Skip to contentSkip to conversation

bitbake.conf: remove DEBUG_PREFIX_MAP from TARGET_LDFLAGS

PATCH In review [email protected] 5 participants 15 messages Jan 22 – Feb 3

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bitbake.conf: remove DEBUG_PREFIX_MAP from TARGET_LDFLAGS

From: Randolph Sapp <[email protected]>

Now that the previous bug affecting binary reproducibility has been addressed [1], we can revert this patch. This will resolve issues with cgo applications becoming unreprodcible.

Currently go considers link arguments to be sacred, meaning any change should produce a different binary output. They ensure this by baking link arguments into the intermediary output, changing the content ID of that step. As such, the marco prefixes inadvertently end up adding build paths to the output binary instead of removing them if they are passed as link arguments to cgo applications.

These paths are later stripped out again, but at this point the content ID of the dependency has changed and thus the build ID of the end application will be affected by the cascade of hash changes. See the upstream bug for more information [2].

This reverts commit fddaecc88979967d0e00e2fafdbaaabec030da9f.

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473
[2] https://github.com/golang/go/issues/77218

---

This resolves the previously reported emptty issues:
https://lists.openembedded.org/g/openembedded-core/message/228549

 meta/conf/bitbake.conf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/conf/bitbake.conf +1 -1
index 88f4d0df69..da873c3f4e 100644
@@ -634,7 +634,7 @@ TARGET_LINK_HASH_STYLE ?= "${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_
 ASNEEDED ?= "-Wl,--as-needed"
 
 export LDFLAGS = "${TARGET_LDFLAGS}"
-TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED} ${DEBUG_PREFIX_MAP}"
+TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}"
 # mips does not support GNU hash style therefore we override
 LINKER_HASH_STYLE:mipsarch:libc-musl = "sysv"
 
Signature · 1 line
2.52.0
#1 lore raw
Quoted 40 lines · Randolph Sapp via lists.openembedded.org

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

From: Randolph Sapp<[email protected]>

Now that the previous bug affecting binary reproducibility has been addressed [1], we can revert this patch. This will resolve issues with cgo applications becoming unreprodcible.

Currently go considers link arguments to be sacred, meaning any change should produce a different binary output. They ensure this by baking link arguments into the intermediary output, changing the content ID of that step. As such, the marco prefixes inadvertently end up adding build paths to the output binary instead of removing them if they are passed as link arguments to cgo applications.

These paths are later stripped out again, but at this point the content ID of the dependency has changed and thus the build ID of the end application will be affected by the cascade of hash changes. See the upstream bug for more information [2].

This reverts commit fddaecc88979967d0e00e2fafdbaaabec030da9f.

[1]https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473 [2]https://github.com/golang/go/issues/77218

---

This resolves the previously reported emptty issues:
https://lists.openembedded.org/g/openembedded-core/message/228549

  meta/conf/bitbake.conf | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
meta/conf/bitbake.conf +0 -0
index 88f4d0df69..da873c3f4e 100644

meta/conf/bitbake.conf TARGET_LINK_HASH_STYLE ?="${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= "-Wl,--as-needed" · line 636

Quoted 4 lines
meta/conf/bitbake.conf +1 -1
@@ -634,7 +634,7 @@ TARGET_LINK_HASH_STYLE ?="${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= "-Wl,--as-needed"
 
  export LDFLAGS = "${TARGET_LDFLAGS}"
-TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED} ${DEBUG_PREFIX_MAP}"
+TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}"

Hi,

After check the related gcc bug and yocto bug,  gcc bug comments 21 ([1])  and yocto bug comments 13 ([2]),

my understanding is that,  when lto is enabled,  even with gcc fix [3],  we still need DEBUG_PREFIX_MAP in LDFLAGS

to make bin reproducible. Also seems Tony's commit [4] is after gcc fix [3].

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473#c21

[2] https://bugzilla.yoctoproject.org/show_bug.cgi?id=14481#c13

[3] https://gcc.gnu.org/g:7cc2df084b7977653a9b59cbc34a9ad500ae619c

[4] https://git.openembedded.org/openembedded-core/commit/?id=fddaecc88979967d0e00e2fafdbaaabec030da9f

@Tony, @Randolph, Could please correct me if my understanding is not right, Thanks.

//Changqing

Quoted 15 lines
meta/conf/bitbake.conf +0 -1
  # mips does not support GNU hash style therefore we override
  LINKER_HASH_STYLE:mipsarch:libc-musl = "sysv"
 
--

2.52.0

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#229857):https://lists.openembedded.org/g/openembedded-core/message/229857 Mute This Topic:https://lists.openembedded.org/mt/117408334/3616873
Group Owner:[email protected] Unsubscribe:https://lists.openembedded.org/g/openembedded-core/unsub [[email protected]]
-=-=-=-=-=-=-=-=-=-=-=-

#2 lore raw
Quoted 63 lines · Changqing Li
Quoted 40 lines · Randolph Sapp via lists.openembedded.org

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

From: Randolph Sapp <[email protected]>

Now that the previous bug affecting binary reproducibility has been addressed [1], we can revert this patch. This will resolve issues with cgo applications becoming unreprodcible.

Currently go considers link arguments to be sacred, meaning any change should produce a different binary output. They ensure this by baking link arguments into the intermediary output, changing the content ID of that step. As such, the marco prefixes inadvertently end up adding build paths to the output binary instead of removing them if they are passed as link arguments to cgo applications.

