On 10/1/18 10:52 PM, Peter Bergner wrote: > Now that we handle conflicts at definitions and the pic hard reg > is set via a copy from the pic pseudo, my PATCH 2 is setup to > handle exactly this scenario (ie, a copy between a pseudo and > a hard reg). I looked at the asm output from a build with both > PATCH 1 and PATCH 2, and yes, it also does not add the conflict > between the pic pseudo and pic hard reg, so our other option to > fix PR87479 is to apply PATCH 2. However, since PATCH 2 handles > the pic pseudo and pic hard reg conflict itself, that means we > don't need the special pic conflict code and it can be removed! > I'm going to update PATCH 2 to remove that pic handling code > and send it through bootstrap and regtesting.
Here is an updated PATCH 2 that adds the generic handling of copies between pseudos and hard regs which obsoletes the special conflict handling of the REAL_PIC_OFFSET_TABLE_REGNUM with the pic pseudo. I have confirmed the assembler generated by this patch for test case pr63534.c matches the code generated before PATCH 1 was committed, so we are successfully removing the copy of the pic pseudo into the pic hard reg with this patch. I'm currently performing bootstrap and regtesting on powerpc64le-linux and x86_64-linux. H.J., could you please test this patch on i686 to verify it doesn't expose any other problems there? Otherwise, I'll take Jeff's suggestion and attempt a build on gcc45, but it sounds like the results will take a while. Is this patch version ok for trunk assuming no regressions are found in the testing mentioned above? Peter gcc/ PR rtl-optimization/86939 PR rtl-optimization/87479 * ira.h (copy_insn_p): New prototype. * ira-lives.c (ignore_reg_for_conflicts): New static variable. (make_hard_regno_dead): Don't add conflicts for register ignore_reg_for_conflicts. (make_object_dead): Likewise. (copy_insn_p): New function. (process_bb_node_lives): Set ignore_reg_for_conflicts for copies. Remove special conflict handling of REAL_PIC_OFFSET_TABLE_REGNUM. * lra-lives.c (ignore_reg_for_conflicts): New static variable. (make_hard_regno_dead): Don't add conflicts for register ignore_reg_for_conflicts. Remove special conflict handling of REAL_PIC_OFFSET_TABLE_REGNUM. Remove now unused argument check_pic_pseudo_p and update callers. (mark_pseudo_dead): Don't add conflicts for register ignore_reg_for_conflicts. (process_bb_lives): Set ignore_reg_for_conflicts for copies. gcc/testsuite/ * gcc.target/powerpc/pr86939.c: New test. Index: gcc/ira.h =================================================================== --- gcc/ira.h (revision 264789) +++ gcc/ira.h (working copy) @@ -210,6 +210,9 @@ extern void ira_adjust_equiv_reg_cost (u /* ira-costs.c */ extern void ira_costs_c_finalize (void); +/* ira-lives.c */ +extern rtx copy_insn_p (rtx_insn *); + /* Spilling static chain pseudo may result in generation of wrong non-local goto code using frame-pointer to address saved stack pointer value after restoring old frame pointer value. The Index: gcc/ira-lives.c =================================================================== --- gcc/ira-lives.c (revision 264789) +++ gcc/ira-lives.c (working copy) @@ -84,6 +84,10 @@ static int *allocno_saved_at_call; supplemental to recog_data. */ static alternative_mask preferred_alternatives; +/* If non-NULL, the source operand of a register to register copy for which + we should