On 15/10/2021 09:48, Richard Biener wrote:
On Tue, 12 Oct 2021, Andre Vieira (lists) wrote:
Hi Richi,
I think this is what you meant, I now hide all the unrolling cost calculations
in the existing target hooks for costs. I did need to adjust 'finish_cost' to
take the loop_vinfo so the target's implementations are able to set the newly
renamed 'suggested_unroll_factor'.
Also added the checks for the epilogue's VF.
Is this more like what you had in mind?
Not exactly (sorry..). For the target hook I think we don't want to
pass vec_info but instead another output parameter like the existing
ones.
vect_estimate_min_profitable_iters should then via
vect_analyze_loop_costing and vect_analyze_loop_2 report the unroll
suggestion to vect_analyze_loop which should then, if the suggestion
was > 1, instead of iterating to the next vector mode run again
with a fixed VF (old VF times suggested unroll factor - there's
min_vf in vect_analyze_loop_2 which we should adjust to
the old VF times two for example and maybe store the suggested
factor as hint) - if it succeeds the result will end up in the
list of considered modes (where we now may have more than one
entry for the same mode but a different VF), we probably want to
only consider more unrolling once.
For simplicity I'd probably set min_vf = max_vf = old VF * suggested
factor, thus take the targets request literally.
Richard.
Hi,
I now pass an output parameter to finish_costs and route it through the
various calls up to vect_analyze_loop. I tried to rework
vect_determine_vectorization_factor and noticed that merely setting
min_vf and max_vf is not enough, we only use these to check whether the
vectorization factor is within range, well actually we only use max_vf
at that stage. We only seem to use 'min_vf' to make sure the
data_references are valid. I am not sure my changes are the most
appropriate here, for instance I am pretty sure the checks for max and
min vf I added in vect_determine_vectorization_factor are currently
superfluous as they will pass by design, but thought they might be good
future proofing?
Also I changed how we compare against max_vf, rather than relying on the
'MAX_VECTORIZATION' I decided to use the estimated_poly_value with
POLY_VALUE_MAX, to be able to bound it further in case we have knowledge
of the VL. I am not entirely about the validity of this change, maybe we
are better off keeping the MAX_VECTORIZATION in place and not making any
changes to max_vf for unrolling.
What do you think?
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index
36519ccc5a58abab483c38d0a6c5f039592bfc7f..9b1e01e9b62050d7e34bc55454771e40bdbdb4cb
100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -15972,8 +15972,8 @@ aarch64_adjust_body_cost (aarch64_vector_costs *costs,
unsigned int body_cost)
/* Implement TARGET_VECTORIZE_FINISH_COST. */
static void
-aarch64_finish_cost (void *data, unsigned *prologue_cost,
- unsigned *body_cost, unsigned *epilogue_cost)
+aarch64_finish_cost (void *data, unsigned *prologue_cost, unsigned *body_cost,
+ unsigned *epilogue_cost, unsigned *suggested_unroll_factor)
{
auto *costs = static_cast<aarch64_vector_costs *> (data);
*prologue_cost = costs->region[vect_prologue];