These paths are later stripped out again, but at this point the content ID of the dependency has changed and thus the build ID of the end application will be affected by the cascade of hash changes. See the upstream bug for more information [2].

This reverts commit fddaecc88979967d0e00e2fafdbaaabec030da9f.

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473
[2] https://github.com/golang/go/issues/77218

---

This resolves the previously reported emptty issues:
https://lists.openembedded.org/g/openembedded-core/message/228549

 meta/conf/bitbake.conf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/conf/bitbake.conf +0 -0
index 88f4d0df69..da873c3f4e 100644

meta/conf/bitbake.conf TARGET_LINK_HASH_STYLE ?= "${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= "-Wl,--as-needed" · line 636

Quoted 4 lines
meta/conf/bitbake.conf +1 -1
@@ -634,7 +634,7 @@ TARGET_LINK_HASH_STYLE ?= "${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= "-Wl,--as-needed"
 
 export LDFLAGS = "${TARGET_LDFLAGS}"
-TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED} ${DEBUG_PREFIX_MAP}"
+TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}"

Hi,

After check the related gcc bug and yocto bug,  gcc bug comments 21 ([1])  and yocto bug comments 13 ([2]),

my understanding is that,  when lto is enabled,  even with gcc fix [3],  we still need DEBUG_PREFIX_MAP in LDFLAGS

to make bin reproducible. Also seems Tony's commit [4] is after gcc fix [3].

[1]  https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473#c21

[2]  https://bugzilla.yoctoproject.org/show_bug.cgi?id=14481#c13

Quoted 10 lines

[3]  https://gcc.gnu.org/g:7cc2df084b7977653a9b59cbc34a9ad500ae619c

[4]
https://git.openembedded.org/openembedded-core/commit/?id=fddaecc88979967d0e00e2fafdbaaabec030da9f

@Tony, @Randolph, Could please correct me if my understanding is not right, Thanks.

Yes, LDFLAGS needs DEBUG_PREFIX_MAP to make LTO builds reproducible, unless something else has changed since 2021 to make it unnecessary. I have not kept up with the issue.

Tony

#3 lore raw
Quoted 74 lines · Tony Battersby
Quoted 63 lines · Changqing Li
Quoted 40 lines · Randolph Sapp via lists.openembedded.org

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

From: Randolph Sapp <[email protected]>

Now that the previous bug affecting binary reproducibility has been addressed [1], we can revert this patch. This will resolve issues with cgo applications becoming unreprodcible.

Currently go considers link arguments to be sacred, meaning any change should produce a different binary output. They ensure this by baking link arguments into the intermediary output, changing the content ID of that step. As such, the marco prefixes inadvertently end up adding build paths to the output binary instead of removing them if they are passed as link arguments to cgo applications.

These paths are later stripped out again, but at this point the content ID of the dependency has changed and thus the build ID of the end application will be affected by the cascade of hash changes. See the upstream bug for more information [2].

This reverts commit fddaecc88979967d0e00e2fafdbaaabec030da9f.

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473
[2] https://github.com/golang/go/issues/77218

---

This resolves the previously reported emptty issues:
https://lists.openembedded.org/g/openembedded-core/message/228549

 meta/conf/bitbake.conf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/conf/bitbake.conf +0 -0
index 88f4d0df69..da873c3f4e 100644

meta/conf/bitbake.conf TARGET_LINK_HASH_STYLE ?= "${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= "-Wl,--as-needed" · line 636

Quoted 4 lines
meta/conf/bitbake.conf +1 -1
@@ -634,7 +634,7 @@ TARGET_LINK_HASH_STYLE ?= "${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= "-Wl,--as-needed"
 
 export LDFLAGS = "${TARGET_LDFLAGS}"
-TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED} ${DEBUG_PREFIX_MAP}"
+TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}"

Hi,

After check the related gcc bug and yocto bug,  gcc bug comments 21 ([1])  and yocto bug comments 13 ([2]),

my understanding is that,  when lto is enabled,  even with gcc fix [3],  we still need DEBUG_PREFIX_MAP in LDFLAGS

to make bin reproducible. Also seems Tony's commit [4] is after gcc fix [3].

[1]  https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473#c21

[2]  https://bugzilla.yoctoproject.org/show_bug.cgi?id=14481#c13

Quoted 10 lines

[3]  https://gcc.gnu.org/g:7cc2df084b7977653a9b59cbc34a9ad500ae619c

[4]
https://git.openembedded.org/openembedded-core/commit/?id=fddaecc88979967d0e00e2fafdbaaabec030da9f

@Tony, @Randolph, Could please correct me if my understanding is not right, Thanks.

Quoted 5 lines

Yes, LDFLAGS needs DEBUG_PREFIX_MAP to make LTO builds reproducible, unless something else has changed since 2021 to make it unnecessary. I have not kept up with the issue.

Tony

I think there has been some change because I tested this locally with the normal reproducible builds selftest and no new packages were added to the failing list. This included oe-core and some layers from meta-openembedded as well.

The example code given in the report is fairly rudimentary, we should have hit the bug if it was still in effect across a majority of packages if my understanding is correct.

I'm also curious to see what golang want's to do about this though. Baking build options into a binary seems a little silly, but I'm sure there was some unusual behavior that led them to that point.

