Hi, As quoted from the function difference_cannot_overflow_p, /* TODO: deeper inspection may be necessary to prove the equality. */ switch (code) { case PLUS_EXPR: return expr_equal_p (e1, offset) || expr_equal_p (e2, offset); case POINTER_PLUS_EXPR: return expr_equal_p (e2, offset);
default: return false; } The overflow check can be improved by using deeper inspection to prove the equality. This patch deals with that by making below two improvements: a) Handles constant cases. b) Uses affine expansion as deeper inspection to check the equality. As a result, functions strip_wrap_conserving_type_conversions and expr_equal_p can be removed now. A test case is also added to illustrate iv elimination opportunity captured by this patch. Thanks, bin 2014-07-17 Bin Cheng <bin.ch...@arm.com> * tree-ssa-loop-ivopts.c (ivopts_data): New field name_expansion. (tree_ssa_iv_optimize_init): Initialize name_expansion. (tree_ssa_iv_optimize_finalize): Free name_expansion. (strip_wrap_conserving_type_conversions, expr_equal_p): Delete. (difference_cannot_overflow_p): New parameter. Handle constant cases. Use affine expansion for equality check. (iv_elimination_compare_lt): Pass new argument. gcc/testsuite/ChangeLog 2014-07-17 Bin Cheng <bin.ch...@arm.com> * gcc.dg/tree-ssa/ivopts-lt-2.c: New test.
Index: gcc/testsuite/gcc.dg/tree-ssa/ivopts-lt-2.c =================================================================== --- gcc/testsuite/gcc.dg/tree-ssa/ivopts-lt-2.c (revision 0) +++ gcc/testsuite/gcc.dg/tree-ssa/ivopts-lt-2.c (revision 0) @@ -0,0 +1,20 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -fdump-tree-ivopts" } */ + +void +f1 (int *p, unsigned int i) +{ + p += i; + do + { + *p = 0; + p += 1; + i++; + } + while (i < 100); +} + +/* { dg-final { scan-tree-dump-times "PHI" 1 "ivopts" } } */ +/* { dg-final { scan-tree-dump-times "PHI <p_" 1 "ivopts"} } */ +/* { dg-final { scan-tree-dump-times "p_\[0-9\]* <" 1 "ivopts" } } */ +/* { dg-final { cleanup-tree-dump "ivopts" } } */ Index: gcc/tree-ssa-loop-ivopts.c =================================================================== --- gcc/tree-ssa-loop-ivopts.c (revision 212387) +++ gcc/tree-ssa-loop-ivopts.c (working copy) @@ -323,6 +323,9 @@ struct ivopts_data /* A bitmap of important candidates. */ bitmap important_candidates; + /* Cache used by tree_to_aff_combination_expand. */ + struct pointer_map_t *name_expansion; + /* The maximum invariant id. */ unsigned max_inv_id; @@ -877,6 +880,7 @@ tree_ssa_iv_optimize_init (struct ivopts_data *dat data->iv_candidates.create (20); data->inv_expr_tab = new hash_table<iv_inv_expr_hasher> (10); data->inv_expr_id = 0; + data->name_expansion = NULL; decl_rtl_to_reset.create (20); } @@ -4449,76 +4453,40 @@ iv_elimination_compare (struct ivopts_data *data, return (exit->flags & EDGE_TRUE_VALUE ? EQ_EXPR : NE_EXPR); } -static tree -strip_wrap_conserving_type_conversions (tree exp) -{ - while (tree_ssa_useless_type_conversion (exp) - && (nowrap_type_p (TREE_TYPE (exp)) - == nowrap_type_p (TREE_TYPE (TREE_OPERAND (exp, 0))))) - exp = TREE_OPERAND (exp, 0); - return exp; -} +/* Returns true if we can prove that BASE - OFFSET does not overflow. For now, + we only check either the case BASE and OFFSET are integer constants, or the + situation that BASE = SOMETHING + OFFSET, where the calculation is performed + in non-wrapping type. For the latter case, we use affine expansion for + further equality check. -/* Walk the SSA form and check whether E == WHAT. Fairly simplistic, we - check for an exact match. */ + TODO: More generally, we could test for the situation that + BASE = SOMETHING + OFFSET' and OFFSET is between OFFSET' and zero. + This would require knowing the sign of OFFSET. */ static bool -expr_equal_p (tree e, tree what) +difference_cannot_overflow_p (struct ivopts_data *data, tree base, tree offset) { - gimple stmt; enum tree_code code; + tree e1, e2; + aff_tree aff_e1, aff_e2, aff_offset; - e = strip_wrap_conserving_type_conversions (e); - what = strip_wrap_conserving_type_conversions (what); - - code = TREE_CODE (what); - if (TREE_TYPE (e) != TREE_TYPE (what)) - return false; - - if (operand_equal_p (e, what, 0)) + /* No overflow if offset is zero. */ + if (offset == integer_zero_node) return true; - if (TREE_CODE (e) != SSA_NAME) - return false; - - stmt = SSA_NAME_DEF_STMT (e); - if (gimple_code (stmt) != GIMPLE_ASSIGN - || gimple_assign_rhs_code (stmt) != code) - return false; - - switch (get_gimple_rhs_class (code)) + /* Overflow can be checked easily for constant values. */ + if (TREE_CODE (base) == INTEGER_CST && TREE_CODE (offset) == INTEGER_CST) { - case GIMPLE_BINARY_RHS: - if (!expr_equal_p (gimple_assign_rhs2 (stmt), TREE_OPERAND (what, 1))) - return false; - /* Fallthru. */ + bool overflow = false; + tree type = TREE_TYPE (base); + signop sign = TYPE_SIGN (type); - case GIMPLE_UNARY_RHS: - case GIMPLE_SINGLE_RHS: - return expr_equal_p (gimple_assign_rhs1 (stmt), TREE_OPERAND (what, 0)); - default: - return false; + wide_int arg2 = wide_int::from (offset, TYPE_PRECISION (type), + TYPE_SIGN (TREE_TYPE (offset))); + (void) wi::sub (base, arg2, sign, &overflow); + return overflow; } -} -/* Returns true if we can prove that BASE - OFFSET does not overflow. For now, - we only detect the situation that BASE = SOMETHING + OFFSET, where the - calculation is performed in non-wrapping type. - - TODO: More generally, we could test for the situation that - BASE = SOMETHING + OFFSET' and OFFSET is between OFFSET' and zero. - This would require knowing the sign of OFFSET. - - Also, we only look for the first addition in the computation of BASE. - More complex analysis would be better, but introducing it just for - this optimization seems like an overkill. */ - -static bool -difference_cannot_overflow_p (tree base, tree offset) -{ - enum tree_code code; - tree e1, e2; - if (!nowrap_type_p (TREE_TYPE (base))) return false; @@ -4547,13 +4515,27 @@ static bool e2 = TREE_OPERAND (base, 1); } - /* TODO: deeper inspection may be necessary to prove the equality. */ + /* Use affine expansion as deeper inspection to prove the equality. */ + tree_to_aff_combination_expand (e2, TREE_TYPE (e2), + &aff_e2, &data->name_expansion); + tree_to_aff_combination_expand (offset, TREE_TYPE (offset), + &aff_offset, &data->name_expansion); + aff_combination_scale (&aff_offset, -1); switch (code) { case PLUS_EXPR: - return expr_equal_p (e1, offset) || expr_equal_p (e2, offset); + aff_combination_add (&aff_e2, &aff_offset); + if (aff_combination_zero_p (&aff_e2)) + return true; + + tree_to_aff_combination_expand (e1, TREE_TYPE (e1), + &aff_e1, &data->name_expansion); + aff_combination_add (&aff_e1, &aff_offset); + return aff_combination_zero_p (&aff_e1); + case POINTER_PLUS_EXPR: - return expr_equal_p (e2, offset); + aff_combination_add (&aff_e2, &aff_offset); + return aff_combination_zero_p (&aff_e2); default: return false; @@ -4677,7 +4659,7 @@ iv_elimination_compare_lt (struct ivopts_data *dat offset = fold_build2 (MULT_EXPR, TREE_TYPE (cand->iv->step), cand->iv->step, fold_convert (TREE_TYPE (cand->iv->step), a)); - if (!difference_cannot_overflow_p (cand->iv->base, offset)) + if (!difference_cannot_overflow_p (data, cand->iv->base, offset)) return false; /* Determine the new comparison operator. */ @@ -6805,6 +6787,7 @@ tree_ssa_iv_optimize_finalize (struct ivopts_data data->iv_candidates.release (); delete data->inv_expr_tab; data->inv_expr_tab = NULL; + free_affine_expand_cache (&data->name_expansion); } /* Returns true if the loop body BODY includes any function calls. */