485487 2024-04-13 14:12:05 +0000 glibc built with -march=x86-64-v3 does not work due to ld.so strcmp 2025-03-11 11:22:56 +0000 1 1 6 Developer tools valgrind memcheck 3.23 GIT Other Linux RESOLVED FIXED https://bugs.kde.org/show_bug.cgi?id=501348 NOR normal --- 1 fweimer jseward fweimer mark sam 0 oldest_to_newest 2311285 0 168460 fweimer 2024-04-13 14:12:05 +0000 Created attachment 168460 valgrind-amd64-redir-strcmp.patch This was originally observed as a glibc test case failure: [PATCH] x86_64: Suppress false positive valgrind error <https://inbox.sourceware.org/libc-alpha/20240226000326.3844791-1-hjl.tools@gmail.com/> However, the analysis was incorrect, and the test failure actually shows that the newly built glibc is incompatible with valgrind. I think it's necessary to add a redirector for strcmp. We see this in Fedora with valgrind-3.22.0-7.fc40.x86_64, which already has one AVX2 inline strcmp fix applied. 2311316 1 mark 2024-04-13 15:08:04 +0000 The proposed patch looks reasonable. Did you test it? Is there a (fedora?) glibc build that triggers this issue? 2311317 2 fweimer 2024-04-13 15:16:33 +0000 (In reply to Mark Wielaard from comment #1) > The proposed patch looks reasonable. Did you test it? Sort of, the valgrind build passed, and the run-time failures are gone. Earlier versions of the patch with an incorrect implementation failed really spectacularly. But I do not write x86-64 assembly for a living (well, not really), so take it with a grain of salt. > Is there a (fedora?) glibc build that triggers this issue? We had a glibc build bug/RPM spec file issue that resulted in rawhide binaries not using x86-64-v3 string functions in ld.so, and I missed the incorrect glibc upstream testsuite change at the time. We have this as a redundant test in glibc.spec that fails the Fedora build, that's how I noticed. We are getting better at not producing a glibc that is totally incompatible with valgrind. But all this means that there is no Fedora build. 8-) I can prepare a rawhide scratch build for you if this helps, though. 2311324 3 fweimer 2024-04-13 15:34:47 +0000 H.J. suggests to use the implementation from <https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/i386/i686/strcmp.S;h=e71d97f2a63a2d3c31e4c1be395c1a1f5e0200ff;hb=HEAD> (adapted for x86-64). I can submit that if there are no licensing concerns. 