From 318a328178a8984bb2a0cc17e8c23ed21015b5fe Mon Sep 17 00:00:00 2001 From: rakdver Date: Thu, 17 Aug 2006 08:22:05 +0000 Subject: [PATCH] PR tree-optimization/27865 * tree-vrp.c (adjust_range_with_scev): Do not use TYPE_{MIN,MAX}_VALUE for pointer types. * tree-scalar-evolution.c (fold_used_pointer_cast, pointer_offset_p, fold_used_pointer, pointer_used_p): New functions. (analyze_scalar_evolution_1): Use fold_used_pointer. * tree-chrec.c (convert_affine_scev): Convert no-op casts correctly. * tree-ssa-loop-ivopts.c (generic_type_for): Return integral type for pointers. git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@116213 138bc75d-0d04-0410-961f-82ee72b054a4 --- gcc/ChangeLog | 12 +++ gcc/tree-chrec.c | 10 ++- gcc/tree-scalar-evolution.c | 183 ++++++++++++++++++++++++++++++++++++++++++++ gcc/tree-ssa-loop-ivopts.c | 6 +- gcc/tree-vrp.c | 40 ++++++---- 5 files changed, 231 insertions(+), 20 deletions(-) diff --git a/gcc/ChangeLog b/gcc/ChangeLog index d2bc8c54cd0..e8c816ad804 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,15 @@ +2006-08-16 Zdenek Dvorak + + PR tree-optimization/27865 + * tree-vrp.c (adjust_range_with_scev): Do not use TYPE_{MIN,MAX}_VALUE + for pointer types. + * tree-scalar-evolution.c (fold_used_pointer_cast, pointer_offset_p, + fold_used_pointer, pointer_used_p): New functions. + (analyze_scalar_evolution_1): Use fold_used_pointer. + * tree-chrec.c (convert_affine_scev): Convert no-op casts correctly. + * tree-ssa-loop-ivopts.c (generic_type_for): Return integral type + for pointers. + 2006-08-17 Paolo Bonzini PR c++/28573 diff --git a/gcc/tree-chrec.c b/gcc/tree-chrec.c index f36fc9b0291..a74a49c3972 100644 --- a/gcc/tree-chrec.c +++ b/gcc/tree-chrec.c @@ -1162,7 +1162,10 @@ convert_affine_scev (struct loop *loop, tree type, -- must_check_src_overflow is true, and the range of TYPE is superset of the range of CT -- i.e., in all cases except if CT signed and TYPE unsigned. - -- both CT and TYPE have the same precision and signedness. */ + -- both CT and TYPE have the same precision and signedness, and we + verify instead that the source does not overflow (this may be + easier than verifying it for the result, as we may use the + information about the semantics of overflow in CT). */ if (must_check_src_overflow) { if (TYPE_UNSIGNED (type) && !TYPE_UNSIGNED (ct)) @@ -1172,7 +1175,10 @@ convert_affine_scev (struct loop *loop, tree type, } else if (TYPE_UNSIGNED (ct) == TYPE_UNSIGNED (type) && TYPE_PRECISION (ct) == TYPE_PRECISION (type)) - must_check_rslt_overflow = false; + { + must_check_rslt_overflow = false; + must_check_src_overflow = true; + } else must_check_rslt_overflow = true; } diff --git a/gcc/tree-scalar-evolution.c b/gcc/tree-scalar-evolution.c index 52b0ba38b67..13cbe42b55f 100644 --- a/gcc/tree-scalar-evolution.c +++ b/gcc/tree-scalar-evolution.c @@ -1718,6 +1718,184 @@ compute_scalar_evolution_in_loop (struct loop *wrto_loop, return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet); } +/* Folds EXPR, if it is a cast to pointer, assuming that the created + polynomial_chrec does not wrap. */ + +static tree +fold_used_pointer_cast (tree expr) +{ + tree op; + tree type, inner_type; + + if (TREE_CODE (expr) != NOP_EXPR && TREE_CODE (expr) != CONVERT_EXPR) + return expr; + + op = TREE_OPERAND (expr, 0); + if (TREE_CODE (op) != POLYNOMIAL_CHREC) + return expr; + + type = TREE_TYPE (expr); + inner_type = TREE_TYPE (op); + + if (!INTEGRAL_TYPE_P (inner_type) + || TYPE_PRECISION (inner_type) != TYPE_PRECISION (type)) + return expr; + + return build_polynomial_chrec (CHREC_VARIABLE (op), + chrec_convert (type, CHREC_LEFT (op), NULL_TREE), + chrec_convert (type, CHREC_RIGHT (op), NULL_TREE)); +} + +/* Returns true if EXPR is an expression corresponding to offset of pointer + in p + offset. */ + +static bool +pointer_offset_p (tree expr) +{ + if (TREE_CODE (expr) == INTEGER_CST) + return true; + + if ((TREE_CODE (expr) == NOP_EXPR || TREE_CODE (expr) == CONVERT_EXPR) + && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (expr, 0)))) + return true; + + return false; +} + +/* EXPR is a scalar evolution of a pointer that is dereferenced or used in + comparison. This means that it must point to a part of some object in + memory, which enables us to argue about overflows and possibly simplify + the EXPR. Returns the simplified value. + + Currently, for + + int i, n; + int *p; + + for (i = -n; i < n; i++) + *(p + i) = ...; + + We generate the following code (assuming that size of int and size_t is + 4 bytes): + + for (i = -n; i < n; i++) + { + size_t tmp1, tmp2; + int *tmp3, *tmp4; + + tmp1 = (size_t) i; (1) + tmp2 = 4 * tmp1; (2) + tmp3 = (int *) tmp2; (3) + tmp4 = p + tmp3; (4) + + *tmp4 = ...; + } + + We in general assume that pointer arithmetics does not overflow (since its + behavior is undefined in that case). One of the problems is that our + translation does not capture this property very well -- (int *) is + considered unsigned, hence the computation in (4) does overflow if i is + negative. + + This impreciseness creates complications in scev analysis. The scalar + evolution of i is [-n, +, 1]. Since int and size_t have the same precision + (in this example), and size_t is unsigned (so we do not care about + overflows), we succeed to derive that scev of tmp1 is [(size_t) -n, +, 1] + and scev of tmp2 is [4 * (size_t) -n, +, 4]. With tmp3, we run into + problem -- [(int *) (4 * (size_t) -n), +, 4] wraps, and since we on several + places assume that this is not the case for scevs with pointer type, we + cannot use this scev for tmp3; hence, its scev is + (int *) [(4 * (size_t) -n), +, 4], and scev of tmp4 is + p + (int *) [(4 * (size_t) -n), +, 4]. Most of the optimizers are unable to + work with scevs of this shape. + + However, since tmp4 is dereferenced, all its values must belong to a single + object, and taking into account that the precision of int * and size_t is + the same, it is impossible for its scev to wrap. Hence, we can derive that + its evolution is [p + (int *) (4 * (size_t) -n), +, 4], which the optimizers + can work with. + + ??? Maybe we should use different representation for pointer arithmetics, + however that is a long-term project with a lot of potential for creating + bugs. */ + +static tree +fold_used_pointer (tree expr) +{ + tree op0, op1, new0, new1; + enum tree_code code = TREE_CODE (expr); + + if (code == PLUS_EXPR + || code == MINUS_EXPR) + { + op0 = TREE_OPERAND (expr, 0); + op1 = TREE_OPERAND (expr, 1); + + if (pointer_offset_p (op1)) + { + new0 = fold_used_pointer (op0); + new1 = fold_used_pointer_cast (op1); + } + else if (code == PLUS_EXPR && pointer_offset_p (op0)) + { + new0 = fold_used_pointer_cast (op0); + new1 = fold_used_pointer (op1); + } + else + return expr; + + if (new0 == op0 && new1 == op1) + return expr; + + if (code == PLUS_EXPR) + expr = chrec_fold_plus (TREE_TYPE (expr), new0, new1); + else + expr = chrec_fold_minus (TREE_TYPE (expr), new0, new1); + + return expr; + } + else + return fold_used_pointer_cast (expr); +} + +/* Returns true if PTR is dereferenced, or used in comparison. */ + +static bool +pointer_used_p (tree ptr) +{ + use_operand_p use_p; + imm_use_iterator imm_iter; + tree stmt, rhs; + struct ptr_info_def *pi = get_ptr_info (ptr); + var_ann_t v_ann = var_ann (SSA_NAME_VAR (ptr)); + + /* Check whether the pointer has a memory tag; if it does, it is + (or at least used to be) dereferenced. */ + if ((pi != NULL && pi->name_mem_tag != NULL) + || v_ann->symbol_mem_tag) + return true; + + FOR_EACH_IMM_USE_FAST (use_p, imm_iter, ptr) + { + stmt = USE_STMT (use_p); + if (TREE_CODE (stmt) == COND_EXPR) + return true; + + if (TREE_CODE (stmt) != MODIFY_EXPR) + continue; + + rhs = TREE_OPERAND (stmt, 1); + if (!COMPARISON_CLASS_P (rhs)) + continue; + + if (TREE_OPERAND (stmt, 0) == ptr + || TREE_OPERAND (stmt, 