On 2021-06-21 16:51, Richard Biener wrote:
On Wed, 9 Jun 2021, guojiufu wrote:On 2021-06-09 17:42, guojiufu via Gcc-patches wrote: > On 2021-06-08 18:13, Richard Biener wrote: >> On Fri, 4 Jun 2021, Jiufu Guo wrote: >> > cut... >>> + gcond *cond = as_a<gcond *> (last); >>> + enum tree_code code = gimple_cond_code (cond); >>> + if (!(code == NE_EXPR >>> + || (code == EQ_EXPR && (e->flags & EDGE_TRUE_VALUE)))) >> >> The NE_EXPR check misses a corresponding && (e->flags & EDGE_FALSE_VALUE) >> check. >> > Thanks, check (e->flags & EDGE_FALSE_VALUE) would be safer. > >>> + continue; >>> + >>> + /* Check if bound is invarant. */ >>> + tree idx = gimple_cond_lhs (cond); >>> + tree bnd = gimple_cond_rhs (cond); >>> + if (expr_invariant_in_loop_p (loop, idx)) >>> + std::swap (idx, bnd); >>> + else if (!expr_invariant_in_loop_p (loop, bnd)) >>> + continue; >>> + >>> + /* Only unsigned type conversion could cause wrap. */ >>> + tree type = TREE_TYPE (idx); >>> + if (!INTEGRAL_TYPE_P (type) || TREE_CODE (idx) != SSA_NAME >>> + || !TYPE_UNSIGNED (type)) >>> + continue; >>> + >>> + /* Avoid to split if bound is MAX/MIN val. */ >>> + tree bound_type = TREE_TYPE (bnd); >>> + if (TREE_CODE (bnd) == INTEGER_CST && INTEGRAL_TYPE_P (bound_type) >>> + && (tree_int_cst_equal (bnd, TYPE_MAX_VALUE (bound_type)) >>> + || tree_int_cst_equal (bnd, TYPE_MIN_VALUE (bound_type)))) >>> + continue; >> >> Note you do not require 'bnd' to be constant and thus at runtime those >> cases still need to be handled correctly. > Yes, bnd is not required to be constant. The above code is filtering the > case > where bnd is const max/min value of the type. So, the code could be updated > as: > if (tree_int_cst_equal (bnd, TYPE_MAX_VALUE (bound_type)) > || tree_int_cst_equal (bnd, TYPE_MIN_VALUE (bound_type)))Yes, and the comment adjusted to "if bound is known to be MAX/MIN val.">> >>> + /* Check if there is possible wrap. */ >>> + class tree_niter_desc niter; >>> + if (!number_of_iterations_exit (loop, e, &niter, false, false)) > cut... >>> + >>> + /* Change if (i != n) to LOOP1:if (i > n) and LOOP2:if (i < n) */ >> >> It now occurs to me that we nowhere check the evolution of IDX >> (split_at_bb_p uses simple_iv for this for example). The transform >> assumes that we will actually hit i == n and that i increments, but >> while you check the control IV from number_of_iterations_exit >> for NE_EXPR that does not guarantee a positive evolution. >> > If I do not correctly reply your question, please point out: > number_of_iterations_exit is similar with simple_iv to invoke > simple_iv_with_niters > which check the evolution, and number_of_iterations_exit check > number_of_iterations_cond > which check no_overflow more accurate, this is one reason I use this > function. > > This transform assumes that the last run hits i==n. > Otherwise, the loop may run infinitely wrap after wrap. > For safe, if the step is 1 or -1, this assumption would be true. I > would add this check.OK.