not add a conflict with the copy's destination operand. */ +static rtx ignore_reg_for_conflicts; + /* Record hard register REGNO as now being live. */ static void make_hard_regno_live (int regno) @@ -101,6 +105,11 @@ make_hard_regno_dead (int regno) { ira_object_t obj = ira_object_id_map[i]; + if (ignore_reg_for_conflicts != NULL_RTX + && REGNO (ignore_reg_for_conflicts) + == (unsigned int) ALLOCNO_REGNO (OBJECT_ALLOCNO (obj))) + continue; + SET_HARD_REG_BIT (OBJECT_CONFLICT_HARD_REGS (obj), regno); SET_HARD_REG_BIT (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj), regno); } @@ -154,12 +163,38 @@ static void make_object_dead (ira_object_t obj) { live_range_t lr; + int ignore_regno = -1; + int last_regno = -1; sparseset_clear_bit (objects_live, OBJECT_CONFLICT_ID (obj)); + /* Check whether any part of IGNORE_REG_FOR_CONFLICTS already conflicts + with OBJ. */ + if (ignore_reg_for_conflicts != NULL_RTX + && REGNO (ignore_reg_for_conflicts) < FIRST_PSEUDO_REGISTER) + { + last_regno = END_REGNO (ignore_reg_for_conflicts); + int src_regno = ignore_regno = REGNO (ignore_reg_for_conflicts); + + while (src_regno < last_regno) + { + if (TEST_HARD_REG_BIT (OBJECT_CONFLICT_HARD_REGS (obj), src_regno)) + { + ignore_regno = last_regno = -1; + break; + } + src_regno++; + } + } + IOR_HARD_REG_SET (OBJECT_CONFLICT_HARD_REGS (obj), hard_regs_live); IOR_HARD_REG_SET (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj), hard_regs_live); + /* If IGNORE_REG_FOR_CONFLICTS did not already conflict with OBJ, make + sure it still doesn't. */ + for (; ignore_regno < last_regno; ignore_regno++) + CLEAR_HARD_REG_BIT (OBJECT_CONFLICT_HARD_REGS (obj), ignore_regno); + lr = OBJECT_LIVE_RANGES (obj); ira_assert (lr != NULL); lr->finish = curr_point; @@ -1022,6 +1057,38 @@ find_call_crossed_cheap_reg (rtx_insn *i return cheap_reg; } +/* Determine whether INSN is a register to register copy of the type where + we do not need to make the source and destiniation registers conflict. + If this is a copy instruction, then return the source reg. Otherwise, + return NULL_RTX. */ +rtx +copy_insn_p (rtx_insn *insn) +{ + rtx set; + + /* Disallow anything other than a simple register to register copy + that has no side effects. */ + if ((set = single_set (insn)) == NULL_RTX + || !REG_P (SET_DEST (set)) + || !REG_P (SET_SRC (set)) + || side_effects_p (set)) + return NULL_RTX; + + int dst_regno = REGNO (SET_DEST (set)); + int src_regno = REGNO (SET_SRC (set)); + machine_mode mode = GET_MODE (SET_DEST (set)); + + /* Computing conflicts for register pairs is difficult to get right, so + for now, disallow it. */ + if ((dst_regno < FIRST_PSEUDO_REGISTER + && hard_regno_nregs (dst_regno, mode) != 1) + || (src_regno < FIRST_PSEUDO_REGISTER + && hard_regno_nregs (src_regno, mode) != 1)) + return NULL_RTX; + + return SET_SRC (set); +} + /* Process insns of the basic block given by its LOOP_TREE_NODE to update allocno live ranges, allocno hard register conflicts, intersected calls, and register pressure info for allocnos for the @@ -1107,22 +1174,7 @@ process_bb_node_lives (ira_loop_tree_nod curr_point); call_p = CALL_P (insn); -#ifdef REAL_PIC_OFFSET_TABLE_REGNUM - int regno; - bool clear_pic_use_conflict_p = false; - /* Processing insn usage in call insn can create conflict - with pic pseudo and pic hard reg and that is wrong. - Check this situation and fix it at the end of the