Re: [PATCH] csa, postreload: Improve csa after recent cselib changes [PR94516]
On Tue, 2020-05-05 at 09:38 +0200, Jakub Jelinek wrote: > Hi! > > On Wed, Apr 22, 2020 at 12:02:12PM -0600, Jeff Law via Gcc-patches wrote: > > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > > > > > 2020-04-20 Jakub Jelinek > > > > > > PR rtl-optimization/94516 > > > * postreload.c (reload_cse_simplify): When replacing sp = sp + const > > > with sp = reg, add REG_EQUAL note with sp + const. > > > * combine-stack-adj.c (try_apply_stack_adjustment): Change return > > > type from int to bool. Add LIVE and OTHER_INSN arguments. Undo > > > postreload sp = sp + const to sp = reg optimization if needed and > > > possible. > > > (combine_stack_adjustments_for_block): Add LIVE argument. Handle > > > reg = sp insn with sp + const REG_EQUAL note. Adjust > > > try_apply_stack_adjustment caller, call > > > df_simulate_initialize_forwards and df_simulate_one_insn_forwards. > > > (combine_stack_adjustments): Allocate and free LIVE bitmap, > > > adjust combine_stack_adjustments_for_block caller. > > I'd probably defer to gcc-11 since this is "just" a missed optimization and > > we're > > getting real close to cutting an RC and I'd hate to introduce new > > instability > > at > > this point. > > Is the patch now ok for stage1? Yes, of course. jeff
Re: [PATCH] csa, postreload: Improve csa after recent cselib changes [PR94516]
Hi! On Wed, Apr 22, 2020 at 12:02:12PM -0600, Jeff Law via Gcc-patches wrote: > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > > > 2020-04-20 Jakub Jelinek > > > > PR rtl-optimization/94516 > > * postreload.c (reload_cse_simplify): When replacing sp = sp + const > > with sp = reg, add REG_EQUAL note with sp + const. > > * combine-stack-adj.c (try_apply_stack_adjustment): Change return > > type from int to bool. Add LIVE and OTHER_INSN arguments. Undo > > postreload sp = sp + const to sp = reg optimization if needed and > > possible. > > (combine_stack_adjustments_for_block): Add LIVE argument. Handle > > reg = sp insn with sp + const REG_EQUAL note. Adjust > > try_apply_stack_adjustment caller, call > > df_simulate_initialize_forwards and df_simulate_one_insn_forwards. > > (combine_stack_adjustments): Allocate and free LIVE bitmap, > > adjust combine_stack_adjustments_for_block caller. > I'd probably defer to gcc-11 since this is "just" a missed optimization and > we're > getting real close to cutting an RC and I'd hate to introduce new instability > at > this point. Is the patch now ok for stage1? Jakub
Re: [PATCH] csa, postreload: Improve csa after recent cselib changes [PR94516]
On Tue, 2020-04-21 at 01:00 +0200, Jakub Jelinek wrote: > Hi! > > This patch addresses a missed optimization caused by the cselib changes. > Already in the past postreload could replace sp = sp + const_int with > sp = regxy if regxy already has the right value, but with the cselib > changes it happens several times more often. It can result in smaller > code, so it seems undesirable to prevent such optimizations, but > unfortunately it can get into the way of stack adjustment coalescing, > where e.g. if we used to have sp = sp + 32; sp = sp - 8;, previously > we'd turn that into sp = sp + 24;, but now postreload optimizes > into sp = r12; sp = sp - 8; and csa gives up. > > The patch just adds a REG_EQUAL note when changing sp = sp + const into > sp = reg, where we remember it was actually a stack adjustment by certain > constant, and the combine-stack-adj changes than make use of those REG_EQUAL > notes, together with LR tracking (csa did enable the note problem, just > didn't simulate each insn) so that we can add the needed clobbers etc. > (taken from the other stack adjustment insn). > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > 2020-04-20 Jakub Jelinek > > PR rtl-optimization/94516 > * postreload.c (reload_cse_simplify): When replacing sp = sp + const > with sp = reg, add REG_EQUAL note with sp + const. > * combine-stack-adj.c (try_apply_stack_adjustment): Change return > type from int to bool. Add LIVE and OTHER_INSN arguments. Undo > postreload sp = sp + const to sp = reg optimization if needed and > possible. > (combine_stack_adjustments_for_block): Add LIVE argument. Handle > reg = sp insn with sp + const REG_EQUAL note. Adjust > try_apply_stack_adjustment caller, call > df_simulate_initialize_forwards and df_simulate_one_insn_forwards. > (combine_stack_adjustments): Allocate and free LIVE bitmap, > adjust combine_stack_adjustments_for_block caller. I'd probably defer to gcc-11 since this is "just" a missed optimization and we're getting real close to cutting an RC and I'd hate to introduce new instability at this point. Jeff >
