https://gcc.gnu.org/g:ef27b91b62c3aa8841c02665dffa8914c742fd37
commit r15-919-gef27b91b62c3aa8841c02665dffa8914c742fd37
Author: liuhongt <hongtao....@intel.com>
Date: Tue Feb 27 15:34:57 2024 +0800
Don't reduce estimated unrolled size for innermost loop.
For the innermost loop, after completely loop unroll, it will most likely
not be able to reduce the body size to 2/3. The current 2/3 reduction
will make some of the larger loops completely unrolled during
cunrolli, which will then result in them not being able to be
vectorized. It also increases the register pressure.
The patch move the 2/3 reduction from estimated_unrolled_size to
tree_unroll_loops_completely.
gcc/ChangeLog:
PR tree-optimization/112325
* tree-ssa-loop-ivcanon.cc (estimated_unrolled_size): Move the
2 / 3 loop body size reduction to ..
(try_unroll_loop_completely): .. here, add it for the check of
body size shrink, and the check of comparison against
param_max_completely_peeled_insns when
(!cunrolli ||loop->inner).
(canonicalize_loop_induction_variables): Add new parameter
cunrolli and pass down.
(tree_unroll_loops_completely_1): Ditto.
(canonicalize_induction_variables): Pass cunrolli as false to
canonicalize_loop_induction_variables.
(tree_unroll_loops_completely): Set cunrolli to true at
beginning and set it to false after CHANGED is true.
gcc/testsuite/ChangeLog:
* gcc.dg/vect/pr112325.c: New test.
Diff:
---
gcc/testsuite/gcc.dg/vect/pr112325.c | 59 ++++++++++++++++++++++++++++++++++++
gcc/tree-ssa-loop-ivcanon.cc | 49 ++++++++++++++++--------------
2 files changed, 86 insertions(+), 22 deletions(-)
diff --git a/gcc/testsuite/gcc.dg/vect/pr112325.c
b/gcc/testsuite/gcc.dg/vect/pr112325.c
new file mode 100644
index 00000000000..71cf4099253
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr112325.c
@@ -0,0 +1,59 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -funroll-loops -fdump-tree-vect-details" } */
+/* { dg-require-effective-target vect_int } */
+/* { dg-additional-options "-mavx2" { target x86_64-*-* i?86-*-* } } */
+
+typedef unsigned short ggml_fp16_t;
+static float table_f32_f16[1 << 16];
+
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+ unsigned short s;
+ __builtin_memcpy(&s, &f, sizeof(unsigned short));
+ return table_f32_f16[s];
+}
+
+typedef struct {
+ ggml_fp16_t d;
+ ggml_fp16_t m;
+ unsigned char qh[4];
+ unsigned char qs[32 / 2];
+} block_q5_1;
+
+typedef struct {
+ float d;
+ float s;
+ char qs[32];
+} block_q8_1;
+
+void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
restrict vx, const void * restrict vy) {
+ const int qk = 32;
+ const int nb = n / qk;
+
+ const block_q5_1 * restrict x = vx;
+ const block_q8_1 * restrict y = vy;
+
+ float sumf = 0.0;
+
+ for (int i = 0; i < nb; i++) {
+ unsigned qh;
+ __builtin_memcpy(&qh, x[i].qh, sizeof(qh));
+
+ int sumi = 0;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const unsigned char xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
+ const unsigned char xh_1 = ((qh >> (j + 12)) ) & 0x10;
+
+ const int x0 = (x[i].qs[j] & 0xF) | xh_0;
+ const int x1 = (x[i].qs[j] >> 4) | xh_1;
+
+ sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
+ }
+
+ sumf += (ggml_lookup_fp16_to_fp32(x[i].d)*y[i].d)*sumi +
ggml_lookup_fp16_to_fp32(x[i].m)*y[i].s;
+ }
+
+ *s = sumf;
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
diff --git a/gcc/tree-ssa-loop-ivcanon.cc b/gcc/tree-ssa-loop-ivcanon.cc
index bf017137260..5ef24a91917 100644
--- a/gcc/tree-ssa-loop-ivcanon.cc
+++ b/gcc/tree-ssa-loop-ivcanon.cc
@@ -437,11 +437,7 @@ tree_estimate_loop_size (class loop *loop, edge exit, edge
edge_to_cancel,
It is (NUNROLL + 1) * size of loop body with taking into account
the fact that in last copy everything after exit conditional
is dead and that some instructions will be eliminated after
- peeling.
-
- Loop body is likely going to simplify further, this is difficult
- to guess, we just decrease the result by 1/3. */
-
+ peeling. */
static unsigned HOST_WIDE_INT
estimated_unrolled_size (struct loop_size *size,
unsigned HOST_WIDE_INT nunroll)
@@ -453,10 +449,6 @@ estimated_unrolled_size (struct loop_size *size,
unr_insns = 0;
unr_insns += size->last_iteration -
size->last_iteration_eliminated_by_peeling;
- unr_insns = unr_insns * 2 / 3;
- if (unr_insns <= 0)
- unr_insns = 1;
-
return unr_insns;
}
@@ -734,7 +726,8 @@ try_unroll_loop_completely (class loop *loop,
edge exit, tree niter, bool may_be_zero,
enum unroll_level ul,
HOST_WIDE_INT maxiter,
- dump_user_location_t locus, bool allow_peel)
+ dump_user_location_t locus, bool allow_peel,
+ bool cunrolli)
{
unsigned HOST_WIDE_INT n_unroll = 0;
bool n_unroll_found = false;
@@ -847,8 +840,9 @@ try_unroll_loop_completely (class loop *loop,
/* If the code is going to shrink, we don't need to be extra
cautious on guessing if the unrolling is going to be
- profitable. */
- if (unr_insns
+ profitable.
