_to_edge_map_create
- (size_t num_basic_blocks);
-static void set_control_dependent_block_to_edge_map_bit
- (control_dependent_block_to_edge_map c, basic_block bb, int edge_index);
-static void control_dependent_block_to_edge_map_free
- (control_dependent_block_to_edge_map c);
-static void find_all_control_dependences
- (struct edge_list *el, dominance_info pdom,
- control_dependent_block_to_edge_map cdbte);
-static void find_control_dependence
- (struct edge_list *el, int edge_index, dominance_info pdom,
- control_dependent_block_to_edge_map cdbte);
-static basic_block find_pdom (dominance_info pdom, basic_block block);
-static int inherently_necessary_register_1 (rtx *current_rtx, void *data);
-static int inherently_necessary_register (rtx current_rtx);
-static int find_inherently_necessary (rtx current_rtx);
-static int propagate_necessity_through_operand (rtx *current_rtx, void *data);
-static void note_inherently_necessary_set (rtx, rtx, void *);
-\f
-/* Unnecessary insns are indicated using insns' in_struct bit. */
-
-/* Indicate INSN is dead-code; returns nothing. */
-#define KILL_INSN(INSN) INSN_DEAD_CODE_P(INSN) = 1
-/* Indicate INSN is necessary, i.e., not dead-code; returns nothing. */
-#define RESURRECT_INSN(INSN) INSN_DEAD_CODE_P(INSN) = 0
-/* Return nonzero if INSN is unnecessary. */
-#define UNNECESSARY_P(INSN) INSN_DEAD_CODE_P(INSN)
-static void mark_all_insn_unnecessary (void);
-/* Execute CODE with free variable INSN for all unnecessary insns in
- an unspecified order, producing no output. */
-#define EXECUTE_IF_UNNECESSARY(INSN, CODE) \
-{ \
- rtx INSN; \
- \
- for (INSN = get_insns (); INSN != NULL_RTX; INSN = NEXT_INSN (INSN)) \
- if (INSN_P (insn) && INSN_DEAD_CODE_P (INSN)) \
- { \
- CODE; \
- } \
-}
-
-/* Find the label beginning block BB. */
-static rtx find_block_label (basic_block bb);
-/* Remove INSN, updating its basic block structure. */
-static void delete_insn_bb (rtx insn);
-\f
-/* Recording which blocks are control dependent on which edges. We
- expect each block to be control dependent on very few edges so we
- use a bitmap for each block recording its edges. An array holds
- the bitmap. Its position 0 entry holds the bitmap for block
- INVALID_BLOCK+1 so that all blocks, including the entry and exit
- blocks can participate in the data structure. */
-
-/* Create a control_dependent_block_to_edge_map, given the number
- NUM_BASIC_BLOCKS of non-entry, non-exit basic blocks, e.g.,
- n_basic_blocks. This memory must be released using
- control_dependent_block_to_edge_map_free (). */
-
-static control_dependent_block_to_edge_map
-control_dependent_block_to_edge_map_create (size_t num_basic_blocks)
-{
- int i;
- control_dependent_block_to_edge_map c
- = xmalloc (sizeof (control_dependent_block_to_edge_map_s));
- c->length = num_basic_blocks - (INVALID_BLOCK+1);
- c->data = xmalloc ((size_t) c->length*sizeof (bitmap));
- for (i = 0; i < c->length; ++i)
- c->data[i] = BITMAP_XMALLOC ();
-
- return c;
-}
-
-/* Indicate block BB is control dependent on an edge with index
- EDGE_INDEX in the mapping C of blocks to edges on which they are
- control-dependent. */
-
-static void
-set_control_dependent_block_to_edge_map_bit (control_dependent_block_to_edge_map c,
- basic_block bb, int edge_index)
-{
- if (bb->index - (INVALID_BLOCK+1) >= c->length)
- abort ();
-
- bitmap_set_bit (c->data[bb->index - (INVALID_BLOCK+1)],
- edge_index);
-}
-
-/* Execute CODE for each edge (given number EDGE_NUMBER within the
- CODE) for which the block containing INSN is control dependent,
- returning no output. CDBTE is the mapping of blocks to edges on
- which they are control-dependent. */
-
-#define EXECUTE_IF_CONTROL_DEPENDENT(CDBTE, INSN, EDGE_NUMBER, CODE) \
- EXECUTE_IF_SET_IN_BITMAP \
- (CDBTE->data[BLOCK_NUM (INSN) - (INVALID_BLOCK+1)], 0, \
- EDGE_NUMBER, CODE)
-
-/* Destroy a control_dependent_block_to_edge_map C. */
-
-static void
-control_dependent_block_to_edge_map_free (control_dependent_block_to_edge_map c)
-{
- int i;
- for (i = 0; i < c->length; ++i)
- BITMAP_XFREE (c->data[i]);
- free (c);
-}
-
-/* Record all blocks' control dependences on all edges in the edge
- list EL, ala Morgan, Section 3.6. The mapping PDOM of blocks to
- their postdominators are used, and results are stored in CDBTE,
- which should be empty. */
-
-static void
-find_all_control_dependences (struct edge_list *el, dominance_info pdom,
- control_dependent_block_to_edge_map cdbte)
-{
- int i;
-
- for (i = 0; i < NUM_EDGES (el); ++i)
- find_control_dependence (el, i, pdom, cdbte);
-}
-
-/* Determine all blocks' control dependences on the given edge with
- edge_list EL index EDGE_INDEX, ala Morgan, Section 3.6. The
- mapping PDOM of blocks to their postdominators are used, and
- results are stored in CDBTE, which is assumed to be initialized
- with zeros in each (block b', edge) position. */
-
-static void
-find_control_dependence (struct edge_list *el, int edge_index,
- dominance_info pdom,
- control_dependent_block_to_edge_map cdbte)
-{
- basic_block current_block;
- basic_block ending_block;
-
- if (INDEX_EDGE_PRED_BB (el, edge_index) == EXIT_BLOCK_PTR)
- abort ();
- ending_block =
- (INDEX_EDGE_PRED_BB (el, edge_index) == ENTRY_BLOCK_PTR)
- ? ENTRY_BLOCK_PTR->next_bb
- : find_pdom (pdom, INDEX_EDGE_PRED_BB (el, edge_index));
-
- for (current_block = INDEX_EDGE_SUCC_BB (el, edge_index);
- current_block != ending_block && current_block != EXIT_BLOCK_PTR;
- current_block = find_pdom (pdom, current_block))
- {
- set_control_dependent_block_to_edge_map_bit (cdbte,
- current_block,
- edge_index);
- }
-}
-\f
-/* Find the immediate postdominator PDOM of the specified basic block
- BLOCK. This function is necessary because some blocks have
- negative numbers. */
-
-static basic_block
-find_pdom (dominance_info pdom, basic_block block)
-{
- if (!block)
- abort ();
- if (block->index == INVALID_BLOCK)
- abort ();
-
- if (block == ENTRY_BLOCK_PTR)
- return ENTRY_BLOCK_PTR->next_bb;
- else if (block == EXIT_BLOCK_PTR)
- return EXIT_BLOCK_PTR;
- else
- {
- basic_block bb = get_immediate_dominator (pdom, block);
- if (!bb)
- return EXIT_BLOCK_PTR;
- return bb;
- }
-}
-
-/* Determine if the given CURRENT_RTX uses a hard register not
- converted to SSA. Returns nonzero only if it uses such a hard
- register. DATA is not used.
-
- The program counter (PC) is not considered inherently necessary
- since code should be position-independent and thus not depend on
- particular PC values. */
-
-static int
-inherently_necessary_register_1 (rtx *current_rtx,
- void *data ATTRIBUTE_UNUSED)
-{
- rtx x = *current_rtx;
-
- if (x == NULL_RTX)
- return 0;
- switch (GET_CODE (x))
- {
- case CLOBBER:
- /* Do not traverse the rest of the clobber. */
- return -1;
- break;
- case PC:
- return 0;
- break;
- case REG:
- if (CONVERT_REGISTER_TO_SSA_P (REGNO (x)) || x == pc_rtx)
- return 0;
- else
- return !0;
- break;
- default:
- return 0;
- break;
- }
-}
-
-/* Return nonzero if the insn CURRENT_RTX is inherently necessary. */
-
-static int
-inherently_necessary_register (rtx current_rtx)
-{
- return for_each_rtx (¤t_rtx,
- &inherently_necessary_register_1, NULL);
-}
-
-
-/* Called via note_stores for each store in an insn. Note whether
- or not a particular store is inherently necessary. Store a
- nonzero value in inherently_necessary_p if such a store is found. */
-
-static void
-note_inherently_necessary_set (rtx dest, rtx set ATTRIBUTE_UNUSED, void *data)
-{
- int *inherently_necessary_set_p = (int *) data;
-
- while (GET_CODE (dest) == SUBREG
- || GET_CODE (dest) == STRICT_LOW_PART
- || GET_CODE (dest) == ZERO_EXTRACT
- || GET_CODE (dest) == SIGN_EXTRACT)
- dest = XEXP (dest, 0);
-
- if (GET_CODE (dest) == MEM
- || GET_CODE (dest) == UNSPEC
- || GET_CODE (dest) == UNSPEC_VOLATILE)
- *inherently_necessary_set_p = 1;
-}
-
-/* Mark X as inherently necessary if appropriate. For example,
- function calls and storing values into memory are inherently
- necessary. This function is to be used with for_each_rtx ().
- Return nonzero iff inherently necessary. */
-
-static int
-find_inherently_necessary (rtx x)
-{
- if (x == NULL_RTX)
- return 0;
- else if (inherently_necessary_register (x))
- return !0;
- else
- switch (GET_CODE (x))
- {
- case CALL_INSN:
- case BARRIER:
- case PREFETCH:
- return !0;
- case CODE_LABEL:
- case NOTE:
- return 0;
- case JUMP_INSN:
- return JUMP_TABLE_DATA_P (x) || computed_jump_p (x) != 0;
- case INSN:
- {
- int inherently_necessary_set = 0;
- note_stores (PATTERN (x),
- note_inherently_necessary_set,
- &inherently_necessary_set);
-
- /* If we found an inherently necessary set or an asm
- instruction, then we consider this insn inherently
- necessary. */
- return (inherently_necessary_set
- || GET_CODE (PATTERN (x)) == ASM_INPUT
- || asm_noperands (PATTERN (x)) >= 0);
- }
- default:
- /* Found an impossible insn type. */
- abort ();
- break;
- }
-}
-
-/* Propagate necessity through REG and SUBREG operands of CURRENT_RTX.
- This function is called with for_each_rtx () on necessary
- instructions. The DATA must be a varray of unprocessed
- instructions. */
-
-static int
-propagate_necessity_through_operand (rtx *current_rtx, void *data)
-{
- rtx x = *current_rtx;
- varray_type *unprocessed_instructions = (varray_type *) data;
-
- if (x == NULL_RTX)
- return 0;
- switch ( GET_CODE (x))
- {
- case REG:
- if (CONVERT_REGISTER_TO_SSA_P (REGNO (x)))
- {
- rtx insn = VARRAY_RTX (ssa_definition, REGNO (x));
- if (insn != NULL_RTX && UNNECESSARY_P (insn))
- {
- RESURRECT_INSN (insn);
- VARRAY_PUSH_RTX (*unprocessed_instructions, insn);
- }
- }
- return 0;
-
- default:
- return 0;
- }
-}
-
-/* Indicate all insns initially assumed to be unnecessary. */
-
-static void
-mark_all_insn_unnecessary (void)
-{
- rtx insn;
- for (insn = get_insns (); insn != NULL_RTX; insn = NEXT_INSN (insn)) {
- if (INSN_P (insn))
- KILL_INSN (insn);
- }
-
-}
-
-/* Find the label beginning block BB, adding one if necessary. */
-
-static rtx
-find_block_label (basic_block bb)
-{
- rtx insn = bb->head;
- if (LABEL_P (insn))
- return insn;
- else
- {
- rtx new_label = emit_label_before (gen_label_rtx (), insn);
- if (insn == bb->head)
- bb->head = new_label;
- return new_label;
- }
-}
-
-/* Remove INSN, updating its basic block structure. */
-
-static void
-delete_insn_bb (rtx insn)
-{
- if (!insn)
- abort ();
-
- /* Do not actually delete anything that is not an INSN.
