Hi,
I previously committed two patches lowering complex address expression for
IVOPT at http://gcc.gnu.org/ml/gcc-patches/2013-11/msg00546.html and
http://gcc.gnu.org/ml/gcc-patches/2013-11/msg01103.html
When I bootstrapping GCC I found there were some peculiar cases like
&MEM[ptr+CST] + xxxx, which should be handled too. This patch consists
below two changes:
1) change in alloc_iv:
Original code lowers top level complex address expressions like
&MEM[ptr+off]. The patch relaxes check condition in order to lower
expressions like &MEM[ptr+off] + xxx, just as the BASE from below dump:
use 2
generic
in statement _595 = &MEM[(void *)&this_prg + 36B] + _594;
at position
type struct gcov_bucket_type *
base (struct gcov_bucket_type *) &MEM[(void *)&this_prg + 36B] +
(sizetype) ((unsigned int) (src_i_683 + -1) * 20)
step 4294967276
base object (void *) &this_prg
related candidates
2) change in tree_to_aff_combination:
The function get_inner_reference returns "&MEM[ptr+off]" as the core for
input like the memory ADDRESS in below dump:
use 2
address
in statement _59 = MEM[(const struct gcov_ctr_summary *)summary_22(D) +
4B].histogram[h_ix_111].min_value;
at position MEM[(const struct gcov_ctr_summary *)summary_22(D) +
4B].histogram[h_ix_111].min_value
type const gcov_type *
base (const gcov_type *) &MEM[(const struct gcov_ctr_summary
*)summary_22(D) + 4B] + 36
step 20
base object (void *) summary_22(D)
related candidates
Which can be further reduced into something like "summary_22(D) + 40B".
This change is necessary for the first one, because I am using
tree_to_aff_combination rather than get_inner_reference_aff now.
Bootstrap and test on x86/x86_64/arm. Is it OK?
Thanks.
bin
2013-11-25 Bin Cheng <[email protected]>
* tree-ssa-loop-ivopts.c (contain_complex_addr_expr): New.
(alloc_iv): Lower more cases by calling contain_complex_addr_expr
and tree_to_aff_combination.
* tree-affine.c (tree_to_aff_combination): Handle &MEM[ptr+CST]
in core part of complex reference.
gcc/testsuite/ChangeLog
2013-11-25 Bin Cheng <[email protected]>
* gcc.dg/tree-ssa/ivopts-lower_base.c: New test.
Index: gcc/testsuite/gcc.dg/tree-ssa/ivopts-lower_base.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/ivopts-lower_base.c (revision 0)
+++ gcc/testsuite/gcc.dg/tree-ssa/ivopts-lower_base.c (revision 0)
@@ -0,0 +1,29 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-ivopts-details" } */
+
+struct tag1
+{
+ int x[100];
+ int y;
+};
+
+struct tag2
+{
+ int a;
+ struct tag1 t1[100];
+ int b;
+};
+
+int
+foo (struct tag2 *t2, int len)
+{
+ int i = 0;
+ for (i = 0; i < len; i++)
+ {
+ (*(struct tag2*)((char *)t2+4)).t1[i].x[len] = len;
+ }
+
+ return 0;
+}
+/* { dg-final { scan-tree-dump-not "base .*&MEM\\\[" "ivopts" } } */
+/* { dg-final { cleanup-tree-dump "ivopts" } } */
Index: gcc/tree-ssa-loop-ivopts.c
===================================================================
--- gcc/tree-ssa-loop-ivopts.c (revision 205087)
+++ gcc/tree-ssa-loop-ivopts.c (working copy)
@@ -924,13 +924,40 @@ determine_base_object (tree expr)
}
}
+/* Return true if complex address expression appears in EXPR. */
+
+static bool
+contain_complex_addr_expr (tree expr)
+{
+ bool res = false;
+
+ STRIP_NOPS (expr);
+ switch (TREE_CODE (expr))
+ {
+ case POINTER_PLUS_EXPR:
+ case PLUS_EXPR:
+ case MINUS_EXPR:
+ res |= contain_complex_addr_expr (TREE_OPERAND (expr, 0));
+ res |= contain_complex_addr_expr (TREE_OPERAND (expr, 1));
+ break;
+
+ case ADDR_EXPR:
+ return (!DECL_P (TREE_OPERAND (expr, 0)));
+
+ default:
+ return false;
+ }
+
+ return res;
+}
+
/* Allocates an induction variable with given initial value BASE and step STEP
for loop LOOP. */
static struct iv *
alloc_iv (tree base, tree step)
{
- tree base_object = base;
+ tree expr = base;
struct iv *iv = XCNEW (struct iv);
gcc_assert (step != NULL_TREE);
@@ -939,21 +966,17 @@ alloc_iv (tree base, tree step)
1) More accurate cost can be computed for address expressions;
2) Duplicate candidates won't be created for bases in different
forms, like &a[0] and &a. */
- STRIP_NOPS (base_object);
- if (TREE_CODE (base_object) == ADDR_EXPR
- && !DECL_P (TREE_OPERAND (base_object, 0)))
+ STRIP_NOPS (expr);
+ if ((TREE_CODE (expr) == ADDR_EXPR && !DECL_P (TREE_OPERAND (expr, 0)))
+ || contain_complex_addr_expr (expr))
{
aff_tree comb;
- double_int size;
- base_object = get_inner_reference_aff (TREE_OPERAND (base_object, 0),
- &comb, &size);
- gcc_assert (base_object != NULL_TREE);
- base_object = build_fold_addr_expr (base_object);
+ tree_to_aff_combination (expr, TREE_TYPE (base), &comb);
base = fold_convert (TREE_TYPE (base), aff_combination_to_tree (&comb));
}
iv->base = base;
- iv->base_object = determine_base_object (base_object);
+ iv->base_object = determine_base_object (base);
iv->step = step;
iv->biv_p = false;
iv->have_use_for = false;
Index: gcc/tree-affine.c
===================================================================
--- gcc/tree-affine.c (revision 205087)
+++ gcc/tree-affine.c (working copy)
@@ -328,7 +328,19 @@ tree_to_aff_combination (tree expr, tree type, aff
double_int::from_uhwi (bitpos / BITS_PER_UNIT));
core = build_fold_addr_expr (core);
if (TREE_CODE (core) == ADDR_EXPR)
- aff_combination_add_elt (comb, core, double_int_one);
+ {
+ /* Handle &MEM[ptr + CST] in core part of complex reference. */
+ if (TREE_CODE (TREE_OPERAND (core, 0)) == MEM_REF)
+ {
+ core = TREE_OPERAND (core, 0);
+ tree_to_aff_combination (TREE_OPERAND (core, 0), type, &tmp);
+ aff_combination_add (comb, &tmp);
+ tree_to_aff_combination (TREE_OPERAND (core, 1), sizetype, &tmp);
+ aff_combination_add (comb, &tmp);
+ }
+ else
+ aff_combination_add_elt (comb, core, double_int_one);
+ }
else
{
tree_to_aff_combination (core, type, &tmp);
