diff mbox series

Make DR_TARGET_ALIGNMENT compile time variable

Message ID 5BE0326C.5090703@arm.com
State New
Headers show
Series Make DR_TARGET_ALIGNMENT compile time variable | expand

Commit Message

Andre Vieira (lists) Nov. 5, 2018, 12:07 p.m. UTC
Hi,

This patch enables targets to describe DR_TARGET_ALIGNMENT as a
compile-time variable.  It does so by turning the variable into a
'poly_uint64'.  This should not affect the current code-generation for
any target.

We hope to use this in the near future for SVE using the
current_vector_size as the preferred target alignment for vectors.  In
fact I have a patch to do just this, but I am still trying to figure out
whether and when it is beneficial to peel for alignment with a runtime
misalignment.  The patch I am working on will change the behavior of
auto-vectorization for SVE when building vector-length agnostic code for
targets that benefit from aligned vector loads/stores.  The patch will
result in  the generation of a runtime computation of misalignment and
the construction of a corresponding mask for the first iteration of the
loop.

I have decided to not offer support for prolog/epilog peeling when the
target alignment is not compile-time constant, as this didn't seem
useful, this is why 'vect_do_peeling' returns early if
DR_TARGET_ALIGNMENT is not constant.

I bootstrapped and tested this on aarch64 and x86 basically
bootstrapping one target that uses this hook and one that doesn't.

Is this OK for trunk?

Cheers,
Andre

2018-11-05  Andre Vieira  <andre.simoesdiasvieira@arm.com>

	* config/aarch64/aarch64.c (aarch64_vectorize_preferred_vector_alignment):
        Change return type to poly_uint64.
        (aarch64_simd_vector_alignment_reachable): Adapt to preferred vector
        alignment being a poly int.
	* doc/tm.texi (TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT): Change return
        type to poly_uint64.
	* target.def (default_preferred_vector_alignment): Likewise.
	* targhooks.c (default_preferred_vector_alignment): Likewise.
	* targhooks.h (default_preferred_vector_alignment): Likewise.
	* tree-vect-data-refs.c
        (vect_calculate_target_alignment): Likewise.
        (vect_compute_data_ref_alignment): Adapt to vector alignment
being a poly int.
	(vect_update_misalignment_for_peel): Likewise.
	(vect_enhance_data_refs_alignment): Likewise.
	(vect_find_same_alignment_drs): Likewise.
	(vect_duplicate_ssa_name_ptr_info): Likewise.
	(vect_setup_realignment): Likewise.
	(vect_can_force_dr_alignment_p): Change alignment parameter type to
poly_uint64.
	* tree-vect-loop-manip.c (get_misalign_in_elems): Learn to construct a mask
        with a compile time variable vector alignment.
	(vect_gen_prolog_loop_niters): Adapt to vector alignment being a poly int.
	(vect_do_peeling): Exit early if vector alignment is not constant.
	* tree-vect-stmts.c (ensure_base_align): Adapt to vector alignment being a
        poly int.
	(vectorizable_store): Likewise.
	(vectorizable_load): Likweise.
	* tree-vectorizer.h (struct dr_vec_info): Make target_alignment field a
        poly_uint64.
	(vect_known_alignment_in_bytes): Adapt to vector alignment being a poly
int.
	(vect_can_force_dr_alignment_p): Change alignment parameter type to
poly_uint64.

Comments

Richard Biener Nov. 5, 2018, 12:41 p.m. UTC | #1
On Mon, Nov 5, 2018 at 1:07 PM Andre Vieira (lists)
<Andre.SimoesDiasVieira@arm.com> wrote:
>
>
> Hi,
>
> This patch enables targets to describe DR_TARGET_ALIGNMENT as a
> compile-time variable.  It does so by turning the variable into a
> 'poly_uint64'.  This should not affect the current code-generation for
> any target.
>
> We hope to use this in the near future for SVE using the
> current_vector_size as the preferred target alignment for vectors.  In
> fact I have a patch to do just this, but I am still trying to figure out
> whether and when it is beneficial to peel for alignment with a runtime
> misalignment.

In fact in most cases I have seen the issue is that it's not visible whether
peeling will be able to align _all_ references and doing peeling only to
align some is hardly beneficial.  To improve things the vectorizer would
have to version the loop for the case where peeling can reach alignment
for a group of DRs and then vectorize one copy with peeling for alignment
and one copy with unaligned accesses.

>  The patch I am working on will change the behavior of
> auto-vectorization for SVE when building vector-length agnostic code for
> targets that benefit from aligned vector loads/stores.  The patch will
> result in  the generation of a runtime computation of misalignment and
> the construction of a corresponding mask for the first iteration of the
> loop.
>
> I have decided to not offer support for prolog/epilog peeling when the
> target alignment is not compile-time constant, as this didn't seem
> useful, this is why 'vect_do_peeling' returns early if
> DR_TARGET_ALIGNMENT is not constant.
>
> I bootstrapped and tested this on aarch64 and x86 basically
> bootstrapping one target that uses this hook and one that doesn't.
>
> Is this OK for trunk?

The patch looks good but I wonder wheter it is really necessary at this
point.

Thanks,
Richard.

