2010-11-16 Chung-Lin Tang 2010-07-21 Richard Henderson gcc/ * config/i386/i386.c (setup_incoming_varargs_64): Emit a simple comparison for avoiding xmm register saves. Emit the xmm register saves explicitly. * config/i386/i386.md (UNSPEC_SSE_PROLOGUE_SAVE): Remove. (UNSPEC_SSE_PROLOGUE_SAVE_LOW): Remove. (sse_prologue_save, sse_prologue_save_insn1, sse_prologue_save_insn): Remove patterns and the associated splitters. 2010-07-22 Richard Henderson gcc/ PR target/45027 * config/i386/i386.c (setup_incoming_varargs_64): Force the use of V4SFmode for the SSE saves; increase stack alignment if needed. 2010-11-16 Chung-Lin Tang Re-merge, backport from mainline: 2010-07-15 Bernd Schmidt gcc/ * postreload.c (last_label_ruid, first_index_reg, last_index_reg): New static variables. (reload_combine_recognize_pattern): New static function, broken out of reload_combine. (reload_combine): Use it. Only initialize first_index_reg and last_index_reg once. 2010-07-17 Bernd Schmidt PR target/42235 gcc/ * postreload.c (reload_cse_move2add): Return bool, true if anything. changed. All callers changed. (move2add_use_add2_insn): Likewise. (move2add_use_add3_insn): Likewise. (reload_cse_regs): If reload_cse_move2add changed anything, rerun reload_combine. (RELOAD_COMBINE_MAX_USES): Bump to 16. (last_jump_ruid): New static variable. (struct reg_use): New members CONTAINING_MEM and RUID. (reg_state): New members ALL_OFFSETS_MATCH and REAL_STORE_RUID. (reload_combine_split_one_ruid, reload_combine_split_ruids, reload_combine_purge_insn_uses, reload_combine_closest_single_use reload_combine_purge_reg_uses_after_ruid, reload_combine_recognize_const_pattern): New static functions. (reload_combine_recognize_pattern): Verify that ALL_OFFSETS_MATCH is true for our reg and that we have available index regs. (reload_combine_note_use): New args RUID and CONTAINING_MEM. All callers changed. Use them to initialize fields in struct reg_use. (reload_combine): Initialize last_jump_ruid. Be careful when to take PREV_INSN of the scanned insn. Update REAL_STORE_RUID fields. Call reload_combine_recognize_const_pattern. (reload_combine_note_store): Update REAL_STORE_RUID field. gcc/testsuite/ * gcc.target/arm/pr42235.c: New test. 2010-07-19 Bernd Schmidt gcc/ * postreload.c (reload_combine_closest_single_use): Ignore the number of uses for DEBUG_INSNs. (fixup_debug_insns): New static function. (reload_combine_recognize_const_pattern): Use it. Don't let the main loop be affected by DEBUG_INSNs. Really disallow moving adds past a jump insn. (reload_combine_recognize_pattern): Don't update use_ruid here. (reload_combine_note_use): Do it here. (reload_combine): Use control_flow_insn_p rather than JUMP_P. 2010-07-20 Bernd Schmidt gcc/ * postreload.c (fixup_debug_insns): Remove arg REGNO. New args FROM and TO. All callers changed. Don't look for tracked uses, just scan the RTL for DEBUG_INSNs and substitute. (reload_combine_recognize_pattern): Call fixup_debug_insns. (reload_combine): Ignore DEBUG_INSNs. 2010-07-22 Bernd Schmidt PR bootstrap/44970 PR middle-end/45009 gcc/ * postreload.c: Include "target.h". (reload_combine_closest_single_use): Don't take DEBUG_INSNs into account. (fixup_debug_insns): Don't copy the rtx. (reload_combine_recognize_const_pattern): DEBUG_INSNs can't have uses. Don't copy when replacing. Call fixup_debug_insns in the case where we merged one add with another. (reload_combine_recognize_pattern): Fail if there aren't any uses. Try harder to determine whether we're picking a valid index register. Don't set store_ruid for an insn we're going to scan in the next iteration. (reload_combine): Remove unused code. (reload_combine_note_use): When updating use information for an old insn, ignore a use that occurs after store_ruid. * Makefile.in (postreload.o): Update dependencies. 2010-07-27 Bernd Schmidt gcc/ * postreload.c (reload_combine_recognize_const_pattern): Move test for limiting the insn movement to the right scope. 2010-07-27 Bernd Schmidt gcc/ * postreload.c (try_replace_in_use): New static function. (reload_combine_recognize_const_pattern): Use it here. Allow substituting into a final add insn, and substituting into a memory reference in an insn that sets the reg. === modified file 'gcc/Makefile.in' --- old/gcc/Makefile.in 2010-11-11 11:34:59 +0000 +++ new/gcc/Makefile.in 2010-11-16 18:05:53 +0000 @@ -3157,7 +3157,7 @@ $(RTL_H) $(REAL_H) $(FLAGS_H) $(EXPR_H) $(OPTABS_H) reload.h $(REGS_H) \ hard-reg-set.h insn-config.h $(BASIC_BLOCK_H) $(RECOG_H) output.h \ $(FUNCTION_H) $(TOPLEV_H) cselib.h $(TM_P_H) $(EXCEPT_H) $(TREE_H) $(MACHMODE_H) \ - $(OBSTACK_H) $(TIMEVAR_H) $(TREE_PASS_H) $(DF_H) $(DBGCNT_H) + $(OBSTACK_H) $(TARGET_H) $(TIMEVAR_H) $(TREE_PASS_H) $(DF_H) $(DBGCNT_H) postreload-gcse.o : postreload-gcse.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \ $(TM_H) $(RTL_H) $(REGS_H) hard-reg-set.h $(FLAGS_H) insn-config.h \ $(RECOG_H) $(EXPR_H) $(BASIC_BLOCK_H) $(FUNCTION_H) output.h $(TOPLEV_H) \ === modified file 'gcc/config/i386/i386.c' --- old/gcc/config/i386/i386.c 2010-09-30 20:24:54 +0000 +++ new/gcc/config/i386/i386.c 2010-11-16 18:05:53 +0000 @@ -6737,12 +6737,8 @@ setup_incoming_varargs_64 (CUMULATIVE_ARGS *cum) { rtx save_area, mem; - rtx label; - rtx label_ref; - rtx tmp_reg; - rtx nsse_reg; alias_set_type set; - int i; + int i, max; /* GPR size of varargs save area. */ if (cfun->va_list_gpr_size) @@ -6752,7 +6748,7 @@ /* FPR size of varargs save area. We don't need it if we don't pass anything in SSE registers. */ - if (cum->sse_nregs && cfun->va_list_fpr_size) + if (TARGET_SSE && cfun->va_list_fpr_size) ix86_varargs_fpr_size = X86_64_SSE_REGPARM_MAX * 16; else ix86_varargs_fpr_size = 0; @@ -6763,10 +6759,11 @@ save_area = frame_pointer_rtx; set = get_varargs_alias_set (); - for (i = cum->regno; - i < X86_64_REGPARM_MAX - && i < cum->regno + cfun->va_list_gpr_size / UNITS_PER_WORD; - i++) + max = cum->regno + cfun->va_list_gpr_size / UNITS_PER_WORD; + if (max > X86_64_REGPARM_MAX) + max = X86_64_REGPARM_MAX; + + for (i = cum->regno; i < max; i++) { mem = gen_rtx_MEM (Pmode, plus_constant (save_area, i * UNITS_PER_WORD)); @@ -6778,62 +6775,42 @@ if (ix86_varargs_fpr_size) { - /* Stack must be aligned to 16byte for FP register save area. */ - if (crtl->stack_alignment_needed < 128) - crtl->stack_alignment_needed = 128; + enum machine_mode smode; + rtx label, test; /* Now emit code to save SSE registers. The AX parameter contains number - of SSE parameter registers used to call this function. We use - sse_prologue_save insn template that produces computed jump across - SSE saves. We need some preparation work to get this working. */ + of SSE parameter registers used to call this function, though all we + actually check here is the zero/non-zero status. */ label = gen_label_rtx (); - label_ref = gen_rtx_LABEL_REF (Pmode, label); - - /* Compute address to jump to : - label - eax*4 + nnamed_sse_arguments*4 Or - label - eax*5 + nnamed_sse_arguments*5 for AVX. */ - tmp_reg = gen_reg_rtx (Pmode); - nsse_reg = gen_reg_rtx (Pmode); - emit_insn (gen_zero_extendqidi2 (nsse_reg, gen_rtx_REG (QImode, AX_REG))); - emit_insn (gen_rtx_SET (VOIDmode, tmp_reg, - gen_rtx_MULT (Pmode, nsse_reg, - GEN_INT (4)))); - - /* vmovaps is one byte longer than movaps. */ - if (TARGET_AVX) - emit_insn (gen_rtx_SET (VOIDmode, tmp_reg, - gen_rtx_PLUS (Pmode, tmp_reg, - nsse_reg))); - - if (cum->sse_regno) - emit_move_insn - (nsse_reg, - gen_rtx_CONST (DImode, - gen_rtx_PLUS (DImode, - label_ref, - GEN_INT (cum->sse_regno - * (TARGET_AVX ? 5 : 4))))); - else - emit_move_insn (nsse_reg, label_ref); - emit_insn (gen_subdi3 (nsse_reg, nsse_reg, tmp_reg)); - - /* Compute address of memory block we save into. We always use pointer - pointing 127 bytes after first byte to store - this is needed to keep - instruction size limited by 4 bytes (5 bytes for AVX) with one - byte displacement. */ - tmp_reg = gen_reg_rtx (Pmode); - emit_insn (gen_rtx_SET (VOIDmode, tmp_reg, - plus_constant (save_area, - ix86_varargs_gpr_size + 127))); - mem = gen_rtx_MEM (BLKmode, plus_constant (tmp_reg, -127)); - MEM_NOTRAP_P (mem) = 1; - set_mem_alias_set (mem, set); - set_mem_align (mem, BITS_PER_WORD); - - /* And finally do the dirty job! */ - emit_insn (gen_sse_prologue_save (mem, nsse_reg, - GEN_INT (cum->sse_regno), label)); + test = gen_rtx_EQ (VOIDmode, gen_rtx_REG (QImode, AX_REG), const0_rtx); + emit_jump_insn (gen_cbranchqi4 (test, XEXP (test, 0), XEXP (test, 1), + label)); + + /* ??? If !TARGET_SSE_TYPELESS_STORES, would we perform better if + we used movdqa (i.e. TImode) instead? Perhaps even better would + be if we could determine the real mode of the data, via a hook + into pass_stdarg. Ignore all that for now. */ + smode = V4SFmode; + if (crtl->stack_alignment_needed < GET_MODE_ALIGNMENT (smode)) + crtl->stack_alignment_needed = GET_MODE_ALIGNMENT (smode); + + max = cum->sse_regno + cfun->va_list_fpr_size / 16; + if (max > X86_64_SSE_REGPARM_MAX) + max = X86_64_SSE_REGPARM_MAX; + + for (i = cum->sse_regno; i < max; ++i) + { + mem = plus_constant (save_area, i * 16 + ix86_varargs_gpr_size); + mem = gen_rtx_MEM (smode, mem); + MEM_NOTRAP_P (mem) = 1; + set_mem_alias_set (mem, set); + set_mem_align (mem, GET_MODE_ALIGNMENT (smode)); + + emit_move_insn (mem, gen_rtx_REG (smode, SSE_REGNO (i))); + } + + emit_label (label); } } === modified file 'gcc/config/i386/i386.md' --- old/gcc/config/i386/i386.md 2010-10-22 04:56:41 +0000 +++ new/gcc/config/i386/i386.md 2010-11-27 15:24:12 +0000 @@ -80,7 +80,6 @@ ; Prologue support (UNSPEC_STACK_ALLOC 11) (UNSPEC_SET_GOT 12) - (UNSPEC_SSE_PROLOGUE_SAVE 13) (UNSPEC_REG_SAVE 14) (UNSPEC_DEF_CFA 15) (UNSPEC_SET_RIP 16) @@ -20252,74 +20251,6 @@ { return ASM_SHORT "0x0b0f"; } [(set_attr "length" "2")]) -(define_expand "sse_prologue_save" - [(parallel [(set (match_operand:BLK 0 "" "") - (unspec:BLK [(reg:DI XMM0_REG) - (reg:DI XMM1_REG) - (reg:DI XMM2_REG) - (reg:DI XMM3_REG) - (reg:DI XMM4_REG) - (reg:DI XMM5_REG) - (reg:DI XMM6_REG) - (reg:DI XMM7_REG)] UNSPEC_SSE_PROLOGUE_SAVE)) - (use (match_operand:DI 1 "register_operand" "")) - (use (match_operand:DI 2 "immediate_operand" "")) - (use (label_ref:DI (match_operand 3 "" "")))])] - "TARGET_64BIT" - "") - -(define_insn "*sse_prologue_save_insn" - [(set (mem:BLK (plus:DI (match_operand:DI 0 "register_operand" "R") - (match_operand:DI 4 "const_int_operand" "n"))) - (unspec:BLK [(reg:DI XMM0_REG) - (reg:DI XMM1_REG) - (reg:DI XMM2_REG) - (reg:DI XMM3_REG) - (reg:DI XMM4_REG) - (reg:DI XMM5_REG) - (reg:DI XMM6_REG) - (reg:DI XMM7_REG)] UNSPEC_SSE_PROLOGUE_SAVE)) - (use (match_operand:DI 1 "register_operand" "r")) - (use (match_operand:DI 2 "const_int_operand" "i")) - (use (label_ref:DI (match_operand 3 "" "X")))] - "TARGET_64BIT - && INTVAL (operands[4]) + X86_64_SSE_REGPARM_MAX * 16 - 16 < 128 - && INTVAL (operands[4]) + INTVAL (operands[2]) * 16 >= -128" -{ - int i; - operands[0] = gen_rtx_MEM (Pmode, - gen_rtx_PLUS (Pmode, operands[0], operands[4])); - /* VEX instruction with a REX prefix will #UD. */ - if (TARGET_AVX && GET_CODE (XEXP (operands[0], 0)) != PLUS) - gcc_unreachable (); - - output_asm_insn ("jmp\t%A1", operands); - for (i = X86_64_SSE_REGPARM_MAX - 1; i >= INTVAL (operands[2]); i--) - { - operands[4] = adjust_address (operands[0], DImode, i*16); - operands[5] = gen_rtx_REG (TImode, SSE_REGNO (i)); - PUT_MODE (operands[4], TImode); - if (GET_CODE (XEXP (operands[0], 0)) != PLUS) - output_asm_insn ("rex", operands); - output_asm_insn ("%vmovaps\t{%5, %4|%4, %5}", operands); - } - (*targetm.asm_out.internal_label) (asm_out_file, "L", - CODE_LABEL_NUMBER (operands[3])); - return ""; -} - [(set_attr "type" "other") - (set_attr "length_immediate" "0") - (set_attr "length_address" "0") - (set (attr "length") - (if_then_else - (eq (symbol_ref "TARGET_AVX") (const_int 0)) - (const_string "34") - (const_string "42"))) - (set_attr "memory" "store") - (set_attr "modrm" "0") - (set_attr "prefix" "maybe_vex") - (set_attr "mode" "DI")]) - (define_expand "prefetch" [(prefetch (match_operand 0 "address_operand" "") (match_operand:SI 1 "const_int_operand" "") === modified file 'gcc/postreload.c' --- old/gcc/postreload.c 2010-11-08 22:08:43 +0000 +++ new/gcc/postreload.c 2010-11-16 18:05:53 +0000 @@ -44,6 +44,7 @@ #include "toplev.h" #include "except.h" #include "tree.h" +#include "target.h" #include "timevar.h" #include "tree-pass.h" #include "df.h" @@ -56,10 +57,10 @@ static int reload_cse_simplify_operands (rtx, rtx); static void reload_combine (void); -static void reload_combine_note_use (rtx *, rtx); +static void reload_combine_note_use (rtx *, rtx, int, rtx); static void reload_combine_note_store (rtx, const_rtx, void *); -static void reload_cse_move2add (rtx); +static bool reload_cse_move2add (rtx); static void move2add_note_store (rtx, const_rtx, void *); /* Call cse / combine like post-reload optimization phases. @@ -67,11 +68,16 @@ void reload_cse_regs (rtx first ATTRIBUTE_UNUSED) { + bool moves_converted; reload_cse_regs_1 (first); reload_combine (); - reload_cse_move2add (first); + moves_converted = reload_cse_move2add (first); if (flag_expensive_optimizations) - reload_cse_regs_1 (first); + { + if (moves_converted) + reload_combine (); + reload_cse_regs_1 (first); + } } /* See whether a single set SET is a noop. */ @@ -660,30 +666,43 @@ /* The maximum number of uses of a register we can keep track of to replace them with reg+reg addressing. */ -#define RELOAD_COMBINE_MAX_USES 6 +#define RELOAD_COMBINE_MAX_USES 16 -/* INSN is the insn where a register has been used, and USEP points to the - location of the register within the rtl. */ -struct reg_use { rtx insn, *usep; }; +/* Describes a recorded use of a register. */ +struct reg_use +{ + /* The insn where a register has been used. */ + rtx insn; + /* Points to the memory reference enclosing the use, if any, NULL_RTX + otherwise. */ + rtx containing_mem; + /* Location of the register withing INSN. */ + rtx *usep; + /* The reverse uid of the insn. */ + int ruid; +}; /* If the register is used in some unknown fashion, USE_INDEX is negative. If it is dead, USE_INDEX is RELOAD_COMBINE_MAX_USES, and STORE_RUID - indicates where it becomes live again. + indicates where it is first set or clobbered. Otherwise, USE_INDEX is the index of the last encountered use of the - register (which is first among these we have seen since we scan backwards), - OFFSET contains the constant offset that is added to the register in - all encountered uses, and USE_RUID indicates the first encountered, i.e. - last, of these uses. + register (which is first among these we have seen since we scan backwards). + USE_RUID indicates the first encountered, i.e. last, of these uses. + If ALL_OFFSETS_MATCH is true, all encountered uses were inside a PLUS + with a constant offset; OFFSET contains this constant in that case. STORE_RUID is always meaningful if we only want to use a value in a register in a different place: it denotes the next insn in the insn - stream (i.e. the last encountered) that sets or clobbers the register. */ + stream (i.e. the last encountered) that sets or clobbers the register. + REAL_STORE_RUID is similar, but clobbers are ignored when updating it. */ static struct { struct reg_use reg_use[RELOAD_COMBINE_MAX_USES]; + rtx offset; int use_index; - rtx offset; int store_ruid; + int real_store_ruid; int use_ruid; + bool all_offsets_match; } reg_state[FIRST_PSEUDO_REGISTER]; /* Reverse linear uid. This is increased in reload_combine while scanning @@ -691,42 +710,548 @@ and the store_ruid / use_ruid fields in reg_state. */ static int reload_combine_ruid; +/* The RUID of the last label we encountered in reload_combine. */ +static int last_label_ruid; + +/* The RUID of the last jump we encountered in reload_combine. */ +static int last_jump_ruid; + +/* The register numbers of the first and last index register. A value of + -1 in LAST_INDEX_REG indicates that we've previously computed these + values and found no suitable index registers. */ +static int first_index_reg = -1; +static int last_index_reg; + #define LABEL_LIVE(LABEL) \ (label_live[CODE_LABEL_NUMBER (LABEL) - min_labelno]) +/* Subroutine of reload_combine_split_ruids, called to fix up a single + ruid pointed to by *PRUID if it is higher than SPLIT_RUID. */ + +static inline void +reload_combine_split_one_ruid (int *pruid, int split_ruid) +{ + if (*pruid > split_ruid) + (*pruid)++; +} + +/* Called when we insert a new insn in a position we've already passed in + the scan. Examine all our state, increasing all ruids that are higher + than SPLIT_RUID by one in order to make room for a new insn. */ + +static void +reload_combine_split_ruids (int split_ruid) +{ + unsigned i; + + reload_combine_split_one_ruid (&reload_combine_ruid, split_ruid); + reload_combine_split_one_ruid (&last_label_ruid, split_ruid); + reload_combine_split_one_ruid (&last_jump_ruid, split_ruid); + + for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) + { + int j, idx = reg_state[i].use_index; + reload_combine_split_one_ruid (®_state[i].use_ruid, split_ruid); + reload_combine_split_one_ruid (®_state[i].store_ruid, split_ruid); + reload_combine_split_one_ruid (®_state[i].real_store_ruid, + split_ruid); + if (idx < 0) + continue; + for (j = idx; j < RELOAD_COMBINE_MAX_USES; j++) + { + reload_combine_split_one_ruid (®_state[i].reg_use[j].ruid, + split_ruid); + } + } +} + +/* Called when we are about to rescan a previously encountered insn with + reload_combine_note_use after modifying some part of it. This clears all + information about uses in that particular insn. */ + +static void +reload_combine_purge_insn_uses (rtx insn) +{ + unsigned i; + + for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) + { + int j, k, idx = reg_state[i].use_index; + if (idx < 0) + continue; + j = k = RELOAD_COMBINE_MAX_USES; + while (j-- > idx) + { + if (reg_state[i].reg_use[j].insn != insn) + { + k--; + if (k != j) + reg_state[i].reg_use[k] = reg_state[i].reg_use[j]; + } + } + reg_state[i].use_index = k; + } +} + +/* Called when we need to forget about all uses of REGNO after an insn + which is identified by RUID. */ + +static void +reload_combine_purge_reg_uses_after_ruid (unsigned regno, int ruid) +{ + int j, k, idx = reg_state[regno].use_index; + if (idx < 0) + return; + j = k = RELOAD_COMBINE_MAX_USES; + while (j-- > idx) + { + if (reg_state[regno].reg_use[j].ruid >= ruid) + { + k--; + if (k != j) + reg_state[regno].reg_use[k] = reg_state[regno].reg_use[j]; + } + } + reg_state[regno].use_index = k; +} + +/* Find the use of REGNO with the ruid that is highest among those + lower than RUID_LIMIT, and return it if it is the only use of this + reg in the insn. Return NULL otherwise. */ + +static struct reg_use * +reload_combine_closest_single_use (unsigned regno, int ruid_limit) +{ + int i, best_ruid = 0; + int use_idx = reg_state[regno].use_index; + struct reg_use *retval; + + if (use_idx < 0) + return NULL; + retval = NULL; + for (i = use_idx; i < RELOAD_COMBINE_MAX_USES; i++) + { + struct reg_use *use = reg_state[regno].reg_use + i; + int this_ruid = use->ruid; + if (this_ruid >= ruid_limit) + continue; + if (this_ruid > best_ruid) + { + best_ruid = this_ruid; + retval = use; + } + else if (this_ruid == best_ruid) + retval = NULL; + } + if (last_label_ruid >= best_ruid) + return NULL; + return retval; +} + +/* After we've moved an add insn, fix up any debug insns that occur + between the old location of the add and the new location. REG is + the destination register of the add insn; REPLACEMENT is the + SET_SRC of the add. FROM and TO specify the range in which we + should make this change on debug insns. */ + +static void +fixup_debug_insns (rtx reg, rtx replacement, rtx from, rtx to) +{ + rtx insn; + for (insn = from; insn != to; insn = NEXT_INSN (insn)) + { + rtx t; + + if (!DEBUG_INSN_P (insn)) + continue; + + t = INSN_VAR_LOCATION_LOC (insn); + t = simplify_replace_rtx (t, reg, replacement); + validate_change (insn, &INSN_VAR_LOCATION_LOC (insn), t, 0); + } +} + +/* Subroutine of reload_combine_recognize_const_pattern. Try to replace REG + with SRC in the insn described by USE, taking costs into account. Return + true if we made the replacement. */ + +static bool +try_replace_in_use (struct reg_use *use, rtx reg, rtx src) +{ + rtx use_insn = use->insn; + rtx mem = use->containing_mem; + bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (use_insn)); + + if (mem != NULL_RTX) + { + addr_space_t as = MEM_ADDR_SPACE (mem); + rtx oldaddr = XEXP (mem, 0); + rtx newaddr = NULL_RTX; + int old_cost = address_cost (oldaddr, GET_MODE (mem), as, speed); + int new_cost; + + newaddr = simplify_replace_rtx (oldaddr, reg, src); + if (memory_address_addr_space_p (GET_MODE (mem), newaddr, as)) + { + XEXP (mem, 0) = newaddr; + new_cost = address_cost (newaddr, GET_MODE (mem), as, speed); + XEXP (mem, 0) = oldaddr; + if (new_cost <= old_cost + && validate_change (use_insn, + &XEXP (mem, 0), newaddr, 0)) + return true; + } + } + else + { + rtx new_set = single_set (use_insn); + if (new_set + && REG_P (SET_DEST (new_set)) + && GET_CODE (SET_SRC (new_set)) == PLUS + && REG_P (XEXP (SET_SRC (new_set), 0)) + && CONSTANT_P (XEXP (SET_SRC (new_set), 1))) + { + rtx new_src; + int old_cost = rtx_cost (SET_SRC (new_set), SET, speed); + + gcc_assert (rtx_equal_p (XEXP (SET_SRC (new_set), 0), reg)); + new_src = simplify_replace_rtx (SET_SRC (new_set), reg, src); + + if (rtx_cost (new_src, SET, speed) <= old_cost + && validate_change (use_insn, &SET_SRC (new_set), + new_src, 0)) + return true; + } + } + return false; +} + +/* Called by reload_combine when scanning INSN. This function tries to detect + patterns where a constant is added to a register, and the result is used + in an address. + Return true if no further processing is needed on INSN; false if it wasn't + recognized and should be handled normally. */ + +static bool +reload_combine_recognize_const_pattern (rtx insn) +{ + int from_ruid = reload_combine_ruid; + rtx set, pat, reg, src, addreg; + unsigned int regno; + struct reg_use *use; + bool must_move_add; + rtx add_moved_after_insn = NULL_RTX; + int add_moved_after_ruid = 0; + int clobbered_regno = -1; + + set = single_set (insn); + if (set == NULL_RTX) + return false; + + reg = SET_DEST (set); + src = SET_SRC (set); + if (!REG_P (reg) + || hard_regno_nregs[REGNO (reg)][GET_MODE (reg)] != 1 + || GET_MODE (reg) != Pmode + || reg == stack_pointer_rtx) + return false; + + regno = REGNO (reg); + + /* We look for a REG1 = REG2 + CONSTANT insn, followed by either + uses of REG1 inside an address, or inside another add insn. If + possible and profitable, merge the addition into subsequent + uses. */ + if (GET_CODE (src) != PLUS + || !REG_P (XEXP (src, 0)) + || !CONSTANT_P (XEXP (src, 1))) + return false; + + addreg = XEXP (src, 0); + must_move_add = rtx_equal_p (reg, addreg); + + pat = PATTERN (insn); + if (must_move_add && set != pat) + { + /* We have to be careful when moving the add; apart from the + single_set there may also be clobbers. Recognize one special + case, that of one clobber alongside the set (likely a clobber + of the CC register). */ + gcc_assert (GET_CODE (PATTERN (insn)) == PARALLEL); + if (XVECLEN (pat, 0) != 2 || XVECEXP (pat, 0, 0) != set + || GET_CODE (XVECEXP (pat, 0, 1)) != CLOBBER + || !REG_P (XEXP (XVECEXP (pat, 0, 1), 0))) + return false; + clobbered_regno = REGNO (XEXP (XVECEXP (pat, 0, 1), 0)); + } + + do + { + use = reload_combine_closest_single_use (regno, from_ruid); + + if (use) + /* Start the search for the next use from here. */ + from_ruid = use->ruid; + + if (use && GET_MODE (*use->usep) == Pmode) + { + bool delete_add = false; + rtx use_insn = use->insn; + int use_ruid = use->ruid; + + /* Avoid moving the add insn past a jump. */ + if (must_move_add && use_ruid <= last_jump_ruid) + break; + + /* If the add clobbers another hard reg in parallel, don't move + it past a real set of this hard reg. */ + if (must_move_add && clobbered_regno >= 0 + && reg_state[clobbered_regno].real_store_ruid >= use_ruid) + break; + + gcc_assert (reg_state[regno].store_ruid <= use_ruid); + /* Avoid moving a use of ADDREG past a point where it is stored. */ + if (reg_state[REGNO (addreg)].store_ruid > use_ruid) + break; + + /* We also must not move the addition past an insn that sets + the same register, unless we can combine two add insns. */ + if (must_move_add && reg_state[regno].store_ruid == use_ruid) + { + if (use->containing_mem == NULL_RTX) + delete_add = true; + else + break; + } + + if (try_replace_in_use (use, reg, src)) + { + reload_combine_purge_insn_uses (use_insn); + reload_combine_note_use (&PATTERN (use_insn), use_insn, + use_ruid, NULL_RTX); + + if (delete_add) + { + fixup_debug_insns (reg, src, insn, use_insn); + delete_insn (insn); + return true; + } + if (must_move_add) + { + add_moved_after_insn = use_insn; + add_moved_after_ruid = use_ruid; + } + continue; + } + } + /* If we get here, we couldn't handle this use. */ + if (must_move_add) + break; + } + while (use); + + if (!must_move_add || add_moved_after_insn == NULL_RTX) + /* Process the add normally. */ + return false; + + fixup_debug_insns (reg, src, insn, add_moved_after_insn); + + reorder_insns (insn, insn, add_moved_after_insn); + reload_combine_purge_reg_uses_after_ruid (regno, add_moved_after_ruid); + reload_combine_split_ruids (add_moved_after_ruid - 1); + reload_combine_note_use (&PATTERN (insn), insn, + add_moved_after_ruid, NULL_RTX); + reg_state[regno].store_ruid = add_moved_after_ruid; + + return true; +} + +/* Called by reload_combine when scanning INSN. Try to detect a pattern we + can handle and improve. Return true if no further processing is needed on + INSN; false if it wasn't recognized and should be handled normally. */ + +static bool +reload_combine_recognize_pattern (rtx insn) +{ + rtx set, reg, src; + unsigned int regno; + + set = single_set (insn); + if (set == NULL_RTX) + return false; + + reg = SET_DEST (set); + src = SET_SRC (set); + if (!REG_P (reg) + || hard_regno_nregs[REGNO (reg)][GET_MODE (reg)] != 1) + return false; + + regno = REGNO (reg); + + /* Look for (set (REGX) (CONST_INT)) + (set (REGX) (PLUS (REGX) (REGY))) + ... + ... (MEM (REGX)) ... + and convert it to + (set (REGZ) (CONST_INT)) + ... + ... (MEM (PLUS (REGZ) (REGY)))... . + + First, check that we have (set (REGX) (PLUS (REGX) (REGY))) + and that we know all uses of REGX before it dies. + Also, explicitly check that REGX != REGY; our life information + does not yet show whether REGY changes in this insn. */ + + if (GET_CODE (src) == PLUS + && reg_state[regno].all_offsets_match + && last_index_reg != -1 + && REG_P (XEXP (src, 1)) + && rtx_equal_p (XEXP (src, 0), reg) + && !rtx_equal_p (XEXP (src, 1), reg) + && reg_state[regno].use_index >= 0 + && reg_state[regno].use_index < RELOAD_COMBINE_MAX_USES + && last_label_ruid < reg_state[regno].use_ruid) + { + rtx base = XEXP (src, 1); + rtx prev = prev_nonnote_insn (insn); + rtx prev_set = prev ? single_set (prev) : NULL_RTX; + rtx index_reg = NULL_RTX; + rtx reg_sum = NULL_RTX; + int i; + + /* Now we need to set INDEX_REG to an index register (denoted as + REGZ in the illustration above) and REG_SUM to the expression + register+register that we want to use to substitute uses of REG + (typically in MEMs) with. First check REG and BASE for being + index registers; we can use them even if they are not dead. */ + if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], regno) + || TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], + REGNO (base))) + { + index_reg = reg; + reg_sum = src; + } + else + { + /* Otherwise, look for a free index register. Since we have + checked above that neither REG nor BASE are index registers, + if we find anything at all, it will be different from these + two registers. */ + for (i = first_index_reg; i <= last_index_reg; i++) + { + if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], i) + && reg_state[i].use_index == RELOAD_COMBINE_MAX_USES + && reg_state[i].store_ruid <= reg_state[regno].use_ruid + && (call_used_regs[i] || df_regs_ever_live_p (i)) + && (!frame_pointer_needed || i != HARD_FRAME_POINTER_REGNUM) + && !fixed_regs[i] && !global_regs[i] + && hard_regno_nregs[i][GET_MODE (reg)] == 1 + && targetm.hard_regno_scratch_ok (i)) + { + index_reg = gen_rtx_REG (GET_MODE (reg), i); + reg_sum = gen_rtx_PLUS (GET_MODE (reg), index_reg, base); + break; + } + } + } + + /* Check that PREV_SET is indeed (set (REGX) (CONST_INT)) and that + (REGY), i.e. BASE, is not clobbered before the last use we'll + create. */ + if (reg_sum + && prev_set + && CONST_INT_P (SET_SRC (prev_set)) + && rtx_equal_p (SET_DEST (prev_set), reg) + && (reg_state[REGNO (base)].store_ruid + <= reg_state[regno].use_ruid)) + { + /* Change destination register and, if necessary, the constant + value in PREV, the constant loading instruction. */ + validate_change (prev, &SET_DEST (prev_set), index_reg, 1); + if (reg_state[regno].offset != const0_rtx) + validate_change (prev, + &SET_SRC (prev_set), + GEN_INT (INTVAL (SET_SRC (prev_set)) + + INTVAL (reg_state[regno].offset)), + 1); + + /* Now for every use of REG that we have recorded, replace REG + with REG_SUM. */ + for (i = reg_state[regno].use_index; + i < RELOAD_COMBINE_MAX_USES; i++) + validate_unshare_change (reg_state[regno].reg_use[i].insn, + reg_state[regno].reg_use[i].usep, + /* Each change must have its own + replacement. */ + reg_sum, 1); + + if (apply_change_group ()) + { + struct reg_use *lowest_ruid = NULL; + + /* For every new use of REG_SUM, we have to record the use + of BASE therein, i.e. operand 1. */ + for (i = reg_state[regno].use_index; + i < RELOAD_COMBINE_MAX_USES; i++) + { + struct reg_use *use = reg_state[regno].reg_use + i; + reload_combine_note_use (&XEXP (*use->usep, 1), use->insn, + use->ruid, use->containing_mem); + if (lowest_ruid == NULL || use->ruid < lowest_ruid->ruid) + lowest_ruid = use; + } + + fixup_debug_insns (reg, reg_sum, insn, lowest_ruid->insn); + + /* Delete the reg-reg addition. */ + delete_insn (insn); + + if (reg_state[regno].offset != const0_rtx) + /* Previous REG_EQUIV / REG_EQUAL notes for PREV + are now invalid. */ + remove_reg_equal_equiv_notes (prev); + + reg_state[regno].use_index = RELOAD_COMBINE_MAX_USES; + return true; + } + } + } + return false; +} + static void reload_combine (void) { - rtx insn, set; - int first_index_reg = -1; - int last_index_reg = 0; + rtx insn, prev; int i; basic_block bb; unsigned int r; - int last_label_ruid; int min_labelno, n_labels; HARD_REG_SET ever_live_at_start, *label_live; - /* If reg+reg can be used in offsetable memory addresses, the main chunk of - reload has already used it where appropriate, so there is no use in - trying to generate it now. */ - if (double_reg_address_ok && INDEX_REG_CLASS != NO_REGS) - return; - /* To avoid wasting too much time later searching for an index register, determine the minimum and maximum index register numbers. */ - for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) - if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], r)) - { - if (first_index_reg == -1) - first_index_reg = r; - - last_index_reg = r; - } - - /* If no index register is available, we can quit now. */ - if (first_index_reg == -1) - return; + if (INDEX_REG_CLASS == NO_REGS) + last_index_reg = -1; + else if (first_index_reg == -1 && last_index_reg == 0) + { + for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) + if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], r)) + { + if (first_index_reg == -1) + first_index_reg = r; + + last_index_reg = r; + } + + /* If no index register is available, we can quit now. Set LAST_INDEX_REG + to -1 so we'll know to quit early the next time we get here. */ + if (first_index_reg == -1) + { + last_index_reg = -1; + return; + } + } /* Set up LABEL_LIVE and EVER_LIVE_AT_START. The register lifetime information is a bit fuzzy immediately after reload, but it's @@ -753,20 +1278,23 @@ } /* Initialize last_label_ruid, reload_combine_ruid and reg_state. */ - last_label_ruid = reload_combine_ruid = 0; + last_label_ruid = last_jump_ruid = reload_combine_ruid = 0; for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) { - reg_state[r].store_ruid = reload_combine_ruid; + reg_state[r].store_ruid = 0; + reg_state[r].real_store_ruid = 0; if (fixed_regs[r]) reg_state[r].use_index = -1; else reg_state[r].use_index = RELOAD_COMBINE_MAX_USES; } - for (insn = get_last_insn (); insn; insn = PREV_INSN (insn)) + for (insn = get_last_insn (); insn; insn = prev) { rtx note; + prev = PREV_INSN (insn); + /* We cannot do our optimization across labels. Invalidating all the use information we have would be costly, so we just note where the label is and then later disable any optimization that would cross it. */ @@ -777,141 +1305,17 @@ if (! fixed_regs[r]) reg_state[r].use_index = RELOAD_COMBINE_MAX_USES; - if (! INSN_P (insn)) + if (! NONDEBUG_INSN_P (insn)) continue; reload_combine_ruid++; - /* Look for (set (REGX) (CONST_INT)) - (set (REGX) (PLUS (REGX) (REGY))) - ... - ... (MEM (REGX)) ... - and convert it to - (set (REGZ) (CONST_INT)) - ... - ... (MEM (PLUS (REGZ) (REGY)))... . - - First, check that we have (set (REGX) (PLUS (REGX) (REGY))) - and that we know all uses of REGX before it dies. - Also, explicitly check that REGX != REGY; our life information - does not yet show whether REGY changes in this insn. */ - set = single_set (insn); - if (set != NULL_RTX - && REG_P (SET_DEST (set)) - && (hard_regno_nregs[REGNO (SET_DEST (set))] - [GET_MODE (SET_DEST (set))] - == 1) - && GET_CODE (SET_SRC (set)) == PLUS - && REG_P (XEXP (SET_SRC (set), 1)) - && rtx_equal_p (XEXP (SET_SRC (set), 0), SET_DEST (set)) - && !rtx_equal_p (XEXP (SET_SRC (set), 1), SET_DEST (set)) - && last_label_ruid < reg_state[REGNO (SET_DEST (set))].use_ruid) - { - rtx reg = SET_DEST (set); - rtx plus = SET_SRC (set); - rtx base = XEXP (plus, 1); - rtx prev = prev_nonnote_nondebug_insn (insn); - rtx prev_set = prev ? single_set (prev) : NULL_RTX; - unsigned int regno = REGNO (reg); - rtx index_reg = NULL_RTX; - rtx reg_sum = NULL_RTX; - - /* Now we need to set INDEX_REG to an index register (denoted as - REGZ in the