[vec-tails] Support loop epilogue vectorization

On Mon, 28 Nov 2016, Yuri Rumyantsev wrote:

> Richard!
> 
> I attached vect dump for hte part of attached test-case which
> illustrated how vectorization of epilogues works through masking:
> #define SIZE 1023
> #define ALIGN 64
> 
> extern int posix_memalign(void **memptr, __SIZE_TYPE__ alignment,
> __SIZE_TYPE__ size) __attribute__((weak));
> extern void free (void *);
> 
> void __attribute__((noinline))
> test_citer (int * __restrict__ a,
>    int * __restrict__ b,
>    int * __restrict__ c)
> {
>   int i;
> 
>   a = (int *)__builtin_assume_aligned (a, ALIGN);
>   b = (int *)__builtin_assume_aligned (b, ALIGN);
>   c = (int *)__builtin_assume_aligned (c, ALIGN);
> 
>   for (i = 0; i < SIZE; i++)
>     c[i] = a[i] + b[i];
> }
> 
> It was compiled with -mavx2 --param vect-epilogues-mask=1 options.
> 
> I did not include in this patch vectorization of low trip-count loops
> since in the original patch additional parameter was introduced:
> +DEFPARAM (PARAM_VECT_SHORT_LOOPS,
> +  "vect-short-loops",
> +  "Enable vectorization of low trip count loops using masking.",
> +  0, 0, 1)
> 
> I assume that this ability can be included very quickly but it
> requires cost model enhancements also.

Comments on the patch itself (as I'm having a closer look again,
I know how it vectorizes the above but I wondered why epilogue
and short-trip loops are not basically the same code path).

Btw, I don't like that the features are behind a --param paywall.
That just means a) nobody will use it, b) it will bit-rot quickly,
c) bugs are well-hidden.

+  if (loop_vinfo && LOOP_VINFO_CAN_BE_MASKED (loop_vinfo)
+      && integer_zerop (nested_in_vect_loop
+                       ? STMT_VINFO_DR_STEP (stmt_info)
+                       : DR_STEP (dr)))
+    {
+      if (dump_enabled_p ())
+       dump_printf_loc (MSG_NOTE, vect_location,
+                        "allow invariant load for masked loop.\n");
+    }

this can test memory_access_type == VMAT_INVARIANT.  Please put
all the checks in a common

  if (loop_vinfo && LOOP_VINFO_CAN_BE_MASKED (loop_vinfo))
    {
       if (memory_access_type == VMAT_INVARIANT)
         {
         }
       else if (...)
         {
            LOOP_VINFO_CAN_BE_MASKED (loop_vinfo) = false;
         }
       else if (..)
...
    }

@@ -6667,6 +6756,15 @@ vectorizable_load (gimple *stmt, 
gimple_stmt_iterator *gsi, gimple **vec_stmt,
       gcc_assert (!nested_in_vect_loop);
       gcc_assert (!STMT_VINFO_GATHER_SCATTER_P (stmt_info));

+      if (loop_vinfo && LOOP_VINFO_CAN_BE_MASKED (loop_vinfo))
+       {
+         if (dump_enabled_p ())
+           dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                            "cannot be masked: grouped access is not"
+                            " supported.");
+         LOOP_VINFO_CAN_BE_MASKED (loop_vinfo) = false;
+      }
+

isn't this already handled by the above?  Or rather the general
disallowance of SLP?

@@ -5730,6 +5792,24 @@ vectorizable_store (gimple *stmt, 
gimple_stmt_iterator *gsi, gimple **vec_stmt,
                            &memory_access_type, &gs_info))
     return false;

+  if (loop_vinfo && LOOP_VINFO_CAN_BE_MASKED (loop_vinfo)
+      && memory_access_type != VMAT_CONTIGUOUS)
+    {
+      LOOP_VINFO_CAN_BE_MASKED (loop_vinfo) = false;
+      if (dump_enabled_p ())
+       dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                        "cannot be masked: unsupported memory access 
type.\n");
+    }
+
+  if (loop_vinfo && LOOP_VINFO_CAN_BE_MASKED (loop_vinfo)
+      && !can_mask_load_store (stmt))
+    {
+      LOOP_VINFO_CAN_BE_MASKED (loop_vinfo) = false;
+      if (dump_enabled_p ())
+       dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                        "cannot be masked: unsupported masked store.\n");
+    }
+

likewise please combine the ifs.

@@ -2354,7 +2401,10 @@ vectorizable_mask_load_store (gimple *stmt, 
gimple_stmt_iterator *gsi,
                                          ptr, vec_mask, vec_rhs);
          vect_finish_stmt_generation (stmt, new_stmt, gsi);
          if (i == 0)
-           STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
+           {
+             STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
+             STMT_VINFO_FIRST_COPY_P (vinfo_for_stmt (new_stmt)) = true;
+           }
          else
            STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
          prev_stmt_info = vinfo_for_stmt (new_stmt);

here you only set the flag, elsewhere you copy DR and VECTYPE as well.