> Cheers,
> Andre
>
> 2018-11-05  Andre Vieira  <andre.simoesdiasvieira@arm.com>
>
>         * config/aarch64/aarch64.c (aarch64_vectorize_preferred_vector_alignment):
>         Change return type to poly_uint64.
>         (aarch64_simd_vector_alignment_reachable): Adapt to preferred vector
>         alignment being a poly int.
>         * doc/tm.texi (TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT): Change return
>         type to poly_uint64.
>         * target.def (default_preferred_vector_alignment): Likewise.
>         * targhooks.c (default_preferred_vector_alignment): Likewise.
>         * targhooks.h (default_preferred_vector_alignment): Likewise.
>         * tree-vect-data-refs.c
>         (vect_calculate_target_alignment): Likewise.
>         (vect_compute_data_ref_alignment): Adapt to vector alignment
> being a poly int.
>         (vect_update_misalignment_for_peel): Likewise.
>         (vect_enhance_data_refs_alignment): Likewise.
>         (vect_find_same_alignment_drs): Likewise.
>         (vect_duplicate_ssa_name_ptr_info): Likewise.
>         (vect_setup_realignment): Likewise.
>         (vect_can_force_dr_alignment_p): Change alignment parameter type to
> poly_uint64.
>         * tree-vect-loop-manip.c (get_misalign_in_elems): Learn to construct a mask
>         with a compile time variable vector alignment.
>         (vect_gen_prolog_loop_niters): Adapt to vector alignment being a poly int.
>         (vect_do_peeling): Exit early if vector alignment is not constant.
>         * tree-vect-stmts.c (ensure_base_align): Adapt to vector alignment being a
>         poly int.
>         (vectorizable_store): Likewise.
>         (vectorizable_load): Likweise.
>         * tree-vectorizer.h (struct dr_vec_info): Make target_alignment field a
>         poly_uint64.
>         (vect_known_alignment_in_bytes): Adapt to vector alignment being a poly
> int.
>         (vect_can_force_dr_alignment_p): Change alignment parameter type to
> poly_uint64.
Andre Vieira (lists) Nov. 9, 2018, 4:08 p.m. UTC | #2
On 05/11/18 12:41, Richard Biener wrote:
> On Mon, Nov 5, 2018 at 1:07 PM Andre Vieira (lists)
> <Andre.SimoesDiasVieira@arm.com> wrote:
>>
>>
>> Hi,
>>
Hi,

Thank you for the quick response! See inline responses below.

>> This patch enables targets to describe DR_TARGET_ALIGNMENT as a
>> compile-time variable.  It does so by turning the variable into a
>> 'poly_uint64'.  This should not affect the current code-generation for
>> any target.
>>
>> We hope to use this in the near future for SVE using the
>> current_vector_size as the preferred target alignment for vectors.  In
>> fact I have a patch to do just this, but I am still trying to figure out
>> whether and when it is beneficial to peel for alignment with a runtime
>> misalignment.
> 
> In fact in most cases I have seen the issue is that it's not visible whether
> peeling will be able to align _all_ references and doing peeling only to
> align some is hardly beneficial.  To improve things the vectorizer would
> have to version the loop for the case where peeling can reach alignment
> for a group of DRs and then vectorize one copy with peeling for alignment
> and one copy with unaligned accesses.


So I have seen code being peeled for alignment even when it only knows
how to align one of a group (only checked 2 or 3) and I think this may
still be beneficial in some cases.  I am more worried about cases where
the number of iterations isn't enough to justify the initial peeling
cost or when the loop isn't memory bound, i.e. very arithmetic heavy
loops.  This is a bigger vectorization problem though, that would
require some kind of cost-model.

> 
>>  The patch I am working on will change the behavior of
>> auto-vectorization for SVE when building vector-length agnostic code for
>> targets that benefit from aligned vector loads/stores.  The patch will
>> result in  the generation of a runtime computation of misalignment and
>> the construction of a corresponding mask for the first iteration of the
>> loop.
>>
>> I have decided to not offer support for prolog/epilog peeling when the
>> target alignment is not compile-time constant, as this didn't seem
>> useful, this is why 'vect_do_peeling' returns early if
>> DR_TARGET_ALIGNMENT is not constant.
>>
>> I bootstrapped and tested this on aarch64 and x86 basically
>> bootstrapping one target that uses this hook and one that doesn't.
>>
>> Is this OK for trunk?
> 
> The patch looks good but I wonder wheter it is really necessary at this
> point.

The goal of this patch is really to enable future work, on it's own it
does nothing.  I am working on a small target-specific patch to enable
this for SVE, but I need to do a bit more analysis and benchmarking to
be able to determine whether its beneficial which I will not be able to
finish before end of stage 1. That is why I split them up and sent this
one upstream to see if I could get the middle-end change in.