As indicated on another thread, we could also just remove this string from Go packages LDFLAGS if that puts everyone at ease.

- Randolph

#4 lore raw
Quoted 76 lines · Randolph Sapp via lists.openembedded.org

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

Quoted 72 lines · Tony Battersby
Quoted 71 lines · Changqing Li
Quoted 40 lines · Randolph Sapp via lists.openembedded.org

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

From: Randolph Sapp<[email protected]>

Now that the previous bug affecting binary reproducibility has been addressed [1], we can revert this patch. This will resolve issues with cgo applications becoming unreprodcible.

Currently go considers link arguments to be sacred, meaning any change should produce a different binary output. They ensure this by baking link arguments into the intermediary output, changing the content ID of that step. As such, the marco prefixes inadvertently end up adding build paths to the output binary instead of removing them if they are passed as link arguments to cgo applications.

These paths are later stripped out again, but at this point the content ID of the dependency has changed and thus the build ID of the end application will be affected by the cascade of hash changes. See the upstream bug for more information [2].

This reverts commit fddaecc88979967d0e00e2fafdbaaabec030da9f.

[1]https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473 [2]https://github.com/golang/go/issues/77218

---

This resolves the previously reported emptty issues:
https://lists.openembedded.org/g/openembedded-core/message/228549

  meta/conf/bitbake.conf | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
meta/conf/bitbake.conf +0 -0
index 88f4d0df69..da873c3f4e 100644

meta/conf/bitbake.conf TARGET_LINK_HASH_STYLE ?="${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= "-Wl,--as-needed" · line 636

Quoted 4 lines
meta/conf/bitbake.conf +1 -1
@@ -634,7 +634,7 @@ TARGET_LINK_HASH_STYLE ?="${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= "-Wl,--as-needed"
 
  export LDFLAGS = "${TARGET_LDFLAGS}"
-TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED} ${DEBUG_PREFIX_MAP}"
+TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}"

Hi,

After check the related gcc bug and yocto bug, gcc bug comments 21 ([1]) and yocto bug comments 13 ([2]),

my understanding is that, when lto is enabled, even with gcc fix [3], we still need DEBUG_PREFIX_MAP in LDFLAGS

to make bin reproducible. Also seems Tony's commit [4] is after gcc fix [3].

[1]https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473#c21

[2]https://bugzilla.yoctoproject.org/show_bug.cgi?id=14481#c13

[3]https://gcc.gnu.org/g:7cc2df084b7977653a9b59cbc34a9ad500ae619c

[4]
https://git.openembedded.org/openembedded-core/commit/?id=fddaecc88979967d0e00e2fafdbaaabec030da9f

@Tony, @Randolph, Could please correct me if my understanding is not right, Thanks.

Quoted 8 lines
Quoted 5 lines

Yes, LDFLAGS needs DEBUG_PREFIX_MAP to make LTO builds reproducible, unless something else has changed since 2021 to make it unnecessary. I have not kept up with the issue.

Tony

I think there has been some change because I tested this locally with the normal reproducible builds selftest and no new packages were added to the failing list. This included oe-core and some layers from meta-openembedded as well.

Got it,  if DEBUG_PREFIX_MAP is not needed any more,  I think remove DEBUG_PREFIX_MAP from LDFLAGS

is a good way to fix the issue.

Quoted 8 lines

The example code given in the report is fairly rudimentary, we should have hit the bug if it was still in effect across a majority of packages if my understanding is correct.

I'm also curious to see what golang want's to do about this though. Baking build options into a binary seems a little silly, but I'm sure there was some unusual behavior that led them to that point.

+1

//Changqing

Quoted 14 lines

As indicated on another thread, we could also just remove this string from Go packages LDFLAGS if that puts everyone at ease.

- Randolph

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#230026):https://lists.openembedded.org/g/openembedded-core/message/230026 Mute This Topic:https://lists.openembedded.org/mt/117408334/3616873
Group Owner:[email protected] Unsubscribe:https://lists.openembedded.org/g/openembedded-core/unsub [[email protected]]
-=-=-=-=-=-=-=-=-=-=-=-

#5 lore raw

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

I wonder if this is related.

Quoted 126 lines · Changqing Li via lists.openembedded.org

On 1/27/26 04:50, Randolph Sapp via lists.openembedded.org wrote:

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

On Mon Jan 26, 2026 at 9:04 AM CST, Tony Battersby wrote:

On 1/26/26 03:39, Changqing Li wrote:

On 1/23/26 03:50, Randolph Sapp via lists.openembedded.org wrote:

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

From: Randolph Sapp <[email protected]> <[email protected]>

Now that the previous bug affecting binary reproducibility has been addressed [1], we can revert this patch. This will resolve issues with cgo applications becoming unreprodcible.

Currently go considers link arguments to be sacred, meaning any change should produce a different binary output. They ensure this by baking link arguments into the intermediary output, changing the content ID of that step. As such, the marco prefixes inadvertently end up adding build paths to the output binary instead of removing them if they are passed as link arguments to cgo applications.

These paths are later stripped out again, but at this point the content ID of the dependency has changed and thus the build ID of the end application will be affected by the cascade of hash changes. See the upstream bug for more information [2].

This reverts commit fddaecc88979967d0e00e2fafdbaaabec030da9f.

