On Mon, Aug 1, 2022 at 5:17 AM Prathamesh Kulkarni
<prathamesh.kulka...@linaro.org> wrote:
>
> On Thu, 21 Jul 2022 at 12:21, Richard Biener <richard.guent...@gmail.com>
> wrote:
> >
> > On Wed, Jul 20, 2022 at 5:36 PM Prathamesh Kulkarni
> > <prathamesh.kulka...@linaro.org> wrote:
> > >
> > > On Mon, 18 Jul 2022 at 11:57, Richard Biener <richard.guent...@gmail.com>
> > > wrote:
> > > >
> > > > On Fri, Jul 15, 2022 at 3:49 PM Prathamesh Kulkarni
> > > > <prathamesh.kulka...@linaro.org> wrote:
> > > > >
> > > > > On Thu, 14 Jul 2022 at 17:22, Richard Sandiford
> > > > > <richard.sandif...@arm.com> wrote:
> > > > > >
> > > > > > Richard Biener <richard.guent...@gmail.com> writes:
> > > > > > > On Thu, Jul 14, 2022 at 9:55 AM Prathamesh Kulkarni
> > > > > > > <prathamesh.kulka...@linaro.org> wrote:
> > > > > > >>
> > > > > > >> On Wed, 13 Jul 2022 at 12:22, Richard Biener
> > > > > > >> <richard.guent...@gmail.com> wrote:
> > > > > > >> >
> > > > > > >> > On Tue, Jul 12, 2022 at 9:12 PM Prathamesh Kulkarni via
> > > > > > >> > Gcc-patches
> > > > > > >> > <gcc-patches@gcc.gnu.org> wrote:
> > > > > > >> > >
> > > > > > >> > > Hi Richard,
> > > > > > >> > > For the following test:
> > > > > > >> > >
> > > > > > >> > > svint32_t f2(int a, int b, int c, int d)
> > > > > > >> > > {
> > > > > > >> > > int32x4_t v = (int32x4_t) {a, b, c, d};
> > > > > > >> > > return svld1rq_s32 (svptrue_b8 (), &v[0]);
> > > > > > >> > > }
> > > > > > >> > >
> > > > > > >> > > The compiler emits following ICE with -O3 -mcpu=generic+sve:
> > > > > > >> > > foo.c: In function ‘f2’:
> > > > > > >> > > foo.c:4:11: error: non-trivial conversion in
> > > > > > >> > > ‘view_convert_expr’
> > > > > > >> > > 4 | svint32_t f2(int a, int b, int c, int d)
> > > > > > >> > > | ^~
> > > > > > >> > > svint32_t
> > > > > > >> > > __Int32x4_t
> > > > > > >> > > _7 = VIEW_CONVERT_EXPR<__Int32x4_t>(_8);
> > > > > > >> > > during GIMPLE pass: forwprop
> > > > > > >> > > dump file: foo.c.109t.forwprop2
> > > > > > >> > > foo.c:4:11: internal compiler error: verify_gimple failed
> > > > > > >> > > 0xfda04a verify_gimple_in_cfg(function*, bool)
> > > > > > >> > > ../../gcc/gcc/tree-cfg.cc:5568
> > > > > > >> > > 0xe9371f execute_function_todo
> > > > > > >> > > ../../gcc/gcc/passes.cc:2091
> > > > > > >> > > 0xe93ccb execute_todo
> > > > > > >> > > ../../gcc/gcc/passes.cc:2145
> > > > > > >> > >
> > > > > > >> > > This happens because, after folding svld1rq_s32 to
> > > > > > >> > > vec_perm_expr, we have:
> > > > > > >> > > int32x4_t v;
> > > > > > >> > > __Int32x4_t _1;
> > > > > > >> > > svint32_t _9;
> > > > > > >> > > vector(4) int _11;
> > > > > > >> > >
> > > > > > >> > > <bb 2> :
> > > > > > >> > > _1 = {a_3(D), b_4(D), c_5(D), d_6(D)};
> > > > > > >> > > v_12 = _1;
> > > > > > >> > > _11 = v_12;
> > > > > > >> > > _9 = VEC_PERM_EXPR <_11, _11, { 0, 1, 2, 3, ... }>;
> > > > > > >> > > return _9;
> > > > > > >> > >
> > > > > > >> > > During forwprop, simplify_permutation simplifies
> > > > > > >> > > vec_perm_expr to
> > > > > > >> > > view_convert_expr,
> > > > > > >> > > and the end result becomes:
> > > > > > >> > > svint32_t _7;
> > > > > > >> > > __Int32x4_t _8;
> > > > > > >> > >
> > > > > > >> > > ;; basic block 2, loop depth 0
> > > > > > >> > > ;; pred: ENTRY
> > > > > > >> > > _8 = {a_2(D), b_3(D), c_4(D), d_5(D)};
> > > > > > >> > > _7 = VIEW_CONVERT_EXPR<__Int32x4_t>(_8);
> > > > > > >> > > return _7;
> > > > > > >> > > ;; succ: EXIT
> > > > > > >> > >
> > > > > > >> > > which causes the error duing verify_gimple since
> > > > > > >> > > VIEW_CONVERT_EXPR
> > > > > > >> > > has incompatible types (svint32_t, int32x4_t).
