multiple errors in gcc/config/rs6000/rs6000.c

[llvmbot]

Bugzilla Link	3099
Resolution	WONTFIX
Resolved on	Jan 07, 2011 12:53
Version	2.4
OS	Linux
Blocks	#1288
Attachments	build script and full log
Reporter	LLVM Bugzilla Contributor
CC	@asl,@nlewycky

Extended Description

I'm trying to cross-compile llvm-gcc for powerpc target on amd64/linux

The compilation stopped when compiling gcc/config/rs6000/rs6000.c

================== log fragment start ==========================
c++ -c -O2 -pipe -DIN_GCC -DCROSS_DIRECTORY_STRUCTURE -W -Wall -Wwrite-strings -pedantic -Wno-long-long -Wno-variadic-macros -Wmissing-format-attribute -DHAVE_CONFIG_H -Wno-unused -DTARGET_NAME="ppc-linux-gnu" -DNDEBUG -I. -I. -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/. -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../include -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../libcpp/include -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../libdecnumber -I../libdecnumber -I/usr/include -DENABLE_LLVM -DNDEBUG -D_GNU_SOURCE -D__STDC_LIMIT_MACROS -I. -I. -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/. -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../include -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../libcpp/include -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../libdecnumber -I../libdecnumber -I/usr/include /tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/llvm-debug.cpp -o llvm-debug.o
gcc -c -O2 -pipe -DIN_GCC -DCROSS_DIRECTORY_STRUCTURE -W -Wall -Wwrite-strings -Wstrict-prototypes -Wmissing-prototypes -pedantic -Wno-long-long -Wno-variadic-macros -Wno-overlength-strings -Wold-style-definition -Wmissing-format-attribute -DHAVE_CONFIG_H -I. -I. -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/. -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../include -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../libcpp/include -I/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/../libdecnumber -I../libdecnumber -I/usr/include -DENABLE_LLVM -DNDEBUG -D_GNU_SOURCE -D__STDC_LIMIT_MACROS
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c -o rs6000.o
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c: In function 'rs6000_override_options':
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:1565: error: 'MASK_MACHO_DYNAMIC_NO_PIC' undeclared (first use in this function)
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:1565: error: (Each undeclared identifier is reported only once
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:1565: error: for each function it appears in.)
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c: In function 'rs6000_emit_prologue':
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:16873: warning: implicit declaration of function 'VECTOR_SAVE_INLINE'
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:16874: error: 'lr_already_set_up_for_pic' undeclared (first use in this function)
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c: In function 'rs6000_output_mi_thunk':
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:18333: warning: implicit declaration of function 'add_compiler_branch_island'
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:18333: warning: assignment makes pointer from integer without a cast
/tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:20026:8: warning: extra tokens at end of #endif directive
In file included from /tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/llvm-gcc4.2-2.4.source/gcc/config/rs6000/rs6000.c:22322:
./gt-rs6000.h: At top level:
./gt-rs6000.h:138: error: 'branch_island_list' undeclared here (not in a function)
make[2]: *** [rs6000.o] Error 1
make[2]: Leaving directory /tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/build/gcc' make[1]: *** [all-gcc] Error 2 make[1]: Leaving directory /tmp/portage/cross-ppc-linux-gnu/gcc-4.2.4-r2/work/build'
================== log fragment end ==========================

IMHO

1. MASK_MACHO_DYNAMIC_NO_PIC is not defined anywhere, though used twice. I looked for it in the complete patch/build tree and in /usr/include

2. "#endif llvm" is present only in this file and breaks C rules and apple/llvm coding style

3. Other errors are present

i enclose ebuild script and full log file

my aim is not to find lots of errors in llvm source code but to compile PowerPC-capable llvm-gcc decompiler on my computer

[asl]

Looks like another bunch of APPLE LOCAL stuff...

my aim is not to find lots of errors in llvm source code but to compile
PowerPC-capable llvm-gcc decompiler on my computer
???? llvm-gcc is compiler, not decompiler.

PS: I think it will require some decent amount of work to bring ppc stuff on non-darwin platforms.