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473
[2] https://github.com/golang/go/issues/77218

---

This resolves the previously reported emptty issues:https://lists.openembedded.org/g/openembedded-core/message/228549

 meta/conf/bitbake.conf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/conf/bitbake.conf +1 -1
index 88f4d0df69..da873c3f4e 100644
@@ -634,7 +634,7 @@ TARGET_LINK_HASH_STYLE ?= "${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= " <$%7B@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ASNEEDED?=>-Wl,--as-needed"
 
 export LDFLAGS = "${TARGET_LDFLAGS}"
-TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED} ${DEBUG_PREFIX_MAP}"
+TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}"
 
 Hi,
 

After check the related gcc bug and yocto bug, gcc bug comments 21 ([1]) and yocto bug comments 13 ([2]),

my understanding is that, when lto is enabled, even with gcc fix [3], we still need DEBUG_PREFIX_MAP in LDFLAGS

to make bin reproducible. Also seems Tony's commit [4] is after gcc fix [3].

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473#c21

[2] https://bugzilla.yoctoproject.org/show_bug.cgi?id=14481#c13

[3] https://gcc.gnu.org/g:7cc2df084b7977653a9b59cbc34a9ad500ae619c

[4] https://git.openembedded.org/openembedded-core/commit/?id=fddaecc88979967d0e00e2fafdbaaabec030da9f

@Tony, @Randolph, Could please correct me if my understanding is not right, Thanks.

Yes, LDFLAGS needs DEBUG_PREFIX_MAP to make LTO builds reproducible, unless something else has changed since 2021 to make it unnecessary. I have not kept up with the issue.

Tony

I think there has been some change because I tested this locally with the normal reproducible builds selftest and no new packages were added to the failing list. This included oe-core and some layers from meta-openembedded as well.

Got it, if DEBUG_PREFIX_MAP is not needed any more, I think remove DEBUG_PREFIX_MAP from LDFLAGS

is a good way to fix the issue.

The example code given in the report is fairly rudimentary, we should have hit the bug if it was still in effect across a majority of packages if my understanding is correct.

I'm also curious to see what golang want's to do about this though. Baking build options into a binary seems a little silly, but I'm sure there was some unusual behavior that led them to that point.

+1

//Changqing

As indicated on another thread, we could also just remove this string from Go packages LDFLAGS if that puts everyone at ease.

- Randolph

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group. View/Reply Online (#230027):
https://lists.openembedded.org/g/openembedded-core/message/230027
Mute This Topic: https://lists.openembedded.org/mt/117408334/1997914
Group Owner: [email protected]
Unsubscribe: https://lists.openembedded.org/g/openembedded-core/unsub [ [email protected]]
-=-=-=-=-=-=-=-=-=-=-=-

#6 lore raw
Quoted 5 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

If that bug is actually still in effect then this requires some change to golang to avoid baking cgo_ldflags into the binaries. Right now we can get away with a go class specific LDFLAGS override if we want to, but that GCC bug may begin to affect go binary reproducibility in the future.

- Randolph

Quoted 121 lines
Quoted 118 lines · Changqing Li via lists.openembedded.org

On 1/27/26 04:50, Randolph Sapp via lists.openembedded.org wrote:

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

On Mon Jan 26, 2026 at 9:04 AM CST, Tony Battersby wrote:

On 1/26/26 03:39, Changqing Li wrote:

On 1/23/26 03:50, Randolph Sapp via lists.openembedded.org wrote:

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

From: Randolph Sapp <[email protected]> <[email protected]>

Now that the previous bug affecting binary reproducibility has been addressed [1], we can revert this patch. This will resolve issues with cgo applications becoming unreprodcible.

Currently go considers link arguments to be sacred, meaning any change should produce a different binary output. They ensure this by baking link arguments into the intermediary output, changing the content ID of that step. As such, the marco prefixes inadvertently end up adding build paths to the output binary instead of removing them if they are passed as link arguments to cgo applications.

These paths are later stripped out again, but at this point the content ID of the dependency has changed and thus the build ID of the end application will be affected by the cascade of hash changes. See the upstream bug for more information [2].

This reverts commit fddaecc88979967d0e00e2fafdbaaabec030da9f.

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473
[2] https://github.com/golang/go/issues/77218

---

This resolves the previously reported emptty issues:https://lists.openembedded.org/g/openembedded-core/message/228549

 meta/conf/bitbake.conf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/conf/bitbake.conf +1 -1
index 88f4d0df69..da873c3f4e 100644
@@ -634,7 +634,7 @@ TARGET_LINK_HASH_STYLE ?= "${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= " <$%7B@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ASNEEDED?=>-Wl,--as-needed"
 
 export LDFLAGS = "${TARGET_LDFLAGS}"
-TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED} ${DEBUG_PREFIX_MAP}"
+TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}"
 
 Hi,
 

After check the related gcc bug and yocto bug, gcc bug comments 21 ([1]) and yocto bug comments 13 ([2]),

my understanding is that, when lto is enabled, even with gcc fix [3], we still need DEBUG_PREFIX_MAP in LDFLAGS

to make bin reproducible. Also seems Tony's commit [4] is after gcc fix [3].

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473#c21

[2] https://bugzilla.yoctoproject.org/show_bug.cgi?id=14481#c13

[3] https://gcc.gnu.org/g:7cc2df084b7977653a9b59cbc34a9ad500ae619c

[4] https://git.openembedded.org/openembedded-core/commit/?id=fddaecc88979967d0e00e2fafdbaaabec030da9f

@Tony, @Randolph, Could please correct me if my understanding is not right, Thanks.

