Patch for context/coroutine assembler for Linux On Power (BE & LE)

[Tony Reix <tony.reix@…>]

Hi,

Tests of context and coroutine libraries produce cores on Linux on Power (LoP), Little Endian and Big Endian.

LoP LE: Ubuntu 14.04 :

"uname -m" --> ppc64le LITTLE_ENDIAN defined.

LoP BE: Fedora 19 :

"uname -m" --> ppc64 BIG_ENDIAN defined.

Linux On Power is only 64bits. No 32bits version.

The source codes to be changed are:

• libs/context/build/Jamfile.v2
• libs/context/src/asm : PPC assembler files jump...S and make...S

jump_ppc64_sysv_elf_gas.S make_ppc64_sysv_elf_gas.S

I am preparing a patch for the PPC assembler files, either as 1 file for each of jump... and make... files (using #ifdef for producing LE or BE code), or 2 files for each (one for LE and one for BE). I don't know yet what is the most appropriate solution: one common file for LE&BE with #ifdef (easier for maintenance), or one file for LE and one other for BE. To be discussed.

Looking at Jamfile.v2 file, I see several cases for PPC :

• power - 32 - elf - gcc
• power - 32 - elf - qcc
• power - 64 - elf - gcc
• power - 64 - elf - qcc

(I'm not an expert of Jam)

I guess that "power - 32 - *" cases should be let as they are, for compatibility with older or different environments.

About "power - 64", the files require changes for running fine on LoP BE & LE. I've done these changes, and tests work fine. However, I don't know if there are some older 64bits PPC environment to keep compatibility with. (I'm also working on a patch for AIX/PPC 32-64, with xcoff format, and with AIX assembler. I'll manage it after LoP patch.)

So, I need some help for finalizing my patch:

• How to name the files ? which code is now different from the existing ones:

jump_ppc64_sysv_elf_gas.S make_ppc64_sysv_elf_gas.S

• Is it possible to use #fidef in these .S files for having one file for LE and BE ?

I thought to use: LITTLE_ENDIAN and BIG_ENDIAN macros. I still have to check that Jam handle them.

• How should I change Jam.v2 ?

I'll provide the modified ppc64 files I've built, in this "ticket".

Regards,

Tony

[olli]

Hello Toni, I suggest to combine little and big endian into one file for make_fcontext and jump_fcontext. Could you adapt your patches according to the code of the develop branch? In this new version the registers will be stored at the current end of the stack.

[Tony Reix <tony.reix@…>]

Replying to olli:

Hello Toni, I suggest to combine little and big endian into one file for make_fcontext and jump_fcontext. Could you adapt your patches according to the code of the develop branch?

Hello Olli, Hummm You seem to say that changes have occured about this subject between 1.55 and now. So, yes, I have to look at develop branch.

Moreover, instead of using the LITTLE_ENDIAN and BIG_ENDIAN, I should use something like:

#if !defined POWERPC64

_CALL_ELF == 2

In this new version the registers will be stored at the current end of the stack.

OK. I have to manage these changes.

Thanks

[Tony Reix <tony.reix@…>]

Looking at build/Jamfile.v2 file, I see changes. Probably also within .S files. I'm now rebuilding my Boost environment on LoP BE & LE with this develop version. So that I can compare and test, and build a new patch, I guess. Tony

[anonymous]

please note that you have to remove the untested.cpp file from build/Jamfile.v2 for the PPC target

[Tony Reix <tony.reix@…>]

I have removed the untested.cpp file for POWER64 from context/build/Jamfile.vé . However, I still have the following error while trying to run the tests of context :

[root@tbx6 build]# cd - /home/reixt/BOOST/modular-boost/status [root@tbx6 status]# ../b2 Performing configuration checks

• symlinks supported : yes (cached)
• hardlinks supported : yes (cached)
• 32-bit : no (cached)
• 64-bit : yes (cached)
• arm : no (cached)
• mips1 : no (cached)
• power : yes (cached)
• lockfree boost::atomic_flag : yes (cached)
• has_icu builds : yes (cached)

...patience... ...patience... ...found 4421 targets... ...updating 13 targets... Jamfile</home/reixt/BOOST/modular-boost/libs/context/build>.gas ../bin.v2/libs/context/build/gcc-4.8.2/debug/address-model-64/architecture-power/link-static/asm/make_ppc64_sysv_elf_gas.o ../libs/context/src/asm/make_ppc64_sysv_elf_gas.S: Messages de l'assembleur: ../libs/context/src/asm/make_ppc64_sysv_elf_gas.S:98: Erreur: pseudo-op inconnu: « .localentry » ../libs/context/src/asm/make_ppc64_sysv_elf_gas.S:102: Erreur: symbole « make_fcontext » est déjà défini

as -o "../bin.v2/libs/context/build/gcc-4.8.2/debug/address-model-64/architecture-power/link-static/asm/make_ppc64_sysv_elf_gas.o" "../libs/context/src/asm/make_ppc64_sysv_elf_gas.S"