-
- We can get here because we only consider INSNs as
- potentially necessary. We leave it to later passes
- to remove unnecessary notes, unused labels, etc. */
- if (! INSN_P (insn))
- return;
-
- delete_insn (insn);
-}
-\f
-/* Perform the dead-code elimination. */
-
-void
-ssa_eliminate_dead_code (void)
-{
- rtx insn;
- basic_block bb;
- /* Necessary instructions with operands to explore. */
- varray_type unprocessed_instructions;
- /* Map element (b,e) is nonzero if the block is control dependent on
- edge. "cdbte" abbreviates control dependent block to edge. */
- control_dependent_block_to_edge_map cdbte;
- /* Element I is the immediate postdominator of block I. */
- dominance_info pdom;
- struct edge_list *el;
-
- /* Initialize the data structures. */
- mark_all_insn_unnecessary ();
- VARRAY_RTX_INIT (unprocessed_instructions, 64,
- "unprocessed instructions");
- cdbte = control_dependent_block_to_edge_map_create (last_basic_block);
-
- /* Prepare for use of BLOCK_NUM (). */
- connect_infinite_loops_to_exit ();
-
- /* Compute control dependence. */
- pdom = calculate_dominance_info (CDI_POST_DOMINATORS);
- el = create_edge_list ();
- find_all_control_dependences (el, pdom, cdbte);
-
- /* Find inherently necessary instructions. */
- for (insn = get_insns (); insn != NULL_RTX; insn = NEXT_INSN (insn))
- if (find_inherently_necessary (insn) && INSN_P (insn))
- {
- RESURRECT_INSN (insn);
- VARRAY_PUSH_RTX (unprocessed_instructions, insn);
- }
-
- /* Propagate necessity using the operands of necessary instructions. */
- while (VARRAY_ACTIVE_SIZE (unprocessed_instructions) > 0)
- {
- rtx current_instruction;
- int edge_number;
-
- current_instruction = VARRAY_TOP_RTX (unprocessed_instructions);
- VARRAY_POP (unprocessed_instructions);
-
- /* Make corresponding control dependent edges necessary. */
- /* Assume the only JUMP_INSN is the block's last insn. It appears
- that the last instruction of the program need not be a
- JUMP_INSN. */
-
- if (INSN_P (current_instruction)
- && !JUMP_TABLE_DATA_P (current_instruction))
- {
- /* Notes and labels contain no interesting operands. */
- EXECUTE_IF_CONTROL_DEPENDENT
- (cdbte, current_instruction, edge_number,
- {
- rtx jump_insn = (INDEX_EDGE_PRED_BB (el, edge_number))->end;
- if (GET_CODE (jump_insn) == JUMP_INSN
- && UNNECESSARY_P (jump_insn))
- {
- RESURRECT_INSN (jump_insn);
- VARRAY_PUSH_RTX (unprocessed_instructions, jump_insn);
- }
- });
-
- /* Propagate through the operands. */
- for_each_rtx (¤t_instruction,
- &propagate_necessity_through_operand,
- &unprocessed_instructions);
-
- /* PHI nodes are somewhat special in that each PHI alternative
- has data and control dependencies. The data dependencies
- are handled via propagate_necessity_through_operand. We
- handle the control dependency here.
-
- We consider the control dependent edges leading to the
- predecessor block associated with each PHI alternative
- as necessary. */
- if (PHI_NODE_P (current_instruction))
- {
- rtvec phi_vec = XVEC (SET_SRC (PATTERN (current_instruction)), 0);
- int num_elem = GET_NUM_ELEM (phi_vec);
- int v;
-
- for (v = num_elem - 2; v >= 0; v -= 2)
- {
- basic_block bb;
-
- bb = BASIC_BLOCK (INTVAL (RTVEC_ELT (phi_vec, v + 1)));
- EXECUTE_IF_CONTROL_DEPENDENT
- (cdbte, bb->end, edge_number,
- {
- rtx jump_insn;
-
- jump_insn = (INDEX_EDGE_PRED_BB (el, edge_number))->end;
- if (((GET_CODE (jump_insn) == JUMP_INSN))
- && UNNECESSARY_P (jump_insn))
- {
- RESURRECT_INSN (jump_insn);
- VARRAY_PUSH_RTX (unprocessed_instructions, jump_insn);
- }
- });
-
- }
- }
- }
- }
-
- /* Remove the unnecessary instructions. */
- EXECUTE_IF_UNNECESSARY (insn,
- {
- if (any_condjump_p (insn))
- {
- basic_block bb = BLOCK_FOR_INSN (insn);
- basic_block pdom_bb = find_pdom (pdom, bb);
- rtx lbl;
- edge e;
-
- /* Egad. The immediate post dominator is the exit block. We
- would like to optimize this conditional jump to jump directly
- to the exit block. That can be difficult as we may not have
- a suitable CODE_LABEL that allows us to fall unmolested into
- the exit block.
-
- So, we just delete the conditional branch by turning it into
- a deleted note. That is safe, but just not as optimal as
- it could be. */
- if (pdom_bb == EXIT_BLOCK_PTR)
- {
- /* Since we're going to just delete the branch, we need
- look at all the edges and remove all those which are not
- a fallthru edge. */
- e = bb->succ;
- while (e)
- {
- edge temp = e;
-
- e = e->succ_next;
- if ((temp->flags & EDGE_FALLTHRU) == 0)
- {
- /* We've found a non-fallthru edge, find any PHI nodes
- at the target and clean them up. */
- if (temp->dest != EXIT_BLOCK_PTR)
- {
- rtx insn
- = first_insn_after_basic_block_note (temp->dest);
-
- while (PHI_NODE_P (insn))
- {
- remove_phi_alternative (PATTERN (insn), temp->src);
- insn = NEXT_INSN (insn);
- }
- }
-
- remove_edge (temp);
- }
- }
-
- /* Now "delete" the conditional jump. */
- PUT_CODE (insn, NOTE);
- NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
- continue;
- }
-
- /* We've found a conditional branch that is unnecessary.
-
- First, remove all outgoing edges from this block, updating
- PHI nodes as appropriate. */
- e = bb->succ;
- while (e)
- {
- edge temp = e;
-
- e = e->succ_next;
-
- if (temp->flags & EDGE_ABNORMAL)
- continue;
-
- /* We found an edge that is not executable. First simplify
- the PHI nodes in the target block. */
- if (temp->dest != EXIT_BLOCK_PTR)
- {
- rtx insn = first_insn_after_basic_block_note (temp->dest);
-
- while (PHI_NODE_P (insn))
- {
- remove_phi_alternative (PATTERN (insn), temp->src);
- insn = NEXT_INSN (insn);
- }
- }
-
- remove_edge (temp);
- }
-
- /* Create an edge from this block to the post dominator.
- What about the PHI nodes at the target? */
- make_edge (bb, pdom_bb, 0);
-
- /* Third, transform this insn into an unconditional
- jump to the label for the immediate postdominator. */
- lbl = find_block_label (pdom_bb);
- SET_SRC (PATTERN (insn)) = gen_rtx_LABEL_REF (VOIDmode, lbl);
- INSN_CODE (insn) = -1;
- JUMP_LABEL (insn) = lbl;
- LABEL_NUSES (lbl)++;
-
- /* A barrier must follow any unconditional jump. Barriers
- are not in basic blocks so this must occur after
- deleting the conditional jump. */
- emit_barrier_after (insn);
- }
- else if (!JUMP_P (insn))
- delete_insn_bb (insn);
- });
-
- /* Remove fake edges from the CFG. */
- remove_fake_edges ();
-
- /* Find any blocks with no successors and ensure they are followed
- by a BARRIER. delete_insn has the nasty habit of deleting barriers
- when deleting insns. */
- FOR_EACH_BB (bb)
- {
- if (bb->succ == NULL)
- {
- rtx next = NEXT_INSN (bb->end);
-
- if (!next || GET_CODE (next) != BARRIER)
- emit_barrier_after (bb->end);
- }
- }
- /* Release allocated memory. */
- for (insn = get_insns (); insn != NULL_RTX; insn = NEXT_INSN (insn)) {
- if (INSN_P (insn))
- RESURRECT_INSN (insn);
- }
-
- if (VARRAY_ACTIVE_SIZE (unprocessed_instructions) != 0)
- abort ();
- control_dependent_block_to_edge_map_free (cdbte);
- free (pdom);
- free_edge_list (el);
-}
+++ /dev/null
-/* Static Single Assignment conversion routines for the GNU compiler.
- Copyright (C) 2000, 2001, 2002, 2003
- Free Software Foundation, Inc.
-
-This file is part of GCC.
-
-GCC is free software; you can redistribute it and/or modify it under
-the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
-version.
-
-GCC is distributed in the hope that it will be useful, but WITHOUT ANY
-WARRANTY; without even the implied warranty of MERCHANTABILITY or
-FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
-for more details.
-
-You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
-
-/* References:
-
- Building an Optimizing Compiler
- Robert Morgan
- Butterworth-Heinemann, 1998
-
- Static Single Assignment Construction
- Preston Briggs, Tim Harvey, Taylor Simpson
- Technical Report, Rice University, 1995
- ftp://ftp.cs.rice.edu/public/preston/optimizer/SSA.ps.gz. */
-
-#include "config.h"
-#include "system.h"
-#include "coretypes.h"
-#include "tm.h"
-
-#include "rtl.h"
-#include "expr.h"
-#include "varray.h"
-#include "partition.h"
-#include "sbitmap.h"
-#include "hashtab.h"
-#include "regs.h"
-#include "hard-reg-set.h"
-#include "flags.h"
-#include "function.h"
-#include "real.h"
-#include "insn-config.h"
-#include "recog.h"
-#include "basic-block.h"
-#include "output.h"
-#include "ssa.h"
-
-/* TODO:
-
- Handle subregs better, maybe. For now, if a reg that's set in a
- subreg expression is duplicated going into SSA form, an extra copy
- is inserted first that copies the entire reg into the duplicate, so
- that the other bits are preserved. This isn't strictly SSA, since
- at least part of the reg is assigned in more than one place (though
- they are adjacent).
-
- ??? What to do about strict_low_part. Probably I'll have to split
- them out of their current instructions first thing.
-
- Actually the best solution may be to have a kind of "mid-level rtl"
- in which the RTL encodes exactly what we want, without exposing a
- lot of niggling processor details. At some later point we lower
- the representation, calling back into optabs to finish any necessary
- expansion. */
-
-/* All pseudo-registers and select hard registers are converted to SSA
- form. When converting out of SSA, these select hard registers are
- guaranteed to be mapped to their original register number. Each
- machine's .h file should define CONVERT_HARD_REGISTER_TO_SSA_P
- indicating which hard registers should be converted.
-
- When converting out of SSA, temporaries for all registers are
- partitioned. The partition is checked to ensure that all uses of
- the same hard register in the same machine mode are in the same
- class. */
-
-/* If conservative_reg_partition is nonzero, use a conservative
- register partitioning algorithm (which leaves more regs after
- emerging from SSA) instead of the coalescing one. This is being
- left in for a limited time only, as a debugging tool until the
- coalescing algorithm is validated. */
-
-static int conservative_reg_partition;
-
-/* This flag is set when the CFG is in SSA form. */
-int in_ssa_form = 0;
-
-/* Element I is the single instruction that sets register I. */
-varray_type ssa_definition;
-
-/* Element I-PSEUDO is the normal register that originated the ssa
- register in question. */
-varray_type ssa_rename_from;
-
-/* Element I is the normal register that originated the ssa
- register in question.
-
- A hash table stores the (register, rtl) pairs. These are each
- xmalloc'ed and deleted when the hash table is destroyed. */
-htab_t ssa_rename_from_ht;
-
-/* The running target ssa register for a given pseudo register.