@@ -15984,6 +15984,9 @@ aarch64_finish_cost (void *data, unsigned
*prologue_cost,
&& costs->vec_flags
&& aarch64_use_new_vector_costs_p ())
*body_cost = aarch64_adjust_body_cost (costs, *body_cost);
+
+ if(suggested_unroll_factor)
+ *suggested_unroll_factor = 1;
}
/* Implement TARGET_VECTORIZE_DESTROY_COST_DATA. */
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index
afc2674d49da370ae0f5ef277df7e9954f303b8e..a48e43879512793907fef946c1575c3ed7f68092
100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -23048,13 +23048,15 @@ ix86_add_stmt_cost (class vec_info *vinfo, void
*data, int count,
/* Implement targetm.vectorize.finish_cost. */
static void
-ix86_finish_cost (void *data, unsigned *prologue_cost,
- unsigned *body_cost, unsigned *epilogue_cost)
+ix86_finish_cost (void *data, unsigned *prologue_cost, unsigned *body_cost,
+ unsigned *epilogue_cost, unsigned *suggested_unroll_factor)
{
unsigned *cost = (unsigned *) data;
*prologue_cost = cost[vect_prologue];
*body_cost = cost[vect_body];
*epilogue_cost = cost[vect_epilogue];
+ if (suggested_unroll_factor)
+ *suggested_unroll_factor = 1;
}
/* Implement targetm.vectorize.destroy_cost_data. */
diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
index
ad81dfb316dff00cde810d6b1edd31fa49d5c1e8..59d30ad6fcd1758383c52e34a0f90a126c501ec3
100644
--- a/gcc/config/rs6000/rs6000.c
+++ b/gcc/config/rs6000/rs6000.c
@@ -5551,8 +5551,8 @@ rs6000_adjust_vect_cost_per_loop (rs6000_cost_data *data)
/* Implement targetm.vectorize.finish_cost. */
static void
-rs6000_finish_cost (void *data, unsigned *prologue_cost,
- unsigned *body_cost, unsigned *epilogue_cost)
+rs6000_finish_cost (void *data, unsigned *prologue_cost, unsigned *body_cost,
+ unsigned *epilogue_cost, unsigned *suggested_unroll_factor)
{
rs6000_cost_data *cost_data = (rs6000_cost_data*) data;
@@ -5578,6 +5578,8 @@ rs6000_finish_cost (void *data, unsigned *prologue_cost,
*prologue_cost = cost_data->cost[vect_prologue];
*body_cost = cost_data->cost[vect_body];
*epilogue_cost = cost_data->cost[vect_epilogue];
+ if (suggested_unroll_factor)
+ *suggested_unroll_factor = 1;
}
/* Implement targetm.vectorize.destroy_cost_data. */
diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi
index
be8148583d8571b0d035b1938db9d056bfd213a8..c584260b02c3e8d4fcd7b31c38321d5f81a71428
100644
--- a/gcc/doc/tm.texi
+++ b/gcc/doc/tm.texi
@@ -6276,11 +6276,12 @@ return value should be viewed as a tentative cost that
may later be
revised.
@end deftypefn
-@deftypefn {Target Hook} void TARGET_VECTORIZE_FINISH_COST (void *@var{data},
unsigned *@var{prologue_cost}, unsigned *@var{body_cost}, unsigned
*@var{epilogue_cost})
+@deftypefn {Target Hook} void TARGET_VECTORIZE_FINISH_COST (void *@var{data},
unsigned *@var{prologue_cost}, unsigned *@var{body_cost}, unsigned
*@var{epilogue_cost}, unsigned *@var{suggested_unroll_factor})
This hook should complete calculations of the cost of vectorizing a loop
or basic block based on @var{data}, and return the prologue, body, and
-epilogue costs as unsigned integers. The default returns the value of
-the three accumulators.
+epilogue costs as unsigned integers. It also asks the backend whether it
+has a @var{suggested_unroll_factor}. The default returns the value of
+the three cost accumulators.