[llvmbot]

PowerPC-capable llvm-gcc decompiler on my computer
???? llvm-gcc is compiler, not decompiler.

if a tool that converts llvm assembler to native assembler is a compiler, then
llvm-gcc is a decompiler. If not, then llc + llvm-gcc is a decompiler, hence llvm-gcc is one-half-of-decompiler

concerning rs6000.c and add_compiler_branch_island problem:

looks like rs6000.c has misplaced #ifdef TARGET_MACH0 directives: reference to to add_compiler_branch_island problem is OUTSIDE of this ifdef, while definition of this function is INSIDE. And original rs6000.c inside gcc-4.2 is free from this defect

PS: I think it will require some decent amount of work to bring ppc stuff on
non-darwin platforms.

has anyone cross-compiled llvm-gcc for powerpc on darwin/x86? There seems to be an old mac with macos available for me, but i don't want to use antique hardware/software unless absolutely necessary

[asl]

PowerPC-capable llvm-gcc decompiler on my computer
???? llvm-gcc is compiler, not decompiler.

if a tool that converts llvm assembler to native assembler is a compiler, then
llvm-gcc is a decompiler. If not, then llc + llvm-gcc is a decompiler, hence
llvm-gcc is one-half-of-decompiler
How so? I don't understand and even don't see the logic here. Have you ever considered looking into documentation?

Look, at least here for description of llvm/llvm-gcc:
http://llvm.org/pubs/2007-03-12-BossaLLVMIntro.pdf

llvm-gcc is a C/C++/ObjC/ObjC++/Ada/Fortran (currently) compiler using LLVM framework components for optimization and codegeneration. Why is it a DE-compiler? There is no way to feed native binaries to llvm-gcc's input: it "understands" only c/c++/... sources.

llc itself is just a backend components of LLVM linked as a standalone tool.

has anyone cross-compiled llvm-gcc for powerpc on darwin/x86? There seems to be
an old mac with macos available for me, but i don't want to use antique
hardware/software unless absolutely necessary
Cross-compilation from darwin/x86 to darwin/powerpc should be definitely ok, especially if you take release tarball (in fact all 'apple style' llvm-gcc builds are made in this manner).

[llvmbot]

Anton's right, most of this is Darwinisms that are escaping. I've checked in (to TOT) some changes that should get you further; please bear in mind I'm unable to test anything in your environment.

MASK_MACHO_DYNAMIC_NO_PIC won't exist on non-darwin targets; it would be defined in options.h, which is constructed as part of the build process and is in the object directory.

VECTOR_SAVE_INLINE describes whether saving the Altivec registers in the prologue should use inline code or a call to an external function. I think it should simply be 1 for non-Darwin systems, can someone with Linux confirm?

I have no idea how longcalls are supposed to work on Linux (that's the "add_compiler_branch_island" area). What I did here is unlikely to work correctly (on Linux) IMO but should allow you to compile.

As far as I can see all references to branch_island_list are inside #if TARGET_MACHO. Please make sure you're including some appropriate definition of TARGET_OBJECT_FORMAT; it's defined in several of the header files. (I'm looking at TOT, it's also possible this was fixed since 4.2.)

I think I've addressed all the problems shown except VECTOR_SAVE_INLINE.

There may well be runtime issues after you get it to build, I don't think anybody has done this before.

[llvmbot]

Cross-compilation from darwin/x86 to darwin/powerpc should be definitely ok,
especially if you take release tarball (in fact all 'apple style' llvm-gcc
release tarball of what? of OS or compiler? If OS then which OS? There are IMHO multiple OS which can be called darwin.

[asl]

Cross-compilation from darwin/x86 to darwin/powerpc should be definitely ok,
especially if you take release tarball (in fact all 'apple style' llvm-gcc
release tarball of what? of OS or compiler? If OS then which OS? There are IMHO
multiple OS which can be called darwin.
Of LLVM surely. Please visit http://llvm.org/releases/download.html#2.4

[llvmbot]

Anton's right, most of this is Darwinisms that are escaping. I've checked in
(to TOT) some changes that should get you further; please bear in mind I'm
TOT should be kinda svn branch? Will you please give direct link. I've googled
for it with no result.

I have no idea how longcalls are supposed to work on Linux (that's the
"add_compiler_branch_island" area). What I did here is unlikely to work
correctly (on Linux) IMO but should allow you to compile.
don't know either, but gcc as powerpc cross compiler compiles and works under
my linux.

and llvm-gcc does not compile. One error i noticed and reported in comment
no.2: definition of add_compiler_branch_island is inside ifdef, while reference is outside

As far as I can see all references to branch_island_list are inside #if
TARGET_MACHO. Please make sure you're including some appropriate definition of
TARGET_OBJECT_FORMAT; it's defined in several of the header files. (I'm
looking at TOT, it's also possible this was fixed since 4.2.)