- (Pseudo registers appear in only one mode.) */
-static rtx *ssa_rename_to_pseudo;
-/* Similar, but for hard registers. A hard register can appear in
- many modes, so we store an equivalent pseudo for each of the
- modes. */
-static rtx ssa_rename_to_hard[FIRST_PSEUDO_REGISTER][NUM_MACHINE_MODES];
-
-/* ssa_rename_from maps pseudo registers to the original corresponding
- RTL. It is implemented as using a hash table. */
-
-typedef struct {
- unsigned int reg;
- rtx original;
-} ssa_rename_from_pair;
-
-struct ssa_rename_from_hash_table_data {
- sbitmap canonical_elements;
- partition reg_partition;
-};
-
-static rtx gen_sequence (void);
-static void ssa_rename_from_initialize (void);
-static rtx ssa_rename_from_lookup (int reg);
-static unsigned int original_register (unsigned int regno);
-static void ssa_rename_from_insert (unsigned int reg, rtx r);
-static void ssa_rename_from_free (void);
-typedef int (*srf_trav) (int regno, rtx r, sbitmap canonical_elements,
- partition reg_partition);
-static void ssa_rename_from_traverse (htab_trav callback_function,
- sbitmap canonical_elements, partition reg_partition);
-/*static Avoid warning message. */ void ssa_rename_from_print (void);
-static int ssa_rename_from_print_1 (void **slot, void *data);
-static hashval_t ssa_rename_from_hash_function (const void * srfp);
-static int ssa_rename_from_equal (const void *srfp1, const void *srfp2);
-static void ssa_rename_from_delete (void *srfp);
-
-static rtx ssa_rename_to_lookup (rtx reg);
-static void ssa_rename_to_insert (rtx reg, rtx r);
-
-/* The number of registers that were live on entry to the SSA routines. */
-static unsigned int ssa_max_reg_num;
-
-/* Local function prototypes. */
-
-struct rename_context;
-
-static inline rtx * phi_alternative (rtx, int);
-static void compute_dominance_frontiers_1 (sbitmap *frontiers,
- dominance_info idom, int bb,
- sbitmap done);
-static void find_evaluations_1 (rtx dest, rtx set, void *data);
-static void find_evaluations (sbitmap *evals, int nregs);
-static void compute_iterated_dominance_frontiers (sbitmap *idfs,
- sbitmap *frontiers,
- sbitmap *evals, int nregs);
-static void insert_phi_node (int regno, int b);
-static void insert_phi_nodes (sbitmap *idfs, sbitmap *evals, int nregs);
-static void create_delayed_rename (struct rename_context *, rtx *);
-static void apply_delayed_renames (struct rename_context *);
-static int rename_insn_1 (rtx *ptr, void *data);
-static void rename_block (int b, dominance_info dom);
-static void rename_registers (int nregs, dominance_info idom);
-
-static inline int ephi_add_node (rtx reg, rtx *nodes, int *n_nodes);
-static int * ephi_forward (int t, sbitmap visited, sbitmap *succ, int *tstack);
-static void ephi_backward (int t, sbitmap visited, sbitmap *pred, rtx *nodes);
-static void ephi_create (int t, sbitmap visited, sbitmap *pred,
- sbitmap *succ, rtx *nodes);
-static void eliminate_phi (edge e, partition reg_partition);
-static int make_regs_equivalent_over_bad_edges (int bb,
- partition reg_partition);
-
-/* These are used only in the conservative register partitioning
- algorithms. */
-static int make_equivalent_phi_alternatives_equivalent
- (int bb, partition reg_partition);
-static partition compute_conservative_reg_partition (void);
-static int record_canonical_element_1 (void **srfp, void *data);
-static int check_hard_regs_in_partition (partition reg_partition);
-
-/* These are used in the register coalescing algorithm. */
-static int coalesce_if_unconflicting (partition p, conflict_graph conflicts,
- int reg1, int reg2);
-static int coalesce_regs_in_copies (basic_block bb, partition p,
- conflict_graph conflicts);
-static int coalesce_reg_in_phi (rtx, int dest_regno, int src_regno,
- void *data);
-static int coalesce_regs_in_successor_phi_nodes (basic_block bb,
- partition p,
- conflict_graph conflicts);
-static partition compute_coalesced_reg_partition (void);
-static int mark_reg_in_phi (rtx *ptr, void *data);
-static void mark_phi_and_copy_regs (regset phi_set);
-
-static int rename_equivalent_regs_in_insn (rtx *ptr, void *data);
-static void rename_equivalent_regs (partition reg_partition);
-
-/* Deal with hard registers. */
-static int conflicting_hard_regs_p (int reg1, int reg2);
-
-/* ssa_rename_to maps registers and machine modes to SSA pseudo registers. */
-
-/* Find the register associated with REG in the indicated mode. */
-
-static rtx
-ssa_rename_to_lookup (rtx reg)
-{
- if (!HARD_REGISTER_P (reg))
- return ssa_rename_to_pseudo[REGNO (reg) - FIRST_PSEUDO_REGISTER];
- else
- return ssa_rename_to_hard[REGNO (reg)][GET_MODE (reg)];
-}
-
-/* Store a new value mapping REG to R in ssa_rename_to. */
-
-static void
-ssa_rename_to_insert (rtx reg, rtx r)
-{
- if (!HARD_REGISTER_P (reg))
- ssa_rename_to_pseudo[REGNO (reg) - FIRST_PSEUDO_REGISTER] = r;
- else
- ssa_rename_to_hard[REGNO (reg)][GET_MODE (reg)] = r;
-}
-
-/* Prepare ssa_rename_from for use. */
-
-static void
-ssa_rename_from_initialize (void)
-{
- /* We use an arbitrary initial hash table size of 64. */
- ssa_rename_from_ht = htab_create (64,
- &ssa_rename_from_hash_function,
- &ssa_rename_from_equal,
- &ssa_rename_from_delete);
-}
-
-/* Find the REG entry in ssa_rename_from. Return NULL_RTX if no entry is
- found. */
-
-static rtx
-ssa_rename_from_lookup (int reg)
-{
- ssa_rename_from_pair srfp;
- ssa_rename_from_pair *answer;
- srfp.reg = reg;
- srfp.original = NULL_RTX;
- answer = htab_find_with_hash (ssa_rename_from_ht, (void *) &srfp, reg);
- return (answer == 0 ? NULL_RTX : answer->original);
-}
-
-/* Find the number of the original register specified by REGNO. If
- the register is a pseudo, return the original register's number.
- Otherwise, return this register number REGNO. */
-
-static unsigned int
-original_register (unsigned int regno)
-{
- rtx original_rtx = ssa_rename_from_lookup (regno);
- return original_rtx != NULL_RTX ? REGNO (original_rtx) : regno;
-}
-
-/* Add mapping from R to REG to ssa_rename_from even if already present. */
-
-static void
-ssa_rename_from_insert (unsigned int reg, rtx r)
-{
- void **slot;
- ssa_rename_from_pair *srfp = xmalloc (sizeof (ssa_rename_from_pair));
- srfp->reg = reg;
- srfp->original = r;
- slot = htab_find_slot_with_hash (ssa_rename_from_ht, (const void *) srfp,
- reg, INSERT);
- if (*slot != 0)
- free ((void *) *slot);
- *slot = srfp;
-}
-
-/* Apply the CALLBACK_FUNCTION to each element in ssa_rename_from.
- CANONICAL_ELEMENTS and REG_PARTITION pass data needed by the only
- current use of this function. */
-
-static void
-ssa_rename_from_traverse (htab_trav callback_function,
- sbitmap canonical_elements, partition reg_partition)
-{
- struct ssa_rename_from_hash_table_data srfhd;
- srfhd.canonical_elements = canonical_elements;
- srfhd.reg_partition = reg_partition;
- htab_traverse (ssa_rename_from_ht, callback_function, (void *) &srfhd);
-}
-
-/* Destroy ssa_rename_from. */
-
-static void
-ssa_rename_from_free (void)
-{
- htab_delete (ssa_rename_from_ht);
-}
-
-/* Print the contents of ssa_rename_from. */
-
-/* static Avoid erroneous error message. */
-void
-ssa_rename_from_print (void)
-{
- printf ("ssa_rename_from's hash table contents:\n");
- htab_traverse (ssa_rename_from_ht, &ssa_rename_from_print_1, NULL);
-}
-
-/* Print the contents of the hash table entry SLOT, passing the unused
- attribute DATA. Used as a callback function with htab_traverse (). */
-
-static int
-ssa_rename_from_print_1 (void **slot, void *data ATTRIBUTE_UNUSED)
-{
- ssa_rename_from_pair * p = *slot;
- printf ("ssa_rename_from maps pseudo %i to original %i.\n",
- p->reg, REGNO (p->original));
- return 1;
-}
-
-/* Given a hash entry SRFP, yield a hash value. */
-
-static hashval_t
-ssa_rename_from_hash_function (const void *srfp)
-{
- return ((const ssa_rename_from_pair *) srfp)->reg;
-}
-
-/* Test whether two hash table entries SRFP1 and SRFP2 are equal. */
-
-static int
-ssa_rename_from_equal (const void *srfp1, const void *srfp2)
-{
- return ssa_rename_from_hash_function (srfp1) ==
- ssa_rename_from_hash_function (srfp2);
-}
-
-/* Delete the hash table entry SRFP. */
-
-static void
-ssa_rename_from_delete (void *srfp)
-{
- free (srfp);
-}
-
-/* Given the SET of a PHI node, return the address of the alternative
- for predecessor block C. */
-
-static inline rtx *
-phi_alternative (rtx set, int c)
-{
- rtvec phi_vec = XVEC (SET_SRC (set), 0);
- int v;
-
- for (v = GET_NUM_ELEM (phi_vec) - 2; v >= 0; v -= 2)
- if (INTVAL (RTVEC_ELT (phi_vec, v + 1)) == c)
- return &RTVEC_ELT (phi_vec, v);
-
- return NULL;
-}
-
-/* Given the SET of a phi node, remove the alternative for predecessor
- block C. Return nonzero on success, or zero if no alternative is
- found for C. */
-
-int
-remove_phi_alternative (rtx set, basic_block block)
-{
- rtvec phi_vec = XVEC (SET_SRC (set), 0);
- int num_elem = GET_NUM_ELEM (phi_vec);
- int v, c;
-
- c = block->index;
- for (v = num_elem - 2; v >= 0; v -= 2)
- if (INTVAL (RTVEC_ELT (phi_vec, v + 1)) == c)
- {
- if (v < num_elem - 2)
- {
- RTVEC_ELT (phi_vec, v) = RTVEC_ELT (phi_vec, num_elem - 2);
- RTVEC_ELT (phi_vec, v + 1) = RTVEC_ELT (phi_vec, num_elem - 1);
- }
- PUT_NUM_ELEM (phi_vec, num_elem - 2);
- return 1;
- }
-
- return 0;
-}
-
-/* For all registers, find all blocks in which they are set.
-
- This is the transform of what would be local kill information that
- we ought to be getting from flow. */
-
-static sbitmap *fe_evals;
-static int fe_current_bb;
-
-static void
-find_evaluations_1 (rtx dest, rtx set ATTRIBUTE_UNUSED,
- void *data ATTRIBUTE_UNUSED)
-{
- if (GET_CODE (dest) == REG
- && CONVERT_REGISTER_TO_SSA_P (REGNO (dest)))
- SET_BIT (fe_evals[REGNO (dest)], fe_current_bb);
-}
-
-static void
-find_evaluations (sbitmap *evals, int nregs)
-{
- basic_block bb;
-
- sbitmap_vector_zero (evals, nregs);
- fe_evals = evals;
-
- FOR_EACH_BB_REVERSE (bb)
- {
- rtx p, last;
-
- fe_current_bb = bb->index;
- p = bb->head;
- last = bb->end;
- while (1)
- {
- if (INSN_P (p))
- note_stores (PATTERN (p), find_evaluations_1, NULL);
-
- if (p == last)
- break;
- p = NEXT_INSN (p);
- }
- }
-}
-
-/* Computing the Dominance Frontier:
-
- As described in Morgan, section 3.5, this may be done simply by
- walking the dominator tree bottom-up, computing the frontier for
- the children before the parent. When considering a block B,
- there are two cases:
-
- (1) A flow graph edge leaving B that does not lead to a child
- of B in the dominator tree must be a block that is either equal
- to B or not dominated by B. Such blocks belong in the frontier
- of B.
-
- (2) Consider a block X in the frontier of one of the children C
- of B. If X is not equal to B and is not dominated by B, it
- is in the frontier of B.
-*/
-
-static void
-compute_dominance_frontiers_1 (sbitmap *frontiers, dominance_info idom,
- int bb, sbitmap done)
-{
- basic_block b = BASIC_BLOCK (bb);
- edge e;
- basic_block c;
-
- SET_BIT (done, bb);
- sbitmap_zero (frontiers[bb]);
-
- /* Do the frontier of the children first. Not all children in the
- dominator tree (blocks dominated by this one) are children in the
- CFG, so check all blocks. */
- FOR_EACH_BB (c)
- if (get_immediate_dominator (idom, c)->index == bb
- && ! TEST_BIT (done, c->index))
- compute_dominance_frontiers_1 (frontiers, idom, c->index, done);
-
- /* Find blocks conforming to rule (1) above. */
- for (e = b->succ; e; e = e->succ_next)
- {
- if (e->dest == EXIT_BLOCK_PTR)
- continue;
- if (get_immediate_dominator (idom, e->dest)->index != bb)
- SET_BIT (frontiers[bb], e->dest->index);
- }
-
- /* Find blocks conforming to rule (2). */
- FOR_EACH_BB (c)
- if (get_immediate_dominator (idom, c)->index == bb)
- {
- int x;
- EXECUTE_IF_SET_IN_SBITMAP (frontiers[c->index], 0, x,
- {
- if (get_immediate_dominator (idom, BASIC_BLOCK (x))->index != bb)
- SET_BIT (frontiers[bb], x);
- });
- }
-}
-
-void
-compute_dominance_frontiers (sbitmap *frontiers, dominance_info idom)
-{
- sbitmap done = sbitmap_alloc (last_basic_block);
- sbitmap_zero (done);
-
- compute_dominance_frontiers_1 (frontiers, idom, 0, done);
-
- sbitmap_free (done);
-}
-
-/* Computing the Iterated Dominance Frontier:
-
- This is the set of merge points for a given register.
-
- This is not particularly intuitive. See section 7.1 of Morgan, in
- particular figures 7.3 and 7.4 and the immediately surrounding text.
