X-Git-Url: http://git.sourceforge.jp/view?p=pf3gnuchains%2Fgcc-fork.git;a=blobdiff_plain;f=gcc%2Fsched-deps.c;h=ad96281b093b0744c0d5a4a21c650b4e69633059;hp=9f20aad331b98c2a4d815382dbc8bdad4b8ec2c6;hb=3da6dfd92b854203218cd1e8f7cc78bd4f4650ba;hpb=501793e320d7c2e201ba7631cbcc5a9010163144 diff --git a/gcc/sched-deps.c b/gcc/sched-deps.c index 9f20aad331b..ad96281b093 100644 --- a/gcc/sched-deps.c +++ b/gcc/sched-deps.c @@ -1,7 +1,7 @@ /* Instruction scheduling pass. This file computes dependencies between instructions. Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, - 1999, 2000, 2001, 2002 Free Software Foundation, Inc. + 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc. Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by, and currently maintained by, Jim Wilson (wilson@cygnus.com) @@ -24,6 +24,8 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA #include "config.h" #include "system.h" +#include "coretypes.h" +#include "tm.h" #include "toplev.h" #include "rtl.h" #include "tm_p.h" @@ -40,6 +42,7 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA #include "sched-int.h" #include "params.h" #include "cselib.h" +#include "df.h" extern char *reg_known_equiv_p; extern rtx *reg_known_value; @@ -51,7 +54,18 @@ static regset_head reg_pending_uses_head; static regset reg_pending_sets; static regset reg_pending_clobbers; static regset reg_pending_uses; -static bool reg_pending_barrier; + +/* The following enumeration values tell us what dependencies we + should use to implement the barrier. We use true-dependencies for + TRUE_BARRIER and anti-dependencies for MOVE_BARRIER. */ +enum reg_pending_barrier_mode +{ + NOT_A_BARRIER = 0, + MOVE_BARRIER, + TRUE_BARRIER +}; + +static enum reg_pending_barrier_mode reg_pending_barrier; /* To speed up the test for duplicate dependency links we keep a record of dependencies created by add_dependence when the average @@ -66,38 +80,36 @@ static bool reg_pending_barrier; has enough entries to represent a dependency on any other insn in the insn chain. All bitmap for true dependencies cache is allocated then the rest two ones are also allocated. */ -static sbitmap *true_dependency_cache; -static sbitmap *anti_dependency_cache; -static sbitmap *output_dependency_cache; +static bitmap_head *true_dependency_cache; +static bitmap_head *anti_dependency_cache; +static bitmap_head *output_dependency_cache; +int cache_size; /* To speed up checking consistency of formed forward insn dependencies we use the following cache. Another possible solution could be switching off checking duplication of insns in forward dependencies. */ #ifdef ENABLE_CHECKING -static sbitmap *forward_dependency_cache; +static bitmap_head *forward_dependency_cache; #endif -static int deps_may_trap_p PARAMS ((rtx)); -static void add_dependence_list PARAMS ((rtx, rtx, enum reg_note)); -static void add_dependence_list_and_free PARAMS ((rtx, rtx *, enum reg_note)); -static void remove_dependence PARAMS ((rtx, rtx)); -static void set_sched_group_p PARAMS ((rtx)); +static int deps_may_trap_p (rtx); +static void add_dependence_list (rtx, rtx, enum reg_note); +static void add_dependence_list_and_free (rtx, rtx *, enum reg_note); +static void set_sched_group_p (rtx); -static void flush_pending_lists PARAMS ((struct deps *, rtx, int, int)); -static void sched_analyze_1 PARAMS ((struct deps *, rtx, rtx)); -static void sched_analyze_2 PARAMS ((struct deps *, rtx, rtx)); -static void sched_analyze_insn PARAMS ((struct deps *, rtx, rtx, rtx)); -static rtx group_leader PARAMS ((rtx)); +static void flush_pending_lists (struct deps *, rtx, int, int); +static void sched_analyze_1 (struct deps *, rtx, rtx); +static void sched_analyze_2 (struct deps *, rtx, rtx); +static void sched_analyze_insn (struct deps *, rtx, rtx, rtx); -static rtx get_condition PARAMS ((rtx)); -static int conditions_mutex_p PARAMS ((rtx, rtx)); +static rtx get_condition (rtx); +static int conditions_mutex_p (rtx, rtx); /* Return nonzero if a load of the memory reference MEM can cause a trap. */ static int -deps_may_trap_p (mem) - rtx mem; +deps_may_trap_p (rtx mem) { rtx addr = XEXP (mem, 0); @@ -112,9 +124,7 @@ deps_may_trap_p (mem) if LIST does not contain INSN. */ rtx -find_insn_list (insn, list) - rtx insn; - rtx list; +find_insn_list (rtx insn, rtx list) { while (list) { @@ -128,8 +138,7 @@ find_insn_list (insn, list) /* Find the condition under which INSN is executed. */ static rtx -get_condition (insn) - rtx insn; +get_condition (rtx insn) { rtx pat = PATTERN (insn); rtx cond; @@ -160,11 +169,10 @@ get_condition (insn) /* Return nonzero if conditions COND1 and COND2 can never be both true. */ static int -conditions_mutex_p (cond1, cond2) - rtx cond1, cond2; +conditions_mutex_p (rtx cond1, rtx cond2) { - if (GET_RTX_CLASS (GET_CODE (cond1)) == '<' - && GET_RTX_CLASS (GET_CODE (cond2)) == '<' + if (COMPARISON_P (cond1) + && COMPARISON_P (cond2) && GET_CODE (cond1) == reverse_condition (GET_CODE (cond2)) && XEXP (cond1, 0) == XEXP (cond2, 0) && XEXP (cond1, 1) == XEXP (cond2, 1)) @@ -173,28 +181,26 @@ conditions_mutex_p (cond1, cond2) } /* Add ELEM wrapped in an INSN_LIST with reg note kind DEP_TYPE to the - LOG_LINKS of INSN, if not already there. DEP_TYPE indicates the type - of dependence that this link represents. */ + LOG_LINKS of INSN, if not already there. DEP_TYPE indicates the + type of dependence that this link represents. The function returns + nonzero if a new entry has been added to insn's LOG_LINK. */ -void -add_dependence (insn, elem, dep_type) - rtx insn; - rtx elem; - enum reg_note dep_type; +int +add_dependence (rtx insn, rtx elem, enum reg_note dep_type) { - rtx link, next; + rtx link; int present_p; rtx cond1, cond2; /* Don't depend an insn on itself. */ if (insn == elem) - return; + return 0; /* We can get a dependency on deleted insns due to optimizations in the register allocation and reloading or due to splitting. Any such dependency is useless and can be ignored. */ if (GET_CODE (elem) == NOTE) - return; + return 0; /* flow.c doesn't handle conditional lifetimes entirely correctly; calls mess up the conditional lifetimes. */ @@ -213,39 +219,7 @@ add_dependence (insn, elem, dep_type) /* Make sure second instruction doesn't affect condition of first instruction if switched. */ && !modified_in_p (cond2, insn)) - return; - } - - /* If elem is part of a sequence that must be scheduled together, then - make the dependence point to the last insn of the sequence. - When HAVE_cc0, it is possible for NOTEs to exist between users and - setters of the condition codes, so we must skip past notes here. - Otherwise, NOTEs are impossible here. */ - next = next_nonnote_insn (elem); - if (next && SCHED_GROUP_P (next) - && GET_CODE (next) != CODE_LABEL) - { - /* Notes will never intervene here though, so don't bother checking - for them. */ - /* Hah! Wrong. */ - /* We must reject CODE_LABELs, so that we don't get confused by one - that has LABEL_PRESERVE_P set, which is represented by the same - bit in the rtl as SCHED_GROUP_P. A CODE_LABEL can never be - SCHED_GROUP_P. */ - - rtx nnext; - while ((nnext = next_nonnote_insn (next)) != NULL - && SCHED_GROUP_P (nnext) - && GET_CODE (nnext) != CODE_LABEL) - next = nnext; - - /* Again, don't depend an insn on itself. */ - if (insn == next) - return; - - /* Make the dependence to NEXT, the last insn of the group, instead - of the original ELEM. */ - elem = next; + return 0; } present_p = 1; @@ -259,7 +233,7 @@ add_dependence (insn, elem, dep_type) elem is a CALL is still required. */ if (GET_CODE (insn) == CALL_INSN && (INSN_BB (elem) != INSN_BB (insn))) - return; + return 0; #endif /* If we already have a dependency for ELEM, then we do not need to @@ -271,19 +245,20 @@ add_dependence (insn, elem, dep_type) if (anti_dependency_cache == NULL || output_dependency_cache == NULL) abort (); - if (TEST_BIT (true_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem))) + if (bitmap_bit_p (&true_dependency_cache[INSN_LUID (insn)], + INSN_LUID (elem))) /* Do nothing (present_set_type is already 0). */ ; - else if (TEST_BIT (anti_dependency_cache[INSN_LUID (insn)], + else if (bitmap_bit_p (&anti_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem))) present_dep_type = REG_DEP_ANTI; - else if (TEST_BIT (output_dependency_cache[INSN_LUID (insn)], + else if (bitmap_bit_p (&output_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem))) present_dep_type = REG_DEP_OUTPUT; - else + else present_p = 0; if (present_p && (int) dep_type >= (int) present_dep_type) - return; + return 0; } #endif @@ -298,12 +273,12 @@ add_dependence (insn, elem, dep_type) if (true_dependency_cache != NULL) { if (REG_NOTE_KIND (link) == REG_DEP_ANTI) - RESET_BIT (anti_dependency_cache[INSN_LUID (insn)], - INSN_LUID (elem)); + bitmap_clear_bit (&anti_dependency_cache[INSN_LUID (insn)], + INSN_LUID (elem)); else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT && output_dependency_cache) - RESET_BIT (output_dependency_cache[INSN_LUID (insn)], - INSN_LUID (elem)); + bitmap_clear_bit (&output_dependency_cache[INSN_LUID (insn)], + INSN_LUID (elem)); else abort (); } @@ -313,25 +288,25 @@ add_dependence (insn, elem, dep_type) one, then change the existing dependence to this type. */ if ((int) dep_type < (int) REG_NOTE_KIND (link)) PUT_REG_NOTE_KIND (link, dep_type); - + #ifdef INSN_SCHEDULING /* If we are adding a dependency to INSN's LOG_LINKs, then note that in the bitmap caches of dependency information. */ if (true_dependency_cache != NULL) { if ((int) REG_NOTE_KIND (link) == 