> 
> Thanks,
> Richard.
> 
>> Cheers,
>> Andre
>>
>> 2018-11-05  Andre Vieira  <andre.simoesdiasvieira@arm.com>
>>
>>         * config/aarch64/aarch64.c (aarch64_vectorize_preferred_vector_alignment):
>>         Change return type to poly_uint64.
>>         (aarch64_simd_vector_alignment_reachable): Adapt to preferred vector
>>         alignment being a poly int.
>>         * doc/tm.texi (TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT): Change return
>>         type to poly_uint64.
>>         * target.def (default_preferred_vector_alignment): Likewise.
>>         * targhooks.c (default_preferred_vector_alignment): Likewise.
>>         * targhooks.h (default_preferred_vector_alignment): Likewise.
>>         * tree-vect-data-refs.c
>>         (vect_calculate_target_alignment): Likewise.
>>         (vect_compute_data_ref_alignment): Adapt to vector alignment
>> being a poly int.
>>         (vect_update_misalignment_for_peel): Likewise.
>>         (vect_enhance_data_refs_alignment): Likewise.
>>         (vect_find_same_alignment_drs): Likewise.
>>         (vect_duplicate_ssa_name_ptr_info): Likewise.
>>         (vect_setup_realignment): Likewise.
>>         (vect_can_force_dr_alignment_p): Change alignment parameter type to
>> poly_uint64.
>>         * tree-vect-loop-manip.c (get_misalign_in_elems): Learn to construct a mask
>>         with a compile time variable vector alignment.
>>         (vect_gen_prolog_loop_niters): Adapt to vector alignment being a poly int.
>>         (vect_do_peeling): Exit early if vector alignment is not constant.
>>         * tree-vect-stmts.c (ensure_base_align): Adapt to vector alignment being a
>>         poly int.
>>         (vectorizable_store): Likewise.
>>         (vectorizable_load): Likweise.
>>         * tree-vectorizer.h (struct dr_vec_info): Make target_alignment field a
>>         poly_uint64.
>>         (vect_known_alignment_in_bytes): Adapt to vector alignment being a poly
>> int.
>>         (vect_can_force_dr_alignment_p): Change alignment parameter type to
>> poly_uint64.
Richard Biener Nov. 12, 2018, 2:03 p.m. UTC | #3
On Fri, Nov 9, 2018 at 5:08 PM Andre Vieira (lists)
<Andre.SimoesDiasVieira@arm.com> wrote:
>
> On 05/11/18 12:41, Richard Biener wrote:
> > On Mon, Nov 5, 2018 at 1:07 PM Andre Vieira (lists)
> > <Andre.SimoesDiasVieira@arm.com> wrote:
> >>
> >>
> >> Hi,
> >>
> Hi,
>
> Thank you for the quick response! See inline responses below.
>
> >> This patch enables targets to describe DR_TARGET_ALIGNMENT as a
> >> compile-time variable.  It does so by turning the variable into a
> >> 'poly_uint64'.  This should not affect the current code-generation for
> >> any target.
> >>
> >> We hope to use this in the near future for SVE using the
> >> current_vector_size as the preferred target alignment for vectors.  In
> >> fact I have a patch to do just this, but I am still trying to figure out
> >> whether and when it is beneficial to peel for alignment with a runtime
> >> misalignment.
> >
> > In fact in most cases I have seen the issue is that it's not visible whether
> > peeling will be able to align _all_ references and doing peeling only to
> > align some is hardly beneficial.  To improve things the vectorizer would
> > have to version the loop for the case where peeling can reach alignment
> > for a group of DRs and then vectorize one copy with peeling for alignment
> > and one copy with unaligned accesses.
>
>
> So I have seen code being peeled for alignment even when it only knows
> how to align one of a group (only checked 2 or 3) and I think this may
> still be beneficial in some cases.  I am more worried about cases where
> the number of iterations isn't enough to justify the initial peeling
> cost or when the loop isn't memory bound, i.e. very arithmetic heavy
> loops.  This is a bigger vectorization problem though, that would
> require some kind of cost-model.
>
> >
> >>  The patch I am working on will change the behavior of
> >> auto-vectorization for SVE when building vector-length agnostic code for
> >> targets that benefit from aligned vector loads/stores.  The patch will
> >> result in  the generation of a runtime computation of misalignment and
> >> the construction of a corresponding mask for the first iteration of the
> >> loop.
> >>
> >> I have decided to not offer support for prolog/epilog peeling when the
> >> target alignment is not compile-time constant, as this didn't seem
> >> useful, this is why 'vect_do_peeling' returns early if
> >> DR_TARGET_ALIGNMENT is not constant.
> >>
> >> I bootstrapped and tested this on aarch64 and x86 basically
> >> bootstrapping one target that uses this hook and one that doesn't.
> >>
> >> Is this OK for trunk?
> >
> > The patch looks good but I wonder wheter it is really necessary at this
> > point.
>
> The goal of this patch is really to enable future work, on it's own it
> does nothing.  I am working on a small target-specific patch to enable
> this for SVE, but I need to do a bit more analysis and benchmarking to
> be able to determine whether its beneficial which I will not be able to
> finish before end of stage 1. That is why I split them up and sent this
> one upstream to see if I could get the middle-end change in.

OK, fine with me then.

Thanks,
Richard.