@@ -2113,6 +2146,20 @@ vectorizable_mask_load_store (gimple *stmt, 
gimple_stmt_iterator *gsi,
               && !useless_type_conversion_p (vectype, rhs_vectype)))
     return false;

+  if (LOOP_VINFO_CAN_BE_MASKED (loop_vinfo))
+    {
+      /* Check that mask conjuction is supported.  */
+      optab tab;
+      tab = optab_for_tree_code (BIT_AND_EXPR, vectype, optab_default);
+      if (!tab || optab_handler (tab, TYPE_MODE (vectype)) == 
CODE_FOR_nothing)
+       {
+         if (dump_enabled_p ())
+           dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                            "cannot be masked: unsupported mask 
operation\n");
+         LOOP_VINFO_CAN_BE_MASKED (loop_vinfo) = false;
+       }
+    }

does this really test whether we can bit-and the mask?  You are
using the vector type of the store (which might be V2DF for example),
also for AVX512 it might be a vector-bool type with integer mode?
Of course we maybe can simply assume mask conjunction is available
(I know no ISA where that would be not true).

+/* Return true if STMT can be converted to masked form.  */
+
+static bool
+can_mask_load_store (gimple *stmt)
+{
+  stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
+  tree vectype, mask_vectype;
+  tree lhs, ref;
+
+  if (!stmt_info)
+    return false;
+  lhs = gimple_assign_lhs (stmt);
+  ref = (TREE_CODE (lhs) == SSA_NAME) ? gimple_assign_rhs1 (stmt) : lhs;
+  if (may_be_nonaddressable_p (ref))
+    return false;
+  vectype = STMT_VINFO_VECTYPE (stmt_info);

You probably modeled this after ifcvt_can_use_mask_load_store but I
don't think checking may_be_nonaddressable_p is necessary (we couldn't
even vectorize such refs).  stmt_info should never be NULL either.
With the check removed tree-ssa-loop-ivopts.h should no longer be
necessary.

+static void
+vect_mask_load_store_stmt (gimple *stmt, tree vectype, tree mask,
+                          data_reference *dr, gimple_stmt_iterator *si)
+{
...
+  addr = force_gimple_operand_gsi (&gsi, build_fold_addr_expr (mem),
+                                  true, NULL_TREE, true,
+                                  GSI_SAME_STMT);
+
+  align = TYPE_ALIGN_UNIT (vectype);
+  if (aligned_access_p (dr))
+    misalign = 0;
+  else if (DR_MISALIGNMENT (dr) == -1)
+    {
+      align = TYPE_ALIGN_UNIT (elem_type);
+      misalign = 0;
+    }
+  else
+    misalign = DR_MISALIGNMENT (dr);
+  set_ptr_info_alignment (get_ptr_info (addr), align, misalign);
+  ptr = build_int_cst (reference_alias_ptr_type (mem),
+                      misalign ? misalign & -misalign : align);

you should simply use

  align = get_object_alignment (mem) / BITS_PER_UNIT;

here rather than trying to be clever.  Eventually you don't need
the DR then (see question above).

+    }
+  gsi_replace (si ? si : &gsi, new_stmt, false);

when you replace the load/store please previously copy VUSE and VDEF
from the original one (we were nearly clean enough to no longer
require a virtual operand rewrite after vectorization...)  Thus

  gimple_set_vuse (new_stmt, gimple_vuse (stmt));
  gimple_set_vdef (new_stmt, gimple_vdef (stmt));

+static void
+vect_mask_loop (loop_vec_info loop_vinfo)
+{
...
+  /* Scan all loop statements to convert vector load/store including 
masked
+     form.  */
+  for (unsigned i = 0; i < loop->num_nodes; i++)
+    {
+      basic_block bb = bbs[i];
+      for (gimple_stmt_iterator si = gsi_start_bb (bb);
+          !gsi_end_p (si); gsi_next (&si))
+       {
+         gimple *stmt = gsi_stmt (si);
+         stmt_vec_info stmt_info = NULL;
+         tree vectype = NULL;
+         data_reference *dr;
+
+         /* Mask load case.  */
+         if (is_gimple_call (stmt)
+             && gimple_call_internal_p (stmt)
+             && gimple_call_internal_fn (stmt) == IFN_MASK_LOAD
+             && !VECTOR_TYPE_P (TREE_TYPE (gimple_call_arg (stmt, 2))))
+           {
...
+             /* Skip invariant loads.  */
+             if (integer_zerop (nested_in_vect_loop_p (loop, stmt)
+                                ? STMT_VINFO_DR_STEP (stmt_info)
+                                : DR_STEP (STMT_VINFO_DATA_REF 
(stmt_info))))
+               continue;

seeing this it would be nice if stmt_info had a flag for whether
the stmt needs masking (and a flag on wheter this is a scalar or a
vectorized stmt).

+         /* Skip hoisted out statements.  */
+         if (!flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
+           continue;

err, you walk stmts in the loop!  Isn't this covered by the above
skipping of 'invariant loads'?

+static gimple *
+vect_mask_reduction_stmt (gimple *stmt, tree mask, gimple *prev)
+{

depending on the reduction operand there are variants that
could get away w/o the VEC_COND_EXPR, like

  S1': tem_4 = d_3 & MASK;
  S2': r_1 = r_2 + tem_4;

which works for plus at least.  More generally doing

  S1': tem_4 = VEC_COND_EXPR<MASK, d_3, neutral operand>
  S2': r_1 = r_2 OP tem_4;

and leaving optimization to & to later opts (& won't work for
AVX512 mask registers I guess).

Good enough for later enhacement of course.

+static void
+vect_gen_ivs_for_masking (loop_vec_info loop_vinfo, vec<tree> *ivs)
+{
...

isn't it enough to always create a single IV and derive the
additional copies by IV + i * { elems, elems, elems ... }?
IVs are expensive -- I'm sure we can optimize the rest of the
scheme further as well but this one looks obvious to me.