@end deftypefn
@deftypefn {Target Hook} void TARGET_VECTORIZE_DESTROY_COST_DATA (void
*@var{data})
diff --git a/gcc/target.def b/gcc/target.def
index
bfa819609c21bd71c0cc585c01dba42534453f47..df0f170ff3378671e802d82a8bce8e153d8cf8fe
100644
--- a/gcc/target.def
+++ b/gcc/target.def
@@ -2078,11 +2078,12 @@ DEFHOOK
(finish_cost,
"This hook should complete calculations of the cost of vectorizing a loop\n\
or basic block based on @var{data}, and return the prologue, body, and\n\
-epilogue costs as unsigned integers. The default returns the value of\n\
-the three accumulators.",
+epilogue costs as unsigned integers. It also asks the backend whether it\n\
+has a @var{suggested_unroll_factor}. The default returns the value of\n\
+the three cost accumulators.",
void,
(void *data, unsigned *prologue_cost, unsigned *body_cost,
- unsigned *epilogue_cost),
+ unsigned *epilogue_cost, unsigned *suggested_unroll_factor),
default_finish_cost)
/* Function to delete target-specific cost modeling data. */
diff --git a/gcc/targhooks.h b/gcc/targhooks.h
index
92d51992e625c2497aa8496b1e2e3d916e5706fd..b9697c366876fe5a8c444ffcf58bdc6b5c33b0ad
100644
--- a/gcc/targhooks.h
+++ b/gcc/targhooks.h
@@ -123,7 +123,8 @@ extern unsigned default_add_stmt_cost (class vec_info *,
void *, int,
enum vect_cost_for_stmt,
class _stmt_vec_info *, tree, int,
enum vect_cost_model_location);
-extern void default_finish_cost (void *, unsigned *, unsigned *, unsigned *);
+extern void default_finish_cost (void *, unsigned *, unsigned *, unsigned *,
+ unsigned *);
extern void default_destroy_cost_data (void *);
/* OpenACC hooks. */
diff --git a/gcc/targhooks.c b/gcc/targhooks.c
index
c9b5208853dbc15706a65d1eb335e28e0564325e..8552d9a0f144e7bcee3f2653f2ea84ea677f80a2
100644
--- a/gcc/targhooks.c
+++ b/gcc/targhooks.c
@@ -1518,13 +1518,18 @@ default_add_stmt_cost (class vec_info *vinfo, void
*data, int count,
/* By default, the cost model just returns the accumulated costs. */
void
-default_finish_cost (void *data, unsigned *prologue_cost,
- unsigned *body_cost, unsigned *epilogue_cost)
+default_finish_cost (void *data,
+ unsigned *prologue_cost, unsigned *body_cost,
+ unsigned *epilogue_cost,
+ unsigned *suggested_unroll_factor)
{
unsigned *cost = (unsigned *) data;
*prologue_cost = cost[vect_prologue];
*body_cost = cost[vect_body];
*epilogue_cost = cost[vect_epilogue];
+ /* Do not unroll. */
+ if (suggested_unroll_factor)
+ *suggested_unroll_factor = 1;
}
/* Free the cost data. */
diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
index
5a5b8da2e771a1dd204f22a6447eba96bb3b352c..1bfe2e4f989143f4415c6c5b4a0b902ef1e00d66
100644
--- a/gcc/tree-vect-loop.c
+++ b/gcc/tree-vect-loop.c
@@ -153,7 +153,8 @@ along with GCC; see the file COPYING3. If not see
http://gcc.gnu.org/projects/tree-ssa/vectorization.html
*/
-static void vect_estimate_min_profitable_iters (loop_vec_info, int *, int *);
+static void vect_estimate_min_profitable_iters (loop_vec_info, int *, int *,
+ unsigned *);
static stmt_vec_info vect_is_simple_reduction (loop_vec_info, stmt_vec_info,
bool *, bool *);
@@ -277,7 +278,8 @@ vect_determine_vf_for_stmt (vec_info *vinfo,
*/
static opt_result
-vect_determine_vectorization_factor (loop_vec_info loop_vinfo)
+vect_determine_vectorization_factor (loop_vec_info loop_vinfo,
+ poly_uint64 min_vf)
{
class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
@@ -354,6 +356,28 @@ vect_determine_vectorization_factor (loop_vec_info
loop_vinfo)
}
}
