From: Aldy Hernandez <aldyh@redhat.com>
To: Andrew MacLeod <amacleod@redhat.com>
Cc: Richard Biener <richard.guenther@gmail.com>,
GCC patches <gcc-patches@gcc.gnu.org>
Subject: Re: [COMMITTED] Convert nonzero mask in irange to wide_int.
Date: Fri, 7 Oct 2022 11:23:22 +0200 [thread overview]
Message-ID: <CAGm3qMWHJLGu7n_Am_znic0z=Tv_9d=eB1=_pLbEO_qOo1-FMQ@mail.gmail.com> (raw)
In-Reply-To: <CAGm3qMW7bfz5mZ=gavAgf4j2jsZcUJLe_NbEWzuFfc81-PDV9w@mail.gmail.com>
[-- Attachment #1: Type: text/plain, Size: 1035 bytes --]
On Wed, Oct 5, 2022 at 12:14 PM Aldy Hernandez <aldyh@redhat.com> wrote:
> However... I don't think I have the stomach to overhaul the allocators
> this late in the release. For this release I may opt to put the
> nonzero mask back in a tree, but have it always be set. The NULL ==
> -1 shortcut was very error prone. The rest of my fixes in this patch
> still apply, as they keep better track of the masks, which we need.
Here is the patch reverting the nonzero mask to trees.
Unfortunately, having the mask always set caused a 10% regression in
VRP, so that's a no go. I've gone back to keeping a NULL mask by
default that semantically means -1. It's not as bad as I thought,
since the code is much cleaner now. This is unfortunate, but a 10%
regression in VRP plus a 1.5% regression in overall compilation is
unacceptable. On the plus side, this is temporary as we're moving
entirely to wide ints next release (with appropriate cache/allocator
changes).
I will commit after a final round of tests finishes.
Thanks.
Aldy
[-- Attachment #2: 0001-Convert-nonzero-mask-back-to-tree.patch --]
[-- Type: text/x-patch, Size: 9368 bytes --]
From 6b6e929b238ff91fee1f133e3ff7adceb3f75660 Mon Sep 17 00:00:00 2001
From: Aldy Hernandez <aldyh@redhat.com>
Date: Fri, 7 Oct 2022 09:57:32 +0200
Subject: [PATCH] Convert nonzero mask back to tree.
Having nonzero masks always set had a performance penalty of 10% in
VRP, so mask==NULL is a shortcut to all bits set.
gcc/ChangeLog:
* value-range.cc (irange::irange_set): Convert nonzero mask to
tree.
(irange::irange_set_anti_range): Same.
(irange::set): Same.
(irange::verify_range): Same.
(irange::contains_p): Same.
(irange::invert): Same.
(irange::set_range_from_nonzero_bits): Same.
(irange::set_nonzero_bits): Same.
(mask_to_wi): Same.
(irange::intersect_nonzero_bits): Same.
(irange::union_nonzero_bits): Same.
* value-range.h (irange::varying_compatible_p): Same.
(gt_ggc_mx): Same.
(gt_pch_nx): Same.
(irange::set_undefined): Same.
(irange::set_varying): Same.
---
gcc/value-range.cc | 85 +++++++++++++++++++++++++++++++++++-----------
gcc/value-range.h | 19 ++++++-----
2 files changed, 77 insertions(+), 27 deletions(-)
diff --git a/gcc/value-range.cc b/gcc/value-range.cc
index 87239fafa77..b4496ea9eea 100644
--- a/gcc/value-range.cc
+++ b/gcc/value-range.cc
@@ -935,7 +935,7 @@ irange::irange_set (tree min, tree max)
m_base[1] = max;
m_num_ranges = 1;
m_kind = VR_RANGE;
- m_nonzero_mask = wi::shwi (-1, TYPE_PRECISION (TREE_TYPE (min)));
+ m_nonzero_mask = NULL;
normalize_kind ();
if (flag_checking)
@@ -1009,7 +1009,7 @@ irange::irange_set_anti_range (tree min, tree max)
}
m_kind = VR_RANGE;
- m_nonzero_mask = wi::shwi (-1, TYPE_PRECISION (TREE_TYPE (min)));
+ m_nonzero_mask = NULL;
normalize_kind ();
if (flag_checking)
@@ -1066,7 +1066,7 @@ irange::set (tree min, tree max, value_range_kind kind)
m_base[0] = min;
m_base[1] = max;
m_num_ranges = 1;
- m_nonzero_mask = wi::shwi (-1, TYPE_PRECISION (TREE_TYPE (min)));
+ m_nonzero_mask = NULL;
return;
}
@@ -1116,7 +1116,7 @@ irange::set (tree min, tree max, value_range_kind kind)
m_base[0] = min;
m_base[1] = max;
m_num_ranges = 1;
- m_nonzero_mask = wi::shwi (-1, TYPE_PRECISION (TREE_TYPE (min)));
+ m_nonzero_mask = NULL;
normalize_kind ();
if (flag_checking)
verify_range ();
@@ -1135,7 +1135,8 @@ irange::verify_range ()
}
if (m_kind == VR_VARYING)
{
- gcc_checking_assert (m_nonzero_mask == -1);
+ gcc_checking_assert (!m_nonzero_mask
+ || wi::to_wide (m_nonzero_mask) == -1);
gcc_checking_assert (m_num_ranges == 1);
gcc_checking_assert (varying_compatible_p ());
return;
@@ -1409,10 +1410,10 @@ irange::contains_p (tree cst) const
gcc_checking_assert (TREE_CODE (cst) == INTEGER_CST);
// See if we can exclude CST based on the nonzero bits.