...failed Jamfile</home/reixt/BOOST/modular-boost/libs/context/build>.gas ../bin.v2/libs/context/build/gcc-4.8.2/debug/address-model-64/architecture-power/link-static/asm/make_ppc64_sysv_elf_gas.o... Jamfile</home/reixt/BOOST/modular-boost/libs/context/build>.gas ../bin.v2/libs/context/build/gcc-4.8.2/debug/address-model-64/architecture-power/link-static/asm/jump_ppc64_sysv_elf_gas.o ../libs/context/src/asm/jump_ppc64_sysv_elf_gas.S: Messages de l'assembleur: ../libs/context/src/asm/jump_ppc64_sysv_elf_gas.S:98: Erreur: pseudo-op inconnu: « .localentry » ../libs/context/src/asm/jump_ppc64_sysv_elf_gas.S:102: Erreur: symbole « jump_fcontext » est déjà défini

as -o "../bin.v2/libs/context/build/gcc-4.8.2/debug/address-model-64/architecture-power/link-static/asm/jump_ppc64_sysv_elf_gas.o" "../libs/context/src/asm/jump_ppc64_sysv_elf_gas.S"

...failed Jamfile</home/reixt/BOOST/modular-boost/libs/context/build>.gas ../bin.v2/libs/context/build/gcc-4.8.2/debug/address-model-64/architecture-power/link-static/asm/jump_ppc64_sysv_elf_gas.o... ...skipped <p../bin.v2/libs/context/build/gcc-4.8.2/debug/link-static>libboost_context.a(clean) for lack of <p../bin.v2/libs/context/build/gcc-4.8.2/debug/address-model-64/architecture-power/link-static>asm/make_ppc64_sysv_elf_gas.o... ...skipped <p../bin.v2/libs/context/build/gcc-4.8.2/debug/link-static>libboost_context.a for lack of <p../bin.v2/libs/context/build/gcc-4.8.2/debug/address-model-64/architecture-power/link-static>asm/make_ppc64_sysv_elf_gas.o... ...skipped <p../bin.v2/libs/context/test/test_context.test/gcc-4.8.2/debug/link-static>test_context for lack of <p../bin.v2/libs/context/build/gcc-4.8.2/debug/link-static>libboost_context.a... ...skipped <p../bin.v2/libs/context/test/test_context.test/gcc-4.8.2/debug/link-static>test_context.run for lack of <p../bin.v2/libs/context/test/test_context.test/gcc-4.8.2/debug/link-static>test_context... ...skipped <p../bin.v2/libs/coroutine/test/test_asymmetric_coroutine.test/gcc-4.8.2/debug/link-static>test_asymmetric_coroutine for lack of <p../bin.v2/libs/context/build/gcc-4.8.2/debug/link-static>libboost_context.a... ...skipped <p../bin.v2/libs/coroutine/test/test_asymmetric_coroutine.test/gcc-4.8.2/debug/link-static>test_asymmetric_coroutine.run for lack of <p../bin.v2/libs/coroutine/test/test_asymmetric_coroutine.test/gcc-4.8.2/debug/link-static>test_asymmetric_coroutine... ...skipped <p../bin.v2/libs/coroutine/test/test_symmetric_coroutine.test/gcc-4.8.2/debug/link-static>test_symmetric_coroutine for lack of <p../bin.v2/libs/context/build/gcc-4.8.2/debug/link-static>libboost_context.a... ...skipped <p../bin.v2/libs/coroutine/test/test_symmetric_coroutine.test/gcc-4.8.2/debug/link-static>test_symmetric_coroutine.run for lack of <p../bin.v2/libs/coroutine/test/test_symmetric_coroutine.test/gcc-4.8.2/debug/link-static>test_symmetric_coroutine... ...failed updating 2 targets... ...skipped 11 targets...

[anonymous]

try to add following to build/Jamfile.v2:

alias asm_context_sources

: asm/make_ppc64_sysv_elf_gas.S

asm/jump_ppc64_sysv_elf_gas.S

: <address-model>64

<architecture>power <binary-format>elf <toolset>gcc

;

[Tony Reix <tony.reix@…>]

Hi anonymous,

There are several errors in the current Boost code for PPC64. (tests below were done in LoP PPC64/BE environment)

1) src/asm/make_ppc64_sysv_elf_gas.S This file contains use of macros, like: #if _CALL_ELF == 2 However, in build/Jamfile.v2, the gas action is defined as: actions gas {

as -o "$(<)" "$(>)"

} and that does not expend the Macros. There is a need of calling cpp.