illustration above) and REG_SUM to the expression - register+register that we want to use to substitute uses of REG - (typically in MEMs) with. First check REG and BASE for being - index registers; we can use them even if they are not dead. */ - if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], regno) - || TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], - REGNO (base))) - { - index_reg = reg; - reg_sum = plus; - } - else - { - /* Otherwise, look for a free index register. Since we have - checked above that neither REG nor BASE are index registers, - if we find anything at all, it will be different from these - two registers. */ - for (i = first_index_reg; i <= last_index_reg; i++) - { - if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], - i) - && reg_state[i].use_index == RELOAD_COMBINE_MAX_USES - && reg_state[i].store_ruid <= reg_state[regno].use_ruid - && hard_regno_nregs[i][GET_MODE (reg)] == 1) - { - index_reg = gen_rtx_REG (GET_MODE (reg), i); - reg_sum = gen_rtx_PLUS (GET_MODE (reg), index_reg, base); - break; - } - } - } - - /* Check that PREV_SET is indeed (set (REGX) (CONST_INT)) and that - (REGY), i.e. BASE, is not clobbered before the last use we'll - create. */ - if (reg_sum - && prev_set - && CONST_INT_P (SET_SRC (prev_set)) - && rtx_equal_p (SET_DEST (prev_set), reg) - && reg_state[regno].use_index >= 0 - && (reg_state[REGNO (base)].store_ruid - <= reg_state[regno].use_ruid)) - { - int i; - - /* Change destination register and, if necessary, the constant - value in PREV, the constant loading instruction. */ - validate_change (prev, &SET_DEST (prev_set), index_reg, 1); - if (reg_state[regno].offset != const0_rtx) - validate_change (prev, - &SET_SRC (prev_set), - GEN_INT (INTVAL (SET_SRC (prev_set)) - + INTVAL (reg_state[regno].offset)), - 1); - - /* Now for every use of REG that we have recorded, replace REG - with REG_SUM. */ - for (i = reg_state[regno].use_index; - i < RELOAD_COMBINE_MAX_USES; i++) - validate_unshare_change (reg_state[regno].reg_use[i].insn, - reg_state[regno].reg_use[i].usep, - /* Each change must have its own - replacement. */ - reg_sum, 1); - - if (apply_change_group ()) - { - /* For every new use of REG_SUM, we have to record the use - of BASE therein, i.e. operand 1. */ - for (i = reg_state[regno].use_index; - i < RELOAD_COMBINE_MAX_USES; i++) - reload_combine_note_use - (&XEXP (*reg_state[regno].reg_use[i].usep, 1), - reg_state[regno].reg_use[i].insn); - - if (reg_state[REGNO (base)].use_ruid - > reg_state[regno].use_ruid) - reg_state[REGNO (base)].use_ruid - = reg_state[regno].use_ruid; - - /* Delete the reg-reg addition. */ - delete_insn (insn); - - if (reg_state[regno].offset != const0_rtx) - /* Previous REG_EQUIV / REG_EQUAL notes for PREV - are now invalid. */ - remove_reg_equal_equiv_notes (prev); - - reg_state[regno].use_index = RELOAD_COMBINE_MAX_USES; - reg_state[REGNO (index_reg)].store_ruid - = reload_combine_ruid; - continue; - } - } - } + if (control_flow_insn_p (insn)) + last_jump_ruid = reload_combine_ruid; + + if (reload_combine_recognize_const_pattern (insn) + || reload_combine_recognize_pattern (insn)) + continue; note_stores (PATTERN (insn), reload_combine_note_store, NULL); @@ -967,7 +1371,8 @@ reg_state[i].use_index = -1; } - reload_combine_note_use (&PATTERN (insn), insn); + reload_combine_note_use (&PATTERN (insn), insn, + reload_combine_ruid, NULL_RTX); for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) { if (REG_NOTE_KIND (note) == REG_INC @@ -976,6 +1381,7 @@ int regno = REGNO (XEXP (note, 0)); reg_state[regno].store_ruid = reload_combine_ruid; + reg_state[regno].real_store_ruid = reload_combine_ruid; reg_state[regno].use_index = -1; } } @@ -985,8 +1391,8 @@ } /* Check if DST is a register or a subreg of a register; if it is, - update reg_state[regno].store_ruid and reg_state[regno].use_index - accordingly. Called via note_stores from reload_combine. */ + update store_ruid, real_store_ruid and use_index in the reg_state + structure accordingly. Called via note_stores from reload_combine. */ static void reload_combine_note_store (rtx dst, const_rtx set, void *data ATTRIBUTE_UNUSED) @@ -1010,14 +1416,14 @@ /* note_stores might have stripped a STRICT_LOW_PART, so we have to be careful with registers / register parts that are not full words. Similarly for ZERO_EXTRACT. */ - if (GET_CODE (set) != SET - || GET_CODE (SET_DEST (set)) == ZERO_EXTRACT + if (GET_CODE (SET_DEST (set)) == ZERO_EXTRACT || GET_CODE (SET_DEST (set)) == STRICT_LOW_PART) { for (i = hard_regno_nregs[regno][mode] - 1 + regno; i >= regno; i--) { reg_state[i].use_index = -1; reg_state[i].store_ruid = reload_combine_ruid; + reg_state[i].real_store_ruid = reload_combine_ruid; } } else @@ -1025,6 +1431,8 @@ for (i = hard_regno_nregs[regno][mode] - 1 + regno; i >= regno; i--) { reg_state[i].store_ruid = reload_combine_ruid; + if (GET_CODE (set) == SET) + reg_state[i].real_store_ruid = reload_combine_ruid; reg_state[i].use_index = RELOAD_COMBINE_MAX_USES; } } @@ -1035,7 +1443,7 @@ *XP is the pattern of INSN, or a part of it. Called from reload_combine, and recursively by itself. */ static void -reload_combine_note_use (rtx *xp, rtx insn) +reload_combine_note_use (rtx *xp, rtx insn, int ruid, rtx containing_mem) { rtx x = *xp; enum rtx_code code = x->code; @@ -1048,7 +1456,7 @@ case SET: if (REG_P (SET_DEST (x))) { - reload_combine_note_use (&SET_SRC (x), insn); + reload_combine_note_use (&SET_SRC (x), insn, ruid, NULL_RTX); return; } break; @@ -1104,6 +1512,11 @@ return; } + /* We may be called to update uses in previously seen insns. + Don't add uses beyond the last store we saw. */ + if (ruid < reg_state[regno].store_ruid) + return; + /* If