- if (m_nonzero_mask != -1)
+ if (m_nonzero_mask)
{
wide_int cstw = wi::to_wide (cst);
- if (cstw != 0 && wi::bit_and (m_nonzero_mask, cstw) == 0)
+ if (cstw != 0 && wi::bit_and (wi::to_wide (m_nonzero_mask), cstw) == 0)
return false;
}
@@ -2776,7 +2777,7 @@ irange::invert ()
signop sign = TYPE_SIGN (ttype);
wide_int type_min = wi::min_value (prec, sign);
wide_int type_max = wi::max_value (prec, sign);
- m_nonzero_mask = wi::shwi (-1, prec);
+ m_nonzero_mask = NULL;
if (m_num_ranges == m_max_ranges
&& lower_bound () != type_min
&& upper_bound () != type_max)
@@ -2878,20 +2879,22 @@ bool
irange::set_range_from_nonzero_bits ()
{
gcc_checking_assert (!undefined_p ());
- unsigned popcount = wi::popcount (m_nonzero_mask);
+ if (!m_nonzero_mask)
+ return false;
+ unsigned popcount = wi::popcount (wi::to_wide (m_nonzero_mask));
// If we have only one bit set in the mask, we can figure out the
// range immediately.
if (popcount == 1)
{
// Make sure we don't pessimize the range.
- if (!contains_p (wide_int_to_tree (type (), m_nonzero_mask)))
+ if (!contains_p (m_nonzero_mask))
return false;
bool has_zero = contains_p (build_zero_cst (type ()));
- wide_int bits = m_nonzero_mask;
- set (type (), bits, bits);
- m_nonzero_mask = bits;
+ tree nz = m_nonzero_mask;
+ set (nz, nz);
+ m_nonzero_mask = nz;
if (has_zero)
{
int_range<2> zero;
@@ -2909,11 +2912,21 @@ irange::set_nonzero_bits (const wide_int_ref &bits)
gcc_checking_assert (!undefined_p ());
unsigned prec = TYPE_PRECISION (type ());
+ if (bits == -1)
+ {
+ m_nonzero_mask = NULL;
+ normalize_kind ();
+ if (flag_checking)
+ verify_range ();
+ return;
+ }
+
// Drop VARYINGs with a nonzero mask to a plain range.
if (m_kind == VR_VARYING && bits != -1)
m_kind = VR_RANGE;
- m_nonzero_mask = wide_int::from (bits, prec, TYPE_SIGN (type ()));
+ wide_int nz = wide_int::from (bits, prec, TYPE_SIGN (type ()));
+ m_nonzero_mask = wide_int_to_tree (type (), nz);
if (set_range_from_nonzero_bits ())
return;
@@ -2937,7 +2950,21 @@ irange::get_nonzero_bits () const
// the mask precisely up to date at all times. Instead, we default
// to -1 and set it when explicitly requested. However, this
// function will always return the correct mask.
- return m_nonzero_mask & get_nonzero_bits_from_range ();
+ if (m_nonzero_mask)
+ return wi::to_wide (m_nonzero_mask) & get_nonzero_bits_from_range ();
+ else
+ return get_nonzero_bits_from_range ();
+}
+
+// Convert tree mask to wide_int. Returns -1 for NULL masks.
+
+inline wide_int
+mask_to_wi (tree mask, tree type)
+{
+ if (mask)
+ return wi::to_wide (mask);
+ else
+ return wi::shwi (-1, TYPE_PRECISION (type));
}
// Intersect the nonzero bits in R into THIS and normalize the range.