Yes, LDFLAGS needs DEBUG_PREFIX_MAP to make LTO builds reproducible, unless something else has changed since 2021 to make it unnecessary. I have not kept up with the issue.

Tony

I think there has been some change because I tested this locally with the normal reproducible builds selftest and no new packages were added to the failing list. This included oe-core and some layers from meta-openembedded as well.

Got it, if DEBUG_PREFIX_MAP is not needed any more, I think remove DEBUG_PREFIX_MAP from LDFLAGS

is a good way to fix the issue.

The example code given in the report is fairly rudimentary, we should have hit the bug if it was still in effect across a majority of packages if my understanding is correct.

I'm also curious to see what golang want's to do about this though. Baking build options into a binary seems a little silly, but I'm sure there was some unusual behavior that led them to that point.

+1

//Changqing

As indicated on another thread, we could also just remove this string from Go packages LDFLAGS if that puts everyone at ease.

- Randolph

#7 lore raw
Quoted 7 lines · Randolph Sapp
Quoted 3 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

Quoted 1 line
Quoted 8 lines
Quoted 1 line

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

I think this is exposing the underlying issue in packages perhaps because they are
not respecting CFLAGS and when it's gone from LDFLAGS the mapping related flags go missing. I have a local patch to fix memstat e.g in master-next, all failures
perhaps need to be looked into one by one and passed cflags from OE env which
might have been missed thus far.

Quoted 149 lines

If that bug is actually still in effect then this requires some change to golang
to avoid baking cgo_ldflags into the binaries. Right now we can get away with a
go class specific LDFLAGS override if we want to, but that GCC bug may begin to
affect go binary reproducibility in the future.

- Randolph

Quoted 1 line

On Mon, Jan 26, 2026 at 6:50 PM Changqing Li via lists.openembedded.org

<

Quoted 7 lines
Quoted 5 lines · [email protected]> wrote:

On 1/27/26 04:50, Randolph Sapp via lists.openembedded.org wrote:

CAUTION: This email comes from a non Wind River email account! Do not click links or open attachments unless you recognize the sender

and know the content is safe.

Quoted 9 lines
Quoted 9 lines · Tony Battersby

On 1/26/26 03:39, Changqing Li wrote:

On 1/23/26 03:50, Randolph Sapp via lists.openembedded.org wrote:

CAUTION: This email comes from a non Wind River email account! Do not click links or open attachments unless you recognize the sender

and know the content is safe.

Quoted 21 lines
Quoted 21 lines

From: Randolph Sapp <[email protected]> <[email protected]>

Now that the previous bug affecting binary reproducibility has been addressed [1], we can revert this patch. This will resolve issues with cgo applications becoming unreprodcible.

Currently go considers link arguments to be sacred, meaning any change should produce a different binary output. They ensure this by baking link arguments into the intermediary output, changing the content ID of that step. As such, the marco prefixes inadvertently end up adding build paths to the output binary instead of removing them if they are passed as link arguments to cgo applications.

These paths are later stripped out again, but at this point the content ID of the dependency has changed and thus the build ID of the end application will be affected by the cascade of hash changes. See the upstream bug for more information [2].

This reverts commit fddaecc88979967d0e00e2fafdbaaabec030da9f.

Quoted 7 lines
Quoted 7 lines

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473
[2] https://github.com/golang/go/issues/77218

---

This resolves the previously reported emptty issues:

https://lists.openembedded.org/g/openembedded-core/message/228549

Quoted 9 lines

meta/conf/bitbake.conf TARGET_LINK_HASH_STYLE ?=

Quoted 9 lines
 meta/conf/bitbake.conf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/conf/bitbake.conf +0 -0
index 88f4d0df69..da873c3f4e 100644
@@ -634,7 +634,7 @@ TARGET_LINK_HASH_STYLE ?=

"${@['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ ASNEEDED ?= " <$%7B@ ['-Wl,--hash-style=gnu',''][d.getVar('LINKER_HASH_ASNEEDED?=>-Wl,--as-needed"

Quoted 3 lines

meta/conf/bitbake.conf TARGET_LINK_HASH_STYLE ?= · line 635

Quoted 3 lines
meta/conf/bitbake.conf +0 -1
 export LDFLAGS = "${TARGET_LDFLAGS}"
-TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}

${DEBUG_PREFIX_MAP}"

Quoted 13 lines
Quoted 13 lines
meta/conf/bitbake.conf +1 -0
+TARGET_LDFLAGS = "-Wl,-O1 ${TARGET_LINK_HASH_STYLE} ${ASNEEDED}"
 
 Hi,
 

After check the related gcc bug and yocto bug, gcc bug comments 21 ([1]) and yocto bug comments 13 ([2]),

my understanding is that, when lto is enabled, even with gcc fix [3], we still need DEBUG_PREFIX_MAP in LDFLAGS

to make bin reproducible. Also seems Tony's commit [4] is after gcc fix [3].

Quoted 7 lines
Quoted 7 lines

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101473#c21

[2] https://bugzilla.yoctoproject.org/show_bug.cgi?id=14481#c13

[3] https://gcc.gnu.org/g:7cc2df084b7977653a9b59cbc34a9ad500ae619c

[4]

https://git.openembedded.org/openembedded-core/commit/?id=fddaecc88979967d0e00e2fafdbaaabec030da9f

Quoted 7 lines
Quoted 7 lines

@Tony, @Randolph, Could please correct me if my understanding is not right, Thanks.

