On Mon, 9 Oct 2023 at 17:05, Richard Sandiford
<richard.sandif...@arm.com> wrote:
>
> Prathamesh Kulkarni <prathamesh.kulka...@linaro.org> writes:
> > Hi,
> > The attached patch attempts to fix PR111648.
> > As mentioned in PR, the issue is when a1 is a multiple of vector
> > length, we end up creating following encoding in result: { base_elem,
> > arg[0], arg[1], ... } (assuming S = 1),
> > where arg is chosen input vector, which is incorrect, since the
> > encoding originally in arg would be: { arg[0], arg[1], arg[2], ... }
> >
> > For the test-case mentioned in PR, vectorizer pass creates
> > VEC_PERM_EXPR<arg0, arg, sel> where:
> > arg0: { -16, -9, -10, -11 }
> > arg1: { -12, -5, -6, -7 }
> > sel = { 3, 4, 5, 6 }
> >
> > arg0, arg1 and sel are encoded with npatterns = 1 and nelts_per_pattern = 3.
> > Since a1 = 4 and arg_len = 4, it ended up creating the result with
> > following encoding:
> > res = { arg0[3], arg1[0], arg1[1] } // npatterns = 1, nelts_per_pattern = 3
> > = { -11, -12, -5 }
> >
> > So for res[3], it used S = (-5) - (-12) = 7
> > And hence computed it as -5 + 7 = 2.
> > instead of selecting arg1[2], ie, -6.
> >
> > The patch tweaks valid_mask_for_fold_vec_perm_cst_p to punt if a1 is a
> > multiple
> > of vector length, so a1 ... ae select elements only from stepped part
> > of the pattern
> > from input vector and return false for this case.
> >
> > Since the vectors are VLS, fold_vec_perm_cst then sets:
> > res_npatterns = res_nelts
> > res_nelts_per_pattern = 1
> > which seems to fix the issue by encoding all the elements.
> >
> > The patch resulted in Case 4 and Case 5 failing from test_nunits_min_2
> > because
> > they used sel = { 0, 0, 1, ... } and {len, 0, 1, ... } respectively,
> > which used a1 = 0, and thus selected arg1[0].
> >
> > I removed Case 4 because it was already covered in test_nunits_min_4,
> > and moved Case 5 to test_nunits_min_4, with sel = { len, 1, 2, ... }
> > and added a new Case 9 to test for this issue.
> >
> > Passes bootstrap+test on aarch64-linux-gnu with and without SVE,
> > and on x86_64-linux-gnu.
> > Does the patch look OK ?
> >
> > Thanks,
> > Prathamesh
> >
> > [PR111648] Fix wrong code-gen due to incorrect VEC_PERM_EXPR folding.
> >
> > gcc/ChangeLog:
> > PR tree-optimization/111648
> > * fold-const.cc (valid_mask_for_fold_vec_perm_cst_p): Punt if a1
> > is a multiple of vector length.
> > (test_nunits_min_2): Remove Case 4 and move Case 5 to ...
> > (test_nunits_min_4): ... here and rename case numbers. Also add
> > Case 9.
> >
> > gcc/testsuite/ChangeLog:
> > PR tree-optimization/111648
> > * gcc.dg/vect/pr111648.c: New test.
> >
> >
> > diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc
> > index 4f8561509ff..c5f421d6b76 100644
> > --- a/gcc/fold-const.cc
> > +++ b/gcc/fold-const.cc
> > @@ -10682,8 +10682,8 @@ valid_mask_for_fold_vec_perm_cst_p (tree arg0, tree
> > arg1,
> > return false;
> > }
> >
> > - /* Ensure that the stepped sequence always selects from the same
> > - input pattern. */
> > + /* Ensure that the stepped sequence always selects from the stepped
> > + part of same input pattern. */
> > unsigned arg_npatterns
> > = ((q1 & 1) == 0) ? VECTOR_CST_NPATTERNS (arg0)
> > : VECTOR_CST_NPATTERNS (arg1);
> > @@ -10694,6 +10694,20 @@ valid_mask_for_fold_vec_perm_cst_p (tree arg0,
> > tree arg1,
> > *reason = "step is not multiple of npatterns";
> > return false;
> > }
> > +
> > + /* If a1 is a multiple of len, it will select base element of input
> > + vector resulting in following encoding:
> > + { base_elem, arg[0], arg[1], ... } where arg is the chosen input
> > + vector. This encoding is not originally present in arg, since it's
> > + defined as:
> > + { arg[0], arg[1], arg[2], ... }. */
> > +
> > + if (multiple_p (a1, arg_len))
> > + {
> > + if (reason)
> > + *reason = "selecting base element of input vector";
> > + return false;
> > + }
>
> That wouldn't catch (for example) cases where a1 == arg_len + 1 and the
> second argument has 2 stepped patterns.