> >
> > Thanks,
> > Richard.
> >
> >> Cheers,
> >> Andre
> >>
> >> 2018-11-05  Andre Vieira  <andre.simoesdiasvieira@arm.com>
> >>
> >>         * config/aarch64/aarch64.c (aarch64_vectorize_preferred_vector_alignment):
> >>         Change return type to poly_uint64.
> >>         (aarch64_simd_vector_alignment_reachable): Adapt to preferred vector
> >>         alignment being a poly int.
> >>         * doc/tm.texi (TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT): Change return
> >>         type to poly_uint64.
> >>         * target.def (default_preferred_vector_alignment): Likewise.
> >>         * targhooks.c (default_preferred_vector_alignment): Likewise.
> >>         * targhooks.h (default_preferred_vector_alignment): Likewise.
> >>         * tree-vect-data-refs.c
> >>         (vect_calculate_target_alignment): Likewise.
> >>         (vect_compute_data_ref_alignment): Adapt to vector alignment
> >> being a poly int.
> >>         (vect_update_misalignment_for_peel): Likewise.
> >>         (vect_enhance_data_refs_alignment): Likewise.
> >>         (vect_find_same_alignment_drs): Likewise.
> >>         (vect_duplicate_ssa_name_ptr_info): Likewise.
> >>         (vect_setup_realignment): Likewise.
> >>         (vect_can_force_dr_alignment_p): Change alignment parameter type to
> >> poly_uint64.
> >>         * tree-vect-loop-manip.c (get_misalign_in_elems): Learn to construct a mask
> >>         with a compile time variable vector alignment.
> >>         (vect_gen_prolog_loop_niters): Adapt to vector alignment being a poly int.
> >>         (vect_do_peeling): Exit early if vector alignment is not constant.
> >>         * tree-vect-stmts.c (ensure_base_align): Adapt to vector alignment being a
> >>         poly int.
> >>         (vectorizable_store): Likewise.
> >>         (vectorizable_load): Likweise.
> >>         * tree-vectorizer.h (struct dr_vec_info): Make target_alignment field a
> >>         poly_uint64.
> >>         (vect_known_alignment_in_bytes): Adapt to vector alignment being a poly
> >> int.
> >>         (vect_can_force_dr_alignment_p): Change alignment parameter type to
> >> poly_uint64.
>
diff mbox series

Patch

diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 4c7790826658539f71f2fd9eb9ea0329081938be..19f084abefd76255d1a4a0b726e51c7025b9cea6 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -14120,7 +14120,7 @@  aarch64_simd_vector_alignment (const_tree type)
 }
 
 /* Implement target hook TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT.  */
-static HOST_WIDE_INT
+static poly_uint64
 aarch64_vectorize_preferred_vector_alignment (const_tree type)
 {
   if (aarch64_sve_data_mode_p (TYPE_MODE (type)))
@@ -14145,9 +14145,11 @@  aarch64_simd_vector_alignment_reachable (const_tree type, bool is_packed)
   /* For fixed-length vectors, check that the vectorizer will aim for
      full-vector alignment.  This isn't true for generic GCC vectors
      that are wider than the ABI maximum of 128 bits.  */
+  poly_uint64 preferred_alignment =
+    aarch64_vectorize_preferred_vector_alignment (type);
   if (TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
-      && (wi::to_widest (TYPE_SIZE (type))
-	  != aarch64_vectorize_preferred_vector_alignment (type)))
+      && maybe_ne (wi::to_widest (TYPE_SIZE (type)),
+		   preferred_alignment))
     return false;
 
   /* Vectors whose size is <= BIGGEST_ALIGNMENT are naturally aligned.  */
diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi
index 0fcf8069b8cc948fcaf5604a1235fe269de7e8f3..328eb43ca2495dd889bc47cf136761381a594cdf 100644
--- a/gcc/doc/tm.texi
+++ b/gcc/doc/tm.texi
@@ -5889,7 +5889,7 @@  For vector memory operations the cost may depend on type (@var{vectype}) and
 misalignment value (@var{misalign}).
 @end deftypefn
 
-@deftypefn {Target Hook} HOST_WIDE_INT TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT (const_tree @var{type})
+@deftypefn {Target Hook} poly_uint64 TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT (const_tree @var{type})
 This hook returns the preferred alignment in bits for accesses to
 vectors of type @var{type} in vectorized code.  This might be less than
 or greater than the ABI-defined value returned by
diff --git a/gcc/target.def b/gcc/target.def
index ad27d352ca40feb92212a733d974aca7ed8e10f1..54bc7a1e179f61bdc3e43b727753e34fb6a3d5a1 100644
--- a/gcc/target.def
+++ b/gcc/target.def
@@ -1802,7 +1802,7 @@  for alignment.\n\
 \n\
 The default hook returns @code{TYPE_ALIGN (@var{type})}, which is\n\
 correct for most targets.",
- HOST_WIDE_INT, (const_tree type),
+ poly_uint64, (const_tree type),
  default_preferred_vector_alignment)
 
 /* Return true if vector alignment is reachable (by peeling N
diff --git a/gcc/targhooks.h b/gcc/targhooks.h
index 176c64d23f534601be5b3534d416bcf3a66cb20a..430c79e772868ec9301fe2a64f90b231799a18c3 100644
--- a/gcc/targhooks.h
+++ b/gcc/targhooks.h
@@ -102,7 +102,7 @@  extern HOST_WIDE_INT constant_alignment_word_strings (const_tree,
 						      HOST_WIDE_INT);
 extern HOST_WIDE_INT default_vector_alignment (const_tree);
 
-extern HOST_WIDE_INT default_preferred_vector_alignment (const_tree);
+extern poly_uint64 default_preferred_vector_alignment (const_tree);
 extern bool default_builtin_vector_alignment_reachable (const_tree, bool);
 extern bool
 default_builtin_support_vector_misalignment (machine_mode mode,
diff --git a/gcc/targhooks.c b/gcc/targhooks.c
index 3d8b3b9d69be8fb1c9bfb052e522ea84b1bdc341..bcbd534e32d7c6c80f4c42ff0f8a04e3d577027d 100644
--- a/gcc/targhooks.c
+++ b/gcc/targhooks.c
@@ -1260,7 +1260,7 @@  default_vector_alignment (const_tree type)
 /* The default implementation of
    TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT.  */
 