0) - SET_BIT (true_dependency_cache[INSN_LUID (insn)], - INSN_LUID (elem)); + bitmap_set_bit (&true_dependency_cache[INSN_LUID (insn)], + INSN_LUID (elem)); else if (REG_NOTE_KIND (link) == REG_DEP_ANTI) - SET_BIT (anti_dependency_cache[INSN_LUID (insn)], - INSN_LUID (elem)); + bitmap_set_bit (&anti_dependency_cache[INSN_LUID (insn)], + INSN_LUID (elem)); else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT) - SET_BIT (output_dependency_cache[INSN_LUID (insn)], - INSN_LUID (elem)); + bitmap_set_bit (&output_dependency_cache[INSN_LUID (insn)], + INSN_LUID (elem)); } #endif - return; - } + return 0; + } /* Might want to check one level of transitivity to save conses. */ link = alloc_INSN_LIST (elem, LOG_LINKS (insn)); @@ -346,21 +321,20 @@ add_dependence (insn, elem, dep_type) if (true_dependency_cache != NULL) { if ((int) dep_type == 0) - SET_BIT (true_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); + bitmap_set_bit (&true_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); else if (dep_type == REG_DEP_ANTI) - SET_BIT (anti_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); + bitmap_set_bit (&anti_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); else if (dep_type == REG_DEP_OUTPUT) - SET_BIT (output_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); + bitmap_set_bit (&output_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)); } #endif + return 1; } /* A convenience wrapper to operate on an entire list. */ static void -add_dependence_list (insn, list, dep_type) - rtx insn, list; - enum reg_note dep_type; +add_dependence_list (rtx insn, rtx list, enum reg_note dep_type) { for (; list; list = XEXP (list, 1)) add_dependence (insn, XEXP (list, 0), dep_type); @@ -369,10 +343,7 @@ add_dependence_list (insn, list, dep_type) /* Similar, but free *LISTP at the same time. */ static void -add_dependence_list_and_free (insn, listp, dep_type) - rtx insn; - rtx *listp; - enum reg_note dep_type; +add_dependence_list_and_free (rtx insn, rtx *listp, enum reg_note dep_type) { rtx list, next; for (list = *listp, *listp = NULL; list ; list = next) @@ -383,102 +354,18 @@ add_dependence_list_and_free (insn, listp, dep_type) } } -/* Remove ELEM wrapped in an INSN_LIST from the LOG_LINKS - of INSN. Abort if not found. */ - -static void -remove_dependence (insn, elem) - rtx insn; - rtx elem; -{ - rtx prev, link, next; - int found = 0; - - for (prev = 0, link = LOG_LINKS (insn); link; link = next) - { - next = XEXP (link, 1); - if (XEXP (link, 0) == elem) - { - if (prev) - XEXP (prev, 1) = next; - else - LOG_LINKS (insn) = next; - -#ifdef INSN_SCHEDULING - /* If we are removing a dependency from the LOG_LINKS list, - make sure to remove it from the cache too. */ - if (true_dependency_cache != NULL) - { - if (REG_NOTE_KIND (link) == 0) - RESET_BIT (true_dependency_cache[INSN_LUID (insn)], - INSN_LUID (elem)); - else if (REG_NOTE_KIND (link) == REG_DEP_ANTI) - RESET_BIT (anti_dependency_cache[INSN_LUID (insn)], - INSN_LUID (elem)); - else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT) - RESET_BIT (output_dependency_cache[INSN_LUID (insn)], - INSN_LUID (elem)); - } -#endif - - free_INSN_LIST_node (link); - - found = 1; - } - else - prev = link; - } - - if (!found) - abort (); - return; -} - -/* Return an insn which represents a SCHED_GROUP, which is - the last insn in the group. */ - -static rtx -group_leader (insn) - rtx insn; -{ - rtx prev; - - do - { - prev = insn; - insn = next_nonnote_insn (insn); - } - while (insn && SCHED_GROUP_P (insn) && (GET_CODE (insn) != CODE_LABEL)); - - return prev; -} - /* Set SCHED_GROUP_P and care for the rest of the bookkeeping that goes along with that. */ static void -set_sched_group_p (insn) - rtx insn; +set_sched_group_p (rtx insn) { - rtx link, prev; + rtx prev; SCHED_GROUP_P (insn) = 1; - /* There may be a note before this insn now, but all notes will - be removed before we actually try to schedule the insns, so - it won't cause a problem later. We must avoid it here though. */ prev = prev_nonnote_insn (insn); - - /* Make a copy of all dependencies on the immediately previous insn, - and add to this insn. This is so that all the dependencies will - apply to the group. Remove an explicit dependence on this insn - as SCHED_GROUP_P now represents it. */ - - if (find_insn_list (prev, LOG_LINKS (insn))) - remove_dependence (insn, prev); - - for (link = LOG_LINKS (prev); link; link = XEXP (link, 1)) - add_dependence (insn, XEXP (link, 0), REG_NOTE_KIND (link)); + add_dependence (insn, prev, REG_DEP_ANTI); } /* Process an insn's memory dependencies. There are four kinds of @@ -497,9 +384,8 @@ set_sched_group_p (insn) so that we can do memory aliasing on it. */ void -add_insn_mem_dependence (deps, insn_list, mem_list, insn, mem) - struct