[Activity] Week 17

[Prathamesh Kulkarni]

== This Week ==
* TCWG-619:

* abe sysroot/libc issue
While building chromium with development build of toolchain on
arm-linux-gnueabihf, I get the following error:
undefined reference to clock_gettime()
Builds fine with release toolchain build.
Symlinking libc in toolchain install directory to
sysroots/arm-linux-gnueabihf appears to work (which is what --tarbin
did).

* PR65778 - v8 fails to build with LTO
- Resolved Invalid

* PR65837 - target specific builtin not available
- created a reduced test case
- There's a weird issue: looks like for lto1, translation unit command
line options,
precede the default ones rather than overriding them.
Assume test.c contains a neon intrinsic:
arm-linux-gnueabihf-gcc -flto -c test.c -mfpu=neon
arm-linux-gnueabihf-gcc -flto test.o
results in target specific builtin not available,
because both -mfpu=neon and -mfpu=vfpv3-d16 is passed
and -mfpu=vfpv3-d16 overrides earlier -mfpu=neon
Passing -mfpu=neon at link time correctly overrides -mfpu=vfpv3-d16.

* PR6778 - ICE in varpool_node::get_constructor.
- Looks like PR65776 appears due to same issue as well.
- It appears error_mark_node gets streamed that causes ICE.
- Triage went wrong, thought it was an error in typeck2.c:check_narrowing().

* PR49551
- The C front-end c/c-decl.c:merge_decls still incorrectly marks the
variable as 'common' for the test-case:
int x = 5;
int x;
However it's not reproducible anymore on trunk (maybe latent?)
- I have an untested patch that resolves the issue in merge_decls().

* Issues
- abe built native gcc on x8_64-unknown-linux-gnu fails to build code
when passed -m32 option (bug 1508).

* Misc
- exams on 20th and 24th april.

* Next week
- Continue working on PR65837, PR65778
- Test patch for PR49551

[Yvan Roux]

== Issue ==
* none

== Progress ==

o Upstream GCC (8/10)
* [TCWG-785] ARM backend insn cleanup
- Validate and re-worked the patch
- About to be submitted upstream
* [TCWG-762] - GCC Maintenance
- PR64208 (iWMMXT LRA bug):
Found hardware to validate the fix (Thanks Riku :)
Validation ongoing
* Lot of time wasted due to lab compromisation

o Misc (2/10)
* Various meetings
* GCC git repository branches cleanup

== Plan ==
- Continue cleaning GCC git repo to prepare our migration
- Continue upstream pending work

[Prathamesh Kulkarni]

On 27 April 2015 at 03:17, Prathamesh Kulkarni

Post by Prathamesh Kulkarni
== This Week ==
* abe sysroot/libc issue
While building chromium with development build of toolchain on
undefined reference to clock_gettime()
Builds fine with release toolchain build.
Symlinking libc in toolchain install directory to
sysroots/arm-linux-gnueabihf appears to work (which is what --tarbin
did).
* PR65778 - v8 fails to build with LTO
- Resolved Invalid
* PR65837 - target specific builtin not available
- created a reduced test case
- There's a weird issue: looks like for lto1, translation unit command
line options,
precede the default ones rather than overriding them.
arm-linux-gnueabihf-gcc -flto -c test.c -mfpu=neon
arm-linux-gnueabihf-gcc -flto test.o
results in target specific builtin not available,
because both -mfpu=neon and -mfpu=vfpv3-d16 is passed
and -mfpu=vfpv3-d16 overrides earlier -mfpu=neon
Passing -mfpu=neon at link time correctly overrides -mfpu=vfpv3-d16.
* PR6778 - ICE in varpool_node::get_constructor.

s/PR6778/PR65858

Post by Prathamesh Kulkarni
- Looks like PR65776 appears due to same issue as well.
- It appears error_mark_node gets streamed that causes ICE.
- Triage went wrong, thought it was an error in typeck2.c:check_narrowing().
* PR49551
- The C front-end c/c-decl.c:merge_decls still incorrectly marks the
int x = 5;
int x;
However it's not reproducible anymore on trunk (maybe latent?)
- I have an untested patch that resolves the issue in merge_decls().
* Issues
- abe built native gcc on x8_64-unknown-linux-gnu fails to build code
when passed -m32 option (bug 1508).
* Misc
- exams on 20th and 24th april.
* Next week
- Continue working on PR65837, PR65778
- Test patch for PR49551