this register is already used in some unknown fashion, we can't do anything. If we decrement the index from zero to -1, we can't store more @@ -1112,29 +1525,34 @@ if (use_index < 0) return; - if (use_index != RELOAD_COMBINE_MAX_USES - 1) - { - /* We have found another use for a register that is already - used later. Check if the offsets match; if not, mark the - register as used in an unknown fashion. */ - if (! rtx_equal_p (offset, reg_state[regno].offset)) - { - reg_state[regno].use_index = -1; - return; - } - } - else + if (use_index == RELOAD_COMBINE_MAX_USES - 1) { /* This is the first use of this register we have seen since we marked it as dead. */ reg_state[regno].offset = offset; - reg_state[regno].use_ruid = reload_combine_ruid; - } + reg_state[regno].all_offsets_match = true; + reg_state[regno].use_ruid = ruid; + } + else + { + if (reg_state[regno].use_ruid > ruid) + reg_state[regno].use_ruid = ruid; + + if (! rtx_equal_p (offset, reg_state[regno].offset)) + reg_state[regno].all_offsets_match = false; + } + reg_state[regno].reg_use[use_index].insn = insn; + reg_state[regno].reg_use[use_index].ruid = ruid; + reg_state[regno].reg_use[use_index].containing_mem = containing_mem; reg_state[regno].reg_use[use_index].usep = xp; return; } + case MEM: + containing_mem = x; + break; + default: break; } @@ -1144,11 +1562,12 @@ for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) { if (fmt[i] == 'e') - reload_combine_note_use (&XEXP (x, i), insn); + reload_combine_note_use (&XEXP (x, i), insn, ruid, containing_mem); else if (fmt[i] == 'E') { for (j = XVECLEN (x, i) - 1; j >= 0; j--) - reload_combine_note_use (&XVECEXP (x, i, j), insn); + reload_combine_note_use (&XVECEXP (x, i, j), insn, ruid, + containing_mem); } } } @@ -1196,9 +1615,10 @@ while REG is known to already have value (SYM + offset). This function tries to change INSN into an add instruction (set (REG) (plus (REG) (OFF - offset))) using the known value. - It also updates the information about REG's known value. */ + It also updates the information about REG's known value. + Return true if we made a change. */ -static void +static bool move2add_use_add2_insn (rtx reg, rtx sym, rtx off, rtx insn) { rtx pat = PATTERN (insn); @@ -1207,6 +1627,7 @@ rtx new_src = gen_int_mode (INTVAL (off) - reg_offset[regno], GET_MODE (reg)); bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); + bool changed = false; /* (set (reg) (plus (reg) (const_int 0))) is not canonical; use (set (reg) (reg)) instead. @@ -1221,13 +1642,13 @@ (reg)), would be discarded. Maybe we should try a truncMN pattern? */ if (INTVAL (off) == reg_offset [regno]) - validate_change (insn, &SET_SRC (pat), reg, 0); + changed = validate_change (insn, &SET_SRC (pat), reg, 0); } else if (rtx_cost (new_src, PLUS, speed) < rtx_cost (src, SET, speed) && have_add2_insn (reg, new_src)) { rtx tem = gen_rtx_PLUS (GET_MODE (reg), reg, new_src); - validate_change (insn, &SET_SRC (pat), tem, 0); + changed = validate_change (insn, &SET_SRC (pat), tem, 0); } else if (sym == NULL_RTX && GET_MODE (reg) != BImode) { @@ -1252,8 +1673,9 @@ gen_rtx_STRICT_LOW_PART (VOIDmode, narrow_reg), narrow_src); - if (validate_change (insn, &PATTERN (insn), - new_set, 0)) + changed = validate_change (insn, &PATTERN (insn), + new_set, 0); + if (changed) break; } } @@ -1263,6 +1685,7 @@ reg_mode[regno] = GET_MODE (reg); reg_symbol_ref[regno] = sym; reg_offset[regno] = INTVAL (off); + return changed; } @@ -1272,9 +1695,10 @@ value (SYM + offset) and change INSN into an add instruction (set (REG) (plus (the found register) (OFF - offset))) if such a register is found. It also updates the information about - REG's known value. */ + REG's known value. + Return true iff we made a change. */ -static void +static bool move2add_use_add3_insn (rtx reg, rtx sym, rtx off, rtx insn) { rtx pat = PATTERN (insn); @@ -1284,6 +1708,7 @@ int min_regno; bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); int i; + bool changed = false; for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) if (reg_set_luid[i] > move2add_last_label_luid @@ -1328,20 +1753,25 @@ GET_MODE (reg)); tem = gen_rtx_PLUS (GET_MODE (reg), tem, new_src); } - validate_change (insn, &SET_SRC (pat), tem, 0); + if (validate_change (insn, &SET_SRC (pat), tem, 0)) + changed = true; } reg_set_luid[regno] = move2add_luid; reg_base_reg[regno] = -1; reg_mode[regno] = GET_MODE (reg); reg_symbol_ref[regno] = sym; reg_offset[regno] = INTVAL (off); + return changed; } -static void +/* Convert move insns with constant inputs to additions if they are cheaper. + Return true if any changes were made. */ +static bool reload_cse_move2add (rtx first) { int i; rtx insn; + bool changed = false; for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--) { @@ -1402,7 +1832,7 @@ && reg_base_reg[regno] < 0 && reg_symbol_ref[regno] == NULL_RTX) { - move2add_use_add2_insn (reg, NULL_RTX, src, insn); + changed |= move2add_use_add2_insn (reg, NULL_RTX, src, insn); continue; } @@ -1463,6 +1893,7 @@ } if (success) delete_insn (insn); + changed |= success; insn = next; reg_mode[regno] = GET_MODE (reg); reg_offset[regno] = @@ -1508,12 +1939,12 @@ && reg_base_reg[regno] < 0 && reg_symbol_ref[regno] != NULL_RTX && rtx_equal_p (sym, reg_symbol_ref[regno])) - move2add_use_add2_insn (reg, sym, off, insn); + changed |= move2add_use_add2_insn (reg, sym, off, insn); /* Otherwise, we have to find a register whose value is sum of sym and some constant value. */ else - move2add_use_add3_insn (reg, sym, off, insn); + changed |= move2add_use_add3_insn (reg, sym, off, insn); continue; } @@ -1568,6 +1999,7 @@ } } } + return changed; } /* SET is a SET or CLOBBER that sets DST. DATA is the insn which