Quoted 7 lines
Quoted 7 lines

Yes, LDFLAGS needs DEBUG_PREFIX_MAP to make LTO builds reproducible, unless something else has changed since 2021 to make it unnecessary. I have not kept up with the issue.

Tony

I think there has been some change because I tested this locally with

the normal

Quoted 1 line
Quoted 1 line

reproducible builds selftest and no new packages were added to the

failing list.

Quoted 3 lines
Quoted 3 lines

This included oe-core and some layers from meta-openembedded as well.

Got it, if DEBUG_PREFIX_MAP is not needed any more, I think remove

DEBUG_PREFIX_MAP

Quoted 5 lines
Quoted 5 lines

from LDFLAGS

is a good way to fix the issue.

The example code given in the report is fairly rudimentary, we should

have hit

Quoted 4 lines
Quoted 4 lines

the bug if it was still in effect across a majority of packages if my understanding is correct.

I'm also curious to see what golang want's to do about this though.

Baking build

Quoted 1 line
Quoted 1 line

options into a binary seems a little silly, but I'm sure there was some

unusual

Quoted 7 lines
Quoted 7 lines

behavior that led them to that point.

+1

//Changqing

As indicated on another thread, we could also just remove this string

from Go

Quoted 9 lines
Quoted 9 lines

packages LDFLAGS if that puts everyone at ease.

- Randolph

#8 lore raw
Quoted 15 lines · Khem Raj via lists.openembedded.org
Quoted 6 lines · Randolph Sapp
Quoted 4 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

Quoted 7 lines
Quoted 1 line

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

Quoted 5 lines

I think this is exposing the underlying issue in packages perhaps because they are not respecting CFLAGS and when it's gone from LDFLAGS the mapping related flags go missing. I have a local patch to fix memstat e.g in master-next, all failures perhaps need to be looked into one by one and passed cflags from OE env which might have been missed thus far.

Just noticed a similar failure with lxc (meta-virtualization), but while the prefix-map options are provided via CFLAGS, it only works when provided via LDFLAGS, probably due LTO (enabled by default by lxc).

Patch available at
https://lists.yoctoproject.org/g/meta-virtualization/message/9553

I guess recipes forcing LTO will probably have similar issues.

Thanks,

Ricardo

#9 lore raw
Quoted 28 lines · Ricardo Salveti via lists.openembedded.org
Quoted 15 lines · Khem Raj via lists.openembedded.org
Quoted 6 lines · Randolph Sapp
Quoted 4 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

Quoted 7 lines
Quoted 1 line

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

Quoted 5 lines

I think this is exposing the underlying issue in packages perhaps because they are not respecting CFLAGS and when it's gone from LDFLAGS the mapping related flags go missing. I have a local patch to fix memstat e.g in master-next, all failures perhaps need to be looked into one by one and passed cflags from OE env which might have been missed thus far.

Just noticed a similar failure with lxc (meta-virtualization), but while the prefix-map options are provided via CFLAGS, it only works when provided via LDFLAGS, probably due LTO (enabled by default by lxc).

Quoted 8 lines

Patch available at
https://lists.yoctoproject.org/g/meta-virtualization/message/9553

I guess recipes forcing LTO will probably have similar issues.

Thanks,

Ricardo

Ran a test locally. Looking at the do_compile log there are actually several compilation steps that do not have the macro prefixes passed as cc arguments.

That LTO bug has actually been addressed, it's just as Khem Raj pointed out, sometimes CFLAGS is ignored and passing the prefixes with LDFLAGS was masking that problem.

- Randolph

#10 lore raw
Quoted 33 lines · Randolph Sapp
Quoted 31 lines · Ricardo Salveti via lists.openembedded.org
Quoted 15 lines · Khem Raj via lists.openembedded.org
Quoted 6 lines · Randolph Sapp
Quoted 4 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

Quoted 7 lines
Quoted 1 line

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

Quoted 5 lines

I think this is exposing the underlying issue in packages perhaps because they are not respecting CFLAGS and when it's gone from LDFLAGS the mapping related flags go missing. I have a local patch to fix memstat e.g in master-next, all failures perhaps need to be looked into one by one and passed cflags from OE env which might have been missed thus far.

Just noticed a similar failure with lxc (meta-virtualization), but while the prefix-map options are provided via CFLAGS, it only works when provided via LDFLAGS, probably due LTO (enabled by default by lxc).

Patch available at
https://lists.yoctoproject.org/g/meta-virtualization/message/9553

Quoted 8 lines
Quoted 1 line

I guess recipes forcing LTO will probably have similar issues.

Ran a test locally. Looking at the do_compile log there are actually several compilation steps that do not have the macro prefixes passed as cc arguments.

That LTO bug has actually been addressed, it's just as Khem Raj pointed out, sometimes CFLAGS is ignored and passing the prefixes with LDFLAGS was masking that problem.

I did look into one of the binaries and it was built with the right cflags options, but didn't investigate too much.

I can also confirm that it works fine after disabling LTO, that is why I think it could be related to it.