+ Move from estimated_unrolled_size to unroll small loops. */
+ if (unr_insns * 2 / 3
/* If there is IV variable that will become constant, we
save one instruction in the loop prologue we do not
account otherwise. */
@@ -919,7 +913,13 @@ try_unroll_loop_completely (class loop *loop,
loop->num);
return false;
}
- else if (unr_insns
+ /* Move 2 / 3 reduction from estimated_unrolled_size, but don't reduce
+ unrolled size for innermost loop.
+ 1) It could increase register pressure.
+ 2) Big loop after completely unroll may not be vectorized
+ by BB vectorizer. */
+ else if ((cunrolli && !loop->inner
+ ? unr_insns : unr_insns * 2 / 3)
> (unsigned) param_max_completely_peeled_insns)
{
if (dump_file && (dump_flags & TDF_DETAILS))
@@ -1227,7 +1227,7 @@ try_peel_loop (class loop *loop,
static bool
canonicalize_loop_induction_variables (class loop *loop,
bool create_iv, enum unroll_level ul,
- bool try_eval, bool allow_peel)
+ bool try_eval, bool allow_peel, bool
cunrolli)
{
edge exit = NULL;
tree niter;
@@ -1314,7 +1314,7 @@ canonicalize_loop_induction_variables (class loop *loop,
dump_user_location_t locus = find_loop_location (loop);
if (try_unroll_loop_completely (loop, exit, niter, may_be_zero, ul,
- maxiter, locus, allow_peel))
+ maxiter, locus, allow_peel, cunrolli))
return true;
if (create_iv
@@ -1358,7 +1358,7 @@ canonicalize_induction_variables (void)
{
changed |= canonicalize_loop_induction_variables (loop,
true, UL_SINGLE_ITER,
- true, false);
+ true, false, false);
}
gcc_assert (!need_ssa_update_p (cfun));
@@ -1392,7 +1392,7 @@ canonicalize_induction_variables (void)
static bool
tree_unroll_loops_completely_1 (bool may_increase_size, bool unroll_outer,
- bitmap father_bbs, class loop *loop)
+ bitmap father_bbs, class loop *loop, bool
cunrolli)
{
class loop *loop_father;
bool changed = false;
@@ -1410,7 +1410,7 @@ tree_unroll_loops_completely_1 (bool may_increase_size,
bool unroll_outer,
if (!child_father_bbs)
child_father_bbs = BITMAP_ALLOC (NULL);
if (tree_unroll_loops_completely_1 (may_increase_size, unroll_outer,
- child_father_bbs, inner))
+ child_father_bbs, inner, cunrolli))
{
bitmap_ior_into (father_bbs, child_father_bbs);
bitmap_clear (child_father_bbs);
@@ -1456,7 +1456,7 @@ tree_unroll_loops_completely_1 (bool may_increase_size,
bool unroll_outer,
ul = UL_NO_GROWTH;
if (canonicalize_loop_induction_variables
- (loop, false, ul, !flag_tree_loop_ivcanon, unroll_outer))
+ (loop, false, ul, !flag_tree_loop_ivcanon, unroll_outer, cunrolli))
{
/* If we'll continue unrolling, we need to propagate constants
within the new basic blocks to fold away induction variable
@@ -1491,6 +1491,7 @@ tree_unroll_loops_completely (bool may_increase_size,
bool unroll_outer)
bool changed;
int iteration = 0;
bool irred_invalidated = false;
+ bool cunrolli = true;
estimate_numbers_of_iterations (cfun);
@@ -1507,10 +1508,15 @@ tree_unroll_loops_completely (bool may_increase_size,
bool unroll_outer)
changed = tree_unroll_loops_completely_1 (may_increase_size,
unroll_outer, father_bbs,
- current_loops->tree_root);
+ current_loops->tree_root,
+ cunrolli);
if (changed)
{
unsigned i;
+ /* For the outer loop, considering that the inner loop is completely
+ unrolled, it would expose more optimization opportunities, so it's
+ better to keep 2/3 reduction of estimated unrolled size. */
+ cunrolli = false;
unloop_loops (loops_to_unloop, loops_to_unloop_nunroll,
edges_to_remove, loop_closed_ssa_invalidated,
@@ -1670,8 +1676,7 @@ pass_complete_unroll::execute (function *fun)
re-peeling the same loop multiple times. */
if (flag_peel_loops)
peeled_loops = BITMAP_ALLOC (NULL);
- unsigned int val = tree_unroll_loops_completely (flag_cunroll_grow_size,
- true);
+ unsigned int val = tree_unroll_loops_completely (flag_cunroll_grow_size,
true);
if (peeled_loops)
{
BITMAP_FREE (peeled_loops);