-HOST_WIDE_INT
+poly_uint64
 default_preferred_vector_alignment (const_tree type)
 {
   return TYPE_ALIGN (type);
diff --git a/gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c
index c4805e724474f5870420f0ea43b552bacf30027b..d3853396094121036d2e98b06517ca143942cfcb 100644
--- a/gcc/tree-vect-data-refs.c
+++ b/gcc/tree-vect-data-refs.c
@@ -809,7 +809,7 @@  vect_record_base_alignments (vec_info *vinfo)
 
 /* Return the target alignment for the vectorized form of DR_INFO.  */
 
-static unsigned int
+static poly_uint64
 vect_calculate_target_alignment (dr_vec_info *dr_info)
 {
   tree vectype = STMT_VINFO_VECTYPE (dr_info->stmt);
@@ -852,10 +852,14 @@  vect_compute_data_ref_alignment (dr_vec_info *dr_info)
   innermost_loop_behavior *drb = vect_dr_behavior (dr_info);
   bool step_preserves_misalignment_p;
 
-  unsigned HOST_WIDE_INT vector_alignment
-    = vect_calculate_target_alignment (dr_info) / BITS_PER_UNIT;
+  poly_uint64 vector_alignment
+    = exact_div (vect_calculate_target_alignment (dr_info), BITS_PER_UNIT);
   DR_TARGET_ALIGNMENT (dr_info) = vector_alignment;
 
+  unsigned HOST_WIDE_INT vect_align_c;
+  if (!vector_alignment.is_constant (&vect_align_c))
+    return;
+
   /* No step for BB vectorization.  */
   if (!loop)
     {
@@ -872,7 +876,7 @@  vect_compute_data_ref_alignment (dr_vec_info *dr_info)
   else if (nested_in_vect_loop_p (loop, stmt_info))
     {
       step_preserves_misalignment_p
-	= (DR_STEP_ALIGNMENT (dr_info->dr) % vector_alignment) == 0;
+	= (DR_STEP_ALIGNMENT (dr_info->dr) % vect_align_c) == 0;
 
       if (dump_enabled_p ())
 	{
@@ -894,7 +898,7 @@  vect_compute_data_ref_alignment (dr_vec_info *dr_info)
     {
       poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
       step_preserves_misalignment_p
-	= multiple_p (DR_STEP_ALIGNMENT (dr_info->dr) * vf, vector_alignment);
+	= multiple_p (DR_STEP_ALIGNMENT (dr_info->dr) * vf, vect_align_c);
 
       if (!step_preserves_misalignment_p && dump_enabled_p ())
 	dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
@@ -913,7 +917,7 @@  vect_compute_data_ref_alignment (dr_vec_info *dr_info)
       base_misalignment = (*entry)->base_misalignment;
     }
 
-  if (drb->offset_alignment < vector_alignment
+  if (drb->offset_alignment < vect_align_c
       || !step_preserves_misalignment_p
       /* We need to know whether the step wrt the vectorized loop is
 	 negative when computing the starting misalignment below.  */
@@ -925,13 +929,13 @@  vect_compute_data_ref_alignment (dr_vec_info *dr_info)
       return;
     }
 
-  if (base_alignment < vector_alignment)
+  if (base_alignment < vect_align_c)
     {
       unsigned int max_alignment;
       tree base = get_base_for_alignment (drb->base_address, &max_alignment);
-      if (max_alignment < vector_alignment
+      if (max_alignment < vect_align_c
 	  || !vect_can_force_dr_alignment_p (base,
-					     vector_alignment * BITS_PER_UNIT))
+					     vect_align_c * BITS_PER_UNIT))
 	{
 	  if (dump_enabled_p ())
 	    dump_printf_loc (MSG_NOTE, vect_location,
@@ -962,8 +966,7 @@  vect_compute_data_ref_alignment (dr_vec_info *dr_info)
 		     * TREE_INT_CST_LOW (drb->step));
 
   unsigned int const_misalignment;
-  if (!known_misalignment (misalignment, vector_alignment,
-			   &const_misalignment))
+  if (!known_misalignment (misalignment, vect_align_c, &const_misalignment))
     {
       if (dump_enabled_p ())
 	dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
@@ -1027,14 +1030,16 @@  vect_update_misalignment_for_peel (dr_vec_info *dr_info,
       return;
     }
 