@@ -3225,12 +3508,32 @@ vect_estimate_min_profitable_iters (loop_vec_info 
loop_vinfo,
   int npeel = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo);
   void *target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo);

+  if (LOOP_VINFO_EPILOGUE_P (loop_vinfo))
+    {
+      /* Currently we don't produce scalar epilogue version in case
+        its masked version is provided.  It means we don't need to
+        compute profitability one more time here.  Just make a
+        masked loop version.  */
+      if (LOOP_VINFO_CAN_BE_MASKED (loop_vinfo)
+         && PARAM_VALUE (PARAM_VECT_EPILOGUES_MASK))
+       {
+         dump_printf_loc (MSG_NOTE, vect_location,
+                          "cost model: mask loop epilogue.\n");
+         LOOP_VINFO_MASK_LOOP (loop_vinfo) = true;
+         *ret_min_profitable_niters = 0;
+         *ret_min_profitable_estimate = 0;
+         return;
+       }
+    }
   /* Cost model disabled.  */
-  if (unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo)))
+  else if (unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo)))
     {
       dump_printf_loc (MSG_NOTE, vect_location, "cost model 
disabled.\n");
       *ret_min_profitable_niters = 0;
       *ret_min_profitable_estimate = 0;
+      if (PARAM_VALUE (PARAM_VECT_EPILOGUES_MASK)
+         && LOOP_VINFO_CAN_BE_MASKED (loop_vinfo))
+       LOOP_VINFO_MASK_LOOP (loop_vinfo) = true;
       return;
     }

the unlimited_cost_model case should come first?  OTOH masking or
not is probably not sth covered by 'unlimited' - that is about
vectorizing or not.  But the above code means that for
epilogue vectorization w/o masking we ignore unlimited_cost_model ()?
That doesn't make sense to me.

Plus if this is short-trip or epilogue vectorization and the
cost model is _not_ unlimited then we dont' want to enable
masking always (if it is possible).  It might be we statically
know the epilogue executes for at most two iterations for example.

I don't see _any_ cost model for vectorizing the epilogue with
masking?  Am I missing something?  A "trivial" cost model
should at least consider the additional IV(s), the mask
compute and the widening and narrowing ops required.


I think the prolog_peeling was fixed during the epilogue vectorization 
review and should no longer be necessary.  Please add
a && ! LOOP_VINFO_MASK_LOOP () to the epilog_peeling init instead
(it should also work for short-trip loop vectorization).

@@ -2022,11 +2291,18 @@ start_over:
       || (max_niter != -1
          && (unsigned HOST_WIDE_INT) max_niter < vectorization_factor))
     {
-      if (dump_enabled_p ())
-       dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
-                        "not vectorized: iteration count smaller than "
-                        "vectorization factor.\n");
-      return false;
+      /* Allow low trip count for loop epilogue we want to mask.  */
+      if (LOOP_VINFO_EPILOGUE_P (loop_vinfo)
+         && PARAM_VALUE (PARAM_VECT_EPILOGUES_MASK))
+       LOOP_VINFO_MASK_LOOP (loop_vinfo) = true;
+      else
+       {
+         if (dump_enabled_p ())

so why do we test only LOOP_VINFO_EPILOGUE_P here?  All the code
I saw sofar would also work for the main loop (but the cost
model is missing).

I am missing testcases.  There's only a single one but we should
have cases covering all kinds of mask IV widths and widen/shorten
masks.

Do you have any numbers on SPEC 2k6 with epilogue vect and/or masking
enabled for an AVX2 machine?

Oh, and I really dislike the --param paywall.

Thanks,
Richard.

