Forgot the patch~
On Wed, Aug 6, 2014 at 10:32 AM, Bin.Cheng <[email protected]> wrote:
> On Fri, Jul 25, 2014 at 8:35 PM, Richard Biener
> <[email protected]> wrote:
>> On Thu, Jul 17, 2014 at 11:08 AM, Bin Cheng <[email protected]> wrote:
>>> 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
>>
>> You add special casing for constants but I don't see any testcases for that.
>> Specifically
>>
>> + /* No overflow if offset is zero. */
>> + if (offset == integer_zero_node)
>> return true;
>>
>> is a bogus check (use integer_zerop). Apart from the special-casing of
>> constants the patch looks good to me.
>
> Hi Richard,
> I modified the patch according to your comments by removing the
> constant case. Re-bootstrap and test on x86_64 and x86. Is this
> version OK?
>
> Thanks,
> bin
>
> 2014-08-06 Bin Cheng <[email protected]>
>
> * 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. Use affine
> expansion for equality check.
> (iv_elimination_compare_lt): Pass new argument.
>
> gcc/testsuite/ChangeLog
> 2014-08-06 Bin Cheng <[email protected]>
>
> * gcc.dg/tree-ssa/ivopts-lt-2.c: New test.
Index: gcc/tree-ssa-loop-ivopts.c
===================================================================
--- gcc/tree-ssa-loop-ivopts.c (revision 213529)
+++ 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;
@@ -876,6 +879,7 @@ 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);
}
@@ -4448,75 +4452,20 @@ 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;
-}
-
-/* Walk the SSA form and check whether E == WHAT. Fairly simplistic, we
- check for an exact match. */
-
-static bool
-expr_equal_p (tree e, tree what)
-{
- gimple stmt;
- enum tree_code code;
-
- 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))
- 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))
- {
- case GIMPLE_BINARY_RHS:
- if (!expr_equal_p (gimple_assign_rhs2 (stmt), TREE_OPERAND (what, 1)))
- return false;
- /* Fallthru. */
-
- case GIMPLE_UNARY_RHS:
- case GIMPLE_SINGLE_RHS:
- return expr_equal_p (gimple_assign_rhs1 (stmt), TREE_OPERAND (what, 0));
- default:
- return false;
- }
-}
-
/* 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.
+ 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)
+difference_cannot_overflow_p (struct ivopts_data *data, tree base, tree offset)
{
enum tree_code code;
tree e1, e2;
+ aff_tree aff_e1, aff_e2, aff_offset;
if (!nowrap_type_p (TREE_TYPE (base)))
return false;
@@ -4546,13 +4495,27 @@ 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;
@@ -4676,7 +4639,7 @@ 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. */
@@ -6801,6 +6764,7 @@ 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. */
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" } } */
