77 lines
2.5 KiB
Diff
77 lines
2.5 KiB
Diff
|
Based on Richi's patch:
|
||
|
|
<https://gcc.gnu.org/pipermail/gcc-patches/2024-April/648725.html>
|
||
|
|
~~
|
||
|
|
The following avoids re-walking and re-combining the instructions
|
||
|
|
between i2 and i3 when the pattern of i2 doesn't change.
|
||
|
|
|
||
|
|
Bootstrap and regtest running ontop of a reversal of
|
||
|
|
r14-9692-g839bc42772ba7a.
|
||
|
|
|
||
|
|
It brings down memory use frmo 9GB to 400MB and compile-time from
|
||
|
|
80s to 3.5s. r14-9692-g839bc42772ba7a does better in both metrics
|
||
|
|
but has shown code generation regressions across acrchitectures.
|
||
|
|
|
||
|
|
PR rtl-optimization/101523
|
||
|
|
* combine.cc (try_combine): When the pattern of i2 doesn't
|
||
|
|
change do not re-start combining at i2 or an earlier insn which
|
||
|
|
had links or notes added.
|
||
|
|
~~
|
||
|
|
But, since the patch affects code generation (for instance,
|
||
|
|
libstdc++-v3/src/c++17/floating_from_chars.o), we limit the bailing out
|
||
|
|
only when I2 hasn't been changed 1000x. I've measured how many times
|
||
|
|
at most is I2 unchanged during a bootstrap + regtest.
|
||
|
|
x86: 134
|
||
|
|
aarch64: 736 (gimple-match-1.cc)
|
||
|
|
s390x: 635 (gimple-match-*)
|
||
|
|
ppc64le: 620 (gimple-match-*)
|
||
|
|
while certain pathological testcases trigger it more than 10,000 times.
|
||
|
|
With the limit in place this patch doesn't affect common code.
|
||
|
|
|
||
|
|
--- a/gcc/combine.cc
|
||
|
|
+++ b/gcc/combine.cc
|
||
|
|
@@ -92,6 +92,11 @@ along with GCC; see the file COPYING3. If not see
|
||
|
|
#include "function-abi.h"
|
||
|
|
#include "rtlanal.h"
|
||
|
|
|
||
|
|
+/* Number of times I2 didn't change in try_combine. Used to prevent a
|
||
|
|
+ combinatorial explosion. */
|
||
|
|
+
|
||
|
|
+static int combine_unchanged;
|
||
|
|
+
|
||
|
|
/* Number of attempts to combine instructions in this function. */
|
||
|
|
|
||
|
|
static int combine_attempts;
|
||
|
|
@@ -1127,6 +1132,7 @@ combine_instructions (rtx_insn *f, unsigned int nregs)
|
||
|
|
return false;
|
||
|
|
|
||
|
|
combine_attempts = 0;
|
||
|
|
+ combine_unchanged = 0;
|
||
|
|
combine_merges = 0;
|
||
|
|
combine_extras = 0;
|
||
|
|
combine_successes = 0;
|
||
|
|
@@ -4196,6 +4201,10 @@ try_combine (rtx_insn *i3, rtx_insn *i2, rtx_insn *i1, rtx_insn *i0,
|
||
|
|
adjust_for_new_dest (i3);
|
||
|
|
}
|
||
|
|
|
||
|
|
+ bool i2_unchanged = false;
|
||
|
|
+ if (rtx_equal_p (newi2pat, PATTERN (i2)))
|
||
|
|
+ i2_unchanged = true;
|
||
|
|
+
|
||
|
|
/* We now know that we can do this combination. Merge the insns and
|
||
|
|
update the status of registers and LOG_LINKS. */
|
||
|
|
|
||
|
|
@@ -4762,6 +4771,13 @@ try_combine (rtx_insn *i3, rtx_insn *i2, rtx_insn *i1, rtx_insn *i0,
|
||
|
|
combine_successes++;
|
||
|
|
undo_commit ();
|
||
|
|
|
||
|
|
+ if (i2_unchanged)
|
||
|
|
+ {
|
||
|
|
+ if (combine_unchanged == 1000)
|
||
|
|
+ return i3;
|
||
|
|
+ ++combine_unchanged;
|
||
|
|
+ }
|
||
|
|
+
|
||
|
|
rtx_insn *ret = newi2pat ? i2 : i3;
|
||
|
|
if (added_links_insn && DF_INSN_LUID (added_links_insn) < DF_INSN_LUID (ret))
|
||
|
|
ret = added_links_insn;
|