For the testcase
double a[16][16][64], y[16][64], x[16][16];
void foo(void)
{
int i, j, k;
for (k = 0; k < 16; ++k)
for (j = 0; j < 64; ++j)
for (i = 0; i < 16; ++i)
{
y[k][j] = y[k][j] + a[k][i][j] * x[k][i];
}
}
loop interchange is performed with -O2 -fno-tree-pre -fno-tree-loop-im
-ftree-loop-linear but not with PRE or lim moving the store/load to/from
y[k][j]. The interchange fails because
/* Validate the resulting matrix. When the transformation
is not valid, reverse to the previous transformation. */
if (!lambda_transform_legal_p (trans, depth, dependence_relations))
returns false which is because constants or SSA names from final values
leak into the access fns like
(compute_affine_dependence
(stmt_a =
pretmp.11_3 = y[k_41][j_40])
(stmt_b =
y[k_41][j_40] = D.1578_20)
(subscript_dependence_tester
(analyze_overlapping_iterations
(chrec_a = {0, +, 1}_2)
(chrec_b = 63)
(analyze_siv_subscript
)
(overlap_iterations_a = [63]
)
(overlap_iterations_b = [0]
)
which confuses dependency analysis here:
else if (!operand_equal_p (access_fn_a, access_fn_b, 0))
{
/* This can be for example an affine vs. constant dependence
(T[i] vs. T[3]) that is not an affine dependence and is
not representable as a distance vector. */
non_affine_dependence_relation (ddr);
return false;
--
Summary: Loop interchange not performed, data dependence analysis
defect
Product: gcc
Version: 4.4.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: rguenth at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36011