> > > > > > >> > >
> > > > > > >> > > The attached patch disables simplification of VEC_PERM_EXPR
> > > > > > >> > > in simplify_permutation, if lhs and rhs have non compatible
> > > > > > >> > > types,
> > > > > > >> > > which resolves ICE, but am not sure if it's the correct
> > > > > > >> > > approach ?
> > > > > > >> >
> > > > > > >> > It for sure papers over the issue. I think the error happens
> > > > > > >> > earlier,
> > > > > > >> > the V_C_E should have been built with the type of the
> > > > > > >> > VEC_PERM_EXPR
> > > > > > >> > which is the type of the LHS. But then you probably run into
> > > > > > >> > the
> > > > > > >> > different sizes ICE (VLA vs constant size). I think for this
> > > > > > >> > case you
> > > > > > >> > want a BIT_FIELD_REF instead of a VIEW_CONVERT_EXPR,
> > > > > > >> > selecting the "low" part of the VLA vector.
> > > > > > >> Hi Richard,
> > > > > > >> Sorry I don't quite follow. In this case, we use VEC_PERM_EXPR to
> > > > > > >> represent dup operation
> > > > > > >> from fixed width to VLA vector. I am not sure how folding it to
> > > > > > >> BIT_FIELD_REF will work.
> > > > > > >> Could you please elaborate ?
> > > > > > >>
> > > > > > >> Also, the issue doesn't seem restricted to this case.
> > > > > > >> The following test case also ICE's during forwprop:
> > > > > > >> svint32_t foo()
> > > > > > >> {
> > > > > > >> int32x4_t v = (int32x4_t) {1, 2, 3, 4};
> > > > > > >> svint32_t v2 = svld1rq_s32 (svptrue_b8 (), &v[0]);
> > > > > > >> return v2;
> > > > > > >> }
> > > > > > >>
> > > > > > >> foo2.c: In function ‘foo’:
> > > > > > >> foo2.c:9:1: error: non-trivial conversion in ‘vector_cst’
> > > > > > >> 9 | }
> > > > > > >> | ^
> > > > > > >> svint32_t
> > > > > > >> int32x4_t
> > > > > > >> v2_4 = { 1, 2, 3, 4 };
> > > > > > >>
> > > > > > >> because simplify_permutation folds
> > > > > > >> VEC_PERM_EXPR< {1, 2, 3, 4}, {1, 2, 3, 4}, {0, 1, 2, 3, ...} >
> > > > > > >> into:
> > > > > > >> vector_cst {1, 2, 3, 4}
> > > > > > >>
> > > > > > >> and it complains during verify_gimple_assign_single because we
> > > > > > >> don't
> > > > > > >> support assignment of vector_cst to VLA vector.
> > > > > > >>
> > > > > > >> I guess the issue really is that currently, only VEC_PERM_EXPR
> > > > > > >> supports lhs and rhs
> > > > > > >> to have vector types with differing lengths, and simplifying it
> > > > > > >> to
> > > > > > >> other tree codes, like above,
> > > > > > >> will result in type errors ?
> > > > > > >
> > > > > > > That might be the case - Richard should know.
> > > > > >
> > > > > > I don't see anything particularly special about VEC_PERM_EXPR here,
> > > > > > or about the VLA vs. VLS thing. We would have the same issue trying
> > > > > > to build a 128-bit vector from 2 64-bit vectors. And there are
> > > > > > other
> > > > > > tree codes whose input types are/can be different from their output
> > > > > > types.
> > > > > >
> > > > > > So it just seems like a normal type correctness issue: a
> > > > > > VEC_PERM_EXPR
> > > > > > of type T needs to be replaced by something of type T. Whether T
> > > > > > has a
> > > > > > constant size or a variable size doesn't matter.