1) == ptr) + return true; + } + + return false; +} + /* Helper recursive function. */ static tree @@ -1766,6 +1944,11 @@ analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res) { case MODIFY_EXPR: res = interpret_rhs_modify_expr (loop, def, TREE_OPERAND (def, 1), type); + + if (POINTER_TYPE_P (type) + && !automatically_generated_chrec_p (res) + && pointer_used_p (var)) + res = fold_used_pointer (res); break; case PHI_NODE: diff --git a/gcc/tree-ssa-loop-ivopts.c b/gcc/tree-ssa-loop-ivopts.c index 0e8fa94b9d6..2ead5495b7e 100644 --- a/gcc/tree-ssa-loop-ivopts.c +++ b/gcc/tree-ssa-loop-ivopts.c @@ -1925,14 +1925,14 @@ strip_offset (tree expr, unsigned HOST_WIDE_INT *offset) } /* Returns variant of TYPE that can be used as base for different uses. - For integer types, we return unsigned variant of the type, which - avoids problems with overflows. For pointer types, we return void *. */ + We return unsigned type with the same precision, which avoids problems + with overflows. */ static tree generic_type_for (tree type) { if (POINTER_TYPE_P (type)) - return ptr_type_node; + return unsigned_type_for (type); if (TYPE_UNSIGNED (type)) return type; diff --git a/gcc/tree-vrp.c b/gcc/tree-vrp.c index 335ae614897..5f532114757 100644 --- a/gcc/tree-vrp.c +++ b/gcc/tree-vrp.c @@ -2015,7 +2015,7 @@ static void adjust_range_with_scev (value_range_t *vr, struct loop *loop, tree stmt, tree var) { - tree init, step, chrec; + tree init, step, chrec, tmin, tmax, min, max, type; enum ev_direction dir; /* TODO. Don't adjust anti-ranges. An anti-range may provide @@ -2049,13 +2049,23 @@ adjust_range_with_scev (value_range_t *vr, struct loop *loop, tree stmt, true)) return; - if (!POINTER_TYPE_P (TREE_TYPE (init)) - && (vr->type == VR_VARYING || vr->type == VR_UNDEFINED)) + type = TREE_TYPE (var); + if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type)) + tmin = lower_bound_in_type (type, type); + else + tmin = TYPE_MIN_VALUE (type); + if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type)) + tmax = upper_bound_in_type (type, type); + else + tmax = TYPE_MAX_VALUE (type); + + if (vr->type == VR_VARYING || vr->type == VR_UNDEFINED) { + min = tmin; + max = tmax; + /* For VARYING or UNDEFINED ranges, just about anything we get from scalar evolutions should be better. */ - tree min = TYPE_MIN_VALUE (TREE_TYPE (init)); - tree max = TYPE_MAX_VALUE (TREE_TYPE (init)); if (dir == EV_DIR_DECREASES) max = init; @@ -2064,7 +2074,8 @@ adjust_range_with_scev (value_range_t *vr, struct loop *loop, tree stmt, /* If we would create an invalid range, then just assume we know absolutely nothing. This may be over-conservative, - but it's clearly safe. */ + but it's clearly safe, and should happen only in unreachable + parts of code, or for invalid programs. */ if (compare_values (min, max) == 1) return; @@ -2072,8 +2083,8 @@ adjust_range_with_scev (value_range_t *vr, struct loop *loop, tree stmt, } else if (vr->type == VR_RANGE) { - tree min = vr->min; - tree max = vr->max; + min = vr->min; + max = vr->max; if (dir == EV_DIR_DECREASES) { @@ -2084,10 +2095,11 @@ adjust_range_with_scev (value_range_t *vr, struct loop *loop, tree stmt, max = init; /* If we just created an invalid range with the minimum - greater than the maximum, take the minimum all the - way to -INF. */ + greater than the maximum, we fail conservatively. + This should happen only in unreachable + parts of code, or for invalid programs. */ if (compare_values (min, max) == 1) - min = TYPE_MIN_VALUE (TREE_TYPE (min)); + return; } } else @@ -2097,11 +2109,9 @@ adjust_range_with_scev (value_range_t *vr, struct loop *loop, tree stmt, { min = init; - /* If we just created an invalid range with the minimum - greater than the maximum, take the maximum all the - way to +INF. */ + /* Again, avoid creating invalid range by failing. */ if (compare_values (min, max) == 1) - max = TYPE_MAX_VALUE (TREE_TYPE (max)); + return; } } -- 2.11.4.GIT