Ah right, thanks for pointing out. In the attached patch I extended the check
so that r1 < arg_npatterns which should check if we are choosing base
elements from any of the patterns in arg (and not just first).
Does that look OK ?
>
> The equivalent condition that handles multiple patterns would
> probably be to reject q1 < arg_npatterns. But that's only necessary if:
Sorry, I don't understand -- we use q1 only for determining which
vector to choose from,
and r1 will give the index for first element ?
>
> (1) the argument has three elements per pattern (i.e. has a stepped
> sequence) and
>
> (2) element 2 - element 1 != element 1 - element 0
>
> I think we should check those to avoid pessimising VLA cases.
Thanks for the suggestions. In attached POC patch (stage-1 tested), I
added the above checks,
does it look in the right direction ? Also, should this patch be the
right fix for PR111754 ?
Thanks,
Prathamesh
>
> Thanks,
> Richard
>
> > }
> >
> > return true;
> > @@ -17425,47 +17439,6 @@ test_nunits_min_2 (machine_mode vmode)
> > tree expected_res[] = { ARG0(0), ARG1(0), ARG0(1), ARG1(1) };
> > validate_res (2, 2, res, expected_res);
> > }
> > -
> > - /* Case 4: mask = {0, 0, 1, ...} // (1, 3)
> > - Test that the stepped sequence of the pattern selects from
> > - same input pattern. Since input vectors have npatterns = 2,
> > - and step (a2 - a1) = 1, step is not a multiple of npatterns
> > - in input vector. So return NULL_TREE. */
> > - {
> > - tree arg0 = build_vec_cst_rand (vmode, 2, 3, 1);
> > - tree arg1 = build_vec_cst_rand (vmode, 2, 3, 1);
> > - poly_uint64 len = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0));
> > -
> > - vec_perm_builder builder (len, 1, 3);
> > - poly_uint64 mask_elems[] = { 0, 0, 1 };
> > - builder_push_elems (builder, mask_elems);
> > -
> > - vec_perm_indices sel (builder, 2, len);
> > - const char *reason;
> > - tree res = fold_vec_perm_cst (TREE_TYPE (arg0), arg0, arg1, sel,
> > - &reason);
> > - ASSERT_TRUE (res == NULL_TREE);
> > - ASSERT_TRUE (!strcmp (reason, "step is not multiple of npatterns"));
> > - }
> > -
> > - /* Case 5: mask = {len, 0, 1, ...} // (1, 3)
> > - Test that stepped sequence of the pattern selects from arg0.
> > - res = { arg1[0], arg0[0], arg0[1], ... } // (1, 3) */
> > - {
> > - tree arg0 = build_vec_cst_rand (vmode, 1, 3, 1);
> > - tree arg1 = build_vec_cst_rand (vmode, 1, 3, 1);
> > - poly_uint64 len = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0));
> > -
> > - vec_perm_builder builder (len, 1, 3);
> > - poly_uint64 mask_elems[] = { len, 0, 1 };
> > - builder_push_elems (builder, mask_elems);
> > -
> > - vec_perm_indices sel (builder, 2, len);
> > - tree res = fold_vec_perm_cst (TREE_TYPE (arg0), arg0, arg1, sel);
> > -
> > - tree expected_res[] = { ARG1(0), ARG0(0), ARG0(1) };
> > - validate_res (1, 3, res, expected_res);
> > - }
> > }
> > }
> >
> > @@ -17528,7 +17501,26 @@ test_nunits_min_4 (machine_mode vmode)
> > validate_res (1, 3, res, expected_res);
> > }
> >
> > - /* Case 4:
> > + /* Case 4: mask = {len, 1, 2, ...} // (1, 3)
> > + Test that stepped sequence of the pattern selects from arg0.
