On Tue, Nov 9, 2021 at 12:47 AM Qing Zhao <qing.z...@oracle.com> wrote:
>
> Hi, I tried both the following patches:
>
> Patch1:
>
> [opc@qinzhao-ol8u3-x86 gcc]$ git diff
> diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c
> index 0cba95411a6..ca49d2b4514 100644
> --- a/gcc/internal-fn.c
> +++ b/gcc/internal-fn.c
> @@ -3073,12 +3073,14 @@ expand_DEFERRED_INIT (internal_fn, gcall *stmt)
> /* If this variable is in a register use expand_assignment.
> For boolean scalars force zero-init. */
> tree init;
> + scalar_int_mode var_mode;
> if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE
> && tree_fits_uhwi_p (var_size)
> && (init_type == AUTO_INIT_PATTERN
> || !is_gimple_reg_type (var_type))
> && int_mode_for_size (tree_to_uhwi (var_size) * BITS_PER_UNIT,
> - 0).exists ())
> + 0).exists (&var_mode)
> + && targetm.scalar_mode_supported_p (var_mode))
> {
> unsigned HOST_WIDE_INT total_bytes = tree_to_uhwi (var_size);
> unsigned char *buf = (unsigned char *) xmalloc (total_bytes);
>
> AND
>
> Patch2:
> diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c
> index 0cba95411a6..7f129655926 100644
> --- a/gcc/internal-fn.c
> +++ b/gcc/internal-fn.c
> @@ -3073,12 +3073,14 @@ expand_DEFERRED_INIT (internal_fn, gcall *stmt)
> /* If this variable is in a register use expand_assignment.
> For boolean scalars force zero-init. */
> tree init;
> + scalar_int_mode var_mode;
> if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE
> && tree_fits_uhwi_p (var_size)
> && (init_type == AUTO_INIT_PATTERN
> || !is_gimple_reg_type (var_type))
> && int_mode_for_size (tree_to_uhwi (var_size) * BITS_PER_UNIT,
> - 0).exists ())
> + 0).exists (&var_mode)
> + && have_insn_for (SET, var_mode))
> {
> unsigned HOST_WIDE_INT total_bytes = tree_to_uhwi (var_size);
> unsigned char *buf = (unsigned char *) xmalloc (total_bytes);
>
> Have the same effect:
>
> 1. Resolved the ICE in gcc11;
> 2. For _Complex long double variables, both return FALSE, as a result, for
> PATTERN initialization of _Complex long double variables, now they are
> initialization with ZEROs instead of FEs.
>
> Let me know you opinion on this, If the above 2 is okay, then I might pick
> the above Patch 1 for the final patch to this issue.
I think zero-initialization is OK, but I'd choose Patch2 for
consistency with what we do in the memcpy
folding.
Richard.
> Thanks.
>
> Qing
>
> > On Nov 8, 2021, at 2:41 AM, Richard Biener <richard.guent...@gmail.com>
> > wrote:
> >
> > On Sat, Nov 6, 2021 at 10:56 AM Jakub Jelinek <ja...@redhat.com> wrote:
> >>
> >> On Fri, Nov 05, 2021 at 05:37:25PM +0000, Qing Zhao wrote:
> >>>> On Nov 5, 2021, at 11:17 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> >>>>
> >>>> On Fri, Nov 05, 2021 at 04:11:36PM +0000, Qing Zhao wrote:
> >>>>> 3076 if (TREE_CODE (TREE_TYPE (lhs)) != BOOLEAN_TYPE
> >>>>> 3077 && tree_fits_uhwi_p (var_size)
> >>>>> 3078 && (init_type == AUTO_INIT_PATTERN
> >>>>> 3079 || !is_gimple_reg_type (var_type))
> >>>>> 3080 && int_mode_for_size (tree_to_uhwi (var_size) *
> >>>>> BITS_PER_UNIT,
> >>>>> 3081 0).exists ())
> >>>>> 3082 {
> >>>>> 3083 unsigned HOST_WIDE_INT total_bytes = tree_to_uhwi
> >>>>> (var_size);
> >>>>> 3084 unsigned char *buf = (unsigned char *) xmalloc
> >>>>> (total_bytes);
> >>>>> 3085 memset (buf, (init_type == AUTO_INIT_PATTERN
> >>>>> 3086 ? INIT_PATTERN_VALUE : 0), total_bytes);
> >>>>> 3087 tree itype = build_nonstandard_integer_type
> >>>>> 3088 (total_bytes * BITS_PER_UNIT, 1);
> >>>>>
> >>>>> The exact failing point is at function
> >>>>> “set_min_and_max_values_for_integral_type”:
> >>>>>
> >>>>> 2851 gcc_assert (precision <= WIDE_INT_MAX_PRECISION);
> >>>>>
> >>>>> For _Complex long double, “precision” is 256.
> >>>>> In GCC11, “WIDE_INT_MAX_PRECISION” is 192, in GCC12, it’s 512.
> >>>>> As a result, the above assertion failed on GCC11.
> >>>>>
> >>>>> I am wondering what’s the best fix for this issue in gcc11?
> >>>>
> >>>> Even for gcc 12 the above is wrong, you can't blindly assume that
> >>>> build_nonstandard_integer_type will work for arbitrary precisions,
> >>>> and even if it works that it will actually work.
> >>>> The fact that such a mode exist is one thing, but
> >>>> targetm.scalar_mode_supported_p should be tested for whether the mode
> >>>> is actually supported.
> >>>
> >>> You mean “int_mode_for_size().exists()” is not enough to make sure
> >>> “build_nonstandard_integer_type” to be valid? We should add
> >>> “targetm.scalar_mode_supported_p” too ?
> >>
> >> Yeah. The former says whether the backend has that mode at all.
> >> But some modes may be there only in some specific patterns but
> >> without support for mov, add, etc. Only for
> >> targetm.scalar_mode_supported_p modes the backend guarantees that
> >> one can use them e.g. in mode attribute and can expect expansion
> >> to expand everything with that mode that is needed in some way.
> >> E.g. only if targetm.scalar_mode_supported_p (TImode) the FEs
> >> support __int128_t type, etc.
> >
> > The memcpy folding code now checks
> >
> > scalar_int_mode mode;
> > if (int_mode_for_size (ilen * 8, 0).exists (&mode)
> > && GET_MODE_SIZE (mode) * BITS_PER_UNIT == ilen * 8
> > && have_insn_for (SET, mode)
> >
> > thus specifically only have_insn_for (SET, mode), which I guess is
> > good enough for this case as well?
> >
> >> Jakub
>