-*/
-
-static void
-compute_iterated_dominance_frontiers (sbitmap *idfs, sbitmap *frontiers,
- sbitmap *evals, int nregs)
-{
- sbitmap worklist;
- int reg, passes = 0;
-
- worklist = sbitmap_alloc (last_basic_block);
-
- for (reg = 0; reg < nregs; ++reg)
- {
- sbitmap idf = idfs[reg];
- int b, changed;
-
- /* Start the iterative process by considering those blocks that
- evaluate REG. We'll add their dominance frontiers to the
- IDF, and then consider the blocks we just added. */
- sbitmap_copy (worklist, evals[reg]);
-
- /* Morgan's algorithm is incorrect here. Blocks that evaluate
- REG aren't necessarily in REG's IDF. Start with an empty IDF. */
- sbitmap_zero (idf);
-
- /* Iterate until the worklist is empty. */
- do
- {
- changed = 0;
- passes++;
- EXECUTE_IF_SET_IN_SBITMAP (worklist, 0, b,
- {
- RESET_BIT (worklist, b);
- /* For each block on the worklist, add to the IDF all
- blocks on its dominance frontier that aren't already
- on the IDF. Every block that's added is also added
- to the worklist. */
- sbitmap_union_of_diff (worklist, worklist, frontiers[b], idf);
- sbitmap_a_or_b (idf, idf, frontiers[b]);
- changed = 1;
- });
- }
- while (changed);
- }
-
- sbitmap_free (worklist);
-
- if (rtl_dump_file)
- {
- fprintf (rtl_dump_file,
- "Iterated dominance frontier: %d passes on %d regs.\n",
- passes, nregs);
- }
-}
-
-/* Insert the phi nodes. */
-
-static void
-insert_phi_node (int regno, int bb)
-{
- basic_block b = BASIC_BLOCK (bb);
- edge e;
- int npred, i;
- rtvec vec;
- rtx phi, reg;
- rtx insn;
- int end_p;
-
- /* Find out how many predecessors there are. */
- for (e = b->pred, npred = 0; e; e = e->pred_next)
- if (e->src != ENTRY_BLOCK_PTR)
- npred++;
-
- /* If this block has no "interesting" preds, then there is nothing to
- do. Consider a block that only has the entry block as a pred. */
- if (npred == 0)
- return;
-
- /* This is the register to which the phi function will be assigned. */
- reg = regno_reg_rtx[regno];
-
- /* Construct the arguments to the PHI node. The use of pc_rtx is just
- a placeholder; we'll insert the proper value in rename_registers. */
- vec = rtvec_alloc (npred * 2);
- for (e = b->pred, i = 0; e ; e = e->pred_next, i += 2)
- if (e->src != ENTRY_BLOCK_PTR)
- {
- RTVEC_ELT (vec, i + 0) = pc_rtx;
- RTVEC_ELT (vec, i + 1) = GEN_INT (e->src->index);
- }
-
- phi = gen_rtx_PHI (VOIDmode, vec);
- phi = gen_rtx_SET (VOIDmode, reg, phi);
-
- insn = first_insn_after_basic_block_note (b);
- end_p = PREV_INSN (insn) == b->end;
- emit_insn_before (phi, insn);
- if (end_p)
- b->end = PREV_INSN (insn);
-}
-
-static void
-insert_phi_nodes (sbitmap *idfs, sbitmap *evals ATTRIBUTE_UNUSED, int nregs)
-{
- int reg;
-
- for (reg = 0; reg < nregs; ++reg)
- if (CONVERT_REGISTER_TO_SSA_P (reg))
- {
- int b;
- EXECUTE_IF_SET_IN_SBITMAP (idfs[reg], 0, b,
- {
- if (REGNO_REG_SET_P (BASIC_BLOCK (b)->global_live_at_start, reg))
- insert_phi_node (reg, b);
- });
- }
-}
-
-/* Rename the registers to conform to SSA.
-
- This is essentially the algorithm presented in Figure 7.8 of Morgan,
- with a few changes to reduce pattern search time in favor of a bit
- more memory usage. */
-
-/* One of these is created for each set. It will live in a list local
- to its basic block for the duration of that block's processing. */
-struct rename_set_data
-{
- struct rename_set_data *next;
- /* This is the SET_DEST of the (first) SET that sets the REG. */
- rtx *reg_loc;
- /* This is what used to be at *REG_LOC. */
- rtx old_reg;
- /* This is the REG that will replace OLD_REG. It's set only
- when the rename data is moved onto the DONE_RENAMES queue. */
- rtx new_reg;
- /* This is what to restore ssa_rename_to_lookup (old_reg) to. It is
- usually the previous contents of ssa_rename_to_lookup (old_reg). */
- rtx prev_reg;
- /* This is the insn that contains all the SETs of the REG. */
- rtx set_insn;
-};
-
-/* This struct is used to pass information to callback functions while
- renaming registers. */
-struct rename_context
-{
- struct rename_set_data *new_renames;
- struct rename_set_data *done_renames;
- rtx current_insn;
-};
-
-/* Queue the rename of *REG_LOC. */
-static void
-create_delayed_rename (struct rename_context *c, rtx *reg_loc)
-{
- struct rename_set_data *r;
- r = xmalloc (sizeof(*r));
-
- if (GET_CODE (*reg_loc) != REG
- || !CONVERT_REGISTER_TO_SSA_P (REGNO (*reg_loc)))
- abort ();
-
- r->reg_loc = reg_loc;
- r->old_reg = *reg_loc;
- r->prev_reg = ssa_rename_to_lookup(r->old_reg);
- r->set_insn = c->current_insn;
- r->next = c->new_renames;
- c->new_renames = r;
-}
-
-/* This is part of a rather ugly hack to allow the pre-ssa regno to be
- reused. If, during processing, a register has not yet been touched,
- ssa_rename_to[regno][machno] will be NULL. Now, in the course of pushing
- and popping values from ssa_rename_to, when we would ordinarily
- pop NULL back in, we pop RENAME_NO_RTX. We treat this exactly the
- same as NULL, except that it signals that the original regno has
- already been reused. */
-#define RENAME_NO_RTX pc_rtx
-
-/* Move all the entries from NEW_RENAMES onto DONE_RENAMES by
- applying all the renames on NEW_RENAMES. */
-
-static void
-apply_delayed_renames (struct rename_context *c)
-{
- struct rename_set_data *r;
- struct rename_set_data *last_r = NULL;
-
- for (r = c->new_renames; r != NULL; r = r->next)
- {
- int new_regno;
-
- /* Failure here means that someone has a PARALLEL that sets
- a register twice (bad!). */
- if (ssa_rename_to_lookup (r->old_reg) != r->prev_reg)
- abort ();
- /* Failure here means we have changed REG_LOC before applying
- the rename. */
- /* For the first set we come across, reuse the original regno. */
- if (r->prev_reg == NULL_RTX && !HARD_REGISTER_P (r->old_reg))
- {
- r->new_reg = r->old_reg;
- /* We want to restore RENAME_NO_RTX rather than NULL_RTX. */
- r->prev_reg = RENAME_NO_RTX;
- }
- else
- r->new_reg = gen_reg_rtx (GET_MODE (r->old_reg));
- new_regno = REGNO (r->new_reg);
- ssa_rename_to_insert (r->old_reg, r->new_reg);
-
- if (new_regno >= (int) ssa_definition->num_elements)
- {
- int new_limit = new_regno * 5 / 4;
- VARRAY_GROW (ssa_definition, new_limit);
- }
-
- VARRAY_RTX (ssa_definition, new_regno) = r->set_insn;
- ssa_rename_from_insert (new_regno, r->old_reg);
- last_r = r;
- }
- if (last_r != NULL)
- {
- last_r->next = c->done_renames;
- c->done_renames = c->new_renames;
- c->new_renames = NULL;
- }
-}
-
-/* Part one of the first step of rename_block, called through for_each_rtx.
- Mark pseudos that are set for later update. Transform uses of pseudos. */
-
-static int
-rename_insn_1 (rtx *ptr, void *data)
-{
- rtx x = *ptr;
- struct rename_context *context = data;
-
- if (x == NULL_RTX)
- return 0;
-
- switch (GET_CODE (x))
- {
- case SET:
- {
- rtx *destp = &SET_DEST (x);
- rtx dest = SET_DEST (x);
-
- /* An assignment to a paradoxical SUBREG does not read from
- the destination operand, and thus does not need to be
- wrapped into a SEQUENCE when translating into SSA form.
- We merely strip off the SUBREG and proceed normally for
- this case. */
- if (GET_CODE (dest) == SUBREG
- && (GET_MODE_SIZE (GET_MODE (dest))
- > GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))
- && GET_CODE (SUBREG_REG (dest)) == REG
- && CONVERT_REGISTER_TO_SSA_P (REGNO (SUBREG_REG (dest))))
- {
- destp = &XEXP (dest, 0);
- dest = XEXP (dest, 0);
- }
-
- /* Some SETs also use the REG specified in their LHS.
- These can be detected by the presence of
- STRICT_LOW_PART, SUBREG, SIGN_EXTRACT, and ZERO_EXTRACT
- in the LHS. Handle these by changing
- (set (subreg (reg foo)) ...)
- into
- (sequence [(set (reg foo_1) (reg foo))
- (set (subreg (reg foo_1)) ...)])
-
- FIXME: Much of the time this is too much. For some constructs
- we know that the output register is strictly an output
- (paradoxical SUBREGs and some libcalls for example).