> > + res = { arg1[0], arg0[1], arg0[2], ... } // (1, 3) */
> > + {
> > + tree arg0 = build_vec_cst_rand (vmode, 1, 3, 1);
> > + tree arg1 = build_vec_cst_rand (vmode, 1, 3, 1);
> > + poly_uint64 len = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0));
> > +
> > + vec_perm_builder builder (len, 1, 3);
> > + poly_uint64 mask_elems[] = { len, 1, 2 };
> > + builder_push_elems (builder, mask_elems);
> > +
> > + vec_perm_indices sel (builder, 2, len);
> > + tree res = fold_vec_perm_cst (TREE_TYPE (arg0), arg0, arg1, sel);
> > +
> > + tree expected_res[] = { ARG1(0), ARG0(1), ARG0(2) };
> > + validate_res (1, 3, res, expected_res);
> > + }
> > +
> > + /* Case 5:
> > sel = { len, 0, 2, ... } // (1, 3)
> > This should return NULL because we cross the input vectors.
> > Because,
> > @@ -17561,7 +17553,7 @@ test_nunits_min_4 (machine_mode vmode)
> > ASSERT_TRUE (!strcmp (reason, "crossed input vectors"));
> > }
> >
> > - /* Case 5: npatterns(arg0) = 4 > npatterns(sel) = 2
> > + /* Case 6: npatterns(arg0) = 4 > npatterns(sel) = 2
> > mask = { 0, len, 1, len + 1, ...} // (2, 2)
> > res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2)
> >
> > @@ -17583,7 +17575,7 @@ test_nunits_min_4 (machine_mode vmode)
> > validate_res (2, 2, res, expected_res);
> > }
> >
> > - /* Case 6: Test combination in sel, where one pattern is dup and
> > other
> > + /* Case 7: Test combination in sel, where one pattern is dup and
> > other
> > is stepped sequence.
> > sel = { 0, 0, 0, 1, 0, 2, ... } // (2, 3)
> > res = { arg0[0], arg0[0], arg0[0],
> > @@ -17605,7 +17597,7 @@ test_nunits_min_4 (machine_mode vmode)
> > validate_res (2, 3, res, expected_res);
> > }
> >
> > - /* Case 7: PR111048: Check that we set arg_npatterns correctly,
> > + /* Case 8: PR111048: Check that we set arg_npatterns correctly,
> > when arg0, arg1 and sel have different number of patterns.
> > arg0 is of shape (1, 1)
> > arg1 is of shape (4, 1)
> > @@ -17634,6 +17626,51 @@ test_nunits_min_4 (machine_mode vmode)
> > ASSERT_TRUE (res == NULL_TREE);
> > ASSERT_TRUE (!strcmp (reason, "step is not multiple of npatterns"));
> > }
> > +
> > + /* Case 9: PR111648 - a1 is multiple of vector length,
> > + which results in incorrect encoding. Verify that we return
> > + NULL for this case.
> > + sel = { base_elem, len, len+1, ... } // (1, 3)
> > + In this case, the single pattern is: { base_elem len, len+1, ...}
> > + Let's assume that base_elem is used for indexing into arg0,
> > + and a1 ... ae chooses elements from arg1.
> > + So res = { arg0[base_elem], arg1[0], arg1[1], ... } // (1, 3)
> > + Which creates an incorrect encoding with S = arg1[1] - arg1[0]
> > + while the original encoding in arg1 is
> > + arg1: { arg1[0], arg1[1], arg1[2], ... }
> > + with S = arg1[2] - arg1[1].
> > +
> > + As a concrete example, for above PR:
> > + arg0: { -16, -9, -10, -11 }
> > + arg1: { -12, -5, -6, -7 }
> > + sel = { 3, 4, 5, 6 }
> > +
> > + arg0, arg1 and sel are encoded with npatterns = 1 and
> > nelts_per_pattern = 3.
> > + Since a1 = 4 and arg_len = 4, it ended up creating the result with
> > + following encoding:
> > + res = { arg0[3], arg1[0], arg1[1] } // (1, 3)
> > + = { -11, -12, -5 }
> > +
> > + So for res[3], it used S = (-5) - (-12) = 7
> > + And hence computed it as -5 + 7 = 2.
> > + instead of arg1[2], ie, -6, which is the correct value.