> Best regards.
> Yuri.
> 
> 
> 2016-11-28 17:39 GMT+03:00 Richard Biener <rguenther@suse.de>:
> > On Thu, 24 Nov 2016, Yuri Rumyantsev wrote:
> >
> >> Hi All,
> >>
> >> Here is the second patch which supports epilogue vectorization using
> >> masking without cost model. Currently it is possible
> >> only with passing parameter "--param vect-epilogues-mask=1".
> >>
> >> Bootstrapping and regression testing did not show any new regression.
> >>
> >> Any comments will be appreciated.
> >
> > Going over the patch the main question is one how it works -- it looks
> > like the decision whether to vectorize & mask the epilogue is made
> > when vectorizing the loop that generates the epilogue rather than
> > in the epilogue vectorization path?
> >
> > That is, I'd have expected to see this handling low-trip count loops
> > by masking?  And thus masking the epilogue simply by it being
> > low-trip count?
> >
> > Richard.
> >
> >> ChangeLog:
> >> 2016-11-24  Yuri Rumyantsev  <ysrumyan@gmail.com>
> >>
> >> * params.def (PARAM_VECT_EPILOGUES_MASK): New.
> >> * tree-vect-data-refs.c (vect_get_new_ssa_name): Support vect_mask_var.
> >> * tree-vect-loop.c: Include insn-config.h, recog.h and alias.h.
> >> (new_loop_vec_info): Add zeroing can_be_masked, mask_loop and
> >> required_mask fields.
> >> (vect_check_required_masks_widening): New.
> >> (vect_check_required_masks_narrowing): New.
> >> (vect_get_masking_iv_elems): New.
> >> (vect_get_masking_iv_type): New.
> >> (vect_get_extreme_masks): New.
> >> (vect_check_required_masks): New.
> >> (vect_analyze_loop_operations): Call vect_check_required_masks if all
> >> statements can be masked.
> >> (vect_analyze_loop_2): Inititalize to zero min_scalar_loop_bound.
> >> Add check that epilogue can be masked with the same vf with issue
> >> fail notes.  Allow epilogue vectorization through masking of low trip
> >> loops. Set to true can_be_masked field before loop operation analysis.
> >> Do not set-up min_scalar_loop_bound for epilogue vectorization through
> >> masking.  Do not peeling for epilogue masking.  Reset can_be_masked
> >> field before repeat analysis.
> >> (vect_estimate_min_profitable_iters): Do not compute profitability
> >> for epilogue masking.  Set up mask_loop filed to true if parameter
> >> PARAM_VECT_EPILOGUES_MASK is non-zero.
> >> (vectorizable_reduction): Add check that statement can be masked.
> >> (vectorizable_induction): Do not support masking for induction.
> >> (vect_gen_ivs_for_masking): New.
> >> (vect_get_mask_index_for_elems): New.
> >> (vect_get_mask_index_for_type): New.
> >> (vect_create_narrowed_masks): New.
> >> (vect_create_widened_masks): New.
> >> (vect_gen_loop_masks): New.
> >> (vect_mask_reduction_stmt): New.
> >> (vect_mask_mask_load_store_stmt): New.
> >> (vect_mask_load_store_stmt): New.
> >> (vect_mask_loop): New.
> >> (vect_transform_loop): Invoke vect_mask_loop if required.
> >> Use div_ceil to recompute upper bounds for masked loops.  Issue
> >> statistics for epilogue vectorization through masking. Do not reduce
> >> vf for masking epilogue.
> >> * tree-vect-stmts.c: Include tree-ssa-loop-ivopts.h.
> >> (can_mask_load_store): New.
> >> (vectorizable_mask_load_store): Check that mask conjuction is
> >> supported.  Set-up first_copy_p field of stmt_vinfo.
> >> (vectorizable_simd_clone_call): Check that simd clone can not be
> >> masked.
> >> (vectorizable_store): Check that store can be masked. Mark the first
> >> copy of generated vector stores and provide it with vectype and the
> >> original data reference.
> >> (vectorizable_load): Check that load can be masked.
> >> (vect_stmt_should_be_masked_for_epilogue): New.
> >> (vect_add_required_mask_for_stmt): New.
> >> (vect_analyze_stmt): Add check on unsupported statements for masking
> >> with printing message.
> >> * tree-vectorizer.h (struct _loop_vec_info): Add new fields
> >> can_be_maske, required_masks, masl_loop.
> >> (LOOP_VINFO_CAN_BE_MASKED): New.
> >> (LOOP_VINFO_REQUIRED_MASKS): New.
> >> (LOOP_VINFO_MASK_LOOP): New.
> >> (struct _stmt_vec_info): Add first_copy_p field.
> >> (STMT_VINFO_FIRST_COPY_P): New.
> >>
> >> gcc/testsuite/
> >>
> >> * gcc.dg/vect/vect-tail-mask-1.c: New test.
> >>
> >> 2016-11-18 18:54 GMT+03:00 Christophe Lyon <christophe.lyon@linaro.org>:
> >> > On 18 November 2016 at 16:46, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
> >> >> It is very strange that this test failed on arm, since it requires
> >> >> target avx2 to check vectorizer dumps:
> >> >>
> >> >> /* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" {
> >> >> target avx2_runtime } } } */