I think I've addressed all the problems shown except VECTOR_SAVE_INLINE.
thank you. Now i know how to define some of the undefined. But IMHO rs6000.c inside llvm-gcc is buggy and someone should correct it

[asl]

Anton's right, most of this is Darwinisms that are escaping. I've checked in
(to TOT) some changes that should get you further; please bear in mind I'm
TOT should be kinda svn branch? Will you please give direct link. I've googled
for it with no result.
TOT = Top Of Tree, just the most recent svn revision

[llvmbot]

Anton's right, most of this is Darwinisms that are escaping. I've checked in
(to TOT) some changes that should get you further; please bear in mind I'm
TOT should be kinda svn branch? Will you please give direct link. I've googled
for it with no result.

As Anton says, Top of Tree. Directions are here (please browse the website a bit, there's lots of useful information):
http://llvm.org/docs/GettingStarted.html#checkout

I have no idea how longcalls are supposed to work on Linux (that's the
"add_compiler_branch_island" area). What I did here is unlikely to work
correctly (on Linux) IMO but should allow you to compile.
don't know either, but gcc as powerpc cross compiler compiles and works under
my linux.

and llvm-gcc does not compile. One error i noticed and reported in comment
no.2: definition of add_compiler_branch_island is inside ifdef, while reference
is outside

Yes, I believe I fixed this one, at least well enough that it will compile, as I said.

As far as I can see all references to branch_island_list are inside #if
TARGET_MACHO. Please make sure you're including some appropriate definition of
TARGET_OBJECT_FORMAT; it's defined in several of the header files. (I'm
looking at TOT, it's also possible this was fixed since 4.2.)

I think I've addressed all the problems shown except VECTOR_SAVE_INLINE.
thank you. Now i know how to define some of the undefined. But IMHO rs6000.c
inside llvm-gcc is buggy and someone should correct it

That's what I'm trying to do. However, I don't have the environment you're trying to set up myself, so I can't test anything, so you (or some other volunteer) are going to have to help. Are things any better with the changes I made? (Note you'll have to do something about VECTOR_SAVE_INLINE, see previous comment.)

[llvmbot]

Anton's right, most of this is Darwinisms that are escaping. I've checked in
(to TOT) some changes that should get you further; please bear in mind I'm
TOT should be kinda svn branch? Will you please give direct link. I've googled
for it with no result.

As Anton says, Top of Tree. Directions are here (please browse the website a
bit, there's lots of useful information):
http://llvm.org/docs/GettingStarted.html#checkout
the GettingStarted.html does not say how to get llvm-gcc via svn, but "Building llvm-gcc from Source" manual suggests "svn co http://llvm.org/svn/llvm-project/llvm-gcc-4.2/trunk ..."

but i don't understand: do i have to leave stable version of llvm with svn llvm-gcc, or take svn versions of both?

i downloaded llvm via svn several minutes ago, and llvm-gcc checkout is in process

i will try to compile today, if download is successful

[llvmbot]

Yes, rs6000.c compiled successfully

build process was unsuccessful (a header file is missing)

it means i must take svn versions of both llvm and llvm-gcc

[asl]

build process was unsuccessful (a header file is missing)
Which header file? Please note, that (in general) you will need a set of header files for target platform for successful build of cross-compiler (mainly for support libraries, etc)

[llvmbot]

ebuild scripts and full log file of llvm-gcc build

[llvmbot]

build process was unsuccessful (a header file is missing)
Which header file? Please note, that (in general) you will need a set of header
files for target platform for successful build of cross-compiler (mainly for
support libraries, etc)

a header file absent in llvm version 2.4 but present in svn version i took today

after upgrading llvm the header problem went away

however, compilation stopped at rs6000.c: undefined VECTOR_SAVE_INLINE and branch_island_list

for details see attachment to comment no. 13

[llvmbot]

preprocessed rs6000.c
looking at preprocessed file i see that branch_island_list only appears in
auto-generated header gt-rs6000.h

misplaced ifdef?