-
- For those cases we are better off not making the false
- dependency. */
- if (GET_CODE (dest) == STRICT_LOW_PART
- || GET_CODE (dest) == SUBREG
- || GET_CODE (dest) == SIGN_EXTRACT
- || GET_CODE (dest) == ZERO_EXTRACT)
- {
- rtx i, reg;
- reg = dest;
-
- while (GET_CODE (reg) == STRICT_LOW_PART
- || GET_CODE (reg) == SUBREG
- || GET_CODE (reg) == SIGN_EXTRACT
- || GET_CODE (reg) == ZERO_EXTRACT)
- reg = XEXP (reg, 0);
-
- if (GET_CODE (reg) == REG
- && CONVERT_REGISTER_TO_SSA_P (REGNO (reg)))
- {
- /* Generate (set reg reg), and do renaming on it so
- that it becomes (set reg_1 reg_0), and we will
- replace reg with reg_1 in the SUBREG. */
-
- struct rename_set_data *saved_new_renames;
- saved_new_renames = context->new_renames;
- context->new_renames = NULL;
- i = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
- for_each_rtx (&i, rename_insn_1, data);
- apply_delayed_renames (context);
- context->new_renames = saved_new_renames;
- }
- }
- else if (GET_CODE (dest) == REG
- && CONVERT_REGISTER_TO_SSA_P (REGNO (dest)))
- {
- /* We found a genuine set of an interesting register. Tag
- it so that we can create a new name for it after we finish
- processing this insn. */
-
- create_delayed_rename (context, destp);
-
- /* Since we do not wish to (directly) traverse the
- SET_DEST, recurse through for_each_rtx for the SET_SRC
- and return. */
- if (GET_CODE (x) == SET)
- for_each_rtx (&SET_SRC (x), rename_insn_1, data);
- return -1;
- }
-
- /* Otherwise, this was not an interesting destination. Continue
- on, marking uses as normal. */
- return 0;
- }
-
- case REG:
- if (CONVERT_REGISTER_TO_SSA_P (REGNO (x))
- && REGNO (x) < ssa_max_reg_num)
- {
- rtx new_reg = ssa_rename_to_lookup (x);
-
- if (new_reg != RENAME_NO_RTX && new_reg != NULL_RTX)
- {
- if (GET_MODE (x) != GET_MODE (new_reg))
- abort ();
- *ptr = new_reg;
- }
- else
- {
- /* Undefined value used, rename it to a new pseudo register so
- that it cannot conflict with an existing register. */
- *ptr = gen_reg_rtx (GET_MODE (x));
- }
- }
- return -1;
-
- case CLOBBER:
- /* There is considerable debate on how CLOBBERs ought to be
- handled in SSA. For now, we're keeping the CLOBBERs, which
- means that we don't really have SSA form. There are a couple
- of proposals for how to fix this problem, but neither is
- implemented yet. */
- {
- rtx dest = XCEXP (x, 0, CLOBBER);
- if (REG_P (dest))
- {
- if (CONVERT_REGISTER_TO_SSA_P (REGNO (dest))
- && REGNO (dest) < ssa_max_reg_num)
- {
- rtx new_reg = ssa_rename_to_lookup (dest);
- if (new_reg != NULL_RTX && new_reg != RENAME_NO_RTX)
- XCEXP (x, 0, CLOBBER) = new_reg;
- }
- /* Stop traversing. */
- return -1;
- }
- else
- /* Continue traversing. */
- return 0;
- }
-
- case PHI:
- /* Never muck with the phi. We do that elsewhere, special-like. */
- return -1;
-
- default:
- /* Anything else, continue traversing. */
- return 0;
- }
-}
-
-static rtx
-gen_sequence (void)
-{
- rtx first_insn = get_insns ();
- rtx result;
- rtx tem;
- int i;
- int len;
-
- /* Count the insns in the chain. */
- len = 0;
- for (tem = first_insn; tem; tem = NEXT_INSN (tem))
- len++;
-
- result = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (len));
-
- for (i = 0, tem = first_insn; tem; tem = NEXT_INSN (tem), i++)
- XVECEXP (result, 0, i) = tem;
-
- return result;
-}
-
-static void
-rename_block (int bb, dominance_info idom)
-{
- basic_block b = BASIC_BLOCK (bb);
- edge e;
- rtx insn, next, last;
- struct rename_set_data *set_data = NULL;
- basic_block c;
-
- /* Step One: Walk the basic block, adding new names for sets and
- replacing uses. */
-
- next = b->head;
- last = b->end;
- do
- {
- insn = next;
- if (INSN_P (insn))
- {
- struct rename_context context;
- context.done_renames = set_data;
- context.new_renames = NULL;
- context.current_insn = insn;
-
- start_sequence ();
- for_each_rtx (&PATTERN (insn), rename_insn_1, &context);
- for_each_rtx (®_NOTES (insn), rename_insn_1, &context);
-
- /* Sometimes, we end up with a sequence of insns that
- SSA needs to treat as a single insn. Wrap these in a
- SEQUENCE. (Any notes now get attached to the SEQUENCE,
- not to the old version inner insn.) */
- if (get_insns () != NULL_RTX)
- {
- rtx seq;
- int i;
-
- emit (PATTERN (insn));
- seq = gen_sequence ();
- /* We really want a SEQUENCE of SETs, not a SEQUENCE
- of INSNs. */
- for (i = 0; i < XVECLEN (seq, 0); i++)
- XVECEXP (seq, 0, i) = PATTERN (XVECEXP (seq, 0, i));
- PATTERN (insn) = seq;
- }
- end_sequence ();
-
- apply_delayed_renames (&context);
- set_data = context.done_renames;
- }
-
- next = NEXT_INSN (insn);
- }
- while (insn != last);
-
- /* Step Two: Update the phi nodes of this block's successors. */
-
- for (e = b->succ; e; e = e->succ_next)
- {
- if (e->dest == EXIT_BLOCK_PTR)
- continue;
-
- insn = first_insn_after_basic_block_note (e->dest);
-
- while (PHI_NODE_P (insn))
- {
- rtx phi = PATTERN (insn);
- rtx reg;
-
- /* Find out which of our outgoing registers this node is
- intended to replace. Note that if this is not the first PHI
- node to have been created for this register, we have to
- jump through rename links to figure out which register
- we're talking about. This can easily be recognized by
- noting that the regno is new to this pass. */
- reg = SET_DEST (phi);
- if (REGNO (reg) >= ssa_max_reg_num)
- reg = ssa_rename_from_lookup (REGNO (reg));
- if (reg == NULL_RTX)
- abort ();
- reg = ssa_rename_to_lookup (reg);
-
- /* It is possible for the variable to be uninitialized on
- edges in. Reduce the arity of the PHI so that we don't
- consider those edges. */
- if (reg == NULL || reg == RENAME_NO_RTX)
- {
- if (! remove_phi_alternative (phi, b))
- abort ();
- }
- else
- {
- /* When we created the PHI nodes, we did not know what mode
- the register should be. Now that we've found an original,
- we can fill that in. */
- if (GET_MODE (SET_DEST (phi)) == VOIDmode)
- PUT_MODE (SET_DEST (phi), GET_MODE (reg));
- else if (GET_MODE (SET_DEST (phi)) != GET_MODE (reg))
- abort ();
-
- *phi_alternative (phi, bb) = reg;
- }
-
- insn = NEXT_INSN (insn);
- }
- }
-
- /* Step Three: Do the same to the children of this block in
- dominator order. */
-
- FOR_EACH_BB (c)
- if (get_immediate_dominator (idom, c)->index == bb)
- rename_block (c->index, idom);
-
- /* Step Four: Update the sets to refer to their new register,
- and restore ssa_rename_to to its previous state. */
-
- while (set_data)
- {
- struct rename_set_data *next;
- rtx old_reg = *set_data->reg_loc;
-
- if (*set_data->reg_loc != set_data->old_reg)
- abort ();
- *set_data->reg_loc = set_data->new_reg;
-
- ssa_rename_to_insert (old_reg, set_data->prev_reg);
-
- next = set_data->next;
- free (set_data);
- set_data = next;
- }
-}
-
-static void
-rename_registers (int nregs, dominance_info idom)
-{
- VARRAY_RTX_INIT (ssa_definition, nregs * 3, "ssa_definition");
- ssa_rename_from_initialize ();
-
- ssa_rename_to_pseudo = alloca (nregs * sizeof(rtx));
- memset (ssa_rename_to_pseudo, 0, nregs * sizeof(rtx));
- memset (ssa_rename_to_hard, 0,
- FIRST_PSEUDO_REGISTER * NUM_MACHINE_MODES * sizeof (rtx));
-
- rename_block (0, idom);
-
- /* ??? Update basic_block_live_at_start, and other flow info
- as needed. */
-
- ssa_rename_to_pseudo = NULL;
-}
-
-/* The main entry point for moving to SSA. */
-
-void
-convert_to_ssa (void)
-{
- /* Element I is the set of blocks that set register I. */
- sbitmap *evals;
-
- /* Dominator bitmaps. */
- sbitmap *dfs;
- sbitmap *idfs;
-
- /* Element I is the immediate dominator of block I. */
- dominance_info idom;
-
- int nregs;
-
- basic_block bb;
-
- /* Don't do it twice. */
- if (in_ssa_form)
- abort ();
-
- /* Need global_live_at_{start,end} up to date. Do not remove any
- dead code. We'll let the SSA optimizers do that. */
- life_analysis (get_insns (), NULL, 0);
-
- idom = calculate_dominance_info (CDI_DOMINATORS);
-
- if (rtl_dump_file)
- {
- fputs (";; Immediate Dominators:\n", rtl_dump_file);
- FOR_EACH_BB (bb)
- fprintf (rtl_dump_file, ";\t%3d = %3d\n", bb->index,
- get_immediate_dominator (idom, bb)->index);
- fflush (rtl_dump_file);
- }
-
- /* Compute dominance frontiers. */
-
- dfs = sbitmap_vector_alloc (last_basic_block, last_basic_block);
- compute_dominance_frontiers (dfs, idom);
-
- if (rtl_dump_file)
- {
- dump_sbitmap_vector (rtl_dump_file, ";; Dominance Frontiers:",
- "; Basic Block", dfs, last_basic_block);
- fflush (rtl_dump_file);
- }
-
- /* Compute register evaluations. */
-
- ssa_max_reg_num = max_reg_num ();
- nregs = ssa_max_reg_num;
- evals = sbitmap_vector_alloc (nregs, last_basic_block);
- find_evaluations (evals, nregs);
-
- /* Compute the iterated dominance frontier for each register. */
-
- idfs = sbitmap_vector_alloc (nregs, last_basic_block);
- compute_iterated_dominance_frontiers (idfs, dfs, evals, nregs);
-
- if (rtl_dump_file)
- {
- dump_sbitmap_vector (rtl_dump_file, ";; Iterated Dominance Frontiers:",
- "; Register", idfs, nregs);
- fflush (rtl_dump_file);
- }
-
- /* Insert the phi nodes. */
-
- insert_phi_nodes (idfs, evals, nregs);
-
- /* Rename the registers to satisfy SSA. */
-
- rename_registers (nregs, idom);
-
- /* All done! Clean up and go home. */
-
- sbitmap_vector_free (dfs);
- sbitmap_vector_free (evals);
- sbitmap_vector_free (idfs);
- in_ssa_form = 1;
-
- reg_scan (get_insns (), max_reg_num (), 1);
- free_dominance_info (idom);
-}
-
-/* REG is the representative temporary of its partition. Add it to the
- set of nodes to be processed, if it hasn't been already. Return the
- index of this register in the node set. */
-
-static inline int
-ephi_add_node (rtx reg, rtx *nodes, int *n_nodes)
-{
- int i;
- for (i = *n_nodes - 1; i >= 0; --i)
- if (REGNO (reg) == REGNO (nodes[i]))
- return i;
-
- nodes[i = (*n_nodes)++] = reg;
- return i;
-}
-
-/* Part one of the topological sort. This is a forward (downward) search
- through the graph collecting a stack of nodes to process. Assuming no
- cycles, the nodes at top of the stack when we are finished will have
- no other dependencies. */
-
-static int *
-ephi_forward (int t, sbitmap visited, sbitmap *succ, int *tstack)
-{
- int s;
-
- SET_BIT (visited, t);
-
- EXECUTE_IF_SET_IN_SBITMAP (succ[t], 0, s,
- {
- if (! TEST_BIT (visited, s))
- tstack = ephi_forward (s, visited, succ, tstack);
- });
-
- *tstack++ = t;
- return tstack;
-}
-
-/* Part two of the topological sort. The is a backward search through
- a cycle in the graph, copying the data forward as we go. */
-
-static void
-ephi_backward (int t, sbitmap visited, sbitmap *pred, rtx *nodes)
-{
- int p;
-
- SET_BIT (visited, t);
-
- EXECUTE_IF_SET_IN_SBITMAP (pred[t], 0, p,
- {
- if (! TEST_BIT (visited, p))
- {
- ephi_backward (p, visited, pred, nodes);
- emit_move_insn (nodes[p], nodes[t]);
- }
- });
-}
-
-/* Part two of the topological sort. Create the copy for a register
- and any cycle of which it is a member. */
-
-static void
-ephi_create (int t, sbitmap visited, sbitmap *pred, sbitmap *succ, rtx *nodes)
-{
- rtx reg_u = NULL_RTX;
- int unvisited_predecessors = 0;
- int p;
-
- /* Iterate through the predecessor list looking for unvisited nodes.
- If there are any, we have a cycle, and must deal with that. At
- the same time, look for a visited predecessor. If there is one,
- we won't need to create a temporary. */
-
- EXECUTE_IF_SET_IN_SBITMAP (pred[t], 0, p,
- {
- if (! TEST_BIT (visited, p))
- unvisited_predecessors = 1;
- else if (!reg_u)
- reg_u = nodes[p];
- });
-
- if (unvisited_predecessors)
- {
- /* We found a cycle. Copy out one element of the ring (if necessary),
- then traverse the ring copying as we go. */
-
- if (!reg_u)
- {
- reg_u = gen_reg_rtx (GET_MODE (nodes[t]));
- emit_move_insn (reg_u, nodes[t]);
- }
-
- EXECUTE_IF_SET_IN_SBITMAP (pred[t], 0, p,
- {
- if (! TEST_BIT (visited, p))
- {
- ephi_backward (p, visited, pred, nodes);
- emit_move_insn (nodes[p], reg_u);
- }
- });
- }
- else
- {
- /* No cycle. Just copy the value from a successor. */
-
- int s;
- EXECUTE_IF_SET_IN_SBITMAP (succ[t], 0, s,
- {
- SET_BIT (visited, t);
- emit_move_insn (nodes[t], nodes[s]);
- return;
- });
- }
-}
-
-/* Convert the edge to normal form. */
-
-static void
-eliminate_phi (edge e, partition reg_partition)
-{
- int n_nodes;
- sbitmap *pred, *succ;
- sbitmap visited;
- rtx *nodes;
- int *stack, *tstack;
- rtx insn;
- int i;
-
- /* Collect an upper bound on the number of registers needing processing. */
-
- insn = first_insn_after_basic_block_note (e->dest);
-
- n_nodes = 0;
- while (PHI_NODE_P (insn))
- {
- insn = next_nonnote_insn (insn);
- n_nodes += 2;
- }
-
- if (n_nodes == 0)
- return;
-
- /* Build the auxiliary graph R(B).
-
- The nodes of the graph are the members of the register partition
- present in Phi(B). There is an edge from FIND(T0)->FIND(T1) for
- each T0 = PHI(...,T1,...), where T1 is for the edge from block C. */
-
- nodes = alloca (n_nodes * sizeof(rtx));
- pred = sbitmap_vector_alloc (n_nodes, n_nodes);
- succ = sbitmap_vector_alloc (n_nodes, n_nodes);
- sbitmap_vector_zero (pred, n_nodes);
- sbitmap_vector_zero (succ, n_nodes);
-
- insn = first_insn_after_basic_block_note (e->dest);
-
- n_nodes = 0;
- for (; PHI_NODE_P (insn); insn = next_nonnote_insn (insn))
- {
- rtx* preg = phi_alternative (PATTERN (insn), e->src->index);
- rtx tgt = SET_DEST (PATTERN (insn));
- rtx reg;
-
- /* There may be no phi alternative corresponding to this edge.