Thanks,

Ricardo

#11 lore raw
Quoted 46 lines · Ricardo de Araujo (Salveti)
Quoted 33 lines · Randolph Sapp
Quoted 31 lines · Ricardo Salveti via lists.openembedded.org
Quoted 15 lines · Khem Raj via lists.openembedded.org
Quoted 6 lines · Randolph Sapp
Quoted 4 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

Quoted 7 lines
Quoted 1 line

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

Quoted 5 lines

I think this is exposing the underlying issue in packages perhaps because they are not respecting CFLAGS and when it's gone from LDFLAGS the mapping related flags go missing. I have a local patch to fix memstat e.g in master-next, all failures perhaps need to be looked into one by one and passed cflags from OE env which might have been missed thus far.

Just noticed a similar failure with lxc (meta-virtualization), but while the prefix-map options are provided via CFLAGS, it only works when provided via LDFLAGS, probably due LTO (enabled by default by lxc).

Patch available at
https://lists.yoctoproject.org/g/meta-virtualization/message/9553

Quoted 8 lines
Quoted 1 line

I guess recipes forcing LTO will probably have similar issues.

Ran a test locally. Looking at the do_compile log there are actually several compilation steps that do not have the macro prefixes passed as cc arguments.

That LTO bug has actually been addressed, it's just as Khem Raj pointed out, sometimes CFLAGS is ignored and passing the prefixes with LDFLAGS was masking that problem.

I did look into one of the binaries and it was built with the right cflags options, but didn't investigate too much.

Quoted 6 lines

I can also confirm that it works fine after disabling LTO, that is why I think it could be related to it.

Thanks,

Ricardo

Oh boy. Looking into this more today, spending plenty of time trying to make sure it wasn't meson or some other argument combination causing issues, I stumbled upon this thread [1]. They report the same issue with debug-prefixes, but they also cite potentially package dependent behavior with LTO. There are quite a few open threads about LTO behavior in general that are concerning.

Maybe this should be reverted and the alternative solution for cgo binary reproducibility should be taken instead [2]? I don't really have the resources to start on compiler development at the moment.

Curious what others have to say, as there seems to be no clean way out right now.

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109805
[2] https://lists.openembedded.org/g/openembedded-core/message/230059

- Randolph

#12 lore raw
Quoted 55 lines · Randolph Sapp
Quoted 42 lines · Ricardo de Araujo (Salveti)
Quoted 32 lines · Randolph Sapp
Quoted 31 lines · Ricardo Salveti via lists.openembedded.org
Quoted 15 lines · Khem Raj via lists.openembedded.org
Quoted 6 lines · Randolph Sapp
Quoted 4 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

Quoted 7 lines
Quoted 1 line

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

Quoted 5 lines

I think this is exposing the underlying issue in packages perhaps because they are not respecting CFLAGS and when it's gone from LDFLAGS the mapping related flags go missing. I have a local patch to fix memstat e.g in master-next, all failures perhaps need to be looked into one by one and passed cflags from OE env which might have been missed thus far.

Just noticed a similar failure with lxc (meta-virtualization), but while the prefix-map options are provided via CFLAGS, it only works when provided via LDFLAGS, probably due LTO (enabled by default by lxc).

Patch available at
https://lists.yoctoproject.org/g/meta-virtualization/message/9553

Quoted 8 lines
Quoted 1 line

I guess recipes forcing LTO will probably have similar issues.

Ran a test locally. Looking at the do_compile log there are actually several compilation steps that do not have the macro prefixes passed as cc arguments.

That LTO bug has actually been addressed, it's just as Khem Raj pointed out, sometimes CFLAGS is ignored and passing the prefixes with LDFLAGS was masking that problem.

Quoted 5 lines

I did look into one of the binaries and it was built with the right cflags options, but didn't investigate too much.

I can also confirm that it works fine after disabling LTO, that is why I think it could be related to it.

Oh boy. Looking into this more today, spending plenty of time trying to make sure it wasn't meson or some other argument combination causing issues, I stumbled upon this thread [1]. They report the same issue with debug-prefixes, but they also cite potentially package dependent behavior with LTO. There are quite a few open threads about LTO behavior in general that are concerning.

Quoted 3 lines

Maybe this should be reverted and the alternative solution for cgo binary reproducibility should be taken instead [2]? I don't really have the resources to start on compiler development at the moment.

Right, it might be indeed better to revert for now as we will probably face similar LTO issues with multiple recipes and layers, since the bug is still valid upstream.

Thanks,

Ricardo

#13 lore raw
Quoted 52 lines · Ricardo de Araujo (Salveti)

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

Quoted 48 lines · Randolph Sapp
Quoted 33 lines · Ricardo de Araujo (Salveti)
Quoted 32 lines · Randolph Sapp
Quoted 24 lines · Ricardo Salveti via lists.openembedded.org
Quoted 9 lines · Khem Raj via lists.openembedded.org
Quoted 8 lines · Randolph Sapp
Quoted 5 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Quoted 8 lines
Quoted 3 lines

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

I think this is exposing the underlying issue in packages perhaps because they are not respecting CFLAGS and when it's gone from LDFLAGS the mapping related flags go missing. I have a local patch to fix memstat e.g in master-next, all failures perhaps need to be looked into one by one and passed cflags from OE env which might have been missed thus far.

Just noticed a similar failure with lxc (meta-virtualization), but while the prefix-map options are provided via CFLAGS, it only works when provided via LDFLAGS, probably due LTO (enabled by default by lxc).