> Thanks so much for pointing out I missed the negative step! > >> Your testcases do not include any negative step examples, but I guess >> the conditions need to be swapped in this case? > > I would add cases and code to support step 1/-1. > >> >> I think you also have to consider the order we split, say with >> >> for (i = start; i != end; ++i) >> { >> push (i); >> if (a[i] != b[i]) >> break; >> } >> >> push (i) calls need to be in the same order for all cases of >> start < end, start == end and start > end (and also cover >> runtime testcases with end == 0 or end == UINT_MAX, likewise >> for start). > I add tests for the above cases. If missing sth, please point out, thanks! > >> >>> + bool inv = expr_invariant_in_loop_p (loop, gimple_cond_lhs (gc)); >>> + enum tree_code up_code = inv ? LT_EXPR : GT_EXPR; >>> + enum tree_code down_code = inv ? GT_EXPR : LT_EXPR; > cut.... > > Thanks again for the very helpful review! > > BR, > Jiufu Guo. Here is the updated patch, thanks for your time! diff --git a/gcc/testsuite/gcc.dg/loop-split1.c b/gcc/testsuite/gcc.dg/loop-split1.c new file mode 100644 index 00000000000..dd2d03a7b96 --- /dev/null +++ b/gcc/testsuite/gcc.dg/loop-split1.c @@ -0,0 +1,101 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -fsplit-loops -fdump-tree-lsplit-details" } */ + +void +foo (int *a, int *b, unsigned l, unsigned n) +{ + while (++l != n) + a[l] = b[l] + 1; +} +void +foo_1 (int *a, int *b, unsigned n) +{ + unsigned l = 0; + while (++l != n) + a[l] = b[l] + 1; +} + +void +foo1 (int *a, int *b, unsigned l, unsigned n) +{ + while (l++ != n) + a[l] = b[l] + 1; +} + +/* No wrap. */ +void +foo1_1 (int *a, int *b, unsigned n) +{ + unsigned l = 0; + while (l++ != n) + a[l] = b[l] + 1; +} + +unsigned +foo2 (char *a, char *b, unsigned l, unsigned n) +{ + while (++l != n) + if (a[l] != b[l]) + break; + + return l; +} + +unsigned +foo2_1 (char *a, char *b, unsigned l, unsigned n) +{ + l = 0; + while (++l != n) + if (a[l] != b[l]) + break; + + return l; +} + +unsigned +foo3 (char *a, char *b, unsigned l, unsigned n) +{ + while (l++ != n) + if (a[l] != b[l]) + break; + + return l; +} + +/* No wrap. */ +unsigned +foo3_1 (char *a, char *b, unsigned l, unsigned n) +{ + l = 0; + while (l++ != n) + if (a[l] != b[l]) + break; + + return l; +} + +void +bar (); +void +foo4 (unsigned n, unsigned i) +{ + do + { + if (i == n) + return; + bar (); + ++i; + } + while (1); +} + +unsigned +find_skip_diff (char *p, char *q, unsigned n, unsigned i) +{ + while (p[i] == q[i] && ++i != n) + p++, q++; + + return i; +} + +/* { dg-final { scan-tree-dump-times "Loop split" 8 "lsplit" } } */ diff --git a/gcc/testsuite/gcc.dg/loop-split2.c b/gcc/testsuite/gcc.dg/loop-split2.c new file mode 100644 index 00000000000..56377e2f2f5 --- /dev/null +++ b/gcc/testsuite/gcc.dg/loop-split2.c @@ -0,0 +1,155 @@ +/* { dg-do run } */ +/* { dg-options "-O3" } */ + +extern void +abort (void); +extern void +exit (int); +void +push (int); + +#define NI __attribute__ ((noinline)) + +void NI +foo (int *a, int *b, unsigned char l, unsigned char n) +{ + while (++l != n) + a[l] = b[l] + 1; +} + +unsigned NI +bar (int *a, int *b, unsigned char l, unsigned char n) +{ + while (l++ != n) + { + push (l); + if (a[l] != b[l]) + break; + push (l + 1); + } + return l; +} + +void NI +foo_1 (int *a, int *b, unsigned char l, unsigned char n) +{ + while (--l != n) + a[l] = b[l] + 1; +} + +unsigned NI +bar_1 (int *a, int *b, unsigned char l, unsigned char n) +{ + while (l-- != n) + { + push (l); + if (a[l] != b[l]) + break; + push (l + 1); + } + + return l; +} + +int a[258]; +int b[258]; +int c[1024]; +static int top = 0; +void +push (int e) +{ + c[top++] = e; +} + +void +reset () +{ + top = 0; + __builtin_memset (c, 0, sizeof (c)); +} + +#define check(a, b) (a == b) + +int +check_c (int *c, int a0, int a1, int a2, int a3, int a4, int a5) +{ + return check (c[0], a0) && check (c[1], a1) && check (c[2], a2) + && check (c[3], a3) && check (c[4], a4) && check (c[5], a5); +} + +int +main () +{ + __builtin_memcpy (b, a, sizeof (a)); + reset (); + if (bar (a, b, 6, 8) != 9 || !check_c (c, 7, 8, 8, 9, 0, 0)) + abort (); + + reset (); + if (bar (a, b, 5, 3) != 4 || !check_c (c, 6, 7, 7, 8, 8, 9) + || !check_c (c + 496, 254, 255, 255, 256, 0, 1)) + abort (); + + reset (); + if (bar (a, b, 6, 6) != 7 || !check_c (c, 0, 0, 0, 0, 0, 0)) + abort (); + + reset ();+ if (bar (a, b, 253, 255) != 0 || !check_c (c, 254, 255, 255, 256, 0, 0))+ abort (); + + reset ();+ if (bar (a, b, 253, 0) != 1 || !check_c (c, 254, 255, 255, 256, 0, 1))+ abort (); + + reset (); + if (bar_1 (a, b, 6, 8) != 7 || !check_c (c, 5, 6, 4, 5, 3, 4)) + abort (); + + reset (); + if (bar_1 (a, b, 5, 3) != 2 || !check_c (c, 4, 5, 3, 4, 0, 0)) + abort (); + + reset (); + if (bar_1 (a, b, 6, 6) != 5) + abort (); + + reset ();+ if (bar_1 (a, b, 2, 255) != 254 || !check_c (c, 1, 2, 0, 1, 255, 256))+ abort (); + + reset (); + if (bar_1 (a, b, 2, 0) != 255 || !check_c (c, 1, 2, 0, 1, 0, 0)) + abort (); + + b[100] += 1; + reset (); + if (bar (a, b, 90, 110) != 100) + abort (); + + reset (); + if (bar (a, b, 110, 105) != 100) + abort (); + + reset (); + if (bar_1 (a, b, 90, 110) != 109) + abort (); + + reset (); + if (bar_1 (a, b, 2, 90) != 100) + abort (); + + foo (a, b, 99, 99); + a[99] = b[99] + 1; + for (int i = 0; i < 256; i++) + if (a[i] != b[i] + 1) + abort (); + + foo_1 (a, b, 99, 99); + a[99] = b[99] + 1; + for (int i = 0; i < 256; i++) + if (a[i] != b[i] + 1) + abort (); + + exit (0); +} diff --git a/gcc/testsuite/gcc.dg/loop-split3.c b/gcc/testsuite/gcc.dg/loop-split3.c new file mode 100644 index 00000000000..ec93ee8bd12 --- /dev/null +++ b/gcc/testsuite/gcc.dg/loop-split3.c @@ -0,0 +1,62 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -fsplit-loops -fdump-tree-lsplit-details" } */ + +void +foo (int *a, int *b, unsigned l, unsigned n) +{ + while (--l != n) + a[l] = b[l] + 1; +} + +void +foo1 (int *a, int *b, unsigned l, unsigned n) +{ + while (l-- != n) + a[l] = b[l] + 1; +} + +unsigned +foo2 (char *a, char *b, unsigned l, unsigned n) +{ + while (--l != n) + if (a[l] != b[l]) + break; + + return l; +} + +unsigned +foo3 (char *a, char *b, unsigned l, unsigned n) +{ + while (l-- != n) + if (a[l] != b[l]) + break; + + return l; +} + +void +bar (); +void +foo4 (unsigned n, unsigned i) +{ + do + { + if (i == n) + return; + bar (); + --i; + } + while (1); +} + +unsigned +find_skip_diff (char *p, char *q, unsigned n, unsigned i) +{ + while (p[i] == q[i] && --i != n) + p--, q--; + + return i; +} + +/* { dg-final { scan-tree-dump-times "Loop split" 6 "lsplit" } } */ diff --git a/gcc/tree-ssa-loop-split.c b/gcc/tree-ssa-loop-split.c index 3a09bbc39e5..e9f23b32186 100644 --- a/gcc/tree-ssa-loop-split.c +++ b/gcc/tree-ssa-loop-split.c @@ -41,6 +41,7 @@ along with GCC; see the file COPYING3. If not see #include "cfghooks.h" #include "gimple-fold.h" #include "gimplify-me.h" +#include "tree-ssa-loop-ivopts.h" /* This file implements two kinds of loop splitting. @@ -229,11 +230,14 @@ easy_exit_values (class loop *loop) conditional). I.e. the second loop can now be entered either via the original entry or via NEW_E, so the entry values of LOOP2 phi nodes are either the original ones or those at the exit - of LOOP1. Insert new phi nodes in LOOP2 pre-header reflecting - this. The loops need to fulfill easy_exit_values(). */ + of LOOP1. Selecting the previous value instead next value as the + exit value of LOOP1 if USE_PREV is true. Insert new phi nodes in + LOOP2 pre-header reflecting this. The loops need to fulfill + easy_exit_values(). */ static void -connect_loop_phis (class loop *loop1, class loop *loop2, edge new_e) +connect_loop_phis (class loop *loop1, class loop *loop2, edge new_e, + bool use_prev = false) { basic_block rest = loop_preheader_edge (loop2)->src; gcc_assert (new_e->dest == rest);@@ -279,7 +283,8 @@ connect_loop_phis (class loop *loop1, class loop *loop2,edge new_e) gphi * newphi = create_phi_node (new_init, rest); add_phi_arg (newphi, init, skip_first, UNKNOWN_LOCATION); - add_phi_arg (newphi, next, new_e, UNKNOWN_LOCATION);+ add_phi_arg (newphi, use_prev ? PHI_RESULT (phi_first) : next, new_e,+ UNKNOWN_LOCATION); SET_USE (op, new_init); } } @@ -1593,6 +1598,252 @@ split_loop_on_cond (struct loop *loop) return do_split; } +/* Check if the LOOP exit branch is like "if (idx != bound)",+ Return the branch edge which exit loop, if wrap may happen on "idx". */+ +static edge +get_ne_cond_branch (struct loop *loop, tree *step) +{ + int i; + edge e; + + auto_vec<edge> edges = get_loop_exit_edges (loop); + FOR_EACH_VEC_ELT (edges, i, e) + { + basic_block bb = e->src; + + /* Check if exit at gcond. */ + gimple *last = last_stmt (bb); + if (!last || gimple_code (last) != GIMPLE_COND) + continue; + gcond *cond = as_a<gcond *> (last);gcond *cont = safe_dyn_cast <gcond *> (last_stmt (bb)); if (!last) continue; is shorter.
Thanks.
