On Fri, 23 Sep 2022, Tamar Christina wrote:
> Hi All,
>
> This is an RFC to figure out how to deal with targets that don't have native
> comparisons against QImode values.
>
> Booleans, at least in C99 and higher are 0-1 valued. This means that we only
> really need to test a single bit. However in RTL we no longer have this
> information available and just have an SImode value (due to the promotion of
> QImode to SImode).
>
> This RFC fixes it by emitting an explicit & 1 during the expansion of the
> conditional branch.
>
> However it's unlikely that we want to do this unconditionally. Most targets
> I've tested seem to have harmless code changes, like x86 changes from testb to
> andl, $1.
>
> So I have two questions:
>
> 1. Should I limit this behind a target macro? Or should I just leave it for
> all
> targets and deal with the fallout.
> 2. How can I tell whether the C99 0-1 values bools are being used or the older
> 0, non-0 variant?
>
> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
> However there are some benign codegen changes on x86, testb changed to andl
> $1.
>
> This pattern occurs more than 120,000 times in SPECCPU 2017 and so is quite
> common.
How does this help a target? Why does RTL nonzerop bits not recover this
information and the desired optimization done later during combine
for example? Why's a SImode compare not OK if there's no QImode compare?
We have undocumented addcc, negcc, etc. patterns, should we have a
andcc pattern for this indicating support for andcc + jump as opposed
to cmpcc + jump?
So - what's the target and what's a testcase?
Thanks,
Richard.
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> * tree.h (tree_zero_one_valued_p): New.
> * dojump.cc (do_jump): Add & 1 if truth type.
>
> --- inline copy of patch --
> diff --git a/gcc/dojump.cc b/gcc/dojump.cc
> index
> 2af0cd1aca3b6af13d5d8799094ee93f18022296..8eaf1be49cd12298e61c6946ae79ca9de6197864
> 100644
> --- a/gcc/dojump.cc
> +++ b/gcc/dojump.cc
> @@ -605,7 +605,17 @@ do_jump (tree exp, rtx_code_label *if_false_label,
> /* Fall through and generate the normal code. */
> default:
> normal:
> - temp = expand_normal (exp);
> + tree cmp = exp;
> + /* If the expression is a truth type then explicitly generate an & 1
> + to indicate to the target that it's a zero-one values type. This
> + allows the target to further optimize the comparison should it
> + choose to. */
> + if (tree_zero_one_valued_p (exp))
> + {
> + type = TREE_TYPE (exp);
> + cmp = build2 (BIT_AND_EXPR, type, exp, build_int_cstu (type, 1));
> + }
> + temp = expand_normal (cmp);
> do_pending_stack_adjust ();
> /* The RTL optimizers prefer comparisons against pseudos. */
> if (GET_CODE (temp) == SUBREG)
> diff --git a/gcc/tree.h b/gcc/tree.h
> index
> 8f8a9660c9e0605eb516de194640b8c1b531b798..be3d2dee82f692e81082cf21c878c10f9fe9e1f1
> 100644
> --- a/gcc/tree.h
> +++ b/gcc/tree.h
> @@ -4690,6 +4690,7 @@ extern tree signed_or_unsigned_type_for (int, tree);
> extern tree signed_type_for (tree);
> extern tree unsigned_type_for (tree);
> extern bool is_truth_type_for (tree, tree);
> +extern bool tree_zero_one_valued_p (tree);
> extern tree truth_type_for (tree);
> extern tree build_pointer_type_for_mode (tree, machine_mode, bool);
> extern tree build_pointer_type (tree);
>
>
>
>
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)