- This indicates that the phi variable is undefined along this
- edge. */
- if (preg == NULL)
- continue;
- reg = *preg;
-
- if (GET_CODE (reg) != REG || GET_CODE (tgt) != REG)
- abort ();
-
- reg = regno_reg_rtx[partition_find (reg_partition, REGNO (reg))];
- tgt = regno_reg_rtx[partition_find (reg_partition, REGNO (tgt))];
- /* If the two registers are already in the same partition,
- nothing will need to be done. */
- if (reg != tgt)
- {
- int ireg, itgt;
-
- ireg = ephi_add_node (reg, nodes, &n_nodes);
- itgt = ephi_add_node (tgt, nodes, &n_nodes);
-
- SET_BIT (pred[ireg], itgt);
- SET_BIT (succ[itgt], ireg);
- }
- }
-
- if (n_nodes == 0)
- goto out;
-
- /* Begin a topological sort of the graph. */
-
- visited = sbitmap_alloc (n_nodes);
- sbitmap_zero (visited);
-
- tstack = stack = alloca (n_nodes * sizeof (int));
-
- for (i = 0; i < n_nodes; ++i)
- if (! TEST_BIT (visited, i))
- tstack = ephi_forward (i, visited, succ, tstack);
-
- sbitmap_zero (visited);
-
- /* As we find a solution to the tsort, collect the implementation
- insns in a sequence. */
- start_sequence ();
-
- while (tstack != stack)
- {
- i = *--tstack;
- if (! TEST_BIT (visited, i))
- ephi_create (i, visited, pred, succ, nodes);
- }
-
- insn = get_insns ();
- end_sequence ();
- insert_insn_on_edge (insn, e);
- if (rtl_dump_file)
- fprintf (rtl_dump_file, "Emitting copy on edge (%d,%d)\n",
- e->src->index, e->dest->index);
-
- sbitmap_free (visited);
-out:
- sbitmap_vector_free (pred);
- sbitmap_vector_free (succ);
-}
-
-/* For basic block B, consider all phi insns which provide an
- alternative corresponding to an incoming abnormal critical edge.
- Place the phi alternative corresponding to that abnormal critical
- edge in the same register class as the destination of the set.
-
- From Morgan, p. 178:
-
- For each abnormal critical edge (C, B),
- if T0 = phi (T1, ..., Ti, ..., Tm) is a phi node in B,
- and C is the ith predecessor of B,
- then T0 and Ti must be equivalent.
-
- Return nonzero iff any such cases were found for which the two
- regs were not already in the same class. */
-
-static int
-make_regs_equivalent_over_bad_edges (int bb, partition reg_partition)
-{
- int changed = 0;
- basic_block b = BASIC_BLOCK (bb);
- rtx phi;
-
- /* Advance to the first phi node. */
- phi = first_insn_after_basic_block_note (b);
-
- /* Scan all the phi nodes. */
- for (;
- PHI_NODE_P (phi);
- phi = next_nonnote_insn (phi))
- {
- edge e;
- int tgt_regno;
- rtx set = PATTERN (phi);
- rtx tgt = SET_DEST (set);
-
- /* The set target is expected to be an SSA register. */
- if (GET_CODE (tgt) != REG
- || !CONVERT_REGISTER_TO_SSA_P (REGNO (tgt)))
- abort ();
- tgt_regno = REGNO (tgt);
-
- /* Scan incoming abnormal critical edges. */
- for (e = b->pred; e; e = e->pred_next)
- if ((e->flags & EDGE_ABNORMAL) && EDGE_CRITICAL_P (e))
- {
- rtx *alt = phi_alternative (set, e->src->index);
- int alt_regno;
-
- /* If there is no alternative corresponding to this edge,
- the value is undefined along the edge, so just go on. */
- if (alt == 0)
- continue;
-
- /* The phi alternative is expected to be an SSA register. */
- if (GET_CODE (*alt) != REG
- || !CONVERT_REGISTER_TO_SSA_P (REGNO (*alt)))
- abort ();
- alt_regno = REGNO (*alt);
-
- /* If the set destination and the phi alternative aren't
- already in the same class... */
- if (partition_find (reg_partition, tgt_regno)
- != partition_find (reg_partition, alt_regno))
- {
- /* ... make them such. */
- if (conflicting_hard_regs_p (tgt_regno, alt_regno))
- /* It is illegal to unify a hard register with a
- different register. */
- abort ();
-
- partition_union (reg_partition,
- tgt_regno, alt_regno);
- ++changed;
- }
- }
- }
-
- return changed;
-}
-
-/* Consider phi insns in basic block BB pairwise. If the set target
- of both insns are equivalent pseudos, make the corresponding phi
- alternatives in each phi corresponding equivalent.
-
- Return nonzero if any new register classes were unioned. */
-
-static int
-make_equivalent_phi_alternatives_equivalent (int bb, partition reg_partition)
-{
- int changed = 0;
- basic_block b = BASIC_BLOCK (bb);
- rtx phi;
-
- /* Advance to the first phi node. */
- phi = first_insn_after_basic_block_note (b);
-
- /* Scan all the phi nodes. */
- for (;
- PHI_NODE_P (phi);
- phi = next_nonnote_insn (phi))
- {
- rtx set = PATTERN (phi);
- /* The regno of the destination of the set. */
- int tgt_regno = REGNO (SET_DEST (PATTERN (phi)));
-
- rtx phi2 = next_nonnote_insn (phi);
-
- /* Scan all phi nodes following this one. */
- for (;
- PHI_NODE_P (phi2);
- phi2 = next_nonnote_insn (phi2))
- {
- rtx set2 = PATTERN (phi2);
- /* The regno of the destination of the set. */
- int tgt2_regno = REGNO (SET_DEST (set2));
-
- /* Are the set destinations equivalent regs? */
- if (partition_find (reg_partition, tgt_regno) ==
- partition_find (reg_partition, tgt2_regno))
- {
- edge e;
- /* Scan over edges. */
- for (e = b->pred; e; e = e->pred_next)
- {
- int pred_block = e->src->index;
- /* Identify the phi alternatives from both phi
- nodes corresponding to this edge. */
- rtx *alt = phi_alternative (set, pred_block);
- rtx *alt2 = phi_alternative (set2, pred_block);
-
- /* If one of the phi nodes doesn't have a
- corresponding alternative, just skip it. */
- if (alt == 0 || alt2 == 0)
- continue;
-
- /* Both alternatives should be SSA registers. */
- if (GET_CODE (*alt) != REG
- || !CONVERT_REGISTER_TO_SSA_P (REGNO (*alt)))
- abort ();
- if (GET_CODE (*alt2) != REG
- || !CONVERT_REGISTER_TO_SSA_P (REGNO (*alt2)))
- abort ();
-
- /* If the alternatives aren't already in the same
- class ... */
- if (partition_find (reg_partition, REGNO (*alt))
- != partition_find (reg_partition, REGNO (*alt2)))
- {
- /* ... make them so. */
- if (conflicting_hard_regs_p (REGNO (*alt), REGNO (*alt2)))
- /* It is illegal to unify a hard register with
- a different register. */
- abort ();
-
- partition_union (reg_partition,
- REGNO (*alt), REGNO (*alt2));
- ++changed;
- }
- }
- }
- }
- }
-
- return changed;
-}
-
-/* Compute a conservative partition of outstanding pseudo registers.
- See Morgan 7.3.1. */
-
-static partition
-compute_conservative_reg_partition (void)
-{
- basic_block bb;
- int changed = 0;
-
- /* We don't actually work with hard registers, but it's easier to
- carry them around anyway rather than constantly doing register
- number arithmetic. */
- partition p =
- partition_new (ssa_definition->num_elements);
-
- /* The first priority is to make sure registers that might have to
- be copied on abnormal critical edges are placed in the same
- partition. This saves us from having to split abnormal critical
- edges. */
- FOR_EACH_BB_REVERSE (bb)
- changed += make_regs_equivalent_over_bad_edges (bb->index, p);
-
- /* Now we have to insure that corresponding arguments of phi nodes
- assigning to corresponding regs are equivalent. Iterate until
- nothing changes. */
- while (changed > 0)
- {
- changed = 0;
- FOR_EACH_BB_REVERSE (bb)
- changed += make_equivalent_phi_alternatives_equivalent (bb->index, p);
- }
-
- return p;
-}
-
-/* The following functions compute a register partition that attempts
- to eliminate as many reg copies and phi node copies as possible by
- coalescing registers. This is the strategy:
-
- 1. As in the conservative case, the top priority is to coalesce
- registers that otherwise would cause copies to be placed on
- abnormal critical edges (which isn't possible).
-
- 2. Figure out which regs are involved (in the LHS or RHS) of
- copies and phi nodes. Compute conflicts among these regs.
-
- 3. Walk around the instruction stream, placing two regs in the
- same class of the partition if one appears on the LHS and the
- other on the RHS of a copy or phi node and the two regs don't
- conflict. The conflict information of course needs to be
- updated.
-
- 4. If anything has changed, there may be new opportunities to
- coalesce regs, so go back to 2.
-*/
-
-/* If REG1 and REG2 don't conflict in CONFLICTS, place them in the
- same class of partition P, if they aren't already. Update
- CONFLICTS appropriately.
-
- Returns one if REG1 and REG2 were placed in the same class but were
- not previously; zero otherwise.
-
- See Morgan figure 11.15. */
-
-static int
-coalesce_if_unconflicting (partition p, conflict_graph conflicts,
- int reg1, int reg2)
-{
- int reg;
-
- /* Work only on SSA registers. */
- if (!CONVERT_REGISTER_TO_SSA_P (reg1) || !CONVERT_REGISTER_TO_SSA_P (reg2))
- return 0;
-
- /* Find the canonical regs for the classes containing REG1 and
- REG2. */
- reg1 = partition_find (p, reg1);
- reg2 = partition_find (p, reg2);
-
- /* If they're already in the same class, there's nothing to do. */
- if (reg1 == reg2)
- return 0;
-
- /* If the regs conflict, our hands are tied. */
- if (conflicting_hard_regs_p (reg1, reg2) ||
- conflict_graph_conflict_p (conflicts, reg1, reg2))
- return 0;
-
- /* We're good to go. Put the regs in the same partition. */
- partition_union (p, reg1, reg2);
-
- /* Find the new canonical reg for the merged class. */
- reg = partition_find (p, reg1);
-
- /* Merge conflicts from the two previous classes. */
- conflict_graph_merge_regs (conflicts, reg, reg1);
- conflict_graph_merge_regs (conflicts, reg, reg2);
-
- return 1;
-}
-
-/* For each register copy insn in basic block BB, place the LHS and
- RHS regs in the same class in partition P if they do not conflict
- according to CONFLICTS.
-
- Returns the number of changes that were made to P.
-
- See Morgan figure 11.14. */
-
-static int
-coalesce_regs_in_copies (basic_block bb, partition p, conflict_graph conflicts)
-{
- int changed = 0;
- rtx insn;
- rtx end = bb->end;
-
- /* Scan the instruction stream of the block. */
- for (insn = bb->head; insn != end; insn = NEXT_INSN (insn))
- {
- rtx pattern;
- rtx src;
- rtx dest;
-
- /* If this isn't a set insn, go to the next insn. */
- if (GET_CODE (insn) != INSN)
- continue;
- pattern = PATTERN (insn);
- if (GET_CODE (pattern) != SET)
- continue;
-
- src = SET_SRC (pattern);
- dest = SET_DEST (pattern);
-
- /* We're only looking for copies. */
- if (GET_CODE (src) != REG || GET_CODE (dest) != REG)
- continue;
-
- /* Coalesce only if the reg modes are the same. As long as
- each reg's rtx is unique, it can have only one mode, so two
- pseudos of different modes can't be coalesced into one.
-
- FIXME: We can probably get around this by inserting SUBREGs
- where appropriate, but for now we don't bother. */
- if (GET_MODE (src) != GET_MODE (dest))
- continue;
-
- /* Found a copy; see if we can use the same reg for both the
- source and destination (and thus eliminate the copy,
- ultimately). */
- changed += coalesce_if_unconflicting (p, conflicts,
- REGNO (src), REGNO (dest));
- }
-
- return changed;
-}
-
-struct phi_coalesce_context
-{
- partition p;
- conflict_graph conflicts;
- int changed;
-};
-
-/* Callback function for for_each_successor_phi. If the set
- destination and the phi alternative regs do not conflict, place
- them in the same partition class. DATA is a pointer to a
- phi_coalesce_context struct. */
-
-static int
-coalesce_reg_in_phi (rtx insn ATTRIBUTE_UNUSED, int dest_regno,
- int src_regno, void *data)
-{
- struct phi_coalesce_context *context =
- (struct phi_coalesce_context *) data;
-
- /* Attempt to use the same reg, if they don't conflict. */
- context->changed
- += coalesce_if_unconflicting (context->p, context->conflicts,
- dest_regno, src_regno);
- return 0;
-}
-
-/* For each alternative in a phi function corresponding to basic block
- BB (in phi nodes in successor block to BB), place the reg in the
- phi alternative and the reg to which the phi value is set into the
- same class in partition P, if allowed by CONFLICTS.
