On Fri, May 10, 2024 at 11:24 AM Aldy Hernandez <[email protected]> wrote:
>
> There are various calls into fold_range() that have the wrong type
> associated with the range temporary used to hold the result. This
> used to work, because we could store either integers or pointers in a
> Value_Range, but is no longer the case with prange's. Now you must
> explicitly state which type of range the temporary will hold before
> storing into it. You can change this at a later time with set_type(),
> but you must always have a type before using the temporary, and it
> must match what fold_range() returns.
>
> This patch adjusts the IPA code to restore the previous functionality,
> so I can re-enable the prange code, but I do question whether the
> previous code was correct. I have added appropriate comments to help
> the maintainers, but someone with more knowledge should revamp this
> going forward.
>
> The basic problem is that pointer comparisons return a boolean, but
> the IPA code is initializing the resulting range as a pointer. This
> wasn't a problem, because fold_range() would previously happily force
> the range into an integer one, and everything would work. But now we
> must initialize the range to an integer before calling into
> fold_range. The thing is, that the failing case sets the result back
> into a pointer, which is just weird but existing behavior. I have
> documented this in the code.
>
> if (!handler
> || !op_res.supports_type_p (vr_type)
> || !handler.fold_range (op_res, vr_type, srcvr, op_vr))
> /* For comparison operators, the type here may be
> different than the range type used in fold_range above.
> For example, vr_type may be a pointer, whereas the type
> returned by fold_range will always be a boolean.
>
> This shouldn't cause any problems, as the set_varying
> below will happily change the type of the range in
> op_res, and then the cast operation in
> ipa_vr_operation_and_type_effects will ultimately leave
> things in the desired type, but it is confusing.
>
> Perhaps the original intent was to use the type of
> op_res here? */
> op_res.set_varying (vr_type);
>
> BTW, this is not to say that the original gimple IR was wrong, but that
> IPA is setting the range type of the result of fold_range() to the type of
> the operands, which does not necessarily match in the case of a
> comparison.
>
> I am just restoring previous behavior here, but I do question whether it
> was right to begin with.
>
> Testing currently in progress on x86-64 and ppc64le with prange enabled.
>
> OK pending tests?
I think this "intermediate" patch is unnecessary and instead the code should
be fixed correctly, avoiding missed-optimization regressions.
Richard.
> gcc/ChangeLog:
>
> PR tree-optimization/114985
> * ipa-cp.cc (ipa_value_range_from_jfunc): Adjust type of op_res.
> (propagate_vr_across_jump_function): Same.
> * ipa-fnsummary.cc (evaluate_conditions_for_known_args): Adjust
> type for res.
> * ipa-prop.h (ipa_type_for_fold_range): New.
> ---
> gcc/ipa-cp.cc | 18 ++++++++++++++++--
> gcc/ipa-fnsummary.cc | 6 +++++-
> gcc/ipa-prop.h | 13 +++++++++++++
> 3 files changed, 34 insertions(+), 3 deletions(-)
>
> diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
> index 5781f50c854..3c395632364 100644
> --- a/gcc/ipa-cp.cc
> +++ b/gcc/ipa-cp.cc
> @@ -1730,7 +1730,7 @@ ipa_value_range_from_jfunc (vrange &vr,
> }
> else
> {
> - Value_Range op_res (vr_type);
> + Value_Range op_res (ipa_type_for_fold_range (operation, vr_type));
> Value_Range res (vr_type);
> tree op = ipa_get_jf_pass_through_operand (jfunc);
> Value_Range op_vr (TREE_TYPE (op));
> @@ -1741,6 +1741,19 @@ ipa_value_range_from_jfunc (vrange &vr,
> if (!handler
> || !op_res.supports_type_p (vr_type)
> || !handler.fold_range (op_res, vr_type, srcvr, op_vr))
> + /* For comparison operators, the type here may be
> + different than the range type used in fold_range above.
> + For example, vr_type may be a pointer, whereas the type
> + returned by fold_range will always be a boolean.