> >> >> /* { dg-final { scan-tree-dump-times "LOOP EPILOGUE VECTORIZED
> >> >> \\(VS=16\\)" 2 "vect" { target avx2_runtime } } } */
> >> >>
> >> >> Could you please clarify what is the reason of the failure?
> >> >
> >> > It's not the scan-dumps that fail, but the execution.
> >> > The test calls abort() for some reason.
> >> >
> >> > It will take me a while to rebuild the test manually in the right
> >> > debug environment to provide you with more traces.
> >> >
> >> >
> >> >
> >> >>
> >> >> Thanks.
> >> >>
> >> >> 2016-11-18 16:20 GMT+03:00 Christophe Lyon <christophe.lyon@linaro.org>:
> >> >>> On 15 November 2016 at 15:41, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
> >> >>>> Hi All,
> >> >>>>
> >> >>>> Here is patch for non-masked epilogue vectoriziation.
> >> >>>>
> >> >>>> Bootstrap and regression testing did not show any new failures.
> >> >>>>
> >> >>>> Is it OK for trunk?
> >> >>>>
> >> >>>> Thanks.
> >> >>>> Changelog:
> >> >>>>
> >> >>>> 2016-11-15  Yuri Rumyantsev  <ysrumyan@gmail.com>
> >> >>>>
> >> >>>> * params.def (PARAM_VECT_EPILOGUES_NOMASK): New.
> >> >>>> * tree-if-conv.c (tree_if_conversion): Make public.
> >> >>>> * * tree-if-conv.h: New file.
> >> >>>> * tree-vect-data-refs.c (vect_analyze_data_ref_dependences) Avoid
> >> >>>> dynamic alias checks for epilogues.
> >> >>>> * tree-vect-loop-manip.c (vect_do_peeling): Return created epilog.
> >> >>>> * tree-vect-loop.c: include tree-if-conv.h.
> >> >>>> (new_loop_vec_info): Add zeroing orig_loop_info field.
> >> >>>> (vect_analyze_loop_2): Don't try to enhance alignment for epilogues.
> >> >>>> (vect_analyze_loop): Add argument ORIG_LOOP_INFO which is not NULL
> >> >>>> if epilogue is vectorized, set up orig_loop_info field of loop_vinfo
> >> >>>> using passed argument.
> >> >>>> (vect_transform_loop): Check if created epilogue should be returned
> >> >>>> for further vectorization with less vf.  If-convert epilogue if
> >> >>>> required. Print vectorization success for epilogue.
> >> >>>> * tree-vectorizer.c (vectorize_loops): Add epilogue vectorization
> >> >>>> if it is required, pass loop_vinfo produced during vectorization of
> >> >>>> loop body to vect_analyze_loop.
> >> >>>> * tree-vectorizer.h (struct _loop_vec_info): Add new field
> >> >>>> orig_loop_info.
> >> >>>> (LOOP_VINFO_ORIG_LOOP_INFO): New.
> >> >>>> (LOOP_VINFO_EPILOGUE_P): New.
> >> >>>> (LOOP_VINFO_ORIG_VECT_FACTOR): New.
> >> >>>> (vect_do_peeling): Change prototype to return epilogue.
> >> >>>> (vect_analyze_loop): Add argument of loop_vec_info type.
> >> >>>> (vect_transform_loop): Return created loop.
> >> >>>>
> >> >>>> gcc/testsuite/
> >> >>>>
> >> >>>> * lib/target-supports.exp (check_avx2_hw_available): New.
> >> >>>> (check_effective_target_avx2_runtime): New.
> >> >>>> * gcc.dg/vect/vect-tail-nomask-1.c: New test.
> >> >>>>
> >> >>>
> >> >>> Hi,
> >> >>>
> >> >>> This new test fails on arm-none-eabi (using default cpu/fpu/mode):
> >> >>>   gcc.dg/vect/vect-tail-nomask-1.c -flto -ffat-lto-objects execution test
> >> >>>   gcc.dg/vect/vect-tail-nomask-1.c execution test
> >> >>>
> >> >>> It does pass on the same target if configured --with-cpu=cortex-a9.
> >> >>>
> >> >>> Christophe
> >> >>>
> >> >>>
> >> >>>
> >> >>>>
> >> >>>> 2016-11-14 20:04 GMT+03:00 Richard Biener <rguenther@suse.de>:
> >> >>>>> On November 14, 2016 4:39:40 PM GMT+01:00, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
> >> >>>>>>Richard,
> >> >>>>>>
> >> >>>>>>I checked one of the tests designed for epilogue vectorization using
> >> >>>>>>patches 1 - 3 and found out that build compiler performs vectorization
> >> >>>>>>of epilogues with --param vect-epilogues-nomask=1 passed:
> >> >>>>>>
> >> >>>>>>$ gcc -Ofast -mavx2 t1.c -S --param vect-epilogues-nomask=1 -o
> >> >>>>>>t1.new-nomask.s -fdump-tree-vect-details
> >> >>>>>>$ grep VECTORIZED -c t1.c.156t.vect
> >> >>>>>>4
> >> >>>>>> Without param only 2 loops are vectorized.
> >> >>>>>>
> >> >>>>>>Should I simply add a part of tests related to this feature or I must
> >> >>>>>>delete all not necessary changes also?
> >> >>>>>
> >> >>>>> Please remove all not necessary changes.
> >> >>>>>
> >> >>>>> Richard.
> >> >>>>>
> >> >>>>>>Thanks.
> >> >>>>>>Yuri.
> >> >>>>>>
> >> >>>>>>2016-11-14 16:40 GMT+03:00 Richard Biener <rguenther@suse.de>:
> >> >>>>>>> On Mon, 14 Nov 2016, Yuri Rumyantsev wrote:
> >> >>>>>>>
> >> >>>>>>>> Richard,
> >> >>>>>>>>
> >> >>>>>>>> In my previous patch I forgot to remove couple lines related to aux
> >> >>>>>>field.
> >> >>>>>>>> Here is the correct updated patch.
> >> >>>>>>>
> >> >>>>>>> Yeah, I noticed.  This patch would be ok for trunk (together with
