From: Jeff Law Date: Fri, 15 Dec 2017 22:11:06 +0000 (-0700) Subject: re PR tree-optimization/81165 (Regression in GCC-8.0.0's optimizer) X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=5806e0622de3356ae291a74a4c0d4c2e0af4c0d6;p=gcc.git re PR tree-optimization/81165 (Regression in GCC-8.0.0's optimizer) PR tree-optimization/81165 * tree-ssa-threadupdate.c (uses_in_bb): New. (estimate_threading_killed_stmts): New. * tree-ssa-threadupdate.h (estimate_threading_killed_stmts): Prototype. * tree-ssa-threadedge.c (record_temporary_equivalences_from_stmts_at_dest): Expand limit when its hit. PR tree-optimization/81165 * gcc.dg/pr81165.c: New. From-SVN: r255724 --- diff --git a/gcc/ChangeLog b/gcc/ChangeLog index 2d53e24b4c1..e5ab1b1a4c7 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,4 +1,14 @@ -2017-12-12 Jeff Law +2017-12-15 Alexandre Oliva + + PR tree-optimization/81165 + * tree-ssa-threadupdate.c (uses_in_bb): New. + (estimate_threading_killed_stmts): New. + * tree-ssa-threadupdate.h (estimate_threading_killed_stmts): Prototype. + * tree-ssa-threadedge.c + (record_temporary_equivalences_from_stmts_at_dest): Expand limit + when its hit. + +2017-12-15 Jeff Law PR tree-optimization/83410 * tree-ssa-threadupdate.c (thread_block_1): Avoid certain jump diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog index 7bbf0e9f09e..bc005b8de26 100644 --- a/gcc/testsuite/ChangeLog +++ b/gcc/testsuite/ChangeLog @@ -1,3 +1,8 @@ +2017-12-15 Alexandre Oliva + + PR tree-optimization/81165 + * gcc.dg/pr81165.c: New. + 2017-12-15 Jakub Jelinek PR c++/83205 diff --git a/gcc/testsuite/gcc.dg/pr81165.c b/gcc/testsuite/gcc.dg/pr81165.c new file mode 100644 index 00000000000..8508d893bed --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr81165.c @@ -0,0 +1,59 @@ +/* { dg-do compile } */ +/* { dg-options "-O3 -fdump-tree-optimized" } */ +/* { dg-final { scan-tree-dump-not " \[/%\] " "optimized" } } */ + +/* Testcase submitted for PR81165, with its main function removed as + it's turned into a compile test. We want to make sure that all of + the divide/remainder computations are removed by tree optimizers. + + We can figure out that we don't need to compute at runtime even the + condition to enter the loop: the initial i==0 would have to be + greater than the sum of two small unsigned values: 1U>>t1 is in the + range 0..1, whereas the char value is bounded by the range 0..127, + being 128 % a positive number (zero would invoke undefined + behavior, so we can assume it doesn't happen). (We know it's + nonnegative because it's 10 times a number that has no more than + the bits for 16, 8 and 1 set.) + + We don't realize that the loop is useless right away: jump + threading helps remove some of the complexity, particularly of the + computation within the loop: t1 is compared with 1, but it can + never be 1. (We could assume as much, since its being 1 would + divide by zero, but we don't.) + + If we don't enter the conditional block, t1 remains at 2; if we do, + it's set to either -1. If we jump thread at the end of the + conditional block, we can figure out the ranges exclude 1 and the + jump body is completely optimized out. However, we used to fail to + consider the block for jump threading due to the amount of + computation in it, without realizing most of it would die in + consequence of the threading. + + We now take the dying code into account when deciding whether or + not to try jump threading. That might enable us to optimize the + function into { if (x2 != 0 || (x1 & 1) == 0) abort (); }. At the + time of this writing, with the patch, we get close, but the test on + x2 only gets as far as ((1 >> x2) == 0). Without the patch, some + of the loop remains. */ + +short x0 = 15; + +void func (){ + volatile int x1 = 1U; + volatile char x2 = 0; + char t0 = 0; + unsigned long t1 = 2LU; + int i = 0; + + if(1>>x2) { + t0 = -1; + t1 = (1&(short)(x1^8U))-1; + } + + while(i > (int)((1U>>t1)+(char)(128%(10*(25LU&(29%x0)))))) { + i += (int)(12L/(1!