deps *deps; - rtx *insn_list, *mem_list, insn, mem; +add_insn_mem_dependence (struct deps *deps, rtx *insn_list, rtx *mem_list, + rtx insn, rtx mem) { rtx link; @@ -511,7 +397,7 @@ add_insn_mem_dependence (deps, insn_list, mem_list, insn, mem) mem = shallow_copy_rtx (mem); XEXP (mem, 0) = cselib_subst_to_values (XEXP (mem, 0)); } - link = alloc_EXPR_LIST (VOIDmode, mem, *mem_list); + link = alloc_EXPR_LIST (VOIDmode, canon_rtx (mem), *mem_list); *mem_list = link; deps->pending_lists_length++; @@ -522,10 +408,8 @@ add_insn_mem_dependence (deps, insn_list, mem_list, insn, mem) dependencies for a read operation, similarly with FOR_WRITE. */ static void -flush_pending_lists (deps, insn, for_read, for_write) - struct deps *deps; - rtx insn; - int for_read, for_write; +flush_pending_lists (struct deps *deps, rtx insn, int for_read, + int for_write) { if (for_write) { @@ -550,10 +434,7 @@ flush_pending_lists (deps, insn, for_read, for_write) destination of X, and reads of everything mentioned. */ static void -sched_analyze_1 (deps, x, insn) - struct deps *deps; - rtx x; - rtx insn; +sched_analyze_1 (struct deps *deps, rtx x, rtx insn) { int regno; rtx dest = XEXP (x, 0); @@ -581,6 +462,19 @@ sched_analyze_1 (deps, x, insn) while (GET_CODE (dest) == STRICT_LOW_PART || GET_CODE (dest) == SUBREG || GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT) { + if (GET_CODE (dest) == STRICT_LOW_PART + || GET_CODE (dest) == ZERO_EXTRACT + || GET_CODE (dest) == SIGN_EXTRACT + || read_modify_subreg_p (dest)) + { + /* These both read and modify the result. We must handle + them as writes to get proper dependencies for following + instructions. We must handle them as reads to get proper + dependencies from this to previous instructions. + Thus we need to call sched_analyze_2. */ + + sched_analyze_2 (deps, XEXP (dest, 0), insn); + } if (GET_CODE (dest) == ZERO_EXTRACT || GET_CODE (dest) == SIGN_EXTRACT) { /* The second and third arguments are values read by this insn. */ @@ -598,7 +492,7 @@ sched_analyze_1 (deps, x, insn) If so, mark all of them just like the first. */ if (regno < FIRST_PSEUDO_REGISTER) { - int i = HARD_REGNO_NREGS (regno, GET_MODE (dest)); + int i = hard_regno_nregs[regno][GET_MODE (dest)]; if (code == SET) { while (--i >= 0) @@ -651,6 +545,7 @@ sched_analyze_1 (deps, x, insn) cselib_lookup (XEXP (t, 0), Pmode, 1); XEXP (t, 0) = cselib_subst_to_values (XEXP (t, 0)); } + t = canon_rtx (t); if (deps->pending_lists_length > MAX_PENDING_LIST_LENGTH) { @@ -704,10 +599,7 @@ sched_analyze_1 (deps, x, insn) /* Analyze the uses of memory and registers in rtx X in INSN. */ static void -sched_analyze_2 (deps, x, insn) - struct deps *deps; - rtx x; - rtx insn; +sched_analyze_2 (struct deps *deps, rtx x, rtx insn) { int i; int j; @@ -723,6 +615,7 @@ sched_analyze_2 (deps, x, insn) { case CONST_INT: case CONST_DOUBLE: + case CONST_VECTOR: case SYMBOL_REF: case CONST: case LABEL_REF: @@ -735,6 +628,9 @@ sched_analyze_2 (deps, x, insn) case CC0: /* User of CC0 depends on immediately preceding insn. */ set_sched_group_p (insn); + /* Don't move CC0 setter to another block (it can set up the + same flag for previous CC0 users which is safe). */ + CANT_MOVE (prev_nonnote_insn (insn)) = 1; return; #endif @@ -743,7 +639,7 @@ sched_analyze_2 (deps, x, insn) int regno = REGNO (x); if (regno < FIRST_PSEUDO_REGISTER) { - int i = HARD_REGNO_NREGS (regno, GET_MODE (x)); + int i = hard_regno_nregs[regno][GET_MODE (x)]; while (--i >= 0) SET_REGNO_REG_SET (reg_pending_uses, regno + i); } @@ -791,6 +687,7 @@ sched_analyze_2 (deps, x, insn) cselib_lookup (XEXP (t, 0), Pmode, 1); XEXP (t, 0) = cselib_subst_to_values (XEXP (t, 0)); } + t = canon_rtx (t); pending = deps->pending_read_insns; pending_mem = deps->pending_read_mems; while (pending) @@ -846,7 +743,7 @@ sched_analyze_2 (deps, x, insn) mode. An insn should not be moved across this even if it only uses pseudo-regs because it might give an incorrectly rounded result. */ if (code != ASM_OPERANDS || MEM_VOLATILE_P (x)) - reg_pending_barrier = true; + reg_pending_barrier = TRUE_BARRIER; /* For all ASM_OPERANDS, we must traverse the vector of input operands. We can not just fall through here since then we would be confused @@ -903,10 +800,7 @@ sched_analyze_2 (deps, x, insn) /* Analyze an INSN with pattern X to find all dependencies. */ static void -sched_analyze_insn (deps, x, insn, loop_notes) - struct deps *deps; - rtx x, insn; - rtx loop_notes; +sched_analyze_insn (struct deps *deps, rtx x, rtx insn, rtx loop_notes) { RTX_CODE code = GET_CODE (x); rtx link; @@ -922,7 +816,15 @@ sched_analyze_insn (deps, x, insn, loop_notes) code = GET_CODE (x); } if (code == SET || code == CLOBBER) - sched_analyze_1 (deps, x, insn); + { + sched_analyze_1 (deps, x, insn); + + /* Bare