> >> >>>>>>> necessary parts from 1 and 2) if all not required parts are removed
> >> >>>>>>> (and you'd add the testcases covering non-masked tail vect).
> >> >>>>>>>
> >> >>>>>>> Thus, can you please produce a single complete patch containing only
> >> >>>>>>> non-masked epilogue vectoriziation?
> >> >>>>>>>
> >> >>>>>>> Thanks,
> >> >>>>>>> Richard.
> >> >>>>>>>
> >> >>>>>>>> Thanks.
> >> >>>>>>>> Yuri.
> >> >>>>>>>>
> >> >>>>>>>> 2016-11-14 15:51 GMT+03:00 Richard Biener <rguenther@suse.de>:
> >> >>>>>>>> > On Fri, 11 Nov 2016, Yuri Rumyantsev wrote:
> >> >>>>>>>> >
> >> >>>>>>>> >> Richard,
> >> >>>>>>>> >>
> >> >>>>>>>> >> I prepare updated 3 patch with passing additional argument to
> >> >>>>>>>> >> vect_analyze_loop as you proposed (untested).
> >> >>>>>>>> >>
> >> >>>>>>>> >> You wrote:
> >> >>>>>>>> >> tw, I wonder if you can produce a single patch containing just
> >> >>>>>>>> >> epilogue vectorization, that is combine patches 1-3 but rip out
> >> >>>>>>>> >> changes only needed by later patches?
> >> >>>>>>>> >>
> >> >>>>>>>> >> Did you mean that I exclude all support for vectorization
> >> >>>>>>epilogues,
> >> >>>>>>>> >> i.e. exclude from 2-nd patch all non-related changes
> >> >>>>>>>> >> like
> >> >>>>>>>> >>
> >> >>>>>>>> >> diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
> >> >>>>>>>> >> index 11863af..32011c1 100644
> >> >>>>>>>> >> --- a/gcc/tree-vect-loop.c
> >> >>>>>>>> >> +++ b/gcc/tree-vect-loop.c
> >> >>>>>>>> >> @@ -1120,6 +1120,12 @@ new_loop_vec_info (struct loop *loop)
> >> >>>>>>>> >>    LOOP_VINFO_PEELING_FOR_GAPS (res) = false;
> >> >>>>>>>> >>    LOOP_VINFO_PEELING_FOR_NITER (res) = false;
> >> >>>>>>>> >>    LOOP_VINFO_OPERANDS_SWAPPED (res) = false;
> >> >>>>>>>> >> +  LOOP_VINFO_CAN_BE_MASKED (res) = false;
> >> >>>>>>>> >> +  LOOP_VINFO_REQUIRED_MASKS (res) = 0;
> >> >>>>>>>> >> +  LOOP_VINFO_COMBINE_EPILOGUE (res) = false;
> >> >>>>>>>> >> +  LOOP_VINFO_MASK_EPILOGUE (res) = false;
> >> >>>>>>>> >> +  LOOP_VINFO_NEED_MASKING (res) = false;
> >> >>>>>>>> >> +  LOOP_VINFO_ORIG_LOOP_INFO (res) = NULL;
> >> >>>>>>>> >
> >> >>>>>>>> > Yes.
> >> >>>>>>>> >
> >> >>>>>>>> >> Did you mean also that new combined patch must be working patch,
> >> >>>>>>i.e.
> >> >>>>>>>> >> can be integrated without other patches?
> >> >>>>>>>> >
> >> >>>>>>>> > Yes.
> >> >>>>>>>> >
> >> >>>>>>>> >> Could you please look at updated patch?
> >> >>>>>>>> >
> >> >>>>>>>> > Will do.
> >> >>>>>>>> >
> >> >>>>>>>> > Thanks,
> >> >>>>>>>> > Richard.
> >> >>>>>>>> >
> >> >>>>>>>> >> Thanks.
> >> >>>>>>>> >> Yuri.
> >> >>>>>>>> >>
> >> >>>>>>>> >> 2016-11-10 15:36 GMT+03:00 Richard Biener <rguenther@suse.de>:
> >> >>>>>>>> >> > On Thu, 10 Nov 2016, Richard Biener wrote:
> >> >>>>>>>> >> >
> >> >>>>>>>> >> >> On Tue, 8 Nov 2016, Yuri Rumyantsev wrote:
> >> >>>>>>>> >> >>
> >> >>>>>>>> >> >> > Richard,
> >> >>>>>>>> >> >> >
> >> >>>>>>>> >> >> > Here is updated 3 patch.
> >> >>>>>>>> >> >> >
> >> >>>>>>>> >> >> > I checked that all new tests related to epilogue
> >> >>>>>>vectorization passed with it.
> >> >>>>>>>> >> >> >
> >> >>>>>>>> >> >> > Your comments will be appreciated.
> >> >>>>>>>> >> >>
> >> >>>>>>>> >> >> A lot better now.  Instead of the ->aux dance I now prefer to
> >> >>>>>>>> >> >> pass the original loops loop_vinfo to vect_analyze_loop as
> >> >>>>>>>> >> >> optional argument (if non-NULL we analyze the epilogue of that
> >> >>>>>>>> >> >> loop_vinfo).  OTOH I remember we mainly use it to get at the
> >> >>>>>>>> >> >> original vectorization factor?  So we can pass down an
> >> >>>>>>(optional)
> >> >>>>>>>> >> >> forced vectorization factor as well?
> >> >>>>>>>> >> >
> >> >>>>>>>> >> > Btw, I wonder if you can produce a single patch containing just
> >> >>>>>>>> >> > epilogue vectorization, that is combine patches 1-3 but rip out
> >> >>>>>>>> >> > changes only needed by later patches?
> >> >>>>>>>> >> >
> >> >>>>>>>> >> > Thanks,
> >> >>>>>>>> >> > Richard.
> >> >>>>>>>> >> >
> >> >>>>>>>> >> >> Richard.
> >> >>>>>>>> >> >>
> >> >>>>>>>> >> >> > 2016-11-08 15:38 GMT+03:00 Richard Biener
> >> >>>>>><rguenther@suse.de>:
> >> >>>>>>>> >> >> > > On Thu, 3 Nov 2016, Yuri Rumyantsev wrote:
> >> >>>>>>>> >> >> > >
> >> >>>>>>>> >> >> > >> Hi Richard,
> >> >>>>>>>> >> >> > >>
> >> >>>>>>>> >> >> > >> I did not understand your last remark:
> >> >>>>>>>> >> >> > >>
> >> >>>>>>>> >> >> > >> > That is, here (and avoid the FOR_EACH_LOOP change):
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > @@ -580,12 +586,21 @@ vectorize_loops (void)
> >> >>>>>>>> >> >> > >> >           && dump_enabled_p ())
> >> >>>>>>>> >> >> > >> >           dump_printf_loc (MSG_OPTIMIZED_LOCATIONS,