=(int)t1)); + } + + if (t0 != -1) __builtin_abort(); + if (t1 != 0L) __builtin_abort(); +} diff --git a/gcc/tree-ssa-threadedge.c b/gcc/tree-ssa-threadedge.c index b7781dc7d60..1fafd7b5932 100644 --- a/gcc/tree-ssa-threadedge.c +++ b/gcc/tree-ssa-threadedge.c @@ -244,7 +244,22 @@ record_temporary_equivalences_from_stmts_at_dest (edge e, expansion, then do not thread through this block. */ stmt_count++; if (stmt_count > max_stmt_count) - return NULL; + { + /* If any of the stmts in the PATH's dests are going to be + killed due to threading, grow the max count + accordingly. */ + if (max_stmt_count + == PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS)) + { + max_stmt_count += estimate_threading_killed_stmts (e->dest); + if (dump_file) + fprintf (dump_file, "threading bb %i up to %i stmts\n", + e->dest->index, max_stmt_count); + } + /* If we're still past the limit, we're done. */ + if (stmt_count > max_stmt_count) + return NULL; + } /* These are temporary ranges, do nto reflect them back into the global range data. */ diff --git a/gcc/tree-ssa-threadupdate.c b/gcc/tree-ssa-threadupdate.c index 63ad8f9c953..b29ffe195c8 100644 --- a/gcc/tree-ssa-threadupdate.c +++ b/gcc/tree-ssa-threadupdate.c @@ -2431,3 +2431,139 @@ register_jump_thread (vec *path) paths.safe_push (path); } + +/* Return how many uses of T there are within BB, as long as there + aren't any uses outside BB. If there are any uses outside BB, + return -1 if there's at most one use within BB, or -2 if there is + more than one use within BB. */ + +static int +uses_in_bb (tree t, basic_block bb) +{ + int uses = 0; + bool outside_bb = false; + + imm_use_iterator iter; + use_operand_p use_p; + FOR_EACH_IMM_USE_FAST (use_p, iter, t) + { + if (is_gimple_debug (USE_STMT (use_p))) + continue; + + if (gimple_bb (USE_STMT (use_p)) != bb) + outside_bb = true; + else + uses++; + + if (outside_bb && uses > 1) + return -2; + } + + if (outside_bb) + return -1; + + return uses; +} + +/* Starting from the final control flow stmt in BB, assuming it will + be removed, follow uses in to-be-removed stmts back to their defs + and count how many defs are to become dead and be removed as + well. */ + +unsigned int +estimate_threading_killed_stmts (basic_block bb) +{ + int killed_stmts = 0; + hash_map ssa_remaining_uses; + auto_vec dead_worklist; + + /* If the block has only two predecessors, threading will turn phi + dsts into either src, so count them as dead stmts. */ + bool drop_all_phis = EDGE_COUNT (bb->preds) == 2; + + if (drop_all_phis) + for (gphi_iterator gsi = gsi_start_phis (bb); + !gsi_end_p (gsi); gsi_next (&gsi)) + { + gphi *phi = gsi.phi (); + tree dst = gimple_phi_result (phi); + + /* We don't count virtual PHIs as stmts in + record_temporary_equivalences_from_phis. */ + if (virtual_operand_p (dst)) + continue; + + killed_stmts++; + } + + if (gsi_end_p (gsi_last_bb (bb))) + return killed_stmts; + + gimple *stmt = gsi_stmt (gsi_last_bb (bb)); + if (gimple_code (stmt) != GIMPLE_COND + && gimple_code (stmt) != GIMPLE_GOTO + && gimple_code (stmt) != GIMPLE_SWITCH) + return killed_stmts; + + /* The control statement is always dead. */ + killed_stmts++; + dead_worklist.quick_push (stmt); + while (!dead_worklist.is_empty ()) + { + stmt = dead_worklist.pop (); + + ssa_op_iter iter; + use_operand_p use_p; + FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) + { + tree t = USE_FROM_PTR (use_p); + gimple *def = SSA_NAME_DEF_STMT (t); + + if (gimple_bb (def) == bb + && (gimple_code (def) != GIMPLE_PHI + || !drop_all_phis) + && !gimple_has_side_effects (def)) + { + int *usesp = ssa_remaining_uses.get (t); + int uses; + + if (usesp) + uses = *usesp; + else + uses = uses_in_bb (t, bb); + + gcc_assert (uses); + + /* Don't bother recording the expected use count if we + won't find any further uses within BB. */ + if (!usesp && (uses < -1 || uses > 1)) + { + usesp = &ssa_remaining_uses.get_or_insert (t); + *usesp = uses; + } + + if (uses < 0) + continue; + + --uses; + if (usesp) + *usesp = uses; + + if (!uses) + { + killed_stmts++; + if (usesp) + ssa_remaining_uses.remove (t); + if (gimple_code (def) != GIMPLE_PHI) + dead_worklist.safe_push (def); + } + } + } + } + + if (dump_file) + fprintf (dump_file, "threading bb %i kills %i stmts\n", + bb->index, killed_stmts); + + return killed_stmts; +} diff --git a/gcc/tree-ssa-threadupdate.h b/gcc/tree-ssa-threadupdate.h index 296eff29ff4..8a3b41d82cd 100644 --- a/gcc/tree-ssa-threadupdate.h +++ b/gcc/tree-ssa-threadupdate.h @@ -47,6 +47,7 @@ extern void remove_jump_threads_including (edge); extern void delete_jump_thread_path (vec *); extern void remove_ctrl_stmt_and_useless_edges (basic_block, basic_block); extern void free_dom_edge_info (edge); +extern unsigned int estimate_threading_killed_stmts (basic_block); enum bb_dom_status {