-  if (known_alignment_for_access_p (dr_info)
+  unsigned HOST_WIDE_INT alignment;
+  if (DR_TARGET_ALIGNMENT (dr_info).is_constant (&alignment)
+      && known_alignment_for_access_p (dr_info)
       && known_alignment_for_access_p (dr_peel_info))
     {
       bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr),
 					    size_zero_node) < 0;
       int misal = DR_MISALIGNMENT (dr_info);
       misal += negative ? -npeel * dr_size : npeel * dr_size;
-      misal &= DR_TARGET_ALIGNMENT (dr_info) - 1;
+      misal &= alignment - 1;
       SET_DR_MISALIGNMENT (dr_info, misal);
       return;
     }
@@ -1678,7 +1683,12 @@  vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
 						    size_zero_node) < 0;
 
 	      vectype = STMT_VINFO_VECTYPE (stmt_info);
-	      unsigned int target_align = DR_TARGET_ALIGNMENT (dr_info);
+	      /* If known_alignment_for_access_p then we have set
+	         DR_MISALIGNMENT which is only done if we know it at compiler
+	         time, so it is safe to assume target alignment is constant.
+	       */
+	      unsigned int target_align =
+		DR_TARGET_ALIGNMENT (dr_info).to_constant ();
 	      unsigned int dr_size = vect_get_scalar_dr_size (dr_info);
 	      mis = (negative
 		     ? DR_MISALIGNMENT (dr_info)
@@ -1955,7 +1965,12 @@  vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
 	      mis = (negative
 		     ? DR_MISALIGNMENT (dr0_info)
 		     : -DR_MISALIGNMENT (dr0_info));
-	      unsigned int target_align = DR_TARGET_ALIGNMENT (dr0_info);
+	      /* If known_alignment_for_access_p then we have set
+	         DR_MISALIGNMENT which is only done if we know it at compiler
+	         time, so it is safe to assume target alignment is constant.
+	       */
+	      unsigned int target_align =
+		DR_TARGET_ALIGNMENT (dr0_info).to_constant ();
 	      npeel = ((mis & (target_align - 1))
 		       / vect_get_scalar_dr_size (dr0_info));
             }
@@ -1995,9 +2010,19 @@  vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
               unsigned max_peel = npeel;
               if (max_peel == 0)
                 {
-		  unsigned int target_align = DR_TARGET_ALIGNMENT (dr0_info);
-		  max_peel = (target_align
-			      / vect_get_scalar_dr_size (dr0_info) - 1);
+		  poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr0_info);
+		  unsigned HOST_WIDE_INT target_align_c;
+		  if (target_align.is_constant (&target_align_c))
+		    max_peel =
+		      target_align_c / vect_get_scalar_dr_size (dr0_info) - 1;
+		  else
+		    {
+		      do_peeling = false;
+		      if (dump_enabled_p ())
+			dump_printf_loc (MSG_NOTE, vect_location,
+			  "Disable peeling, max peels set and vector"
+			  " alignment unknown\n");
+		    }
                 }
               if (max_peel > max_allowed_peel)
                 {
@@ -2236,11 +2261,18 @@  vect_find_same_alignment_drs (vec_info *vinfo, data_dependence_relation *ddr)
   if (maybe_ne (diff, 0))
     {
       /* Get the wider of the two alignments.  */
-      unsigned int align_a = (vect_calculate_target_alignment (dr_info_a)
-			      / BITS_PER_UNIT);
-      unsigned int align_b = (vect_calculate_target_alignment (dr_info_b)
-			      / BITS_PER_UNIT);
-      unsigned int max_align = MAX (align_a, align_b);
+      poly_uint64 align_a =
+	exact_div (vect_calculate_target_alignment (dr_info_a),
+		   BITS_PER_UNIT);
+      poly_uint64 align_b =
+	exact_div (vect_calculate_target_alignment (dr_info_b),
+		   BITS_PER_UNIT);
+      unsigned HOST_WIDE_INT align_a_c, align_b_c;
+      if (!align_a.is_constant (&align_a_c)
+	  || !align_b.is_constant (&align_b_c))
+	return;
+
+      unsigned HOST_WIDE_INT max_align = MAX (align_a_c, align_b_c);
 
       /* Require the gap to be a multiple of the larger vector alignment.  */
       if (!multiple_p (diff, max_align))
@@ -4296,7 +4328,8 @@  vect_duplicate_ssa_name_ptr_info (tree name, dr_vec_info *dr_info)
     mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (name));
   else
     set_ptr_info_alignment (SSA_NAME_PTR_INFO (name),
-			    DR_TARGET_ALIGNMENT (dr_info), misalign);
+			    known_alignment (DR_TARGET_ALIGNMENT (dr_info)),
+			    misalign);
 }
 
 /* Function vect_create_addr_base_for_vector_ref.
@@ -5340,10 +5373,13 @@  vect_setup_realignment (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 	new_temp = copy_ssa_name (ptr);
       else
 	new_temp = make_ssa_name (TREE_TYPE (ptr));
-      unsigned int align = DR_TARGET_ALIGNMENT (dr_info);
+      poly_uint64 align = DR_TARGET_ALIGNMENT (dr_info);
+      tree type = TREE_TYPE (ptr);
       new_stmt = gimple_build_assign
 		   (new_temp, BIT_AND_EXPR, ptr,
-		    build_int_cst (TREE_TYPE (ptr), -(HOST_WIDE_INT) align));
+		    fold_build2 (MINUS_EXPR, type,
+				 build_int_cst (type, 0),
+				 build_int_cst (type, align)));
       new_bb = gsi_insert_on_edge_immediate (pe, new_stmt);
       gcc_assert (!new_bb);
       data_ref
@@ -6226,7 +6262,7 @@  vect_record_grouped_load_vectors (stmt_vec_info stmt_info,
    on ALIGNMENT bit boundary.  */
 