> >> >>>>>>vect_location,
> >> >>>>>>>> >> >> > >> >                            "loop vectorized\n");
> >> >>>>>>>> >> >> > >> > -       vect_transform_loop (loop_vinfo);
> >> >>>>>>>> >> >> > >> > +       new_loop = vect_transform_loop (loop_vinfo);
> >> >>>>>>>> >> >> > >> >         num_vectorized_loops++;
> >> >>>>>>>> >> >> > >> >        /* Now that the loop has been vectorized, allow
> >> >>>>>>it to be unrolled
> >> >>>>>>>> >> >> > >> >           etc.  */
> >> >>>>>>>> >> >> > >> >      loop->force_vectorize = false;
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > +       /* Add new loop to a processing queue.  To make
> >> >>>>>>it easier
> >> >>>>>>>> >> >> > >> > +          to match loop and its epilogue vectorization
> >> >>>>>>in dumps
> >> >>>>>>>> >> >> > >> > +          put new loop as the next loop to process.
> >> >>>>>>*/
> >> >>>>>>>> >> >> > >> > +       if (new_loop)
> >> >>>>>>>> >> >> > >> > +         {
> >> >>>>>>>> >> >> > >> > +           loops.safe_insert (i + 1, new_loop->num);
> >> >>>>>>>> >> >> > >> > +           vect_loops_num = number_of_loops (cfun);
> >> >>>>>>>> >> >> > >> > +         }
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > simply dispatch to a vectorize_epilogue (loop_vinfo,
> >> >>>>>>new_loop)
> >> >>>>>>>> >> >> > >> f> unction which will set up stuff properly (and also
> >> >>>>>>perform
> >> >>>>>>>> >> >> > >> > the if-conversion of the epilogue there).
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > That said, if we can get in non-masked epilogue
> >> >>>>>>vectorization
> >> >>>>>>>> >> >> > >> > separately that would be great.
> >> >>>>>>>> >> >> > >>
> >> >>>>>>>> >> >> > >> Could you please clarify your proposal.
> >> >>>>>>>> >> >> > >
> >> >>>>>>>> >> >> > > When a loop was vectorized set things up to immediately
> >> >>>>>>vectorize
> >> >>>>>>>> >> >> > > its epilogue, avoiding changing the loop iteration and
> >> >>>>>>avoiding
> >> >>>>>>>> >> >> > > the re-use of ->aux.
> >> >>>>>>>> >> >> > >
> >> >>>>>>>> >> >> > > Richard.
> >> >>>>>>>> >> >> > >
> >> >>>>>>>> >> >> > >> Thanks.
> >> >>>>>>>> >> >> > >> Yuri.
> >> >>>>>>>> >> >> > >>
> >> >>>>>>>> >> >> > >> 2016-11-02 15:27 GMT+03:00 Richard Biener
> >> >>>>>><rguenther@suse.de>:
> >> >>>>>>>> >> >> > >> > On Tue, 1 Nov 2016, Yuri Rumyantsev wrote:
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> >> Hi All,
> >> >>>>>>>> >> >> > >> >>
> >> >>>>>>>> >> >> > >> >> I re-send all patches sent by Ilya earlier for review
> >> >>>>>>which support
> >> >>>>>>>> >> >> > >> >> vectorization of loop epilogues and loops with low
> >> >>>>>>trip count. We
> >> >>>>>>>> >> >> > >> >> assume that the only patch -
> >> >>>>>>vec-tails-07-combine-tail.patch - was not
> >> >>>>>>>> >> >> > >> >> approved by Jeff.
> >> >>>>>>>> >> >> > >> >>
> >> >>>>>>>> >> >> > >> >> I did re-base of all patches and performed
> >> >>>>>>bootstrapping and
> >> >>>>>>>> >> >> > >> >> regression testing that did not show any new failures.
> >> >>>>>>Also all
> >> >>>>>>>> >> >> > >> >> changes related to new vect_do_peeling algorithm have
> >> >>>>>>been changed
> >> >>>>>>>> >> >> > >> >> accordingly.
> >> >>>>>>>> >> >> > >> >>
> >> >>>>>>>> >> >> > >> >> Is it OK for trunk?
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > I would have prefered that the series up to
> >> >>>>>>-03-nomask-tails would
> >> >>>>>>>> >> >> > >> > _only_ contain epilogue loop vectorization changes but
> >> >>>>>>unfortunately
> >> >>>>>>>> >> >> > >> > the patchset is oddly separated.
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > I have a comment on that part nevertheless:
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > @@ -1608,7 +1614,10 @@ vect_enhance_data_refs_alignment
> >> >>>>>>(loop_vec_info
> >> >>>>>>>> >> >> > >> > loop_vinfo)
> >> >>>>>>>> >> >> > >> >    /* Check if we can possibly peel the loop.  */
> >> >>>>>>>> >> >> > >> >    if (!vect_can_advance_ivs_p (loop_vinfo)
> >> >>>>>>>> >> >> > >> >        || !slpeel_can_duplicate_loop_p (loop,
> >> >>>>>>single_exit (loop))
> >> >>>>>>>> >> >> > >> > -      || loop->inner)
> >> >>>>>>>> >> >> > >> > +      || loop->inner
> >> >>>>>>>> >> >> > >> > +      /* Required peeling was performed in prologue
> >> >>>>>>and
> >> >>>>>>>> >> >> > >> > +        is not required for epilogue.  */
> >> >>>>>>>> >> >> > >> > +      || LOOP_VINFO_EPILOGUE_P (loop_vinfo))
> >> >>>>>>>> >> >> > >> >      do_peeling = false;
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> >    if (do_peeling
> >> >>>>>>>> >> >> > >> > @@ -1888,7 +1897,10 @@ vect_enhance_data_refs_alignment
> >> >>>>>>(loop_vec_info