@@ -2948,10 +2975,20 @@ irange::intersect_nonzero_bits (const irange &r)
{
gcc_checking_assert (!undefined_p () && !r.undefined_p ());
+ if (!m_nonzero_mask && !r.m_nonzero_mask)
+ {
+ normalize_kind ();
+ if (flag_checking)
+ verify_range ();
+ return false;
+ }
+
bool changed = false;
- if (m_nonzero_mask != r.m_nonzero_mask)
+ tree t = type ();
+ if (mask_to_wi (m_nonzero_mask, t) != mask_to_wi (r.m_nonzero_mask, t))
{
- m_nonzero_mask = get_nonzero_bits () & r.get_nonzero_bits ();
+ wide_int nz = get_nonzero_bits () & r.get_nonzero_bits ();
+ m_nonzero_mask = wide_int_to_tree (t, nz);
if (set_range_from_nonzero_bits ())
return true;
changed = true;
@@ -2970,10 +3007,20 @@ irange::union_nonzero_bits (const irange &r)
{
gcc_checking_assert (!undefined_p () && !r.undefined_p ());
+ if (!m_nonzero_mask && !r.m_nonzero_mask)
+ {
+ normalize_kind ();
+ if (flag_checking)
+ verify_range ();
+ return false;
+ }
+
bool changed = false;
- if (m_nonzero_mask != r.m_nonzero_mask)
+ tree t = type ();
+ if (mask_to_wi (m_nonzero_mask, t) != mask_to_wi (r.m_nonzero_mask, t))
{
- m_nonzero_mask = get_nonzero_bits () | r.get_nonzero_bits ();
+ wide_int nz = get_nonzero_bits () | r.get_nonzero_bits ();
+ m_nonzero_mask = wide_int_to_tree (t, nz);
// No need to call set_range_from_nonzero_bits, because we'll
// never narrow the range. Besides, it would cause endless
// recursion because of the union_ in
diff --git a/gcc/value-range.h b/gcc/value-range.h
index b06ca7477cd..484f911bd90 100644
--- a/gcc/value-range.h
+++ b/gcc/value-range.h
@@ -215,7 +215,7 @@ private:
bool intersect (const wide_int& lb, const wide_int& ub);
unsigned char m_num_ranges;
unsigned char m_max_ranges;
- wide_int m_nonzero_mask;
+ tree m_nonzero_mask;
tree *m_base;
};
@@ -683,11 +683,11 @@ irange::varying_compatible_p () const
if (INTEGRAL_TYPE_P (t))
return (wi::to_wide (l) == wi::min_value (prec, sign)
&& wi::to_wide (u) == wi::max_value (prec, sign)
- && m_nonzero_mask == -1);
+ && (!m_nonzero_mask || wi::to_wide (m_nonzero_mask) == -1));
if (POINTER_TYPE_P (t))
return (wi::to_wide (l) == 0
&& wi::to_wide (u) == wi::max_value (prec, sign)
- && m_nonzero_mask == -1);
+ && (!m_nonzero_mask || wi::to_wide (m_nonzero_mask) == -1));
return true;
}
@@ -754,6 +754,8 @@ gt_ggc_mx (irange *x)
gt_ggc_mx (x->m_base[i * 2]);
gt_ggc_mx (x->m_base[i * 2 + 1]);
}
+ if (x->m_nonzero_mask)
+ gt_ggc_mx (x->m_nonzero_mask);
}
inline void
@@ -764,6 +766,8 @@ gt_pch_nx (irange *x)
gt_pch_nx (x->m_base[i * 2]);
gt_pch_nx (x->m_base[i * 2 + 1]);
}
+ if (x->m_nonzero_mask)
+ gt_pch_nx (x->m_nonzero_mask);
}
inline void
@@ -774,6 +778,8 @@ gt_pch_nx (irange *x, gt_pointer_operator op, void *cookie)
op (&x->m_base[i * 2], NULL, cookie);
op (&x->m_base[i * 2 + 1], NULL, cookie);
}
+ if (x->m_nonzero_mask)
+ op (&x->m_nonzero_mask, NULL, cookie);
}
template<unsigned N>
@@ -868,6 +874,7 @@ irange::set_undefined ()
{
m_kind = VR_UNDEFINED;
m_num_ranges = 0;
+ m_nonzero_mask = NULL;
}
inline void
@@ -875,11 +882,7 @@ irange::set_varying (tree type)
{
m_kind = VR_VARYING;
m_num_ranges = 1;
-
- if (type == error_mark_node)
- m_nonzero_mask = wi::shwi (-1, 1);
- else
- m_nonzero_mask = wi::shwi (-1, TYPE_PRECISION (type));
+ m_nonzero_mask = NULL;
if (INTEGRAL_TYPE_P (type))
{
--
2.37.1
prev parent reply other threads:[~2022-10-07 9:23 UTC|newest]
Thread overview: 11+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-10-04 7:35 Aldy Hernandez
2022-10-04 7:55 ` Richard Biener
2022-10-04 11:28 ` Aldy Hernandez
2022-10-04 12:13 ` Aldy Hernandez
2022-10-04 13:27 ` Andrew MacLeod
2022-10-04 14:30 ` Aldy Hernandez
2022-10-04 14:34 ` Richard Biener
2022-10-04 15:14 ` Aldy Hernandez
2022-10-04 15:42 ` Andrew MacLeod
2022-10-05 10:14 ` Aldy Hernandez
2022-10-07 9:23 ` Aldy Hernandez [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CAGm3qMWHJLGu7n_Am_znic0z=Tv_9d=eB1=_pLbEO_qOo1-FMQ@mail.gmail.com' \
--to=aldyh@redhat.com \
--cc=amacleod@redhat.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=richard.guenther@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).