I did not use invert_tree_comparison in the code, because it would generate+ enum tree_code code = gimple_cond_code (cond); + if (!((code == NE_EXPR && (e->flags & EDGE_FALSE_VALUE)) + || (code == EQ_EXPR && (e->flags & EDGE_TRUE_VALUE)))) + continue; + + /* Check if bound is invarant. */ + tree idx = gimple_cond_lhs (cond); + tree bnd = gimple_cond_rhs (cond); + if (expr_invariant_in_loop_p (loop, idx)) + std::swap (idx, bnd); + else if (!expr_invariant_in_loop_p (loop, bnd)) + continue; + + /* Only unsigned type conversion could cause wrap. */ + tree type = TREE_TYPE (idx); + if (!INTEGRAL_TYPE_P (type) || TREE_CODE (idx) != SSA_NAME + || !TYPE_UNSIGNED (type)) + continue; + + /* Avoid to split if bound is MAX/MIN val. */ + tree bound_type = TREE_TYPE (bnd); + if (tree_int_cst_equal (bnd, TYPE_MAX_VALUE (bound_type)) + || tree_int_cst_equal (bnd, TYPE_MIN_VALUE (bound_type))) + continue; + + /* Check if there is possible wrap. */ + class tree_niter_desc niter; + if (!number_of_iterations_exit (loop, e, &niter, false, false)) + continue; + if (niter.control.no_overflow) + return NULL; + if (niter.cmp != NE_EXPR) + continue; + if (!integer_onep (niter.control.step) + && !integer_minus_onep (niter.control.step)) + continue; + *step = niter.control.step; ++ /* If exit edge is just before the empty latch, it is easy to link+ the split loops: just jump from the exit edge of one loop to the + header of new loop. */ + if (single_pred_p (loop->latch) + && single_pred_edge (loop->latch)->src == bb + && empty_block_p (loop->latch)) + return e; ++ /* If exit edge is at end of header, and header contains i++ or ++i,+ only, it is simple to link the split loops: jump from the end of+ one loop header to the new loop header, and use unchanged PHI result+ of the first loop as the entry PHI value of the second loop. */ + if (bb == loop->header) + { + /* Only one phi. */ + gphi_iterator psi = gsi_start_phis (bb); + if (gsi_end_p (psi)) + continue; + gphi *phi = psi.phi (); + gsi_next (&psi); + if (!gsi_end_p (psi)) + continue; + + /* Check it is ++i or ++i */ + tree next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); + tree prev = PHI_RESULT (phi); + if (idx != prev && idx != next) + continue; ++ gimple_stmt_iterator gsi = gsi_start_nondebug_after_labels_bb (bb);+ if (gsi_end_p (gsi)) + continue; + gimple *s1 = gsi_stmt (gsi); + if (!is_gimple_assign (s1) || gimple_assign_lhs (s1) != next + || gimple_assign_rhs1 (s1) != prev) + continue; + + gsi_next_nondebug (&gsi); + if (!gsi_end_p (gsi) && gsi_stmt (gsi) == cond) + return e; + } + } + + return NULL; +} ++/* Split the LOOP with NE_EXPR into two loops with GT_EXPR and LT_EXPR. */+ +static bool +split_ne_loop (struct loop *loop, edge cond_e, tree step) +{ + initialize_original_copy_tables (); + + struct loop *loop2 = loop_version (loop, boolean_true_node, NULL, + profile_probability::always (), + profile_probability::never (), + profile_probability::always (), + profile_probability::always (), true); + + gcc_assert (loop2); + update_ssa (TODO_update_ssa); + + basic_block loop2_cond_exit_bb = get_bb_copy (cond_e->src); + free_original_copy_tables (); + + gcond *gc = as_a<gcond *> (last_stmt (cond_e->src)); + gcond *dup_gc = as_a<gcond *> (last_stmt (loop2_cond_exit_bb)); + + /* Invert edges for gcond. */ + if (gimple_cond_code (gc) == EQ_EXPR) + { + auto invert_edge = [](basic_block bb) { + edge out = EDGE_SUCC (bb, 0); + edge in = EDGE_SUCC (bb, 1); + if (in->flags & EDGE_TRUE_VALUE) + std::swap (in, out); + in->flags |= EDGE_TRUE_VALUE; + in->flags &= ~EDGE_FALSE_VALUE; + out->flags |= EDGE_FALSE_VALUE; + out->flags &= ~EDGE_TRUE_VALUE; + }; + + invert_edge (gimple_bb (gc)); + invert_edge (gimple_bb (dup_gc)); + } + + /* Change if (i != n) to LOOP1:if (i > n) and LOOP2:if (i < n) */ + bool inv = expr_invariant_in_loop_p (loop, gimple_cond_lhs (gc)); + if (tree_int_cst_sign_bit (step)) + inv = !inv; + enum tree_code first_loop_code = inv ? LT_EXPR : GT_EXPR; + enum tree_code second_loop_code = inv ? GT_EXPR : LT_EXPR;You could handle gimple_cond_code (gc) == EQ_EXPR via if (gimple_cond_code (gc) == EQ_EXPR) { first_loop_code = invert_tree_comparison (first_loop_code, false);second_loop_code = invert_tree_comparison (second_loop_code, false);} that looks simpler than the lambda dance with inverting the edge flags.
