On May 21, 2015 5:28:14 PM GMT+02:00, Jakub Jelinek <ja...@redhat.com> wrote:
>Hi!
>
>We ICE on the following testcase at -O3 on x86_64-linux, because
>gimple folding attempts to simplify FLOAT_EXPR conversion of
>signed V4SI to V4SF feeding FIX_TRUNC_EXPR to unsigned V4SI
>into a FIX_TRUNC_EXPR with unsigned V4SI lhs and signed V4SI rhs1,
>which is invalid GIMPLE.
>All the other simplifications in the same iterator block don't
>optimize anything for vector types, and I can't find out any case
>where something like this would be beneficial for vector types.
>These days we represent source level casts of vectors to same sized
>integers as VIEW_CONVERT_EXPR, which isn't handled in here,
>and *_prec doesn't really mean what it tests for vector types
>(it is log2 of number of elements), vector integer or float widening
>is not represented using convert/float/fix_trunc, but using
>VEC_PERM_EXPR,
>VEC_UNPACK*_{LO,HI}_EXPR etc.
>I've bootstrapped/regtested with a logging variant and if
>(inside_vec || inter_vec || final_vec) is true, we (mis)optimize
>anything only on the testcase included in the patch and on
>gfortran.dg/stfunc_4.f90 testcase, in both cases it is
>V4SI -> V4SF -> V4SI, which we really shouldn't be optimizing,
>because SF mode obviously can't represent all integers exactly.
>
>So, this patch disables optimizing vectors.
>Ok for trunk/5.2 if bootstrap/regtest succeeds?
OK.
Thanks,
Richard.
>For 4.9/4.8 a similar patch will be needed, but to
>fold-const.c/tree-ssa-forwprop.c instead of match.pd.
>
>2015-05-21 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/66233
> * match.pd (ocvt (icvt@1 @0)): Don't handle vector types.
> Simplify.
>
> * gcc.c-torture/execute/pr66233.c: New test.
>
>--- gcc/match.pd.jj 2015-05-19 15:53:43.000000000 +0200
>+++ gcc/match.pd 2015-05-21 16:21:35.627916502 +0200
>@@ -730,16 +730,12 @@ (define_operator_list inverted_tcc_compa
> (for integers). Avoid this if the final type is a pointer since
> then we sometimes need the middle conversion. Likewise if the
> final type has a precision not equal to the size of its mode. */
>- (if (((inter_int && inside_int)
>- || (inter_float && inside_float)
>- || (inter_vec && inside_vec))
>+ (if (((inter_int && inside_int) || (inter_float && inside_float))
>+ && (final_int || final_float)
> && inter_prec >= inside_prec
>- && (inter_float || inter_vec
>- || inter_unsignedp == inside_unsignedp)
>- && ! (final_prec != GET_MODE_PRECISION (element_mode (type))
>- && element_mode (type) == element_mode (inter_type))
>- && ! final_ptr
>- && (! final_vec || inter_prec == inside_prec))
>+ && (inter_float || inter_unsignedp == inside_unsignedp)
>+ && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
>+ && TYPE_MODE (type) == TYPE_MODE (inter_type)))
> (ocvt @0))
>
> /* If we have a sign-extension of a zero-extended value, we can
>--- gcc/testsuite/gcc.c-torture/execute/pr66233.c.jj 2015-05-21
>17:13:32.639713225 +0200
>+++ gcc/testsuite/gcc.c-torture/execute/pr66233.c 2015-05-21
>17:10:57.000000000 +0200
>@@ -0,0 +1,22 @@
>+/* PR tree-optimization/66233 */
>+
>+unsigned int v[8];
>+
>+__attribute__((noinline, noclone)) void
>+foo (void)
>+{
>+ int i;
>+ for (i = 0; i < 8; i++)
>+ v[i] = (float) i;
>+}
>+
>+int
>+main ()
>+{
>+ unsigned int i;
>+ foo ();
>+ for (i = 0; i < 8; i++)
>+ if (v[i] != i)
>+ __builtin_abort ();
>+ return 0;
>+}
>
> Jakub