 bool
-vect_can_force_dr_alignment_p (const_tree decl, unsigned int alignment)
+vect_can_force_dr_alignment_p (const_tree decl, poly_uint64 alignment)
 {
   if (!VAR_P (decl))
     return false;
@@ -6236,9 +6272,9 @@  vect_can_force_dr_alignment_p (const_tree decl, unsigned int alignment)
     return false;
 
   if (TREE_STATIC (decl))
-    return (alignment <= MAX_OFILE_ALIGNMENT);
+    return (known_le (alignment, MAX_OFILE_ALIGNMENT));
   else
-    return (alignment <= MAX_STACK_ALIGNMENT);
+    return (known_le (alignment, (unsigned HOST_WIDE_INT) MAX_STACK_ALIGNMENT));
 }
 
 
diff --git a/gcc/tree-vect-loop-manip.c b/gcc/tree-vect-loop-manip.c
index 1d1d1147696fefebf94fb1b844af65b8088c1ecb..71e9397a166d0818cdf4fdc6fddc3465be787ea7 100644
--- a/gcc/tree-vect-loop-manip.c
+++ b/gcc/tree-vect-loop-manip.c
@@ -1561,8 +1561,9 @@  get_misalign_in_elems (gimple **seq, loop_vec_info loop_vinfo)
   stmt_vec_info stmt_info = dr_info->stmt;
   tree vectype = STMT_VINFO_VECTYPE (stmt_info);
 
-  unsigned int target_align = DR_TARGET_ALIGNMENT (dr_info);
-  gcc_assert (target_align != 0);
+  poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info);
+  unsigned HOST_WIDE_INT target_align_c;
+  tree target_align_minus_1;
 
   bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr),
 					size_zero_node) < 0;
@@ -1572,7 +1573,18 @@  get_misalign_in_elems (gimple **seq, loop_vec_info loop_vinfo)
   tree start_addr = vect_create_addr_base_for_vector_ref (stmt_info, seq,
 							  offset);
   tree type = unsigned_type_for (TREE_TYPE (start_addr));
-  tree target_align_minus_1 = build_int_cst (type, target_align - 1);
+  if (target_align.is_constant (&target_align_c))
+    target_align_minus_1 = build_int_cst (type, target_align_c - 1);
+  else
+    {
+      tree vla = build_int_cst (type, target_align);
+      tree vla_align = fold_build2 (BIT_AND_EXPR, type, vla,
+				    fold_build2 (MINUS_EXPR, type,
+						 build_int_cst (type, 0), vla));
+      target_align_minus_1 = fold_build2 (MINUS_EXPR, type, vla_align,
+					  build_int_cst (type, 1));
+    }
+
   HOST_WIDE_INT elem_size
     = int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
   tree elem_size_log = build_int_cst (type, exact_log2 (elem_size));
@@ -1631,7 +1643,7 @@  vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo,
   tree iters, iters_name;
   stmt_vec_info stmt_info = dr_info->stmt;
   tree vectype = STMT_VINFO_VECTYPE (stmt_info);
-  unsigned int target_align = DR_TARGET_ALIGNMENT (dr_info);
+  poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info);
 
   if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0)
     {
@@ -1650,8 +1662,12 @@  vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo,
       tree type = TREE_TYPE (misalign_in_elems);
       HOST_WIDE_INT elem_size
 	= int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
-      HOST_WIDE_INT align_in_elems = target_align / elem_size;
-      tree align_in_elems_minus_1 = build_int_cst (type, align_in_elems - 1);
+      /* We only do prolog peeling if the target alignment is known at compile
+         time.  */
+      poly_uint64 align_in_elems =
+	exact_div (target_align, elem_size);
+      tree align_in_elems_minus_1 =
+	build_int_cst (type, align_in_elems - 1);
       tree align_in_elems_tree = build_int_cst (type, align_in_elems);
 
       /* Create:  (niters_type) ((align_in_elems - misalign_in_elems)
@@ -1666,7 +1682,11 @@  vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo,
 			     misalign_in_elems);
       iters = fold_build2 (BIT_AND_EXPR, type, iters, align_in_elems_minus_1);
       iters = fold_convert (niters_type, iters);
-      *bound = align_in_elems - 1;
+      unsigned HOST_WIDE_INT align_in_elems_c;
+      if (align_in_elems.is_constant (&align_in_elems_c))
+	*bound = align_in_elems_c - 1;
+      else
+	*bound = -1;
     }
 
   if (dump_enabled_p ())
@@ -2404,6 +2424,13 @@  vect_do_peeling (loop_vec_info loop_vinfo, tree niters, tree nitersm1,
   profile_probability prob_prolog, prob_vector, prob_epilog;
   int estimated_vf;
   int prolog_peeling = 0;
+  /* We currently do not support prolog peeling if the target alignment is not
+     known at compile time.  'vect_gen_prolog_loop_niters' depends on the
+     target alignment being constant.  */
+  dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
+  if (dr_info && !DR_TARGET_ALIGNMENT (dr_info).is_constant ())
+    return NULL;
+
   if (!vect_use_loop_mask_for_alignment_p (loop_vinfo))
     prolog_peeling = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo);
 
diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
index 8108d52a3cb97bcbb4174a5b6b7b61947030b5ae..e1d2c6c8bd3441ebab30459b38d1c7cb134bb320 100644
--- a/gcc/tree-vect-stmts.c
+++ b/gcc/tree-vect-stmts.c
@@ -6131,8 +6131,10 @@  ensure_base_align (dr_vec_info *dr_info)
     {
       tree base_decl = dr_info->base_decl;
 