+ /* Apply the suggested unrolling factor, this was determined by the backend
+ during finish_cost the first time we ran the analyzis for this
+ vector mode. */
+ if (loop_vinfo->suggested_unroll_factor > 1)
+ {
+ poly_uint64 unrolled_vf
+ = vectorization_factor * loop_vinfo->suggested_unroll_factor;
+ unsigned HOST_WIDE_INT max_vf = estimated_poly_value (unrolled_vf,
+ POLY_VALUE_MAX);
+ /* Make sure the unrolled vectorization factor fits the min and max
+ vectorization factor. */
+ if (max_vf <= LOOP_VINFO_MAX_VECT_FACTOR (loop_vinfo)
+ && known_le (unrolled_vf, min_vf))
+ vectorization_factor = unrolled_vf;
+ else if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "Unrolling failed due to unroll factor not fitting in"
+ " range of min and max vectorization factor:"
+ " [%d, %d]\n",
+ min_vf, max_vf);
+ }
+
/* TODO: Analyze cost. Decide if worth while to vectorize. */
if (dump_enabled_p ())
{
@@ -828,6 +852,7 @@ _loop_vec_info::_loop_vec_info (class loop *loop_in,
vec_info_shared *shared)
skip_main_loop_edge (nullptr),
skip_this_loop_edge (nullptr),
reusable_accumulators (),
+ suggested_unroll_factor (1),
max_vectorization_factor (0),
mask_skip_niters (NULL_TREE),
rgroup_compare_type (NULL_TREE),
@@ -1829,7 +1854,8 @@ vect_known_niters_smaller_than_vf (loop_vec_info
loop_vinfo)
definitely no, or -1 if it's worth retrying. */
static int
-vect_analyze_loop_costing (loop_vec_info loop_vinfo)
+vect_analyze_loop_costing (loop_vec_info loop_vinfo,
+ unsigned *suggested_unroll_factor)
{
class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
unsigned int assumed_vf = vect_vf_for_cost (loop_vinfo);
@@ -1863,7 +1889,8 @@ vect_analyze_loop_costing (loop_vec_info loop_vinfo)
int min_profitable_iters, min_profitable_estimate;
vect_estimate_min_profitable_iters (loop_vinfo, &min_profitable_iters,
- &min_profitable_estimate);
+ &min_profitable_estimate,
+ suggested_unroll_factor);
if (min_profitable_iters < 0)
{
@@ -2128,10 +2155,16 @@ vect_determine_partial_vectors_and_peeling
(loop_vec_info loop_vinfo,
vectors to the epilogue, with the main loop continuing to operate
on full vectors.
+ If we are unrolling we also do not want to use partial vectors. This
+ is to avoid the overhead of generating multiple masks and also to
+ avoid having to execute entire iterations of FALSE masked instructions
+ when dealing with one or less full iterations.
+
??? We could then end up failing to use partial vectors if we
decide to peel iterations into a prologue, and if the main loop
then ends up processing fewer than VF iterations. */
- if (param_vect_partial_vector_usage == 1
+ if ((param_vect_partial_vector_usage == 1
+ || loop_vinfo->suggested_unroll_factor > 1)
&& !LOOP_VINFO_EPILOGUE_P (loop_vinfo)
&& !vect_known_niters_smaller_than_vf (loop_vinfo))
LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P (loop_vinfo) = true;
@@ -2198,13 +2231,16 @@ vect_determine_partial_vectors_and_peeling
(loop_vec_info loop_vinfo,
for it. The different analyses will record information in the
loop_vec_info struct. */
static opt_result
-vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool &fatal, unsigned *n_stmts)
+vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool &fatal, unsigned *n_stmts,
+ unsigned *suggested_unroll_factor,
+ poly_uint64 min_vf = 2)
{
opt_result ok = opt_result::success ();
int res;
unsigned int max_vf = MAX_VECTORIZATION_FACTOR;
- poly_uint64 min_vf = 2;
loop_vec_info orig_loop_vinfo = NULL;