> > > > > >
> > > > > > > If so your type check
> > > > > > > is still too late, you should instead recognize that we are
> > > > > > > permuting
> > > > > > > a VLA vector and then refuse to go any of the non-VEC_PERM
> > > > > > > generating
> > > > > > > paths - that probably means just allowing the code ==
> > > > > > > VEC_PERM_EXPR
> > > > > > > case and not any of the CTOR/CST/VIEW_CONVERT_EXPR cases?
> > > > > >
> > > > > > Yeah. If we're talking about the match.pd code, I think only:
> > > > > >
> > > > > > (if (sel.series_p (0, 1, 0, 1))
> > > > > > { op0; }
> > > > > > (if (sel.series_p (0, 1, nelts, 1))
> > > > > > { op1; }
> > > > > >
> > > > > > need a type compatibility check. For fold_vec_perm I think
> > > > > > we should just rearrange:
> > > > > >
> > > > > > gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (type), nelts)
> > > > > > && known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0)),
> > > > > > nelts)
> > > > > > && known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg1)),
> > > > > > nelts));
> > > > > > if (TREE_TYPE (TREE_TYPE (arg0)) != TREE_TYPE (type)
> > > > > > || TREE_TYPE (TREE_TYPE (arg1)) != TREE_TYPE (type))
> > > > > > return NULL_TREE;
> > > > > >
> > > > > > so that the assert comes after the early-out.
> > > > > >
> > > > > > It would be good at some point to relax fold_vec_perm to cases
> > > > > > where the
> > > > > > outputs are a different length from the inputs, since the
> > > > > > all-constant
> > > > > > VEC_PERM_EXPR above could be folded to a VECTOR_CST.
> > > > > Hi,
> > > > > For the above case, I think the issue is that simplify_permutation
> > > > > uses TREE_TYPE (arg0) for res_type,
> > > > > while it should now use type for lhs.
> > > > >
> > > > > /* Shuffle of a constructor. */
> > > > > bool ret = false;
> > > > > tree res_type = TREE_TYPE (arg0);
> > > > > tree opt = fold_ternary (VEC_PERM_EXPR, res_type, arg0, arg1,
> > > > > op2);
> > > > >
> > > > > Using, res_type = TREE_TYPE (gimple_get_lhs (stmt)),
> > > > > resolves the ICE, and emits all constant VEC_PERM_EXPR:
> > > > >
> > > > > v2_4 = VEC_PERM_EXPR <{ 1, 2, 3, 4 }, { 1, 2, 3, 4 }, { 0, 1, 2, 3,
> > > > > ... }>;
> > > > > return v2_4;
> > > > >
> > > > > Does the patch look OK to commit after bootstrap+test ?
> > > >
> > > > Ok with using gimple_assign_lhs (stmt) instead of gimple_get_lhs (stmt).
> > > Hi,
> > > I committed the patch but unfortunately it caused PR106360.
> > > The issue is that for slp-reduc-sad-2.c on ppc64le,
> > > simplify_permutation sees the following during forwprop4:
> > >
> > > _78 = (void *) ivtmp.21_73;
> > > _92 = MEM <unsigned long> [(uint8_t *)_78];
> > > _91 = {_92, 0};
> > > vect__1.6_90 = VIEW_CONVERT_EXPR<vector(16) unsigned char>(_91);
> > > _88 = MEM <unsigned long> [(uint8_t *)_78 + 16B];
> > > _87 = {_88, 0};
> > > vect__1.7_86 = VIEW_CONVERT_EXPR<vector(16) unsigned char>(_87);
> > > vect__1.8_85 = VEC_PERM_EXPR <vect__1.6_90, vect__1.7_86, { 0, 1, 2,
> > > 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23 }>;
> > >
> > > So for,
> > > tree res_type = TREE_TYPE (gimple_assign_lhs (stmt));
> > > tree opt = fold_ternary (VEC_PERM_EXPR, res_type, arg0, arg1, op2);
> > >
> > > we have:
> > > res_type = V16QI
> > > arg0 = {_92, 0}
> > > arg1 = {_88, 0}
> > > op2 = {0, 2}
> > >
> > > and thus we hit the following assert in fold_vec_perm:
> > >
> > > gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (type), nelts)
> > > && known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0)), nelts)
> > > && known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg1)),
> > > nelts));
> > >
> > > since nelts == 2, and TYPE_VECTOR_SUBPARTS (type) == 16.