> > + Ensure that valid_mask_for_fold_vec_perm_cst returns false for this
> > case. */
> > + {
> > + tree arg0 = build_vec_cst_rand (vmode, 1, 3);
> > + tree arg1 = build_vec_cst_rand (vmode, 1, 3);
> > + poly_uint64 len = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0));
> > +
> > + vec_perm_builder builder (len, 1, 3);
> > + poly_uint64 mask_elems[] = { 0, len, len+1 };
> > + builder_push_elems (builder, mask_elems);
> > +
> > + vec_perm_indices sel (builder, 2, len);
> > + const char *reason;
> > + tree res = fold_vec_perm_cst (TREE_TYPE (arg0), arg0, arg1, sel,
> > &reason);
> > + ASSERT_TRUE (res == NULL_TREE);
> > + ASSERT_TRUE (!strcmp (reason,
> > + "selecting base element of input vector"));
> > + }
> > }
> > }
> >
> > diff --git a/gcc/testsuite/gcc.dg/vect/pr111648.c
> > b/gcc/testsuite/gcc.dg/vect/pr111648.c
> > new file mode 100644
> > index 00000000000..093e2b02654
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/vect/pr111648.c
> > @@ -0,0 +1,23 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O2 -fdump-tree-optimized" } */
> > +
> > +int a;
> > +int *b = &a;
> > +static int **c = &b;
> > +static int d;
> > +short e;
> > +short f;
> > +
> > +_Bool foo ()
> > +{
> > + f = -21;
> > + for (; f < -5; f++) {
> > + e = f ^ 3;
> > + d = *b;
> > + **c = e;
> > + }
> > +
> > + return d == -6;
> > +}
> > +
> > +/* { dg-final { scan-tree-dump "return 1" "optimized" } } */
diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc
index 4f8561509ff..ae914cbd880 100644
--- a/gcc/fold-const.cc
+++ b/gcc/fold-const.cc
@@ -10607,6 +10607,25 @@ vec_cst_ctor_to_array (tree arg, unsigned int nelts,
tree *elts)
return true;
}
+static poly_int64
+vector_cst_elt_poly_index (tree arg, poly_uint64 index)
+{
+ int q;
+ poly_uint64 r;
+ poly_uint64 len = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg));
+
+ if (!can_div_trunc_p (index, len, &q, &r))
+ return NULL_TREE;
+
+ unsigned HOST_WIDE_INT i;
+ if (!r.is_constant (&i))
+ return NULL_TREE;
+
+ tree elem = vector_cst_elt (arg, i);
+ gcc_assert (elem != NULL_TREE);
+ return tree_to_poly_int64 (elem);
+}
+
/* Helper routine for fold_vec_perm_cst to check if SEL is a suitable
mask for VLA vec_perm folding.
REASON if specified, will contain the reason why SEL is not suitable.
@@ -10684,9 +10703,8 @@ valid_mask_for_fold_vec_perm_cst_p (tree arg0, tree
arg1,
/* Ensure that the stepped sequence always selects from the same
input pattern. */
- unsigned arg_npatterns
- = ((q1 & 1) == 0) ? VECTOR_CST_NPATTERNS (arg0)
- : VECTOR_CST_NPATTERNS (arg1);
+ tree arg = ((q1 & 1) == 0) ? arg0 : arg1;
+ unsigned arg_npatterns = VECTOR_CST_NPATTERNS (arg);
if (!multiple_p (step, arg_npatterns))
{
@@ -10694,6 +10712,21 @@ valid_mask_for_fold_vec_perm_cst_p (tree arg0, tree
arg1,
*reason = "step is not multiple of npatterns";
return false;
}
+
+ /* Ensure that a1 only selects from stepped part of the pattern from arg,
+ if the pattern is not a natural stepped sequence, ie,
+ ((a2 - a1) != (a1 - a0)). */
+
+ poly_int64 arg_elem0 = vector_cst_elt_poly_index (arg, sel[pattern]);
+ poly_int64 arg_elem1 = vector_cst_elt_poly_index (arg, a1);
+ poly_int64 arg_elem2 = vector_cst_elt_poly_index (arg, a2);
+ if (!known_eq (arg_elem2 - arg_elem1, arg_elem1 - arg_elem0)
+ && known_lt (r1, arg_npatterns))
+ {
+ if (reason)
+ *reason = "choosing base element from input vector";
+ return false;
+ }
}
return true;