clobber insns are used for letting life analysis, reg-stack + and others know that a value is dead. Depend on the last call + instruction so that reg-stack won't get confused. */ + if (code == CLOBBER) + add_dependence_list (insn, deps->last_function_call, REG_DEP_OUTPUT); + } else if (code == PARALLEL) { int i; @@ -957,7 +859,7 @@ sched_analyze_insn (deps, x, insn, loop_notes) sched_analyze_2 (deps, XEXP (link, 0), insn); } if (find_reg_note (insn, REG_SETJMP, NULL)) - reg_pending_barrier = true; + reg_pending_barrier = MOVE_BARRIER; } if (GET_CODE (insn) == JUMP_INSN) @@ -965,16 +867,29 @@ sched_analyze_insn (deps, x, insn, loop_notes) rtx next; next = next_nonnote_insn (insn); if (next && GET_CODE (next) == BARRIER) - reg_pending_barrier = true; + reg_pending_barrier = TRUE_BARRIER; else { rtx pending, pending_mem; - regset_head tmp; - INIT_REG_SET (&tmp); + regset_head tmp_uses, tmp_sets; + INIT_REG_SET (&tmp_uses); + INIT_REG_SET (&tmp_sets); + + (*current_sched_info->compute_jump_reg_dependencies) + (insn, &deps->reg_conditional_sets, &tmp_uses, &tmp_sets); + /* Make latency of jump equal to 0 by using anti-dependence. */ + EXECUTE_IF_SET_IN_REG_SET (&tmp_uses, 0, i, + { + struct deps_reg *reg_last = &deps->reg_last[i]; + add_dependence_list (insn, reg_last->sets, REG_DEP_ANTI); + add_dependence_list (insn, reg_last->clobbers, REG_DEP_ANTI); + reg_last->uses_length++; + reg_last->uses = alloc_INSN_LIST (insn, reg_last->uses); + }); + IOR_REG_SET (reg_pending_sets, &tmp_sets); - (*current_sched_info->compute_jump_reg_dependencies) (insn, &tmp); - IOR_REG_SET (reg_pending_uses, &tmp); - CLEAR_REG_SET (&tmp); + CLEAR_REG_SET (&tmp_uses); + CLEAR_REG_SET (&tmp_sets); /* All memory writes and volatile reads must happen before the jump. Non-volatile reads must happen before the jump iff @@ -1022,7 +937,7 @@ sched_analyze_insn (deps, x, insn, loop_notes) || INTVAL (XEXP (link, 0)) == NOTE_INSN_LOOP_END || INTVAL (XEXP (link, 0)) == NOTE_INSN_EH_REGION_BEG || INTVAL (XEXP (link, 0)) == NOTE_INSN_EH_REGION_END) - reg_pending_barrier = true; + reg_pending_barrier = MOVE_BARRIER; link = XEXP (link, 1); } @@ -1031,22 +946,28 @@ sched_analyze_insn (deps, x, insn, loop_notes) } /* If this instruction can throw an exception, then moving it changes - where block boundaries fall. This is mighty confusing elsewhere. + where block boundaries fall. This is mighty confusing elsewhere. Therefore, prevent such an instruction from being moved. */ if (can_throw_internal (insn)) - reg_pending_barrier = true; + reg_pending_barrier = MOVE_BARRIER; /* Add dependencies if a scheduling barrier was found. */ if (reg_pending_barrier) { + /* In the case of barrier the most added dependencies are not + real, so we use anti-dependence here. */ if (GET_CODE (PATTERN (insn)) == COND_EXEC) { EXECUTE_IF_SET_IN_REG_SET (&deps->reg_last_in_use, 0, i, { struct deps_reg *reg_last = &deps->reg_last[i]; add_dependence_list (insn, reg_last->uses, REG_DEP_ANTI); - add_dependence_list (insn, reg_last->sets, 0); - add_dependence_list (insn, reg_last->clobbers, 0); + add_dependence_list + (insn, reg_last->sets, + reg_pending_barrier == TRUE_BARRIER ? 0 : REG_DEP_ANTI); + add_dependence_list + (insn, reg_last->clobbers, + reg_pending_barrier == TRUE_BARRIER ? 0 : REG_DEP_ANTI); }); } else @@ -1056,8 +977,12 @@ sched_analyze_insn (deps, x, insn, loop_notes) struct deps_reg *reg_last = &deps->reg_last[i]; add_dependence_list_and_free (insn, ®_last->uses, REG_DEP_ANTI); - add_dependence_list_and_free (insn, ®_last->sets, 0); - add_dependence_list_and_free (insn, ®_last->clobbers, 0); + add_dependence_list_and_free + (insn, ®_last->sets, + reg_pending_barrier == TRUE_BARRIER ? 0 : REG_DEP_ANTI); + add_dependence_list_and_free + (insn, ®_last->clobbers, + reg_pending_barrier == TRUE_BARRIER ? 0 : REG_DEP_ANTI); reg_last->uses_length = 0; reg_last->clobbers_length = 0; }); @@ -1071,7 +996,8 @@ sched_analyze_insn (deps, x, insn, loop_notes) } flush_pending_lists (deps, insn, true, true); - reg_pending_barrier = false; + CLEAR_REG_SET (&deps->reg_conditional_sets); + reg_pending_barrier = NOT_A_BARRIER; } else { @@ -1102,6 +1028,7 @@ sched_analyze_insn (deps, x, insn, loop_notes) add_dependence_list (insn, reg_last->clobbers, REG_DEP_OUTPUT); add_dependence_list (insn, reg_last->uses, REG_DEP_ANTI); reg_last->sets = alloc_INSN_LIST (insn, reg_last->sets); + SET_REGNO_REG_SET (&deps->reg_conditional_sets, i); }); } else @@ -1117,8 +1044,6 @@ sched_analyze_insn (deps, x, insn, loop_notes) EXECUTE_IF_SET_IN_REG_SET (reg_pending_clobbers, 0, i, { struct deps_reg *reg_last = &deps->reg_last[i]; - add_dependence_list (insn, reg_last->sets, REG_DEP_OUTPUT); - add_dependence_list (insn, reg_last->uses, REG_DEP_ANTI); if (reg_last->uses_length > MAX_PENDING_LIST_LENGTH || reg_last->clobbers_length > MAX_PENDING_LIST_LENGTH) { @@ -1128,6 +1053,7 @@ sched_analyze_insn (deps, x, insn, loop_notes) REG_DEP_ANTI); add_dependence_list_and_free (insn, ®_last->clobbers, REG_DEP_OUTPUT); + reg_last->sets = alloc_INSN_LIST (insn, reg_last->sets); reg_last->clobbers_length = 0; reg_last->uses_length = 0; } @@ -1151,6 +1077,7 @@ sched_analyze_insn (deps, x, insn, loop_notes) reg_last->sets = alloc_INSN_LIST (insn, reg_last->sets); reg_last->uses_length = 0; reg_last->clobbers_length = 0; + CLEAR_REGNO_REG_SET (&deps->reg_conditional_sets, i); }); } @@ -1162,6 +1089,16 @@ sched_analyze_insn (deps, x, insn, loop_notes) CLEAR_REG_SET (reg_pending_clobbers); CLEAR_REG_SET (reg_pending_sets); + /* If we are currently in a libcall scheduling group, then mark the + current insn as being in a scheduling group and that it can not + be moved into a different basic block. */ + + if (deps->libcall_block_tail_insn) + { + set_sched_group_p (insn); + CANT_MOVE (insn) = 1; + } + /* If a post-call group is still open, see if it should remain so. This insn must be a simple move of a hard reg to a pseudo or vice-versa. @@ -1190,7 +1127,13 @@ sched_analyze_insn (deps, x, insn, loop_notes) tmp = SET_SRC (set); if (GET_CODE (tmp) == SUBREG) tmp = SUBREG_REG (tmp); - if (GET_CODE (tmp) == REG) + if ((GET_CODE (tmp) == PLUS + || GET_CODE (tmp) == MINUS) + && GET_CODE (XEXP (tmp, 0)) == REG + && REGNO (XEXP (tmp, 0)) == STACK_POINTER_REGNUM + && dest_regno == STACK_POINTER_REGNUM) + src_regno = STACK_POINTER_REGNUM; + else if (GET_CODE (tmp) == REG) src_regno = REGNO (tmp); else goto end_call_group; @@ -1213,26 +1156,23 @@ sched_analyze_insn (deps, x, insn, loop_notes) for every dependency. */ void -sched_analyze (deps, head, tail) - struct deps *deps; - rtx head, tail; +sched_analyze (struct deps *deps, rtx head, rtx tail) { rtx insn; rtx loop_notes = 0; if (current_sched_info->use_cselib) - cselib_init (); + cselib_init (true); for (insn = head;; insn = NEXT_INSN (insn)) { + rtx link, end_seq, r0, set; + if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN) { /* Clear out the stale LOG_LINKS from flow. */ free_INSN_LIST_list (&LOG_LINKS (insn)); - /* Clear out stale SCHED_GROUP_P. */ - SCHED_GROUP_P (insn) = 0; - /* Make each JUMP_INSN a scheduling barrier for memory references. */ if (GET_CODE (insn) == JUMP_INSN) @@ -1251,9 +1191,6 @@ sched_analyze (deps, head, tail) { int i; - /* Clear out stale SCHED_GROUP_P. */ - SCHED_GROUP_P (insn) = 0; - CANT_MOVE (insn) = 1; /* Clear out the stale LOG_LINKS from flow. */ @@ -1263,7 +1200,7 @@ sched_analyze (deps, head, tail) { /* This is setjmp. Assume that all registers, not just hard registers, may be clobbered by this call. */ - reg_pending_barrier = true; + reg_pending_barrier = MOVE_BARRIER; } else { @@ -1274,8 +1211,12 @@ sched_analyze (deps, head, tail) SET_REGNO_REG_SET (reg_pending_sets, i); SET_REGNO_REG_SET (reg_pending_uses, i); } - /* Other call-clobbered hard regs may be clobbered. */ - else if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i)) + /* Other call-clobbered hard regs may be clobbered. + Since we only have a choice between 'might be clobbered' + and 'definitely not clobbered', we must include all + partly call-clobbered registers here. */ + else if (HARD_REGNO_CALL_PART_CLOBBERED (i, reg_raw_mode[i]) + || TEST_HARD_REG_BIT (regs_invalidated_by_call, i)) SET_REGNO_REG_SET (reg_pending_clobbers, i); /* We don't know what set of fixed registers might be used by the function, but it is certain that the stack pointer @@ -1320,17 +1261,7 @@ sched_analyze (deps, head, tail) /* See comments on reemit_notes as to why we do this. ??? Actually, the reemit_notes just say what is done, not why. */ - else if (GET_CODE (insn) == NOTE - && (NOTE_LINE_NUMBER (insn) == NOTE_INSN_RANGE_BEG - || NOTE_LINE_NUMBER (insn) == NOTE_INSN_RANGE_END)) - { - loop_notes = alloc_EXPR_LIST (REG_SAVE_NOTE, NOTE_RANGE_INFO (insn), - loop_notes); - loop_notes = alloc_EXPR_LIST (REG_SAVE_NOTE, - GEN_INT (NOTE_LINE_NUMBER (insn)), - loop_notes); - } - else if (GET_CODE (insn) == NOTE + if (GET_CODE (insn) == NOTE && (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG || NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END || NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_BEG @@ -1342,7 +1273,7 @@ sched_analyze (deps, head, tail) || NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_END) rtx_region = GEN_INT (NOTE_EH_HANDLER (insn)); else - rtx_region = GEN_INT (0); + rtx_region = const0_rtx; loop_notes = alloc_EXPR_LIST (REG_SAVE_NOTE, rtx_region, @@ -1355,6 +1286,46 @@ sched_analyze (deps, head, tail) if (current_sched_info->use_cselib) cselib_process_insn (insn); + + /* Now that we have completed