2) Adding the lines you gave me in previous comment still leads to an error (both on BE and LE): [root@tbx6 build]# pwd /home/reixt/BOOST/modular-boost/libs/context/build [root@tbx6 build]# bjam warning: mismatched versions of Boost.Build engine and core warning: Boost.Build engine (bjam) is 2011.12.01 warning: Boost.Build core (at /home/reixt/BOOST/modular-boost/tools/build/src) is 2013.05-svn Performing configuration checks

• symlinks supported : yes
• hardlinks supported : yes

/home/reixt/BOOST/modular-boost/tools/build/src/tools/link.jam:314: in link-recursively ERROR: rule "READLINK" unknown in module "link". ........

Maybe I have some mis-match in my environment. Or I'm not building in the correct place ?

3) src/asm/make_ppc64_sysv_elf_gas.S contains PPC 32bits assembler instructions, and it mixes PPC 32bits conventions with 64bits. So it cannot work.

I need some clarifications.

Thanks

Tony

[olli]

1) you could try the modified Jamfile -> the assembler files are passed to the preprocessor before applied to gas.

2) I guess it's a problem of the boost.build system

3) could you identify which instructions must be corrected and generate a patch for it?

[olli]

preprocessor before gas

[Tony Reix <tony.reix@…>]

Hi,

I'm nuts about complex assembler stuff. However I'm helped by an expert. He has found the following:

1) First, it seems that the API of jump_fcontext() has been changed (from 1.55 to today).

intptr_t jump_fcontext( fcontext_t * ofc, fcontext_t const* nfc, intptr_t vp, bool preserve_fpu = true);

=>

intptr_t jump_fcontext( fcontext_t * ofc, fcontext_t nfc, intptr_t vp, bool preserve_fpu = true);

2) It seems that you have made changes for Power that are not valid for Linux on Power (LoP), both BE and LE.

So, we propose the following changes (see files that I gonna attach to the ticket) that are valid to Linux on PowerPC Big Endian or Little Endian (there is only 64bits systems, no 32bits systems):

• make_fcontext :
  • use of 64bits opcodes (std instead of stw)
  • about ELFv1 case: we grab the address and TOC of the context function.

• jump_fcontext :
  • about ELFv1 case: we save/restore the TOC.
  • we don't save/restore R13
  • about ELFv2: the address of the context function is passed through R12

4) We use the Macros:

_CALL_ELF == 2 _CALL_LINUX

3) These jump/make files must be preprocessed by cpp, in order to process the Macros. I'm not sure that it works now with Jamfile.v2. My colleague tested his code in a much simpler environment. I gonna check asap.

We have tested the files I gonna attach both on LoP BE and LE. Only 64bits.

jump_ppc64_sysv_elf_gas.S make_ppc64_sysv_elf_gas.S

In order to prevent confusion, I suggest that either you rename these 2 files another way (jump/make_ppc64_linux_elf_gas.S) or you mix all versions together (which may be difficult to maintain). I would suggest first solution. However, I do not yet master what the Jamfile.v2 of context lib does.

Regards,

Tony

[Tony Reix <tony.reix@…>]

Assembler code for Linux on Power BE/LE PPC64

[Tony Reix <tony.reix@…>]

Assembler code for Linux on Power BE/LE PPC64

[Tony Reix <tony.reix@…>]

I've attached the 2 new files, named:

jump_ppc64_linux_elf_gas.S make_ppc64_linux_elf_gas.S

(but use the name that seems appropriate for you).

[olli]

Hi Tony, thx for the paches. I've changed the Jamfile - the asm files will be passed to the preprocessor before GNU as. Could you pull the last changes from git repo (branch develop) and execute the unit-tests (sub-directory test; call b2/bjam at command line)? best, Oliver

[Tony Reix <tony.reix@…>]

Hi Oliver,

I've pulled the last changes in context directory from "develop" to my Ubuntu PC, and I've copied the files on my tests machines. However, I still have an error when running bjam in "modular-boost/libs/context/build" (on my test machines and on my PC), though it worked fine with v1.55 :

tony@tony2:~/BOOST/modular-boost/libs/context/build$ bjam

• symlinks supported : yes
• hardlinks supported : yes

/home/tony/BOOST/modular-boost/tools/build/src/tools/link.jam:314: in link-recursively ERROR: rule "READLINK" unknown in module "link".

Any idea about why this does not work ?