-
- Return the number of changes that were made to P.
-
- See Morgan figure 11.14. */
-
-static int
-coalesce_regs_in_successor_phi_nodes (basic_block bb, partition p,
- conflict_graph conflicts)
-{
- struct phi_coalesce_context context;
- context.p = p;
- context.conflicts = conflicts;
- context.changed = 0;
-
- for_each_successor_phi (bb, &coalesce_reg_in_phi, &context);
-
- return context.changed;
-}
-
-/* Compute and return a partition of pseudos. Where possible,
- non-conflicting pseudos are placed in the same class.
-
- The caller is responsible for deallocating the returned partition. */
-
-static partition
-compute_coalesced_reg_partition (void)
-{
- basic_block bb;
- int changed = 0;
- regset_head phi_set_head;
- regset phi_set = &phi_set_head;
-
- partition p =
- partition_new (ssa_definition->num_elements);
-
- /* The first priority is to make sure registers that might have to
- be copied on abnormal critical edges are placed in the same
- partition. This saves us from having to split abnormal critical
- edges (which can't be done). */
- FOR_EACH_BB_REVERSE (bb)
- make_regs_equivalent_over_bad_edges (bb->index, p);
-
- INIT_REG_SET (phi_set);
-
- do
- {
- conflict_graph conflicts;
-
- changed = 0;
-
- /* Build the set of registers involved in phi nodes, either as
- arguments to the phi function or as the target of a set. */
- CLEAR_REG_SET (phi_set);
- mark_phi_and_copy_regs (phi_set);
-
- /* Compute conflicts. */
- conflicts = conflict_graph_compute (phi_set, p);
-
- /* FIXME: Better would be to process most frequently executed
- blocks first, so that most frequently executed copies would
- be more likely to be removed by register coalescing. But any
- order will generate correct, if non-optimal, results. */
- FOR_EACH_BB_REVERSE (bb)
- {
- changed += coalesce_regs_in_copies (bb, p, conflicts);
- changed +=
- coalesce_regs_in_successor_phi_nodes (bb, p, conflicts);
- }
-
- conflict_graph_delete (conflicts);
- }
- while (changed > 0);
-
- FREE_REG_SET (phi_set);
-
- return p;
-}
-
-/* Mark the regs in a phi node. PTR is a phi expression or one of its
- components (a REG or a CONST_INT). DATA is a reg set in which to
- set all regs. Called from for_each_rtx. */
-
-static int
-mark_reg_in_phi (rtx *ptr, void *data)
-{
- rtx expr = *ptr;
- regset set = (regset) data;
-
- switch (GET_CODE (expr))
- {
- case REG:
- SET_REGNO_REG_SET (set, REGNO (expr));
- /* Fall through. */
- case CONST_INT:
- case PHI:
- return 0;
- default:
- abort ();
- }
-}
-
-/* Mark in PHI_SET all pseudos that are used in a phi node -- either
- set from a phi expression, or used as an argument in one. Also
- mark regs that are the source or target of a reg copy. Uses
- ssa_definition. */
-
-static void
-mark_phi_and_copy_regs (regset phi_set)
-{
- unsigned int reg;
-
- /* Scan the definitions of all regs. */
- for (reg = 0; reg < VARRAY_SIZE (ssa_definition); ++reg)
- if (CONVERT_REGISTER_TO_SSA_P (reg))
- {
- rtx insn = VARRAY_RTX (ssa_definition, reg);
- rtx pattern;
- rtx src;
-
- if (insn == NULL
- || (GET_CODE (insn) == NOTE
- && NOTE_LINE_NUMBER (insn) == NOTE_INSN_DELETED))
- continue;
- pattern = PATTERN (insn);
- /* Sometimes we get PARALLEL insns. These aren't phi nodes or
- copies. */
- if (GET_CODE (pattern) != SET)
- continue;
- src = SET_SRC (pattern);
-
- if (GET_CODE (src) == REG)
- {
- /* It's a reg copy. */
- SET_REGNO_REG_SET (phi_set, reg);
- SET_REGNO_REG_SET (phi_set, REGNO (src));
- }
- else if (GET_CODE (src) == PHI)
- {
- /* It's a phi node. Mark the reg being set. */
- SET_REGNO_REG_SET (phi_set, reg);
- /* Mark the regs used in the phi function. */
- for_each_rtx (&src, mark_reg_in_phi, phi_set);
- }
- /* ... else nothing to do. */
- }
-}
-
-/* Rename regs in insn PTR that are equivalent. DATA is the register
- partition which specifies equivalences. */
-
-static int
-rename_equivalent_regs_in_insn (rtx *ptr, void* data)
-{
- rtx x = *ptr;
- partition reg_partition = (partition) data;
-
- if (x == NULL_RTX)
- return 0;
-
- switch (GET_CODE (x))
- {
- case REG:
- if (CONVERT_REGISTER_TO_SSA_P (REGNO (x)))
- {
- unsigned int regno = REGNO (x);
- unsigned int new_regno = partition_find (reg_partition, regno);
- rtx canonical_element_rtx = ssa_rename_from_lookup (new_regno);
-
- if (canonical_element_rtx != NULL_RTX &&
- HARD_REGISTER_P (canonical_element_rtx))
- {
- if (REGNO (canonical_element_rtx) != regno)
- *ptr = canonical_element_rtx;
- }
- else if (regno != new_regno)
- {
- rtx new_reg = regno_reg_rtx[new_regno];
- if (GET_MODE (x) != GET_MODE (new_reg))
- abort ();
- *ptr = new_reg;
- }
- }
- return -1;
-
- case PHI:
- /* No need to rename the phi nodes. We'll check equivalence
- when inserting copies. */
- return -1;
-
- default:
- /* Anything else, continue traversing. */
- return 0;
- }
-}
-
-/* Record the register's canonical element stored in SRFP in the
- canonical_elements sbitmap packaged in DATA. This function is used
- as a callback function for traversing ssa_rename_from. */
-
-static int
-record_canonical_element_1 (void **srfp, void *data)
-{
- unsigned int reg = ((ssa_rename_from_pair *) *srfp)->reg;
- sbitmap canonical_elements =
- ((struct ssa_rename_from_hash_table_data *) data)->canonical_elements;
- partition reg_partition =
- ((struct ssa_rename_from_hash_table_data *) data)->reg_partition;
-
- SET_BIT (canonical_elements, partition_find (reg_partition, reg));
- return 1;
-}
-
-/* For each class in the REG_PARTITION corresponding to a particular
- hard register and machine mode, check that there are no other
- classes with the same hard register and machine mode. Returns
- nonzero if this is the case, i.e., the partition is acceptable. */
-
-static int
-check_hard_regs_in_partition (partition reg_partition)
-{
- /* CANONICAL_ELEMENTS has a nonzero bit if a class with the given register
- number and machine mode has already been seen. This is a
- problem with the partition. */
- sbitmap canonical_elements;
- int element_index;
- int already_seen[FIRST_PSEUDO_REGISTER][NUM_MACHINE_MODES];
- int reg;
- int mach_mode;
-
- /* Collect a list of canonical elements. */
- canonical_elements = sbitmap_alloc (max_reg_num ());
- sbitmap_zero (canonical_elements);
- ssa_rename_from_traverse (&record_canonical_element_1,
- canonical_elements, reg_partition);
-
- /* We have not seen any hard register uses. */
- for (reg = 0; reg < FIRST_PSEUDO_REGISTER; ++reg)
- for (mach_mode = 0; mach_mode < NUM_MACHINE_MODES; ++mach_mode)
- already_seen[reg][mach_mode] = 0;
-
- /* Check for classes with the same hard register and machine mode. */
- EXECUTE_IF_SET_IN_SBITMAP (canonical_elements, 0, element_index,
- {
- rtx hard_reg_rtx = ssa_rename_from_lookup (element_index);
- if (hard_reg_rtx != NULL_RTX &&
- HARD_REGISTER_P (hard_reg_rtx) &&
- already_seen[REGNO (hard_reg_rtx)][GET_MODE (hard_reg_rtx)] != 0)
- /* Two distinct partition classes should be mapped to the same
- hard register. */
- return 0;
- });
-
- sbitmap_free (canonical_elements);
-
- return 1;
-}
-
-/* Rename regs that are equivalent in REG_PARTITION. Also collapse
- any SEQUENCE insns. */
-
-static void
-rename_equivalent_regs (partition reg_partition)
-{
- basic_block b;
-
- FOR_EACH_BB_REVERSE (b)
- {
- rtx next = b->head;
- rtx last = b->end;
- rtx insn;
-
- do
- {
- insn = next;
- if (INSN_P (insn))
- {
- for_each_rtx (&PATTERN (insn),
- rename_equivalent_regs_in_insn,
- reg_partition);
- for_each_rtx (®_NOTES (insn),
- rename_equivalent_regs_in_insn,
- reg_partition);
-
- if (GET_CODE (PATTERN (insn)) == SEQUENCE)
- {
- rtx s = PATTERN (insn);
- int slen = XVECLEN (s, 0);
- int i;
-
- if (slen <= 1)
- abort ();
-
- PATTERN (insn) = XVECEXP (s, 0, slen-1);
- for (i = 0; i < slen - 1; i++)
- emit_insn_before (XVECEXP (s, 0, i), insn);
- }
- }
-
- next = NEXT_INSN (insn);
- }
- while (insn != last);
- }
-}
-
-/* The main entry point for moving from SSA. */
-
-void
-convert_from_ssa (void)
-{
- basic_block b, bb;
- partition reg_partition;
- rtx insns = get_insns ();
-
- /* Need global_live_at_{start,end} up to date. There should not be
- any significant dead code at this point, except perhaps dead
- stores. So do not take the time to perform dead code elimination.
-
- Register coalescing needs death notes, so generate them. */
- life_analysis (insns, NULL, PROP_DEATH_NOTES);
-
- /* Figure out which regs in copies and phi nodes don't conflict and
- therefore can be coalesced. */
- if (conservative_reg_partition)
- reg_partition = compute_conservative_reg_partition ();
- else
- reg_partition = compute_coalesced_reg_partition ();
-
- if (!check_hard_regs_in_partition (reg_partition))
- /* Two separate partitions should correspond to the same hard
- register but do not. */
- abort ();
-
- rename_equivalent_regs (reg_partition);
-
- /* Eliminate the PHI nodes. */
- FOR_EACH_BB_REVERSE (b)
- {
- edge e;
-
- for (e = b->pred; e; e = e->pred_next)
- if (e->src != ENTRY_BLOCK_PTR)
- eliminate_phi (e, reg_partition);
- }
-
- partition_delete (reg_partition);
-
- /* Actually delete the PHI nodes. */
- FOR_EACH_BB_REVERSE (bb)
- {
- rtx insn = bb->head;
-
- while (1)
- {
- /* If this is a PHI node delete it. */
- if (PHI_NODE_P (insn))
- {
- if (insn == bb->end)
- bb->end = PREV_INSN (insn);
- insn = delete_insn (insn);
- }
- /* Since all the phi nodes come at the beginning of the
- block, if we find an ordinary insn, we can stop looking
- for more phi nodes. */
- else if (INSN_P (insn))
- break;
- /* If we've reached the end of the block, stop. */
- else if (insn == bb->end)
- break;
- else
- insn = NEXT_INSN (insn);
- }
- }
-
- /* Commit all the copy nodes needed to convert out of SSA form. */
- commit_edge_insertions ();
-
- in_ssa_form = 0;
-
- count_or_remove_death_notes (NULL, 1);
-
- /* Deallocate the data structures. */
- ssa_definition = 0;
- ssa_rename_from_free ();
-}
-
-/* Scan phi nodes in successors to BB. For each such phi node that
- has a phi alternative value corresponding to BB, invoke FN. FN
- is passed the entire phi node insn, the regno of the set
- destination, the regno of the phi argument corresponding to BB,
- and DATA.