> >> >>>>>>>> >> >> > >> > loop_vinfo)
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> >    do_versioning =
> >> >>>>>>>> >> >> > >> >         optimize_loop_nest_for_speed_p (loop)
> >> >>>>>>>> >> >> > >> > -       && (!loop->inner); /* FORNOW */
> >> >>>>>>>> >> >> > >> > +       && (!loop->inner) /* FORNOW */
> >> >>>>>>>> >> >> > >> > +        /* Required versioning was performed for the
> >> >>>>>>>> >> >> > >> > +          original loop and is not required for
> >> >>>>>>epilogue.  */
> >> >>>>>>>> >> >> > >> > +       && !LOOP_VINFO_EPILOGUE_P (loop_vinfo);
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> >    if (do_versioning)
> >> >>>>>>>> >> >> > >> >      {
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > please do that check in the single caller of this
> >> >>>>>>function.
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > Otherwise I still dislike the new ->aux use and I
> >> >>>>>>believe that simply
> >> >>>>>>>> >> >> > >> > passing down info from the processed parent would be
> >> >>>>>>_much_ cleaner.
> >> >>>>>>>> >> >> > >> > That is, here (and avoid the FOR_EACH_LOOP change):
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > @@ -580,12 +586,21 @@ vectorize_loops (void)
> >> >>>>>>>> >> >> > >> >             && dump_enabled_p ())
> >> >>>>>>>> >> >> > >> >            dump_printf_loc (MSG_OPTIMIZED_LOCATIONS,
> >> >>>>>>vect_location,
> >> >>>>>>>> >> >> > >> >                             "loop vectorized\n");
> >> >>>>>>>> >> >> > >> > -       vect_transform_loop (loop_vinfo);
> >> >>>>>>>> >> >> > >> > +       new_loop = vect_transform_loop (loop_vinfo);
> >> >>>>>>>> >> >> > >> >         num_vectorized_loops++;
> >> >>>>>>>> >> >> > >> >         /* Now that the loop has been vectorized, allow
> >> >>>>>>it to be unrolled
> >> >>>>>>>> >> >> > >> >            etc.  */
> >> >>>>>>>> >> >> > >> >         loop->force_vectorize = false;
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > +       /* Add new loop to a processing queue.  To make
> >> >>>>>>it easier
> >> >>>>>>>> >> >> > >> > +          to match loop and its epilogue vectorization
> >> >>>>>>in dumps
> >> >>>>>>>> >> >> > >> > +          put new loop as the next loop to process.
> >> >>>>>>*/
> >> >>>>>>>> >> >> > >> > +       if (new_loop)
> >> >>>>>>>> >> >> > >> > +         {
> >> >>>>>>>> >> >> > >> > +           loops.safe_insert (i + 1, new_loop->num);
> >> >>>>>>>> >> >> > >> > +           vect_loops_num = number_of_loops (cfun);
> >> >>>>>>>> >> >> > >> > +         }
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > simply dispatch to a vectorize_epilogue (loop_vinfo,
> >> >>>>>>new_loop)
> >> >>>>>>>> >> >> > >> > function which will set up stuff properly (and also
> >> >>>>>>perform
> >> >>>>>>>> >> >> > >> > the if-conversion of the epilogue there).
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > That said, if we can get in non-masked epilogue
> >> >>>>>>vectorization
> >> >>>>>>>> >> >> > >> > separately that would be great.
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > I'm still torn about all the rest of the stuff and
> >> >>>>>>question its
> >> >>>>>>>> >> >> > >> > usability (esp. merging the epilogue with the main
> >> >>>>>>vector loop).
> >> >>>>>>>> >> >> > >> > But it has already been approved ... oh well.
> >> >>>>>>>> >> >> > >> >
> >> >>>>>>>> >> >> > >> > Thanks,
> >> >>>>>>>> >> >> > >> > Richard.
> >> >>>>>>>> >> >> > >>
> >> >>>>>>>> >> >> > >>
> >> >>>>>>>> >> >> > >
> >> >>>>>>>> >> >> > > --
> >> >>>>>>>> >> >> > > Richard Biener <rguenther@suse.de>
> >> >>>>>>>> >> >> > > SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard,
> >> >>>>>>Graham Norton, HRB 21284 (AG Nuernberg)
> >> >>>>>>>> >> >> >
> >> >>>>>>>> >> >>
> >> >>>>>>>> >> >>
> >> >>>>>>>> >> >
> >> >>>>>>>> >> > --
> >> >>>>>>>> >> > Richard Biener <rguenther@suse.de>
> >> >>>>>>>> >> > SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham
> >> >>>>>>Norton, HRB 21284 (AG Nuernberg)
> >> >>>>>>>> >>
> >> >>>>>>>> >
> >> >>>>>>>> > --
> >> >>>>>>>> > Richard Biener <rguenther@suse.de>
> >> >>>>>>>> > SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham
> >> >>>>>>Norton, HRB 21284 (AG Nuernberg)
> >> >>>>>>>>
> >> >>>>>>>
> >> >>>>>>> --
> >> >>>>>>> Richard Biener <rguenther@suse.de>
> >> >>>>>>> SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham
> >> >>>>>>Norton, HRB 21284 (AG Nuernberg)
> >> >>>>>
> >> >>>>>
> >>
> >
> > --
> > Richard Biener <rguenther@suse.de>
> > SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)
>


[vec-tails] Support loop epilogue vectorization

Commit Message

Comments

Patch