2311344 4 fweimer 2024-04-13 17:50:10 +0000 Based on a further suggestion from H.J., I put the strcmp implementation posted here into glibc and ran its strcmp test suite against it. It passes. For such a simple implementation it should have really good coverage. So I think the posted implementation should be okay. 2311346 5 mark 2024-04-13 17:56:03 +0000 (In reply to Florian Weimer from comment #4) > Based on a further suggestion from H.J., I put the strcmp implementation > posted here into glibc and ran its strcmp test suite against it. It passes. > For such a simple implementation it should have really good coverage. So I > think the posted implementation should be okay. Yes. The license would be fine. But simpler is actually better for valgrind. The code will run on the virtual machine and will be instrumented by tools. I'll add it to the next fedora rawhide valgrind build so we can test it a bit more. 2311368 6 mark 2024-04-13 19:27:14 +0000 (In reply to Mark Wielaard from comment #5) > I'll add it to the next fedora rawhide valgrind build so we can test it a > bit more. There is now valgrind-3.22.0-8.fc41 with this and some other (x86-64-v3) fixes to try out. https://bodhi.fedoraproject.org/updates/FEDORA-2024-7579427715 2311486 7 mark 2024-04-14 09:29:56 +0000 There is now also an fedora eln build: https://koji.fedoraproject.org/koji/buildinfo?buildID=2437715 Since eln is build with -march=x86-64-v3 this might be a good testing environment. Please let me know if you are able to build/test a glibc against one of the above packages with this patch applied. 2311581 8 fweimer 2024-04-14 17:20:11 +0000 The ELN glibc scratch build succeeded: https://src.fedoraproject.org/rpms/glibc/pull-request/92#comment-194579 There's a TMT failure, but it's unrelated (potential kernel robust mutex issue). This test does not run under valgrind anyway. (That POWER10 work fixed some glibc.spec issues, which exposed the pre-existing x86-64-v3 bug.) 2311603 9 mark 2024-04-14 20:19:54 +0000 commit dad279b80c1da4fb2c842cca9783ad0e49339ddc Author: Florian Weimer <fweimer@redhat.com> Date: Sun Apr 14 21:56:03 2024 +0200 add_hardwired_spec for ld-linux-x86-64.so.2 strcmp glibc built with -march=x86-64-v3 does not work due to ld.so strcmp https://bugs.kde.org/show_bug.cgi?id=485487 168460 2024-04-13 14:12:05 +0000 2024-04-13 14:12:05 +0000 valgrind-amd64-redir-strcmp.patch valgrind-amd64-redir-strcmp.patch text/plain 2181 fweimer ZGlmZiAtLWdpdCBhL2NvcmVncmluZC9tX3JlZGlyLmMgYi9jb3JlZ3JpbmQvbV9yZWRpci5jCmlu ZGV4IGNlZjI0MWI0Zjg1YWFkMjMuLmNlNjA3M2E2MjE5MGE2MzkgMTAwNjQ0Ci0tLSBhL2NvcmVn cmluZC9tX3JlZGlyLmMKKysrIGIvY29yZWdyaW5kL21fcmVkaXIuYwpAQCAtMTQxMCw2ICsxNDEw LDE1IEBAIHZvaWQgVkdfKHJlZGlyX2luaXRpYWxpc2UpICggdm9pZCApCiAgICAgICAgICBjb21w bGFpbl9hYm91dF9zdHJpcHBlZF9nbGliY19sZHNvCiAjICAgICAgICBlbmRpZgogICAgICAgKTsg