-
- If FN ever returns nonzero, stops immediately and returns this
- value. Otherwise, returns zero. */
-
-int
-for_each_successor_phi (basic_block bb, successor_phi_fn fn, void *data)
-{
- edge e;
-
- if (bb == EXIT_BLOCK_PTR)
- return 0;
-
- /* Scan outgoing edges. */
- for (e = bb->succ; e != NULL; e = e->succ_next)
- {
- rtx insn;
-
- basic_block successor = e->dest;
- if (successor == ENTRY_BLOCK_PTR
- || successor == EXIT_BLOCK_PTR)
- continue;
-
- /* Advance to the first non-label insn of the successor block. */
- insn = first_insn_after_basic_block_note (successor);
-
- if (insn == NULL)
- continue;
-
- /* Scan phi nodes in the successor. */
- for ( ; PHI_NODE_P (insn); insn = NEXT_INSN (insn))
- {
- int result;
- rtx phi_set = PATTERN (insn);
- rtx *alternative = phi_alternative (phi_set, bb->index);
- rtx phi_src;
-
- /* This phi function may not have an alternative
- corresponding to the incoming edge, indicating the
- assigned variable is not defined along the edge. */
- if (alternative == NULL)
- continue;
- phi_src = *alternative;
-
- /* Invoke the callback. */
- result = (*fn) (insn, REGNO (SET_DEST (phi_set)),
- REGNO (phi_src), data);
-
- /* Terminate if requested. */
- if (result != 0)
- return result;
- }
- }
-
- return 0;
-}
-
-/* Assuming the ssa_rename_from mapping has been established, yields
- nonzero if 1) only one SSA register of REG1 and REG2 comes from a
- hard register or 2) both SSA registers REG1 and REG2 come from
- different hard registers. */
-
-static int
-conflicting_hard_regs_p (int reg1, int reg2)
-{
- int orig_reg1 = original_register (reg1);
- int orig_reg2 = original_register (reg2);
- if (HARD_REGISTER_NUM_P (orig_reg1) && HARD_REGISTER_NUM_P (orig_reg2)
- && orig_reg1 != orig_reg2)
- return 1;
- if (HARD_REGISTER_NUM_P (orig_reg1) && !HARD_REGISTER_NUM_P (orig_reg2))
- return 1;
- if (!HARD_REGISTER_NUM_P (orig_reg1) && HARD_REGISTER_NUM_P (orig_reg2))
- return 1;
-
- return 0;
-}
+++ /dev/null
-/* Static Single Assignment (SSA) definitions for GCC
- Copyright (C) 2000, 2001, 2003 Free Software Foundation, Inc.
- Written by Jeffrey D. Oldham <oldham@codesourcery.com>.
-
-This file is part of GCC.
-
-GCC is free software; you can redistribute it and/or modify it under
-the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
-version.
-
-GCC is distributed in the hope that it will be useful, but WITHOUT ANY
-WARRANTY; without even the implied warranty of MERCHANTABILITY or
-FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
-for more details.
-
-You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
-
-
-/* Main SSA routines. */
-extern void convert_to_ssa (void);
-extern void convert_from_ssa (void);
-typedef int (*successor_phi_fn) (rtx, int, int, void *);
-extern int for_each_successor_phi (basic_block bb, successor_phi_fn,
- void *);
-void compute_dominance_frontiers (sbitmap *frontiers, dominance_info idom);
-extern int remove_phi_alternative (rtx, basic_block);
-
-
-/* Optimizations. */
-/* In ssa-dce.c */
-extern void ssa_eliminate_dead_code (void);
-
-/* In ssa-ccp.c */
-extern void ssa_const_prop (void);
-
-
-/* SSA definitions and uses. */
-/* This flag is set when the CFG is in SSA form. */
-extern int in_ssa_form;
-
-/* Element I is the single instruction that sets register I. */
-extern GTY(()) varray_type ssa_definition;
-
-/* Element I is an INSN_LIST of instructions that use register I. */
-extern varray_type ssa_uses;
-
-
-/* Specify which hard registers should be converted. */
-
-/* All pseudo-registers (having register number >=
- FIRST_PSEUDO_REGISTER) and hard registers satisfying
- CONVERT_HARD_REGISTER_TO_SSA_P are converted to SSA form. */
-
-/* Given a hard register number REG_NO, return nonzero if and only if
- the register should be converted to SSA. */
-
-#ifndef CONVERT_HARD_REGISTER_TO_SSA_P
-#define CONVERT_HARD_REGISTER_TO_SSA_P(REG_NO) (0) /* default of no hard registers */
-#endif /* CONVERT_HARD_REGISTER_TO_SSA_P */
-
-/* Given a register number REG_NO, return nonzero if and only if the
- register should be converted to SSA. */
-
-#define CONVERT_REGISTER_TO_SSA_P(REG_NO) \
- ((!HARD_REGISTER_NUM_P (REG_NO)) || \
- (CONVERT_HARD_REGISTER_TO_SSA_P (REG_NO)))
+2003-11-20 Richard Henderson <rth@redhat.com>
+
+ * gcc.dg/20020201-2.c: Remove.
+ * gcc.dg/20020201-4.c: Remove.
+ * gcc.dg/20020304-1.c: Remove.
+
2003-11-19 Nathanael Nerode <neroden@gcc.gnu.org>
* gcc.dg/cpp/trad/xwin1.c: New test case.
+++ /dev/null
-/* This testcase caused ICE because gcc was not able to add instructions
- on edge from ENTRY block successor to itself. */
-/* { dg-do compile } */
-/* { dg-options "-O3 -fssa" } */
-
-struct A { int a1; int a2; };
-struct B { long int b[32]; };
-
-extern int bar (struct B *, struct A *);
-
-int
-foo (struct B x)
-{
- struct A a, b;
- struct B c;
- int d;
-
- while (1)
- {
- a.a1 = 0;
- a.a2 = 0;
- b = a;
- c = x;
- d = bar (&c, &b);
- if (d >= 0)
- return d;
- }
-
- return 0;
-}
+++ /dev/null
-/* This testcase failed because recog_for_combine used to pass a different
- pattern than contained in insn to recog. */
-/* { dg-do compile } */
-/* { dg-options "-O2 -fssa -fssa-ccp" } */
-/* { dg-options "-O2 -march=i686 -fssa -fssa-ccp" { target i?86-*-* } } */
-
-extern int bar (char *);
-
-int
-foo (void)
-{
- char b[512];
-
- bar (b);
- return __builtin_strlen (b);
-}
+++ /dev/null
-/* { dg-do compile } */
-/* { dg-options "-O -fssa -fssa-ccp" } */
-
-double a[10][35], b[10][8];
-int c, c, d, e, f, g, h;
-
-int foo ()
-{
- int i, j, k, l;
-
- if (c > 10)
- c = 10;
-
- for (j = 0; j < c; j++)
- {
- k = 0;
- for (l = 0; l < h; l++)
- {
- if (d != 5)
- return -1;
- k = l * g;
- a[j][k] = (double) e; k++;
- a[j][k] = (double) f; k++;
- }
- for (i = 0;i < 35; i++)
- {
- if (a[j][i] >= 0.9)
- a[j][i] = 0.9;
- if (a[j][i] <= 0.1)
- a[j][i] = 0.1;
- }
- k = 0;
- b[j][k] = (double) e; k++;
- b[j][k] = (double) f; k++;
- }
- return 0;
-}
DEFTIMEVAR (TV_REORDER_BLOCKS , "reorder blocks")
DEFTIMEVAR (TV_SHORTEN_BRANCH , "shorten branches")
DEFTIMEVAR (TV_REG_STACK , "reg stack")
-DEFTIMEVAR (TV_TO_SSA , "convert to SSA")
-DEFTIMEVAR (TV_SSA_CCP , "SSA CCP")
-DEFTIMEVAR (TV_SSA_DCE , "SSA aggressive DCE")
-DEFTIMEVAR (TV_FROM_SSA , "convert from SSA")
DEFTIMEVAR (TV_FINAL , "final")
DEFTIMEVAR (TV_SYMOUT , "symout")
#include "regs.h"
#include "timevar.h"
#include "diagnostic.h"
-#include "ssa.h"
#include "params.h"
#include "reload.h"
#include "dwarf2asm.h"
/* Rest of compilation helper functions. */
static bool rest_of_handle_inlining (tree);
-static rtx rest_of_handle_ssa (tree, rtx);
static void rest_of_handle_cse (tree, rtx);
static void rest_of_handle_cse2 (tree, rtx);
static void rest_of_handle_gcse (tree, rtx);
DFI_sibling,
DFI_eh,
DFI_jump,
- DFI_ssa,
- DFI_ssa_ccp,
- DFI_ssa_dce,
- DFI_ussa,
DFI_null,
DFI_cse,
DFI_addressof,
Remaining -d letters:
- " m q "
- " JK O Q Y "
+ " e m q "
+ " JK O Q WXY "
*/
static struct dump_file_info dump_file[DFI_MAX] =
{ "sibling", 'i', 0, 0, 0 },
{ "eh", 'h', 0, 0, 0 },
{ "jump", 'j', 0, 0, 0 },
- { "ssa", 'e', 1, 0, 0 },
- { "ssaccp", 'W', 1, 0, 0 },
- { "ssadce", 'X', 1, 0, 0 },
- { "ussa", 'e', 1, 0, 0 }, /* Yes, duplicate enable switch. */
{ "null", 'u', 0, 0, 0 },
{ "cse", 's', 0, 0, 0 },
{ "addressof", 'F', 0, 0, 0 },
/* Nonzero means put zero initialized data in the bss section. */
int flag_zero_initialized_in_bss = 1;
-/* Enable SSA. */
-int flag_ssa = 0;
-
-/* Enable ssa conditional constant propagation. */
-int flag_ssa_ccp = 0;
-
-/* Enable ssa aggressive dead code elimination. */
-int flag_ssa_dce = 0;
-
/* Tag all structures with __attribute__(packed). */
int flag_pack_struct = 0;
{"dump-unnumbered", &flag_dump_unnumbered, 1 },
{"instrument-functions", &flag_instrument_function_entry_exit, 1 },
{"zero-initialized-in-bss", &flag_zero_initialized_in_bss, 1 },
- {"ssa", &flag_ssa, 1 },
- {"ssa-ccp", &flag_ssa_ccp, 1 },
- {"ssa-dce", &flag_ssa_dce, 1 },
{"leading-underscore", &flag_leading_underscore, 1 },
{"ident", &flag_no_ident, 0 },
{ "peephole2", &flag_peephole2, 1 },
return (bool) DECL_EXTERNAL (decl);
}
-/* Rest of compilation helper to convert the rtl to SSA form. */
-static rtx
-rest_of_handle_ssa (tree decl, rtx insns)
-{
- timevar_push (TV_TO_SSA);
- open_dump_file (DFI_ssa, decl);
-
- cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_PRE_LOOP);
- convert_to_ssa ();
-
- close_dump_file (DFI_ssa, print_rtl_with_bb, insns);
- timevar_pop (TV_TO_SSA);
-
- /* Perform sparse conditional constant propagation, if requested. */
- if (flag_ssa_ccp)
- {
- timevar_push (TV_SSA_CCP);
- open_dump_file (DFI_ssa_ccp, decl);
-
- ssa_const_prop ();
-
- close_dump_file (DFI_ssa_ccp, print_rtl_with_bb, get_insns ());
- timevar_pop (TV_SSA_CCP);
- }
-
- /* It would be useful to cleanup the CFG at this point, but block
- merging and possibly other transformations might leave a PHI
- node in the middle of a basic block, which is a strict no-no. */
-
- /* The SSA implementation uses basic block numbers in its phi
- nodes. Thus, changing the control-flow graph or the basic
- blocks, e.g., calling find_basic_blocks () or cleanup_cfg (),
- may cause problems. */
-
- if (flag_ssa_dce)
- {
- /* Remove dead code. */
-
- timevar_push (TV_SSA_DCE);
- open_dump_file (DFI_ssa_dce, decl);
-
- insns = get_insns ();
- ssa_eliminate_dead_code ();
-
- close_dump_file (DFI_ssa_dce, print_rtl_with_bb, insns);
- timevar_pop (TV_SSA_DCE);
- }
-
- /* Convert from SSA form. */
-
- timevar_push (TV_FROM_SSA);
- open_dump_file (DFI_ussa, decl);
-
- convert_from_ssa ();
- /* New registers have been created. Rescan their usage. */
- reg_scan (insns, max_reg_num (), 1);
-
- close_dump_file (DFI_ussa, print_rtl_with_bb, insns);
- timevar_pop (TV_FROM_SSA);
-
- ggc_collect ();
-
- return insns;
-}
-
/* Try to identify useless null pointer tests and delete them. */
static void
rest_of_handle_null_pointer (tree decl, rtx insns)
if (rtl_dump_and_exit || flag_syntax_only || DECL_DEFER_OUTPUT (decl))
goto exit_rest_of_compilation;
- /* Long term, this should probably move before the jump optimizer too,
- but I didn't want to disturb the rtl_dump_and_exit and related
- stuff at this time. */
- if (optimize > 0 && flag_ssa)
- insns = rest_of_handle_ssa (decl, insns);
-
timevar_push (TV_JUMP);
if (optimize)
extern int flag_unroll_all_loops;
extern int flag_unswitch_loops;
extern int flag_cprop_registers;
-extern int flag_ssa;
-extern int flag_ssa_ccp;
-extern int flag_ssa_dce;
extern int time_report;
extern int flag_new_regalloc;
projects / gcc.git / commitdiff