From 770e8279194dd8752beaf16b0959f7c0071b407e Mon Sep 17 00:00:00 2001 From: Avi Kivity Date: Mon, 21 Oct 2024 19:12:22 +0300 Subject: [PATCH] build: disable Seastar exception hack In [1], Seastar started to bypass a lock in libgcc's exception throwing mechanism to allow scalability on large machines. The problem is documented in [2] and reported as fixed. In [3], testing results on a 2s96c192t machine are reported. The problem appears indeed fixed with gcc 14's runtime (which we use, even though we build with clang). Given the new results, we can safely drop the exception scalability hack. As [1] states that the hack causes the loss of a translation cache, we may gain some performance this way. With that, we disable the cache by defining some random macros. [1] https://github.com/scylladb/seastar/464f5e3ae43b366b05573018fc46321863bf2fae [2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71744 [3] https://github.com/scylladb/seastar/issues/2479#issuecomment-2427098413 --- configure.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/configure.py b/configure.py index 0a4393261af6..071b736210d0 100755 --- a/configure.py +++ b/configure.py @@ -1662,6 +1662,9 @@ def dynamic_linker_option(): curdir = os.getcwd() user_cflags = args.user_cflags + f" -ffile-prefix-map={curdir}=." +# Since gcc 13, libgcc doesn't need the exception workaround +user_cflags += ' -DNO_EXCEPTION_INTERCEPT -DSEASTAR_NO_EXCEPTION_HACK' + if args.target != '': user_cflags += ' -march=' + args.target