[PATCH] csa, postreload: Improve csa after recent cselib changes [PR94516]
Hi! This patch addresses a missed optimization caused by the cselib changes. Already in the past postreload could replace sp = sp + const_int with sp = regxy if regxy already has the right value, but with the cselib changes it happens several times more often. It can result in smaller code, so it seems undesirable to prevent such optimizations, but unfortunately it can get into the way of stack adjustment coalescing, where e.g. if we used to have sp = sp + 32; sp = sp - 8;, previously we'd turn that into sp = sp + 24;, but now postreload optimizes into sp = r12; sp = sp - 8; and csa gives up. The patch just adds a REG_EQUAL note when changing sp = sp + const into sp = reg, where we remember it was actually a stack adjustment by certain constant, and the combine-stack-adj changes than make use of those REG_EQUAL notes, together with LR tracking (csa did enable the note problem, just didn't simulate each insn) so that we can add the needed clobbers etc. (taken from the other stack adjustment insn). Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2020-04-20 Jakub Jelinek PR rtl-optimization/94516 * postreload.c (reload_cse_simplify): When replacing sp = sp + const with sp = reg, add REG_EQUAL note with sp + const. * combine-stack-adj.c (try_apply_stack_adjustment): Change return type from int to bool. Add LIVE and OTHER_INSN arguments. Undo postreload sp = sp + const to sp = reg optimization if needed and possible. (combine_stack_adjustments_for_block): Add LIVE argument. Handle reg = sp insn with sp + const REG_EQUAL note. Adjust try_apply_stack_adjustment caller, call df_simulate_initialize_forwards and df_simulate_one_insn_forwards. (combine_stack_adjustments): Allocate and free LIVE bitmap, adjust combine_stack_adjustments_for_block caller. --- gcc/postreload.c.jj 2020-04-19 12:10:19.042877054 +0200 +++ gcc/postreload.c2020-04-20 12:16:01.672848537 +0200 @@ -97,6 +97,16 @@ reload_cse_simplify (rtx_insn *insn, rtx if (NO_FUNCTION_CSE && CALL_P (insn)) return false; + /* Remember if this insn has been sp += const_int. */ + rtx sp_set = set_for_reg_notes (insn); + rtx sp_addend = NULL_RTX; + if (sp_set + && SET_DEST (sp_set) == stack_pointer_rtx + && GET_CODE (SET_SRC (sp_set)) == PLUS + && XEXP (SET_SRC (sp_set), 0) == stack_pointer_rtx + && CONST_INT_P (XEXP (SET_SRC (sp_set), 1))) +sp_addend = XEXP (SET_SRC (sp_set), 1); + if (GET_CODE (body) == SET) { int count = 0; @@ -180,6 +190,15 @@ reload_cse_simplify (rtx_insn *insn, rtx reload_cse_simplify_operands (insn, testreg); } + /* If sp += const_int insn is changed into sp = reg;, add REG_EQUAL + note so that the stack_adjustments pass can undo it if beneficial. */ + if (sp_addend + && SET_DEST (sp_set) == stack_pointer_rtx + && REG_P (SET_SRC (sp_set))) +set_dst_reg_note (insn, REG_EQUAL, + gen_rtx_PLUS (Pmode, stack_pointer_rtx, + sp_addend), stack_pointer_rtx); + done: return (EDGE_COUNT (insn_bb->succs) != insn_bb_succs); } --- gcc/combine-stack-adj.c.jj 2020-01-12 11:54:36.221416305 +0100 +++ gcc/combine-stack-adj.c 2020-04-20 14:52:47.294323991 +0200 @@ -70,9 +70,10 @@ static rtx single_set_for_csa (rtx_insn static void free_csa_reflist (struct csa_reflist *); static struct csa_reflist *record_one_stack_ref (rtx_insn *, rtx *, struct csa_reflist *); -static int try_apply_stack_adjustment (rtx_insn *, struct csa_reflist *, - HOST_WIDE_INT, HOST_WIDE_INT); -static void combine_stack_adjustments_for_block (basic_block); +static bool try_apply_stack_adjustment (rtx_insn *, struct csa_reflist *, + HOST_WIDE_INT, HOST_WIDE_INT, + bitmap, rtx_insn *); +static void combine_stack_adjustments_for_block (basic_block, bitmap); /* Main entry point for stack adjustment combination. */ @@ -81,9 +82,12 @@ static void combine_stack_adjustments (void) { basic_block bb; + bitmap live = BITMAP_ALLOC (_obstack); FOR_EACH_BB_FN (bb, cfun) -combine_stack_adjustments_for_block (bb); +combine_stack_adjustments_for_block (bb, live); + + BITMAP_FREE (live); } /* Recognize a MEM of the form (sp) or (plus sp const). */ @@ -219,18 +223,60 @@ no_unhandled_cfa (rtx_insn *insn) as each of the memories and stack references in REFLIST. Return true on success. */ -static int +static bool try_apply_stack_adjustment (rtx_insn *insn, struct csa_reflist *reflist, - HOST_WIDE_INT new_adjust, HOST_WIDE_INT delta) + HOST_WIDE_INT new_adjust, HOST_WIDE_INT delta, + bitmap live, rtx_insn *other_insn) { struct