-      unsigned int align_base_to
-	= DR_TARGET_ALIGNMENT (dr_info) * BITS_PER_UNIT;
+      // We should only be able to increase the alignment of a base object if
+      // we know what its new alignment should be at compile time.
+      unsigned HOST_WIDE_INT align_base_to =
+	DR_TARGET_ALIGNMENT (dr_info).to_constant () * BITS_PER_UNIT;
 
       if (decl_in_symtab_p (base_decl))
 	symtab_node::get (base_decl)->increase_alignment (align_base_to);
@@ -7080,7 +7082,8 @@  vectorizable_store (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 	  stmt_vec_info next_stmt_info = first_stmt_info;
 	  for (i = 0; i < vec_num; i++)
 	    {
-	      unsigned align, misalign;
+	      unsigned misalign;
+	      unsigned HOST_WIDE_INT align;
 
 	      tree final_mask = NULL_TREE;
 	      if (loop_masks)
@@ -7121,7 +7124,7 @@  vectorizable_store (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 		   vect_permute_store_chain().  */
 		vec_oprnd = result_chain[i];
 
-	      align = DR_TARGET_ALIGNMENT (first_dr_info);
+	      align = known_alignment (DR_TARGET_ALIGNMENT (first_dr_info));
 	      if (aligned_access_p (first_dr_info))
 		misalign = 0;
 	      else if (DR_MISALIGNMENT (first_dr_info) == -1)
@@ -8309,7 +8312,8 @@  vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 		case dr_aligned:
 		case dr_unaligned_supported:
 		  {
-		    unsigned int align, misalign;
+		    unsigned int misalign;
+		    unsigned HOST_WIDE_INT align;
 
 		    if (memory_access_type == VMAT_GATHER_SCATTER)
 		      {
@@ -8329,7 +8333,8 @@  vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 			break;
 		      }
 
-		    align = DR_TARGET_ALIGNMENT (dr_info);
+		    align =
+		      known_alignment (DR_TARGET_ALIGNMENT (first_dr_info));
 		    if (alignment_support_scheme == dr_aligned)
 		      {
 			gcc_assert (aligned_access_p (first_dr_info));
@@ -8396,7 +8401,10 @@  vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 		      ptr = copy_ssa_name (dataref_ptr);
 		    else
 		      ptr = make_ssa_name (TREE_TYPE (dataref_ptr));
-		    unsigned int align = DR_TARGET_ALIGNMENT (first_dr_info);
+		    // For explicit realign the target alignment should be
+		    // known at compile time.
+		    unsigned HOST_WIDE_INT align =
+		      DR_TARGET_ALIGNMENT (first_dr_info).to_constant ();
 		    new_stmt = gimple_build_assign
 				 (ptr, BIT_AND_EXPR, dataref_ptr,
 				  build_int_cst
@@ -8440,7 +8448,10 @@  vectorizable_load (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
 		      new_temp = copy_ssa_name (dataref_ptr);
 		    else
 		      new_temp = make_ssa_name (TREE_TYPE (dataref_ptr));
-		    unsigned int align = DR_TARGET_ALIGNMENT (first_dr_info);
+		    // We should only be doing this if we know the target
+		    // alignment at compile time.
+		    unsigned HOST_WIDE_INT align =
+		      DR_TARGET_ALIGNMENT (first_dr_info).to_constant ();
 		    new_stmt = gimple_build_assign
 		      (new_temp, BIT_AND_EXPR, dataref_ptr,
 		       build_int_cst (TREE_TYPE (dataref_ptr),
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index 988456808318dabd0058f6b0d038f8c272e75c6b..d649108593fe8409276e3ac13c7f3fbe366eb3c4 100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -773,7 +773,7 @@  struct dr_vec_info {
   int misalignment;
   /* The byte alignment that we'd ideally like the reference to have,
      and the value that misalignment is measured against.  */
-  int target_alignment;
+  poly_uint64 target_alignment;
   /* If true the alignment of base_decl needs to be increased.  */
   bool base_misaligned;
   tree base_decl;
@@ -1299,7 +1299,7 @@  vect_known_alignment_in_bytes (dr_vec_info *dr_info)
   if (DR_MISALIGNMENT (dr_info) == DR_MISALIGNMENT_UNKNOWN)
     return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info->dr)));
   if (DR_MISALIGNMENT (dr_info) == 0)
-    return DR_TARGET_ALIGNMENT (dr_info);
+    return known_alignment (DR_TARGET_ALIGNMENT (dr_info));
   return DR_MISALIGNMENT (dr_info) & -DR_MISALIGNMENT (dr_info);
 }
 
@@ -1518,7 +1518,7 @@  extern opt_result vect_get_vector_types_for_stmt (stmt_vec_info, tree *,
 extern opt_tree vect_get_mask_type_for_stmt (stmt_vec_info);
 
 /* In tree-vect-data-refs.c.  */
-extern bool vect_can_force_dr_alignment_p (const_tree, unsigned int);
+extern bool vect_can_force_dr_alignment_p (const_tree, poly_uint64);
 extern enum dr_alignment_support vect_supportable_dr_alignment
                                            (dr_vec_info *, bool);
 extern tree vect_get_smallest_scalar_type (stmt_vec_info, HOST_WIDE_INT *,