ICAKKyAgICAgIGFkZF9oYXJkd2lyZWRfc3BlYygKKyAgICAgICAgICJsZC1saW51eC14ODYtNjQu c28uMiIsICJzdHJjbXAiLAorICAgICAgICAgKEFkZHIpJlZHXyhhbWQ2NF9saW51eF9SRURJUl9G T1Jfc3RyY21wKSwKKyMgICAgICAgIGlmbmRlZiBHTElCQ19NQU5EQVRPUllfU1RSTEVOX1JFRElS RUNUCisgICAgICAgICBOVUxMCisjICAgICAgICBlbHNlCisgICAgICAgICBjb21wbGFpbl9hYm91 dF9zdHJpcHBlZF9nbGliY19sZHNvCisjICAgICAgICBlbmRpZgorICAgICAgKTsKICAgIH0KIAog IyAgZWxpZiBkZWZpbmVkKFZHUF9wcGMzMl9saW51eCkKZGlmZiAtLWdpdCBhL2NvcmVncmluZC9t X3RyYW1wb2xpbmUuUyBiL2NvcmVncmluZC9tX3RyYW1wb2xpbmUuUwppbmRleCBkYTk2OTcyMzIz ODI0YmRiLi43MTZhZmIyMTIxYjg0YjU3IDEwMDY0NAotLS0gYS9jb3JlZ3JpbmQvbV90cmFtcG9s aW5lLlMKKysrIGIvY29yZWdyaW5kL21fdHJhbXBvbGluZS5TCkBAIC0yMTgsNiArMjE4LDI5IEBA IFZHXyhhbWQ2NF9saW51eF9SRURJUl9GT1Jfc3RybGVuKToKIC5MZm5FNToKIC5zaXplIFZHXyhh bWQ2NF9saW51eF9SRURJUl9GT1Jfc3RybGVuKSwgLi1WR18oYW1kNjRfbGludXhfUkVESVJfRk9S X3N0cmxlbikKIAorLyogVGhlcmUncyBubyBwYXJ0aWN1bGFyIHJlYXNvbiB0aGF0IHRoaXMgbmVl ZHMgdG8gYmUgaGFuZHdyaXR0ZW4KKyAgIGFzc2VtYmx5LCBidXQgc2luY2UgdGhhdCdzIHdoYXQg dGhpcyBmaWxlIGNvbnRhaW5zLCBoZXJlJ3MgYQorICAgc2ltcGxlIHN0cmNtcC4pCisqLworLmds b2JhbCBWR18oYW1kNjRfbGludXhfUkVESVJfRk9SX3N0cmNtcCkKKy50eXBlICAgVkdfKGFtZDY0 X2xpbnV4X1JFRElSX0ZPUl9zdHJjbXApLCBAZnVuY3Rpb24KK1ZHXyhhbWQ2NF9saW51eF9SRURJ Ul9GT1Jfc3RyY21wKToKKwl4b3JsCSVlY3gsICVlY3gKKzE6CisJbW92emJsCSglcmRpLCAlcmN4 KSwgJWVheAorCW1vdnpibAkoJXJzaSwgJXJjeCksICVlZHgKKwl0ZXN0YgklYWwsICVhbAorCWpu ZQkyZgorCW1vdmwJJWVkeCwgJWVheAorCW5lZ2wJJWVheAorCXJldAorMjoKKwlpbmNxCSVyY3gK KwlzdWJsCSVlZHgsICVlYXgKKwlqZQkxYgorCXJldAorLnNpemUgVkdfKGFtZDY0X2xpbnV4X1JF RElSX0ZPUl9zdHJjbXApLCAuLVZHXyhhbWQ2NF9saW51eF9SRURJUl9GT1Jfc3RyY21wKQorCiAu Z2xvYmFsIFZHXyhhbWQ2NF9saW51eF9SRURJUl9GT1JfaW5kZXgpCiAudHlwZSAgIFZHXyhhbWQ2 NF9saW51eF9SRURJUl9GT1JfaW5kZXgpLCBAZnVuY3Rpb24KIFZHXyhhbWQ2NF9saW51eF9SRURJ Ul9GT1JfaW5kZXgpOgpkaWZmIC0tZ2l0IGEvY29yZWdyaW5kL3B1Yl9jb3JlX3RyYW1wb2xpbmUu aCBiL2NvcmVncmluZC9wdWJfY29yZV90cmFtcG9saW5lLmgKaW5kZXggNTRjNTc1YTcyMzNhOTQ1 MS4uYzExMDYwMTM5ZTU3NjYxNiAxMDA2NDQKLS0tIGEvY29yZWdyaW5kL3B1Yl9jb3JlX3RyYW1w b2xpbmUuaAorKysgYi9jb3JlZ3JpbmQvcHViX2NvcmVfdHJhbXBvbGluZS5oCkBAIC03Nyw2ICs3 Nyw3IEBAIGV4dGVybiBBZGRyIFZHXyhhbWQ2NF9saW51eF9SRURJUl9GT1JfdmdldHRpbWVvZmRh eSk7CiBleHRlcm4gQWRkciBWR18oYW1kNjRfbGludXhfUkVESVJfRk9SX3Z0aW1lKTsKIGV4dGVy biBBZGRyIFZHXyhhbWQ2NF9saW51eF9SRURJUl9GT1JfdmdldGNwdSk7CiBleHRlcm4gVUludCBW R18oYW1kNjRfbGludXhfUkVESVJfRk9SX3N0cmxlbikoIHZvaWQqICk7CitleHRlcm4gVUludCBW R18oYW1kNjRfbGludXhfUkVESVJfRk9SX3N0cmNtcCkoIHZvaWQqLCB2b2lkKiApOwogZXh0ZXJu IENoYXIqIFZHXyhhbWQ2NF9saW51eF9SRURJUl9GT1JfaW5kZXgpICggY29uc3QgQ2hhciosIElu dCApOwogI2VuZGlmCiAK