> +
> + This shouldn't cause any problems, as the set_varying
> + below will happily change the type of the range in
> + op_res, and then the cast operation in
> + ipa_vr_operation_and_type_effects will ultimately leave
> + things in the desired type, but it is confusing.
> +
> + Perhaps the original intent was to use the type of
> + op_res here? */
> op_res.set_varying (vr_type);
>
> if (ipa_vr_operation_and_type_effects (res,
> @@ -2540,7 +2553,7 @@ propagate_vr_across_jump_function (cgraph_edge *cs,
> ipa_jump_func *jfunc,
> {
> tree op = ipa_get_jf_pass_through_operand (jfunc);
> Value_Range op_vr (TREE_TYPE (op));
> - Value_Range op_res (param_type);
> + Value_Range op_res (ipa_type_for_fold_range (operation,
> param_type));
> range_op_handler handler (operation);
>
> ipa_range_set_and_normalize (op_vr, op);
> @@ -2549,6 +2562,7 @@ propagate_vr_across_jump_function (cgraph_edge *cs,
> ipa_jump_func *jfunc,
> || !ipa_supports_p (operand_type)
> || !handler.fold_range (op_res, operand_type,
> src_lats->m_value_range.m_vr, op_vr))
> + /* See note in ipa_value_range_from_jfunc. */
> op_res.set_varying (param_type);
>
> ipa_vr_operation_and_type_effects (vr,
> diff --git a/gcc/ipa-fnsummary.cc b/gcc/ipa-fnsummary.cc
> index 07a853f78e3..4deba2394f5 100644
> --- a/gcc/ipa-fnsummary.cc
> +++ b/gcc/ipa-fnsummary.cc
> @@ -502,7 +502,8 @@ evaluate_conditions_for_known_args (struct cgraph_node
> *node,
> if (vr.varying_p () || vr.undefined_p ())
> break;
>
> - Value_Range res (op->type);
> + Value_Range res (ipa_type_for_fold_range (op->code,
> + op->type));
> if (!op->val[0])
> {
> Value_Range varying (op->type);
> @@ -511,6 +512,7 @@ evaluate_conditions_for_known_args (struct cgraph_node
> *node,
> if (!handler
> || !res.supports_type_p (op->type)
> || !handler.fold_range (res, op->type, vr, varying))
> + /* See note in ipa_value_from_jfunc. */
> res.set_varying (op->type);
> }
> else if (!op->val[1])
> @@ -525,9 +527,11 @@ evaluate_conditions_for_known_args (struct cgraph_node
> *node,
> || !handler.fold_range (res, op->type,
> op->index ? op0 : vr,
> op->index ? vr : op0))
> + /* See note in ipa_value_from_jfunc. */
> res.set_varying (op->type);
> }
> else
> + /* See note in ipa_value_from_jfunc. */
> res.set_varying (op->type);
> vr = res;
> }
> diff --git a/gcc/ipa-prop.h b/gcc/ipa-prop.h
> index 93d1b87b1f7..8493dd19b92 100644
> --- a/gcc/ipa-prop.h
> +++ b/gcc/ipa-prop.h
> @@ -1277,6 +1277,19 @@ ipa_range_set_and_normalize (vrange &r, tree val)
> r.set (val, val);
> }
>
> +/* Return the resulting type for a fold_range() operation for OP and
> + TYPE. */
> +
> +inline tree
> +ipa_type_for_fold_range (tree_code op, tree type)
> +{
> + /* Comparisons return boolean regardless of their input operands. */
> + if (TREE_CODE_CLASS (op) == tcc_comparison)
> + return boolean_type_node;
> +
> + return type;
> +}
> +
> bool ipa_return_value_range (Value_Range &range, tree decl);
> void ipa_record_return_value_range (Value_Range val);
> bool ipa_jump_functions_equivalent_p (ipa_jump_func *jf1, ipa_jump_func
> *jf2);
> --
> 2.45.0
>