insn - processing. */ - if (call_p && pic_offset_table_rtx != NULL_RTX - && (regno = REGNO (pic_offset_table_rtx)) >= FIRST_PSEUDO_REGISTER - && (a = ira_curr_regno_allocno_map[regno]) != NULL) - clear_pic_use_conflict_p - = (find_regno_fusage (insn, USE, REAL_PIC_OFFSET_TABLE_REGNUM) - && ! TEST_HARD_REG_BIT (OBJECT_CONFLICT_HARD_REGS - (ALLOCNO_OBJECT (a, 0)), - REAL_PIC_OFFSET_TABLE_REGNUM)); -#endif + ignore_reg_for_conflicts = copy_insn_p (insn); /* Mark each defined value as live. We need to do this for unused values because they still conflict with quantities @@ -1275,20 +1327,9 @@ process_bb_node_lives (ira_loop_tree_nod } } -#ifdef REAL_PIC_OFFSET_TABLE_REGNUM - if (clear_pic_use_conflict_p) - { - regno = REGNO (pic_offset_table_rtx); - a = ira_curr_regno_allocno_map[regno]; - CLEAR_HARD_REG_BIT (OBJECT_CONFLICT_HARD_REGS (ALLOCNO_OBJECT (a, 0)), - REAL_PIC_OFFSET_TABLE_REGNUM); - CLEAR_HARD_REG_BIT (OBJECT_TOTAL_CONFLICT_HARD_REGS - (ALLOCNO_OBJECT (a, 0)), - REAL_PIC_OFFSET_TABLE_REGNUM); - } -#endif curr_point++; } + ignore_reg_for_conflicts = NULL_RTX; if (bb_has_eh_pred (bb)) for (j = 0; ; ++j) Index: gcc/lra-lives.c =================================================================== --- gcc/lra-lives.c (revision 264789) +++ gcc/lra-lives.c (working copy) @@ -96,6 +96,10 @@ static bitmap_head temp_bitmap; /* Pool for pseudo live ranges. */ static object_allocator<lra_live_range> lra_live_range_pool ("live ranges"); +/* If non-NULL, the source operand of a register to register copy for which + we should not add a conflict with the copy's destination operand. */ +static rtx ignore_reg_for_conflicts; + /* Free live range list LR. */ static void free_live_range_list (lra_live_range_t lr) @@ -239,11 +243,9 @@ make_hard_regno_live (int regno) /* Process the definition of hard register REGNO. This updates hard_regs_live, START_DYING and conflict hard regs for living - pseudos. Conflict hard regs for the pic pseudo is not updated if - REGNO is REAL_PIC_OFFSET_TABLE_REGNUM and CHECK_PIC_PSEUDO_P is - true. */ + pseudos. */ static void -make_hard_regno_dead (int regno, bool check_pic_pseudo_p ATTRIBUTE_UNUSED) +make_hard_regno_dead (int regno) { lra_assert (regno < FIRST_PSEUDO_REGISTER); if (! TEST_HARD_REG_BIT (hard_regs_live, regno)) @@ -251,13 +253,12 @@ make_hard_regno_dead (int regno, bool ch sparseset_set_bit (start_dying, regno); unsigned int i; EXECUTE_IF_SET_IN_SPARSESET (pseudos_live, i) -#ifdef REAL_PIC_OFFSET_TABLE_REGNUM - if (! check_pic_pseudo_p - || regno != REAL_PIC_OFFSET_TABLE_REGNUM - || pic_offset_table_rtx == NULL - || i != REGNO (pic_offset_table_rtx)) -#endif + { + if (ignore_reg_for_conflicts != NULL_RTX + && REGNO (ignore_reg_for_conflicts) == i) + continue; SET_HARD_REG_BIT (lra_reg_info[i].conflict_hard_regs, regno); + } CLEAR_HARD_REG_BIT (hard_regs_live, regno); if (fixed_regs[regno] || TEST_HARD_REG_BIT (hard_regs_spilled_into, regno)) { @@ -294,14 +295,41 @@ static void mark_pseudo_dead (int regno, int point) { lra_live_range_t p; + int ignore_regno = -1; + int last_regno = -1; lra_assert (regno >= FIRST_PSEUDO_REGISTER); lra_assert (sparseset_bit_p (pseudos_live, regno)); sparseset_clear_bit (pseudos_live, regno); sparseset_set_bit (start_dying, regno); + /* Check whether any part of IGNORE_REG_FOR_CONFLICTS already conflicts + with REGNO. */ + if (ignore_reg_for_conflicts != NULL_RTX + && REGNO (ignore_reg_for_conflicts) < FIRST_PSEUDO_REGISTER) + { + last_regno = END_REGNO (ignore_reg_for_conflicts); + int src_regno = ignore_regno = REGNO (ignore_reg_for_conflicts); + + while (src_regno < last_regno) + { + if (TEST_HARD_REG_BIT (lra_reg_info[regno].conflict_hard_regs, + src_regno)) + { + ignore_regno = last_regno = -1; + break; + } + src_regno++; + } + } + IOR_HARD_REG_SET (lra_reg_info[regno].conflict_hard_regs, hard_regs_live); + /* If IGNORE_REG_FOR_CONFLICTS did not already conflict with REGNO, make + sure it still doesn't. */ + for (; ignore_regno < last_regno; ignore_regno++) + CLEAR_HARD_REG_BIT (lra_reg_info[regno].conflict_hard_regs, ignore_regno); + if (complete_info_p || lra_get_regno_hard_regno (regno) < 0) { p = lra_reg_info[regno].live_ranges; @@ -350,7 +378,7 @@ mark_regno_dead (int regno, machine_mode if (regno < FIRST_PSEUDO_REGISTER) { for (last = end_hard_regno (mode, regno); regno < last; regno++) - make_hard_regno_dead (regno, false); + make_hard_regno_dead (regno); } else { @@ -747,6 +775,7 @@ process_bb_lives (basic_block bb, int &c } call_p = CALL_P (curr_insn); + ignore_reg_for_conflicts = copy_insn_p (curr_insn); src_regno = (set != NULL_RTX && REG_P (SET_SRC (set)) ? REGNO (SET_SRC (set)) : -1); dst_regno = (set != NULL_RTX && REG_P (SET_DEST (set)) @@ -858,14 +887,13 @@ process_bb_lives (basic_block bb, int &c for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) if (reg->type == OP_OUT && ! reg_early_clobber_p (reg, n_alt) && ! reg->subreg_p) - make_hard_regno_dead (reg->regno, false); + make_hard_regno_dead (reg->regno); if (curr_id->arg_hard_regs != NULL) for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++) if (regno >= FIRST_PSEUDO_REGISTER) - /* It is a clobber. Don't create conflict of used - REAL_PIC_OFFSET_TABLE_REGNUM and the pic pseudo. */ - make_hard_regno_dead (regno - FIRST_PSEUDO_REGISTER, true); + /* It is a clobber. */ + make_hard_regno_dead (regno - FIRST_PSEUDO_REGISTER); if (call_p) { @@ -925,8 +953,7 @@ process_bb_lives (basic_block bb, int &c make_hard_regno_live (reg->regno); if (curr_id->arg_hard_regs != NULL) - /* Make argument hard registers live. Don't create conflict - of used REAL_PIC_OFFSET_TABLE_REGNUM and the pic pseudo. */ + /* Make argument hard registers live. */ for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++) if (regno < FIRST_PSEUDO_REGISTER) make_hard_regno_live (regno); @@ -954,7 +981,7 @@ process_bb_lives (basic_block bb, int &c if (reg2->type != OP_OUT && reg2->regno == reg->regno) break; if (reg2 == NULL) - make_hard_regno_dead (reg->regno, false); + make_hard_regno_dead (reg->regno); } if (need_curr_point_incr) @@ -989,6 +1016,7 @@ process_bb_lives (basic_block bb, int &c EXECUTE_IF_SET_IN_SPARSESET (unused_set, j) add_reg_note (curr_insn, REG_UNUSED, regno_reg_rtx[j]); } + ignore_reg_for_conflicts = NULL_RTX; if (bb_has_eh_pred (bb)) for (j = 0; ; ++j) Index: gcc/testsuite/gcc.target/powerpc/pr86939.c =================================================================== --- gcc/testsuite/gcc.target/powerpc/pr86939.c (nonexistent) +++ gcc/testsuite/gcc.target/powerpc/pr86939.c (working copy) @@ -0,0 +1,12 @@ +/* { dg-do compile { target { powerpc*-*-* } } } */ +/* { dg-options "-O2" } */ + +typedef long (*fptr_t) (void); +long +func (fptr_t *p) +{ + if (p) + return (*p) (); + return 0; +} +/* { dg-final { scan-assembler-not {mr %?r?12,} } } */