GE_EXPR/LE_EXPR which look like containing EQ_EXPR :)I would update patch to use invert_tree_comparison which is simpler and shorter.
+ gimple_cond_set_code (gc, first_loop_code); + gimple_cond_set_code (dup_gc, second_loop_code); + + /* Link the exit cond edge to new loop. */ + gcond *break_cond = as_a<gcond *> (gimple_copy (gc)); + edge pred_e = single_pred_edge (loop->latch); + bool simple_loop+ = pred_e && pred_e->src == cond_e->src && empty_block_p (loop->latch);+ if (simple_loop) + gimple_cond_set_code (break_cond, second_loop_code); + else + gimple_cond_make_true (break_cond); + + basic_block break_bb = split_edge (cond_e); + gimple_stmt_iterator gsi = gsi_last_bb (break_bb); + gsi_insert_after (&gsi, break_cond, GSI_NEW_STMT); + + edge to_exit = single_succ_edge (break_bb);+ edge to_new_loop = make_edge (break_bb, loop_preheader_edge (loop2)->src,0); + to_new_loop->flags |= EDGE_TRUE_VALUE; + to_exit->flags |= EDGE_FALSE_VALUE; + to_exit->flags &= ~EDGE_FALLTHRU; + to_exit->probability = cond_e->probability; + to_new_loop->probability = to_exit->probability.invert (); + + update_ssa (TODO_update_ssa);I've re-organized the pass to perform a single TODO_update_ssa at the very end, please do not update SSA form here, nor+ connect_loop_phis (loop, loop2, to_new_loop, !simple_loop); + + rewrite_into_loop_closed_ssa_1 (NULL, 0, SSA_OP_USE, loop);re-write into loop-closed SSA.
Thanks!
+ if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, ";; Loop split on wrap index.\n"); + + return true; +} ++/* Checks if LOOP contains a suitable NE_EXPR conditional block to split.+L_H: + if (i!=N) + S; + else + goto EXIT; + i++; + goto L_H; + +The "i!=N" is like "i>N || i<N", then it could be transformed to: + +L_H: + if (i>N) + S; + else + goto EXIT; + i++; + goto L_H; +L1_H: + if (i<N) + S; + else + goto EXIT; + i++; + goto L1_H; + +The loop with "i<N" is in favor of both GIMPLE and RTL passes. */ + +static bool +split_loop_on_ne_cond (class loop *loop) +{ + tree step; + edge branch_edge = get_ne_cond_branch (loop, &step); + if (!branch_edge) + return false; + + int num = 0; + basic_block *bbs = get_loop_body (loop); + for (unsigned i = 0; i < loop->num_nodes; i++)+ num += estimate_num_insns_seq (bb_seq (bbs[i]), &eni_size_weights);Since the motivation is to make data-refs analyzable after the transformif there are no datarefs the transform only increases code-size. In particular I would look for calls which will be not analyzable. Since we're looking at each stmt above that could be embedded here. As said earlier once num exceeds param_max_peeled_insns you can stop the above loop. So heuristically I'd do sth like
Right, if no following optimization enabled, it only increases code-size and increase branches overhead, and may not gain performance.
for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p
(gsi); gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
if (is_gimple_debug (stmt))
continue;
if (gimple_has_side_effects (stmt))
{
free (bbs);
return false;
}
num += estimate_num_insns (stmt, &eni_size_weights);
if (num > param_max_peeled_insns)
{
free (bbs);
return false;
}
if (gimple_vuse (stmt))
any_dr = true;
}
if (!any_dr)
{
free (bbs);
return false;
}
There's also still the issue that the transformed loop will fail
number of iteration analysis for the loop that iterates until
the IV wraps. That's a blocker for the acceptance of this transform.
Understand, I saw you opened PR101145. Thanks for your comments! BR, Jiufu Guo.
Richard.