+ if (*suggested_unroll_factor > 1)
+ max_vf = estimated_poly_value (min_vf, POLY_VALUE_MAX);
/* If we are dealing with an epilogue then orig_loop_vinfo points to the
loop_vec_info of the first vectorized loop. */
@@ -2308,11 +2344,12 @@ vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool
&fatal, unsigned *n_stmts)
return ok;
}
if (max_vf != MAX_VECTORIZATION_FACTOR
- && maybe_lt (max_vf, min_vf))
+ && loop_vinfo->suggested_unroll_factor == 1
+ && max_vf < estimated_poly_value (min_vf, POLY_VALUE_MAX))
return opt_result::failure_at (vect_location, "bad data dependence.\n");
LOOP_VINFO_MAX_VECT_FACTOR (loop_vinfo) = max_vf;
- ok = vect_determine_vectorization_factor (loop_vinfo);
+ ok = vect_determine_vectorization_factor (loop_vinfo, min_vf);
if (!ok)
{
if (dump_enabled_p ())
@@ -2321,7 +2358,9 @@ vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool
&fatal, unsigned *n_stmts)
return ok;
}
if (max_vf != MAX_VECTORIZATION_FACTOR
- && maybe_lt (max_vf, LOOP_VINFO_VECT_FACTOR (loop_vinfo)))
+ && loop_vinfo->suggested_unroll_factor == 1
+ && max_vf < estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
+ POLY_VALUE_MAX))
return opt_result::failure_at (vect_location, "bad data dependence.\n");
/* Compute the scalar iteration cost. */
@@ -2547,7 +2586,7 @@ start_over:
return ok;
/* Check the costings of the loop make vectorizing worthwhile. */
- res = vect_analyze_loop_costing (loop_vinfo);
+ res = vect_analyze_loop_costing (loop_vinfo, suggested_unroll_factor);
if (res < 0)
{
ok = opt_result::failure_at (vect_location,
@@ -2879,6 +2918,122 @@ vect_joust_loop_vinfos (loop_vec_info new_loop_vinfo,
return true;
}
+/* Determine whether we can unroll this loop. */
+
+static bool
+vect_can_unroll (loop_vec_info loop_vinfo)
+{
+ stmt_vec_info stmt_info;
+ unsigned i;
+ poly_uint64 vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+
+ if (known_le (vectorization_factor, 1U))
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "will not unroll loop with a VF of 1 or less\n");
+ return false;
+ }
+
+ FOR_EACH_VEC_ELT (loop_vinfo->stmt_vec_infos, i, stmt_info)
+ {
+ if (STMT_VINFO_IN_PATTERN_P (stmt_info)
+ || !STMT_VINFO_RELEVANT_P (stmt_info)
+ || stmt_info->vectype == NULL_TREE)
+ continue;
+ /* Do not unroll loops with negative steps as it is unlikely that
+ vectorization will succeed due to the way we deal with negative steps
+ in loads and stores in 'get_load_store_type'. */
+ if (stmt_info->dr_aux.dr
+ && !STMT_VINFO_GATHER_SCATTER_P (stmt_info))
+ {
+ dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info);
+ tree step = vect_dr_behavior (loop_vinfo, dr_info)->step;
+ if (TREE_CODE (step) == INTEGER_CST
+ && tree_int_cst_compare (step, size_zero_node) < 0)
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "could not unroll due to negative step\n");
+ return false;
+ }
+ }
+
+ if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
+ {
+ auto red_info = info_for_reduction (loop_vinfo, stmt_info);
+ if (STMT_VINFO_REDUC_TYPE (red_info) != TREE_CODE_REDUCTION)
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "could not unroll loop with reduction due to "
+ "non TREE_CODE_REDUCTION\n");
+ return false;
+ }
+ }
+ }
+
+ return true;
+}
+
+
+/* Try to unroll the current loop. First determine the unrolling factor using
+ the analysis done for the current vector mode. Then re-analyze the loop for
+ the given unrolling factor and the current vector mode. */
+
+static opt_loop_vec_info