> > >
> > > If we revert the committed patch so we pass res_type = TREE_TYPE (arg0)
> > > instead,
> > > it simplifies the above VEC_PERM_EXPR to VIEW_CONVERT_EXPR:
> > > _92 = MEM <unsigned long> [(uint8_t *)_78];
> > > _88 = MEM <unsigned long> [(uint8_t *)_78 + 16B];
> > > _5 = {_92, _88};
> > > vect__1.8_85 = VIEW_CONVERT_EXPR<vector(16) unsigned char>(_5);
> > >
> > > I suppose it's legal to cast vector of one type to another as long as
> > > sizes match ?
> > >
> > > IIUC, the above VIEW_CONVERT_EXPR will result in:
> > > vect__1.8_85 = { (uint8_t) _92, 0, 0, 0, 0, 0, 0, 0, (uint8_t) _88, 0,
> > > 0, 0, 0, 0, 0, 0 } ?
> > >
> > > In the attached patch, it restores res_type to TREE_TYPE (arg0), and
> > > checks
> > > if lhs_type and res_type differ but have same size, and in that case emit:
> > > lhs = VIEW_CONVERT_EXPR<lhs_type> (opt),
> > > instead of:
> > > lhs = VIEW_CONVERT_EXPR<op1 type> (opt)
> > > where opt is result of folding VEC_PERM_EXPR<res_type, arg0, arg1, op2>
> > >
> > > Does it look OK ?
> >
> > Definitely the original change was bogus.
> >
> > + if (!operand_equal_p (TYPE_SIZE (lhs_type), TYPE_SIZE (res_type)))
> > + return 0;
> >
> > just repeats your very original change though ... I'll note that
> > fold_ternary will
> > ICE on now valid VEC_PERM_EXPRs so we should fix it, possibly by
> > returning NULL_TREE on cases it does not handle.
> >
> > I think what should be done is, in the
> >
> > /* If there are any VIEW_CONVERT_EXPRs found when finding permutation
> > operands source, check whether it's valid to transform and prepare
> > the required new operands. */
> > if (code == VIEW_CONVERT_EXPR || code2 == VIEW_CONVERT_EXPR)
> > {
> > ...
> >
> > path also transform the expected result type. It should remain V_C_E
> > compatible
> > to TREE_TYPE (lhs) but get a new element type.
> >
> > But as said,
> >
> > tree
> > fold_vec_perm (tree type, tree arg0, tree arg1, const vec_perm_indices &sel)
> > {
> > unsigned int i;
> > unsigned HOST_WIDE_INT nelts;
> > bool need_ctor = false;
> >
> > if (!sel.length ().is_constant (&nelts))
> > return NULL_TREE;
> > gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (type), nelts)
> > && known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0)), nelts)
> > && known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg1)), nelts));
> >
> > ^^^ this doesn't match what we allow for VEC_PERM_EXPRs now and fold_ternary
> > doesn't guard according to those asserts (I think we should extend
> > fold_vec_perm
> > to support the new constraints).
> Hi Richard,
> Thanks for the suggestions and sorry for late reply. I reverted the change to
> simplify_permutaton which resolved the ppc64le case ICE.
>
> The attached patch extends fold_vec_perm to handle vectors with
> differing lengths.
> For,
> lhs = vec_perm_expr<arg0, arg1, mask>,
> the patch:
> (a) asserts lhs and mask have same vector length.
> (b) asserts arg0 and arg1 have same vector length.
> (c) returns NULL_TREE if element type differs for lhs, arg0 and arg1.
> (d) if len(lhs) > len(arg0), then the patch allows permuting arg0, arg1
> if the mask has npatterns == len(arg0) and nelts_per_pattern == 1.
> The intent is to permute arg0 and arg1, and then to dup elements in result
> to target vector length.
> So for eg:
> vec_perm_expr< {1, 2, 3, 4}, {5, 6, 7, 8}, {0, 4, 1, 5, ...}>
> will result in vla vector {1, 5, 2, 6, ....} with {1, 5, 2, 6}
> replicated thru-out.
> Does it look OK ?
>
> With the patch, we don't ICE for either of the aarch64 tests above.
