https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85863
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |wrong-code
Status|UNCONFIRMED |ASSIGNED
Last reconfirmed| |2018-05-22
CC| |jakub at gcc dot gnu.org
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Target Milestone|--- |9.0
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
gamess works fine here (x86_64). The test that triggers is
if (maybe_ne (TYPE_VECTOR_SUBPARTS (type),
TYPE_VECTOR_SUBPARTS (op0_type)))
{
error ("invalid vector comparison resulting type");
debug_generic_expr (type);
return true;
that means the IL is bogus and the issue likely latent. Reduced testcase:
SUBROUTINE SOBOOK(MHSO,HSOMAX)
IMPLICIT DOUBLE PRECISION(A-H,O-Z)
COMPLEX*16 HSOT,HSO1(2)
PARAMETER (ZERO=0.0D+00,TWO=2.0D+00)
DIMENSION SOL1(3,2),SOL2(3)
SQRT2=SQRT(TWO)
DO IH=1,MHSO
IF(MS.EQ.0) THEN
HSO1(IH) = DCMPLX(ZERO,-SOL1(3,IH))
HSOT = DCMPLX(ZERO,-SOL2(3))
ELSE
HSO1(IH) = DCMPLX(-SOL1(2,IH),SOL1(1,IH))/SQRT2
HSOT = DCMPLX(-SOL2(2),SOL2(1))/SQRT2
ENDIF
ENDDO
HSOT=HSOT+HSO1(1)
HSOMAX=MAX(HSOMAX,ABS(HSOT))
RETURN
END
The bogus comparison is built here:
0x0000000001378d87 in vectorizable_condition (stmt=
<gimple_assign 0x7ffff69aa4e0>, gsi=0x7fffffffcf30,
vec_stmt=0x7fffffffcbf0, reduc_def=<tree 0x0>, reduc_index=0,
slp_node=0x2928520, cost_vec=0x0) at /tmp/trunk/gcc/tree-vect-stmts.c:9000
9000 vec_cond_lhs, vec_cond_rhs);
Value returned is $7 = (tree_node *) 0x7ffff69cc870
(gdb) l
8995 else
8996 {
8997 vec_cond_rhs = vec_oprnds1[i];
8998 if (bitop1 == NOP_EXPR)
8999 vec_compare = build2 (cond_code, vec_cmp_type,
9000 vec_cond_lhs, vec_cond_rhs);
9001 else
9002 {
9003 new_temp = make_ssa_name (vec_cmp_type);
9004 if (bitop1 == BIT_NOT_EXPR)
so we face a conversion that isn't NOP. Which looks somewhat weird.
The stmt we try to vectorize is
_ifc__74 = ms_36(D) == 0 ? _ifc__75 : _ifc__76;
where _ifc__75 is DFmode and the comparison is a SImode comparison.
Nothing in vectorizable_condition seems to verify we need no
widening/shortening
of the comparison result... I guess pattern recognition usually deals with
this but doesn't in this case. Ah, and it is that we run into
the following vect_is_simple_cond code:
/* Invariant comparison. */
if (! *comp_vectype)
{
tree scalar_type = TREE_TYPE (lhs);
/* If we can widen the comparison to match vectype do so. */
if (INTEGRAL_TYPE_P (scalar_type)
&& tree_int_cst_lt (TYPE_SIZE (scalar_type),
TYPE_SIZE (TREE_TYPE (vectype))))
scalar_type = build_nonstandard_integer_type
(tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))),
TYPE_UNSIGNED (scalar_type));
*comp_vectype = get_vectype_for_scalar_type (scalar_type);
which widens the vector type. But nothing in code-generation deals
with this in the SLP case since only vect_get_vec_def_for_operand
gets the extra invariant vector type argument but vect_get_slp_defs
knows nothing about such requirement.
Jakub ventured last in this area of the code but obviously a patch like
the following would help to reflect that reality. I'm testing that patch.
Index: gcc/tree-vect-stmts.c
===================================================================
--- gcc/tree-vect-stmts.c (revision 260499)
+++ gcc/tree-vect-stmts.c (working copy)
@@ -8661,7 +8661,7 @@ vect_is_simple_cond (tree cond, vec_info
*comp_vectype = vectype1 ? vectype1 : vectype2;
/* Invariant comparison. */
- if (! *comp_vectype)
+ if (! *comp_vectype && vectype)
{
tree scalar_type = TREE_TYPE (lhs);
/* If we can widen the comparison to match vectype do so. */
@@ -8773,7 +8773,7 @@ vectorizable_condition (gimple *stmt, gi
else_clause = gimple_assign_rhs3 (stmt);
if (!vect_is_simple_cond (cond_expr, stmt_info->vinfo,
- &comp_vectype, &dts[0], vectype)
+ &comp_vectype, &dts[0], slp_node ? NULL : vectype)
|| !comp_vectype)
return false;