+vect_try_unrolling (loop_vec_info loop_vinfo, unsigned *n_stmts,
+ unsigned suggested_unroll_factor)
+{
+ DUMP_VECT_SCOPE ("vect_try_unrolling");
+
+ if (suggested_unroll_factor == 1)
+ return opt_loop_vec_info::failure_at (vect_location,
+ "*** Target determined unrolling is"
+ " not profitable.\n");
+
+ if (!vect_can_unroll (loop_vinfo))
+ return opt_loop_vec_info::failure_at (vect_location,
+ "*** Can not unroll this loop.\n");
+
+ loop_vec_info unrolled_vinfo
+ = opt_loop_vec_info::success (vect_analyze_loop_form (loop_vinfo->loop,
+ loop_vinfo->shared));
+ unrolled_vinfo->vector_mode = loop_vinfo->vector_mode;
+
+ /* Use the suggested_unrolling_factor that was returned at the target's
+ TARGET_VECTORIZE_FINISH_COST hook. */
+ unrolled_vinfo->suggested_unroll_factor = suggested_unroll_factor;
+ poly_uint64 unrolled_vf
+ = LOOP_VINFO_VECT_FACTOR (loop_vinfo) * suggested_unroll_factor;
+
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "***** unrolling factor %d chosen for vector mode %s,"
+ "re-trying analyzis...\n",
+ suggested_unroll_factor,
+ GET_MODE_NAME (unrolled_vinfo->vector_mode));
+ bool unrolling_fatal = false;
+ if (vect_analyze_loop_2 (unrolled_vinfo, unrolling_fatal, n_stmts,
+ &suggested_unroll_factor,
+ unrolled_vf)
+ && known_ne (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
+ LOOP_VINFO_VECT_FACTOR (unrolled_vinfo)))
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "unrolling succeeded with factor = %d\n",
+ suggested_unroll_factor);
+ unrolled_vinfo->loop->aux = NULL;
+ return opt_loop_vec_info::success (unrolled_vinfo);
+ }
+
+ loop_vinfo->loop->aux = NULL;
+ return opt_loop_vec_info::failure_at (vect_location,
+ "unrolling failed with factor = %d\n",
+ suggested_unroll_factor);
+}
+
/* If LOOP_VINFO is already a main loop, return it unmodified. Otherwise
try to reanalyze it as a main loop. Return the loop_vinfo on success
and null on failure. */
@@ -2902,8 +3057,16 @@ vect_reanalyze_as_main_loop (loop_vec_info loop_vinfo,
unsigned int *n_stmts)
main_loop_vinfo->vector_mode = loop_vinfo->vector_mode;
bool fatal = false;
- bool res = vect_analyze_loop_2 (main_loop_vinfo, fatal, n_stmts);
+ unsigned suggested_unroll_factor = 1;
+ bool res = vect_analyze_loop_2 (main_loop_vinfo, fatal, n_stmts,
+ &suggested_unroll_factor);
loop->aux = NULL;
+ opt_loop_vec_info unrolled_vinfo
+ = opt_loop_vec_info::success (vect_try_unrolling (main_loop_vinfo, n_stmts,
+ suggested_unroll_factor));
+ if (unrolled_vinfo)
+ main_loop_vinfo = unrolled_vinfo;
+
if (!res)
{
if (dump_enabled_p ())
@@ -2960,6 +3123,7 @@ vect_analyze_loop (class loop *loop, vec_info_shared
*shared)
bool vect_epilogues = false;
opt_result res = opt_result::success ();
unsigned HOST_WIDE_INT simdlen = loop->simdlen;
+ unsigned suggested_unroll_factor = 1;
while (1)
{
/* Check the CFG characteristics of the loop (nesting, entry/exit). */
@@ -3007,7 +3171,8 @@ vect_analyze_loop (class loop *loop, vec_info_shared
*shared)
if (vect_epilogues)
LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo) = first_loop_vinfo;
- res = vect_analyze_loop_2 (loop_vinfo, fatal, &n_stmts);
+ res = vect_analyze_loop_2 (loop_vinfo, fatal, &n_stmts,
+ &suggested_unroll_factor);
if (mode_i == 0)
autodetected_vector_mode = loop_vinfo->vector_mode;
if (dump_enabled_p ())
@@ -3038,6 +3203,18 @@ vect_analyze_loop (class loop *loop, vec_info_shared
*shared)
if (res)
{
+ /* Only try unrolling main loops. */
+ if (!LOOP_VINFO_EPILOGUE_P (loop_vinfo))
+ {