Anyway: If I run ./b2 at top and then run ../b2 in status directory, that works fine, and I've been able to check that all the context and coroutine tests are "passed", both on LoP LE (Ubuntu 14.04) and LOP BE (Fedora 19).

Jamfile</home/tony/BOOST/modular-boost/libs/context/build>.gas ../bin.v2/libs/context/build/gcc-4.8/debug/address-model-64/architecture-power/link-static/asm/make_ppc64_sysv_elf_gas.o Jamfile</home/tony/BOOST/modular-boost/libs/context/build>.gas ../bin.v2/libs/context/build/gcc-4.8/debug/address-model-64/architecture-power/link-static/asm/jump_ppc64_sysv_elf_gas.o

So, that seems perfect for Linux On Power in LE and BE modes ! :)

However, please check that it does not break other PPC64 environment.

Regards,

Tony

[olli]

Replying to Tony Reix <tony.reix@…>:

Hi Oliver,

I've pulled the last changes in context directory from "develop" to my Ubuntu PC, and I've copied the files on my tests machines. However, I still have an error when running bjam in "modular-boost/libs/context/build" (on my test machines and on my PC), though it worked fine with v1.55 :

tony@tony2:~/BOOST/modular-boost/libs/context/build$ bjam

• symlinks supported : yes
• hardlinks supported : yes

/home/tony/BOOST/modular-boost/tools/build/src/tools/link.jam:314: in link-recursively ERROR: rule "READLINK" unknown in module "link".

Any idea about why this does not work ?

no idea - boost/tools/build/src/tools/link.jam is part of the boost.build system and I'm not an expert of boost.build. maybe you should add an bug for boost.build?

Anyway: If I run ./b2 at top and then run ../b2 in status directory, that works fine, and I've been able to check that all the context and coroutine tests are "passed", both on LoP LE (Ubuntu 14.04) and LOP BE (Fedora 19).

Jamfile</home/tony/BOOST/modular-boost/libs/context/build>.gas ../bin.v2/libs/context/build/gcc-4.8/debug/address-model-64/architecture-power/link-static/asm/make_ppc64_sysv_elf_gas.o Jamfile</home/tony/BOOST/modular-boost/libs/context/build>.gas ../bin.v2/libs/context/build/gcc-4.8/debug/address-model-64/architecture-power/link-static/asm/jump_ppc64_sysv_elf_gas.o

So, that seems perfect for Linux On Power in LE and BE modes ! :)

that's really nice - I'll merge the changes to branch master

However, please check that it does not break other PPC64 environment.

unfortunately I've no PPC64 system and boost does not have PPC64 regression tests for this platform.

best, Oliver

[Steven Watanabe]

Replying to Tony Reix <tony.reix@…>:

tony@tony2:~/BOOST/modular-boost/libs/context/build$ bjam

• symlinks supported : yes
• hardlinks supported : yes

/home/tony/BOOST/modular-boost/tools/build/src/tools/link.jam:314: in link-recursively ERROR: rule "READLINK" unknown in module "link".

Any idea about why this does not work ?

Your b2 is out of date. Please rebuild it. READLINK is a new builtin rule which is needed to support modular boost.

[Tony Reix <tony.reix@…>]

Replying to steven_watanabe:

Replying to Tony Reix <tony.reix@…>:

tony@tony2:~/BOOST/modular-boost/libs/context/build$ bjam

• symlinks supported : yes
• hardlinks supported : yes

/home/tony/BOOST/modular-boost/tools/build/src/tools/link.jam:314: in link-recursively ERROR: rule "READLINK" unknown in module "link".

Any idea about why this does not work ?

Your b2 is out of date. Please rebuild it. READLINK is a new builtin rule which is needed to support modular boost.

Hummm AFAIK, rebuilding b2 is done by: 1) remove it, 2) launch again ./bootstrap.sh . I've done this and it produced exactly the same file b2. Is there a special process I'm not aware of for rebuilding b2 ? Thx

[ad@…]

I have a regression from 1.55 final to 1.56 develop and this change

gcc-4.8.2 => target-os=windows => ok clang-darwin-4.2.1 => target-os=darwin => failure

"cpp" is not known (I may symlink cpp to clang.exe) but after "as" doesn't know the switch -am

[ad@…]

I have a regression from 1.55 final to 1.56 develop and this change

gcc-4.8.2 => target-os=windows => ok
gcc-4.8.2 => target-os=linux => ok
clang-darwin-4.2.1 => target-os=darwin => failure

"cpp" is not known (I may symlink cpp to clang.exe) but after "as" doesn't know the switch -am

[olli]

is already fixed

[class101 <ad@…>]

Thank you olli , indeed fixed, posted my results [1]

[1] ​https://svn.boost.org/trac/boost/ticket/9774#comment:4