Quoted 4 lines

Patch available at
https://lists.yoctoproject.org/g/meta-virtualization/message/9553

I guess recipes forcing LTO will probably have similar issues.

Ran a test locally. Looking at the do_compile log there are actually several compilation steps that do not have the macro prefixes passed as cc arguments.

Quoted 8 lines
Quoted 3 lines

That LTO bug has actually been addressed, it's just as Khem Raj pointed out, sometimes CFLAGS is ignored and passing the prefixes with LDFLAGS was masking that problem.

I did look into one of the binaries and it was built with the right cflags options, but didn't investigate too much.

I can also confirm that it works fine after disabling LTO, that is why I think it could be related to it.

Oh boy. Looking into this more today, spending plenty of time trying to make sure it wasn't meson or some other argument combination causing issues, I stumbled upon this thread [1]. They report the same issue with debug-prefixes, but they also cite potentially package dependent behavior with LTO. There are quite a few open threads about LTO behavior in general that are concerning.

Quoted 6 lines
Quoted 3 lines

Maybe this should be reverted and the alternative solution for cgo binary reproducibility should be taken instead [2]? I don't really have the resources to start on compiler development at the moment.

Right, it might be indeed better to revert for now as we will probably face similar LTO issues with multiple recipes and layers, since the bug is still valid upstream.

I did a world build with this patch + LTO enabled with our distro,  there are many failures, and it should be related to

this change.  I did not check them all one by one, just check some of them.  Eg:  cups(oe-core)

Maybe we can only append DEBUG_PREFIX_MAP in TARGET_LDFLAGS when LTO enabled. I will send a patch for this later.

So we don't need to revert this change,  and but we also need to take the following commit:

https://lists.openembedded.org/g/openembedded-core/message/230059

Regards

Changqing

Quoted 4 lines

Thanks,

Ricardo

#14 lore raw
Quoted 74 lines · Changqing Li via lists.openembedded.org
Quoted 52 lines · Ricardo de Araujo (Salveti)

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

Quoted 48 lines · Randolph Sapp
Quoted 33 lines · Ricardo de Araujo (Salveti)
Quoted 32 lines · Randolph Sapp
Quoted 24 lines · Ricardo Salveti via lists.openembedded.org
Quoted 9 lines · Khem Raj via lists.openembedded.org
Quoted 8 lines · Randolph Sapp
Quoted 5 lines · Khem Raj via lists.openembedded.org

I am seeing some additional diagnostics see

https://autobuilder.yoctoproject.org/valkyrie/api/v2/logs/5119153/raw_inline

I wonder if this is related.

Man, the mail clients are having fun with this thread.

Quoted 8 lines
Quoted 3 lines

Yeah, I'm able to reproduce that issue locally now. Not sure why it's this package subset that's acting up, but it is directly related to the LDFLAGS change unfortunately.

I think this is exposing the underlying issue in packages perhaps because they are not respecting CFLAGS and when it's gone from LDFLAGS the mapping related flags go missing. I have a local patch to fix memstat e.g in master-next, all failures perhaps need to be looked into one by one and passed cflags from OE env which might have been missed thus far.

Just noticed a similar failure with lxc (meta-virtualization), but while the prefix-map options are provided via CFLAGS, it only works when provided via LDFLAGS, probably due LTO (enabled by default by lxc).

Quoted 4 lines

Patch available at
https://lists.yoctoproject.org/g/meta-virtualization/message/9553

I guess recipes forcing LTO will probably have similar issues.

Ran a test locally. Looking at the do_compile log there are actually several compilation steps that do not have the macro prefixes passed as cc arguments.

Quoted 8 lines
Quoted 3 lines

That LTO bug has actually been addressed, it's just as Khem Raj pointed out, sometimes CFLAGS is ignored and passing the prefixes with LDFLAGS was masking that problem.

I did look into one of the binaries and it was built with the right cflags options, but didn't investigate too much.

I can also confirm that it works fine after disabling LTO, that is why I think it could be related to it.

Oh boy. Looking into this more today, spending plenty of time trying to make sure it wasn't meson or some other argument combination causing issues, I stumbled upon this thread [1]. They report the same issue with debug-prefixes, but they also cite potentially package dependent behavior with LTO. There are quite a few open threads about LTO behavior in general that are concerning.

Quoted 6 lines
Quoted 3 lines

Maybe this should be reverted and the alternative solution for cgo binary reproducibility should be taken instead [2]? I don't really have the resources to start on compiler development at the moment.

Right, it might be indeed better to revert for now as we will probably face similar LTO issues with multiple recipes and layers, since the bug is still valid upstream.

I did a world build with this patch + LTO enabled with our distro, there are many failures, and it should be related to

this change.  I did not check them all one by one, just check some of them.  Eg:  cups(oe-core)

Maybe we can only append DEBUG_PREFIX_MAP in TARGET_LDFLAGS when LTO enabled. I will send a patch for this later.

So we don't need to revert this change,  and but we also need to take the following commit:

Quoted 6 lines

https://lists.openembedded.org/g/openembedded-core/message/230059

Regards

Changqing

Thanks Changqing, the LTO + DEBUG_PREFIX_MAP is fine if we know ahead of time or are setting that flag ourselves. Unfortunately some things like lxc force LTO with meson project flags. Tricks like that makes things a little messy.

- Randolph