+ opt_loop_vec_info unrolled_vinfo =
+ vect_try_unrolling (loop_vinfo, &n_stmts,
+ suggested_unroll_factor);
+ if (unrolled_vinfo)
+ loop_vinfo = unrolled_vinfo;
+ /* Reset suggested_unroll_factor for next loop_vinfo. */
+ suggested_unroll_factor = 1;
+ }
+
LOOP_VINFO_VECTORIZABLE_P (loop_vinfo) = 1;
vectorized_loops++;
@@ -3056,13 +3233,26 @@ vect_analyze_loop (class loop *loop, vec_info_shared
*shared)
/* Keep trying to roll back vectorization attempts while the
loop_vec_infos they produced were worse than this one. */
vec<loop_vec_info> &vinfos = first_loop_vinfo->epilogue_vinfos;
+ poly_uint64 vinfo_vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+ poly_uint64 first_vinfo_vf
+ = LOOP_VINFO_VECT_FACTOR (first_loop_vinfo);
while (!vinfos.is_empty ()
+ && (known_lt (vinfo_vf, first_vinfo_vf)
+ || (LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)
+ && maybe_eq (vinfo_vf, first_vinfo_vf)))
&& vect_joust_loop_vinfos (loop_vinfo, vinfos.last ()))
{
gcc_assert (vect_epilogues);
delete vinfos.pop ();
}
+ /* Check if we may want to replace the current first_loop_vinfo
+ with the new loop, but only if they have different vector
+ modes. If they have the same vector mode this means the main
+ loop is an unrolled loop and we are trying to vectorize the
+ epilogue using the same vector mode but with a lower
+ vectorization factor. */
if (vinfos.is_empty ()
+ && loop_vinfo->vector_mode != first_loop_vinfo->vector_mode
&& vect_joust_loop_vinfos (loop_vinfo, first_loop_vinfo))
{
loop_vec_info main_loop_vinfo
@@ -3105,14 +3295,34 @@ vect_analyze_loop (class loop *loop, vec_info_shared
*shared)
/* For now only allow one epilogue loop. */
&& first_loop_vinfo->epilogue_vinfos.is_empty ())
{
- first_loop_vinfo->epilogue_vinfos.safe_push (loop_vinfo);
- poly_uint64 th = LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo);
- gcc_assert (!LOOP_REQUIRES_VERSIONING (loop_vinfo)
- || maybe_ne (lowest_th, 0U));
- /* Keep track of the known smallest versioning
- threshold. */
- if (ordered_p (lowest_th, th))
- lowest_th = ordered_min (lowest_th, th);
+ /* Ensure the epilogue has a smaller VF than the main loop or
+ uses predication and has the same VF. */
+ if (known_lt (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
+ LOOP_VINFO_VECT_FACTOR (first_loop_vinfo))
+ || (LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)
+ && maybe_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
+ LOOP_VINFO_VECT_FACTOR (first_loop_vinfo))))
+ {
+ first_loop_vinfo->epilogue_vinfos.safe_push (loop_vinfo);
+ poly_uint64 th = LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo);
+ gcc_assert (!LOOP_REQUIRES_VERSIONING (loop_vinfo)
+ || maybe_ne (lowest_th, 0U));
+ /* Keep track of the known smallest versioning
+ threshold. */
+ if (ordered_p (lowest_th, th))
+ lowest_th = ordered_min (lowest_th, th);
+ }
+ else
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "***** Will not use %s mode as an"
+ " epilogue, since it leads to an higher"
+ " vectorization factor than main loop\n",
+ GET_MODE_NAME (loop_vinfo->vector_mode));
+ delete loop_vinfo;
+ loop_vinfo = opt_loop_vec_info::success (NULL);
+ }
}
else
{
@@ -3153,13 +3363,32 @@ vect_analyze_loop (class loop *loop, vec_info_shared
*shared)
/* Handle the case that the original loop can use partial
vectorization, but want to only adopt it for the epilogue.
- The retry should be in the same mode as original. */
+ The retry should be in the same mode as original.