> For,
> svint32_t f1()
> {
> int32x4_t v = {1, 2, 3, 4};
> return svld1rq_s32 (svptrue_b8 (), &v[0]);
> }
>
> optimized dump shows:
> svint32_t f1 ()
> {
> int32x4_t v;
>
> <bb 2> :
> return { 1, 2, 3, 4, ... };
>
> }
>
> code-gen:
> f1:
> .LFB3900:
> .cfi_startproc
> ptrue p0.b, all
> adrp x0, .LC0
> add x0, x0, :lo12:.LC0
> ld1rqw z0.s, p0/z, [x0]
> ret
> .LC0:
> .word 1
> .word 2
> .word 3
> .word 4
>
> I guess for this particular case, we could use index instead.
>
> For,
> svint32_t f2(int a, int b, int c, int d)
> {
> int32x4_t v = {a, b, c, d};
> return svld1rq_s32 (svptrue_b8 (), &v[0]);
> }
>
> optimized dump shows:
> svint32_t f2 (int a, int b, int c, int d)
> {
> svint32_t _6;
>
> <bb 2> [local count: 1073741824]:
> _6 = {a_1(D), b_2(D), c_3(D), d_4(D), ... };
> return _6;
>
> The code-gen seems pretty bad however:
> f2:
> .LFB3901:
> .cfi_startproc
> addvl sp, sp, #-4
> .cfi_escape 0xf,0x9,0x8f,0,0x92,0x2e,0,0x8,0x20,0x1e,0x22
> ptrue p0.b, all
> addvl x4, sp, #3
> mov z0.b, #0
> st1w z0.s, p0, [sp, #3, mul vl]
> str w0, [x4]
> addvl x0, sp, #2
> ld1w z0.s, p0/z, [sp, #3, mul vl]
> st1w z0.s, p0, [sp, #2, mul vl]
> str w1, [x0, 4]
> addvl x0, sp, #1
> ld1w z0.s, p0/z, [sp, #2, mul vl]
> st1w z0.s, p0, [sp, #1, mul vl]
> str w2, [x0, 8]
> ld1w z0.s, p0/z, [sp, #1, mul vl]
> st1w z0.s, p0, [sp]
> str w3, [sp, 12]
> ld1w z0.s, p0/z, [sp]
> addvl sp, sp, #4
> .cfi_def_cfa_offset 0
> ret
>
> I will try to address code-gen issues in follow up patches.
> Bootstrapped+tested on x64_64-linux-gnu and aarch64-linux-gnu.
/* If result vector has greater length than input vector,
+ then allow permuting two vectors as long as:
+ a) sel.nelts_per_pattern == 1
+ b) sel.npatterns == len of input vector.
+ The intent is to permute input vectors, and
+ dup the elements in resulting vector to target vector length. */
+
+ if (maybe_gt (TYPE_VECTOR_SUBPARTS (type),
+ TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0))))
+ {
+ nelts = sel.encoding ().npatterns ();
+ if (sel.encoding ().nelts_per_pattern () != 1
+ || (!known_eq (nelts, TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg0)))))
+ return NULL_TREE;
+ }
so the only case you add is non-VLA to VLA and there
explicitely only the case of a period that's same as the
element count in the input vectors.
@@ -2602,6 +2602,9 @@ dump_generic_node (pretty_printer *pp, tree
node, int spc, dump_flags_t flags,
pp_space (pp);
}
}
+ if (VECTOR_TYPE_P (TREE_TYPE (node))
+ && !TYPE_VECTOR_SUBPARTS (TREE_TYPE (node)).is_constant ())
+ pp_string (pp, ", ... ");
pp_right_brace (pp);
btw, I do wonder if VLA CONSTRUCTORs are a "thing"? Are they?
I had hoped that you would make
tree *in_elts = XALLOCAVEC (tree, nelts * 2);
if (!vec_cst_ctor_to_array (arg0, nelts, in_elts)
|| !vec_cst_ctor_to_array (arg1, nelts, in_elts + nelts))
return NULL_TREE;
VLA agnostic, thus support for example permuting
{ 0, 2, 4, 8, ... }, { 1, 3, 5, 7 ... }, { 0, 4, 1, 5 ... }
as { 0, 1, 2, 3, ... }, etc.
that should be entirely doable, to somebody familiar with VLA and the APIs
even more so.
Richard.
> Thanks,
> Prathamesh
>
> >
> > Richard.
> >
> > > Thanks,
> > > Prathamesh
> > >
> > > >
> > > > > I will try to address the folding for above VEC_PERM_EXPR in
> > > > > follow-up patch.
> > > > >
> > > > > Thanks,
> > > > > Prathamesh
> > > > > >
> > > > > > Thanks,
> > > > > > Richard
