* tree-vectorizer.h (DR_MISALIGNMENT): Cast aux to integer.

	(SET_DR_MISALIGNMENT): New.
	* tree-vect-analyze.c (vect_compute_data_ref_alignment,
	vect_update_misalignment_for_peel, vect_enhance_data_refs_alignment):
	Use SET_DR_MISALIGNMENT.
	* tree-predcom.c (split_data_refs_to_components): Cast dr->aux from
	pointer.
	* tree-data-ref.c (create_data_ref, compute_all_dependences,
	find_loop_nest): Export.
	* tree-data-ref.h (struct data_reference): Change aux field to pointer.
	(create_data_ref, compute_all_dependences, find_loop_nest): Declare.
	* tree-ssa-loop-prefetch.c: Include tree-data-ref.h.
	(L1_CACHE_SIZE_BYTES, L2_CACHE_SIZE_BYTES, NONTEMPORAL_FRACTION):
	New macros.
	(struct mem_ref): Add field reuse_distance.
	(find_or_create_group, record_ref): Use XNEW instead of xcalloc.
	Initialize reuse_distance field.
	(issue_prefetch_ref): Select temporality of prefetch according to
	reuse_distance.
	(volume_of_references, volume_of_dist_vector, add_subscript_strides,
	self_reuse_distance, determine_loop_nest_reuse): New functions.
	(loop_prefetch_arrays): Call determine_loop_nest_reuse.
	(tree_ssa_prefetch_arrays): Dump L2 cache size.
	* Makefile.in (tree-ssa-loop-prefetch.o): Add TREE_DATA_REF_H
	dependency.

	* gcc.dg/tree-ssa/prefetch-6.c: New test.



git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@125172 138bc75d-0d04-0410-961f-82ee72b054a4

1 files changed, 54 insertions, 0 deletions

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/prefetch-6.c b/gcc/testsuite/gcc.dg/tree-ssa/prefetch-6.c
new file mode 100644
index 00000000000..44f65bffb20
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/prefetch-6.c
@@ -0,0 +1,54 @@
+/* { dg-do compile { target i?86-*-* x86_64-*-* } } */
+/* { dg-require-effective-target ilp32 } */
+/* { dg-options "-O2 -fprefetch-loop-arrays -march=athlon -msse2 -mfpmath=sse --param simultaneous-prefetches=100 -fdump-tree-aprefetch-details" } */
+
+#define N 1000
+#define K 900
+
+double a[N][N];
+
+double test(void)
+{
+  unsigned i, j;
+  double sum = 0;
+
+  /* Here, we should use non-temporal prefetch instruction.  */
+  for (i = 0; i < K; i++)
+    for (j = 0; j < K; j++)
+      sum += a[i][j];
+
+  /* Here, we should not use non-temporal prefetch instruction, since the
+     value of a[i+10][j] is reused in L2 cache.  */
+  for (i = 0; i < K; i++)
+    for (j = 0; j < K; j++)
+      sum += a[i][j] * a[i + 10][j];
+
+  /* Here, we should use non-temporal prefetch instruction, since the
+     value of a[i+100][j] is too far to be reused in L2 cache.  */
+  for (i = 0; i < K; i++)
+    for (j = 0; j < K; j++)
+      sum += a[i][j] * a[i + 100][j];
+
+  /* Here, temporal prefetches should be used, since the volume of the
+     memory accesses is smaller than L2 cache.  */
+  for (i = 0; i < 100; i++)
+    for (j = 0; j < 100; j++)
+      sum += a[i][j] * a[i + 100][j];
+
+  /* Temporal prefetches should be used here (even though the accesses to
+     a[j][i] are independent, the same cache line is almost always hit
+     every N iterations).  */
+  for (i = 0; i < N; i++)
+    for (j = 0; j < N; j++)
+      sum += a[j][i];
+
+  return sum;
+}
+
+/* { dg-final { scan-tree-dump-times "Issued prefetch" 5 "aprefetch" } } */
+/* { dg-final { scan-tree-dump-times "Issued nontemporal prefetch" 3 "aprefetch" } } */
+
+/* { dg-final { scan-assembler-times "prefetcht" 5 } } */
+/* { dg-final { scan-assembler-times "prefetchnta" 3 } } */
+
+/* { dg-final { cleanup-tree-dump "aprefetch" } } */