+ Also handle the case where we have unrolled the main loop and want to
+ retry all vector modes again for the epilogues, since the VF is now
+ at least twice as high as the current vector mode. */
if (vect_epilogues
&& loop_vinfo
- && LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P (loop_vinfo))
+ && (LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P (loop_vinfo)
+ || loop_vinfo->suggested_unroll_factor > 1))
{
- gcc_assert (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
+ gcc_assert ((LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
+ || loop_vinfo->suggested_unroll_factor > 1)
&& !LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo));
+ /* If we are unrolling, try all VECTOR_MODES for the epilogue. */
+ if (loop_vinfo->suggested_unroll_factor > 1)
+ {
+ next_vector_mode = vector_modes[0];
+ mode_i = 1;
+
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "***** Re-trying analysis with vector mode"
+ " %s for epilogues after unrolling.\n",
+ GET_MODE_NAME (next_vector_mode));
+ continue;
+ }
+
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"***** Re-trying analysis with same vector mode"
@@ -3862,7 +4091,8 @@ vect_get_known_peeling_cost (loop_vec_info loop_vinfo,
int peel_iters_prologue,
static void
vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo,
int *ret_min_profitable_niters,
- int *ret_min_profitable_estimate)
+ int *ret_min_profitable_estimate,
+ unsigned *suggested_unroll_factor)
{
int min_profitable_iters;
int min_profitable_estimate;
@@ -4222,8 +4452,9 @@ vect_estimate_min_profitable_iters (loop_vec_info
loop_vinfo,
}
/* Complete the target-specific cost calculations. */
- finish_cost (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo), &vec_prologue_cost,
- &vec_inside_cost, &vec_epilogue_cost);
+ finish_cost (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo),
+ &vec_prologue_cost, &vec_inside_cost, &vec_epilogue_cost,
+ suggested_unroll_factor);
vec_outside_cost = (int)(vec_prologue_cost + vec_epilogue_cost);
@@ -7212,7 +7443,8 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
participating. */
if (ncopies > 1
&& (STMT_VINFO_RELEVANT (stmt_info) <= vect_used_only_live)
- && reduc_chain_length == 1)
+ && reduc_chain_length == 1
+ && loop_vinfo->suggested_unroll_factor == 1)
single_defuse_cycle = true;
if (single_defuse_cycle || lane_reduc_code_p)
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index
024a1c38a2342246d7891db1de5f1d6e6458d5dd..a8a6c6a19ed4c98144f9097467c59386fdbe8233
100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -5418,8 +5418,8 @@ vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo,
}
while (vi < li_vector_costs.length ()
&& li_vector_costs[vi].first == vl);
- finish_cost (vect_target_cost_data, &vec_prologue_cost,
- &vec_inside_cost, &vec_epilogue_cost);
+ finish_cost (vect_target_cost_data, &vec_prologue_cost, &vec_inside_cost,
+ &vec_epilogue_cost);
destroy_cost_data (vect_target_cost_data);
vec_outside_cost = vec_prologue_cost + vec_epilogue_cost;
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index
c4c5678e7f1abafc25c465319dbacf3ef50f0ae9..8b182cd34e7d6a8d9e55a9c1003900b8216a952f
100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -621,6 +621,13 @@ public:
about the reductions that generated them. */
hash_map<tree, vect_reusable_accumulator> reusable_accumulators;
+ /* The number of times that the target suggested we unroll the vector loop
+ in order to promote more ILP. This value will be used to re-analyze the
+ loop for vectorization and if successful the value will be folded into
+ vectorization_factor (and therefore exactly divides
+ vectorization_factor). */
+ unsigned int suggested_unroll_factor;
+
/* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
if there is no particular limit. */
unsigned HOST_WIDE_INT max_vectorization_factor;
@@ -1571,9 +1578,11 @@ add_stmt_cost (vec_info *vinfo, void *data,
stmt_info_for_cost *i)
static inline void
finish_cost (void *data, unsigned *prologue_cost,
- unsigned *body_cost, unsigned *epilogue_cost)
+ unsigned *body_cost, unsigned *epilogue_cost,
+ unsigned *suggested_unroll_factor = NULL)
{
- targetm.vectorize.finish_cost (data, prologue_cost, body_cost,
epilogue_cost);
+ targetm.vectorize.finish_cost (data, prologue_cost, body_cost, epilogue_cost,
+ suggested_unroll_factor);
}
/* Alias targetm.vectorize.destroy_cost_data. */