handling INSN, check and see if it is + a CLOBBER beginning a libcall block. If it is, record the + end of the libcall sequence. + + We want to schedule libcall blocks as a unit before reload. While + this restricts scheduling, it preserves the meaning of a libcall + block. + + As a side effect, we may get better code due to decreased register + pressure as well as less chance of a foreign insn appearing in + a libcall block. */ + if (!reload_completed + /* Note we may have nested libcall sequences. We only care about + the outermost libcall sequence. */ + && deps->libcall_block_tail_insn == 0 + /* The sequence must start with a clobber of a register. */ + && GET_CODE (insn) == INSN + && GET_CODE (PATTERN (insn)) == CLOBBER + && (r0 = XEXP (PATTERN (insn), 0), GET_CODE (r0) == REG) + && GET_CODE (XEXP (PATTERN (insn), 0)) == REG + /* The CLOBBER must also have a REG_LIBCALL note attached. */ + && (link = find_reg_note (insn, REG_LIBCALL, NULL_RTX)) != 0 + && (end_seq = XEXP (link, 0)) != 0 + /* The insn referenced by the REG_LIBCALL note must be a + simple nop copy with the same destination as the register + mentioned in the clobber. */ + && (set = single_set (end_seq)) != 0 + && SET_DEST (set) == r0 && SET_SRC (set) == r0 + /* And finally the insn referenced by the REG_LIBCALL must + also contain a REG_EQUAL note and a REG_RETVAL note. */ + && find_reg_note (end_seq, REG_EQUAL, NULL_RTX) != 0 + && find_reg_note (end_seq, REG_RETVAL, NULL_RTX) != 0) + deps->libcall_block_tail_insn = XEXP (link, 0); + + /* If we have reached the end of a libcall block, then close the + block. */ + if (deps->libcall_block_tail_insn == insn) + deps->libcall_block_tail_insn = 0; + if (insn == tail) { if (current_sched_info->use_cselib) @@ -1365,17 +1336,52 @@ sched_analyze (deps, head, tail) abort (); } + +/* The following function adds forward dependence (FROM, TO) with + given DEP_TYPE. The forward dependence should be not exist before. */ + +void +add_forward_dependence (rtx from, rtx to, enum reg_note dep_type) +{ + rtx new_link; + +#ifdef ENABLE_CHECKING + /* If add_dependence is working properly there should never + be notes, deleted insns or duplicates in the backward + links. Thus we need not check for them here. + + However, if we have enabled checking we might as well go + ahead and verify that add_dependence worked properly. */ + if (GET_CODE (from) == NOTE + || INSN_DELETED_P (from) + || (forward_dependency_cache != NULL + && bitmap_bit_p (&forward_dependency_cache[INSN_LUID (from)], + INSN_LUID (to))) + || (forward_dependency_cache == NULL + && find_insn_list (to, INSN_DEPEND (from)))) + abort (); + if (forward_dependency_cache != NULL) + bitmap_bit_p (&forward_dependency_cache[INSN_LUID (from)], + INSN_LUID (to)); +#endif + + new_link = alloc_INSN_LIST (to, INSN_DEPEND (from)); + + PUT_REG_NOTE_KIND (new_link, dep_type); + + INSN_DEPEND (from) = new_link; + INSN_DEP_COUNT (to) += 1; +} + /* Examine insns in the range [ HEAD, TAIL ] and Use the backward dependences from LOG_LINKS to build forward dependences in INSN_DEPEND. */ void -compute_forward_dependences (head, tail) - rtx head, tail; +compute_forward_dependences (rtx head, rtx tail) { rtx insn, link; rtx next_tail; - enum reg_note dep_type; next_tail = NEXT_INSN (tail); for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) @@ -1383,44 +1389,8 @@ compute_forward_dependences (head, tail) if (! INSN_P (insn)) continue; - insn = group_leader (insn); - for (link = LOG_LINKS (insn); link; link = XEXP (link, 1)) - { - rtx x = group_leader (XEXP (link, 0)); - rtx new_link; - - if (x != XEXP (link, 0)) - continue; - -#ifdef ENABLE_CHECKING - /* If add_dependence is working properly there should never - be notes, deleted insns or duplicates in the backward - links. Thus we need not check for them here. - - However, if we have enabled checking we might as well go - ahead and verify that add_dependence worked properly. */ - if (GET_CODE (x) == NOTE - || INSN_DELETED_P (x) - || (forward_dependency_cache != NULL - && TEST_BIT (forward_dependency_cache[INSN_LUID (x)], - INSN_LUID (insn))) - || (forward_dependency_cache == NULL - && find_insn_list (insn, INSN_DEPEND (x)))) - abort (); - if (forward_dependency_cache != NULL) - SET_BIT (forward_dependency_cache[INSN_LUID (x)], - INSN_LUID (insn)); -#endif - - new_link = alloc_INSN_LIST (insn, INSN_DEPEND (x)); - - dep_type = REG_NOTE_KIND (link); - PUT_REG_NOTE_KIND (new_link, dep_type); - - INSN_DEPEND (x) = new_link; - INSN_DEP_COUNT (insn) += 1; - } + add_forward_dependence (XEXP (link, 0), insn, REG_NOTE_KIND (link)); } } @@ -1428,15 +1398,14 @@ compute_forward_dependences (head, tail) n_bbs is the number of region blocks. */ void -init_deps (deps) - struct deps *deps; +init_deps (struct deps *deps) { int max_reg = (reload_completed ? FIRST_PSEUDO_REGISTER : max_reg_num ()); deps->max_reg = max_reg; - deps->reg_last = (struct deps_reg *) - xcalloc (max_reg, sizeof (struct deps_reg)); + deps->reg_last = xcalloc (max_reg, sizeof (struct deps_reg)); INIT_REG_SET (&deps->reg_last_in_use); + INIT_REG_SET (&deps->reg_conditional_sets); deps->pending_read_insns = 0; deps->pending_read_mems = 0; @@ -1448,13 +1417,13 @@ init_deps (deps) deps->last_function_call = 0; deps->sched_before_next_call = 0; deps->in_post_call_group_p = false; + deps->libcall_block_tail_insn = 0; } /* Free insn lists found in DEPS. */ void -free_deps (deps) - struct deps *deps; +free_deps (struct deps *deps) { int i; @@ -1465,27 +1434,30 @@ free_deps (deps) free_INSN_LIST_list (&deps->last_pending_memory_flush); /* Without the EXECUTE_IF_SET, this loop is executed max_reg * nr_regions - times. For a test case with 42000 regs and 8000 small basic blocks, + times. For a testcase with 42000 regs and 8000 small basic blocks, this loop accounted for nearly 60% (84 sec) of the total -O2 runtime. */ EXECUTE_IF_SET_IN_REG_SET (&deps->reg_last_in_use, 0, i, { struct deps_reg *reg_last = &deps->reg_last[i]; - free_INSN_LIST_list (®_last->uses); - free_INSN_LIST_list (®_last->sets); - free_INSN_LIST_list (®_last->clobbers); + if (reg_last->uses) + free_INSN_LIST_list (®_last->uses); + if (reg_last->sets) + free_INSN_LIST_list (®_last->sets); + if (reg_last->clobbers) + free_INSN_LIST_list (®_last->clobbers); }); CLEAR_REG_SET (&deps->reg_last_in_use); + CLEAR_REG_SET (&deps->reg_conditional_sets); free (deps->reg_last); } /* If it is profitable to use them, initialize caches for tracking - dependency informatino. LUID is the number of insns to be scheduled, + dependency information. LUID is the number of insns to be scheduled, it is used in the estimate of profitability. */ void -init_dependency_caches (luid) - int luid; +init_dependency_caches (int luid) { /* ?!? We could save some memory by computing a per-region luid mapping which could reduce both the number of vectors in the cache and the size @@ -1495,34 +1467,52 @@ init_dependency_caches (luid) what we consider "very high". */ if (luid / n_basic_blocks > 100 * 5) { - true_dependency_cache = sbitmap_vector_alloc (luid, luid); - sbitmap_vector_zero (true_dependency_cache, luid); - anti_dependency_cache = sbitmap_vector_alloc (luid, luid); - sbitmap_vector_zero (anti_dependency_cache, luid); - output_dependency_cache = sbitmap_vector_alloc (luid, luid); - sbitmap_vector_zero (output_dependency_cache, luid); + int i; + true_dependency_cache = xmalloc (luid * sizeof (bitmap_head)); + anti_dependency_cache = xmalloc (luid * sizeof (bitmap_head)); + output_dependency_cache = xmalloc (luid * sizeof (bitmap_head)); +#ifdef ENABLE_CHECKING + forward_dependency_cache = xmalloc (luid * sizeof (bitmap_head)); +#endif + for (i = 0; i < luid; i++) + { + bitmap_initialize (&true_dependency_cache[i], 0); + bitmap_initialize (&anti_dependency_cache[i], 0); + bitmap_initialize (&output_dependency_cache[i], 0); #ifdef ENABLE_CHECKING - forward_dependency_cache = sbitmap_vector_alloc (luid, luid); - sbitmap_vector_zero (forward_dependency_cache, luid); + bitmap_initialize (&forward_dependency_cache[i], 0); #endif + } + cache_size = luid; } } /* Free the caches allocated in init_dependency_caches. */ void -free_dependency_caches () +free_dependency_caches (void) { if (true_dependency_cache) { - sbitmap_vector_free (true_dependency_cache); + int i; + + for (i = 0; i < cache_size; i++) + { + bitmap_clear (&true_dependency_cache[i]); + bitmap_clear (&anti_dependency_cache[i]); + bitmap_clear (&output_dependency_cache[i]); +#ifdef ENABLE_CHECKING + bitmap_clear (&forward_dependency_cache[i]); +#endif + } + free (true_dependency_cache); true_dependency_cache = NULL; - sbitmap_vector_free (anti_dependency_cache); + free (anti_dependency_cache); anti_dependency_cache = NULL; - sbitmap_vector_free (output_dependency_cache); + free (output_dependency_cache); output_dependency_cache = NULL; #ifdef ENABLE_CHECKING - sbitmap_vector_free (forward_dependency_cache); + free (forward_dependency_cache); forward_dependency_cache = NULL; #endif } @@ -1532,18 +1522,18 @@ free_dependency_caches () code. */ void -init_deps_global () +init_deps_global (void) { reg_pending_sets = INITIALIZE_REG_SET (reg_pending_sets_head); reg_pending_clobbers = INITIALIZE_REG_SET (reg_pending_clobbers_head); reg_pending_uses = INITIALIZE_REG_SET (reg_pending_uses_head); - reg_pending_barrier = false; + reg_pending_barrier = NOT_A_BARRIER; } /* Free everything used by the dependency analysis code. */ void -finish_deps_global () +finish_deps_global (void) { FREE_REG_SET (reg_pending_sets); FREE_REG_SET (reg_pending_clobbers);