[BUG] NDK r23 gives confusing error message for `-fno-integrated-as` #1569

bog-dan-ro · 2021-08-24T12:01:27Z

Description

Steps to reproduce:

create a simple .c file $ cat test.c :

int main() {return 0;}

Try to compile it:

$ /path/to/ndk/23.0.7599858/toolchains/llvm/prebuilt/linux-x86_64/bin/clang test.c --target=armv7a-linux-androideabi28 -fno-integrated-as -pipe -D_FILE_OFFSET_BITS=64 -c
/usr/bin/as: unrecognized option '-EL'
clang-12: error: assembler command failed with exit code 1 (use -v to see invocation)

Without -fno-integrated-as it compiles just fine.

I need to pass -fno-integrated-as to workaround pixman issues: https://gitlab.freedesktop.org/pixman/pixman/-/issues/45

The text was updated successfully, but these errors were encountered:

enh-google · 2021-08-24T16:40:37Z

yeah, -fno-integrated-as isn't supported in r23. (i don't whether we should configure things so that if you do use it [and don't manually specify yasm or whatever] we run a script or something that says "-fno-integrated-as is not supported" rather than just quietly running /usr/bin/as which is just misleading.)

we have some docs here about how to fix assembler: https://android.googlesource.com/platform/ndk/+/master/docs/ClangMigration.md#assembler-issues

the .func error in particular is just dead code for gas too, so that's easy. i didn't see the .if error when we fixed all of Android, so i don't have any specific advice to give there (but in general, and especially since the source file is already .S rather than .s, "just use the C preprocessor instead" was useful to me!).

does it work if you move all the .if stuff below the push? the only somewhat-relevant references to this error i found on the internet implied that there are issues before the first instruction is emitted, though they seemed to be .ifs that involved calculations relative to . that doesn't apply here. (that's also supposed to have been fixed in 2019, so we should have that fix already.)

enh-google · 2021-08-24T18:01:06Z

wait... it's just occurred to me that you're talking about r23, which still has gas, not r24, which is the first release that doesn't.

so for r23, you've probably just got a misconfigured build that's calling the wrong as. (but for r24 you'll need code that can actually build with LLVM.)

rprichard · 2021-08-24T20:35:58Z

r23 still has gas, but not with the filename that Clang expects.

r22 has:

toolchains/${target}/prebuilt/${host}/${triple}/bin/as
toolchains/${target}/prebuilt/${host}/bin/${triple}-as
toolchains/llvm/prebuilt/${host}/${triple}/bin/as
toolchains/llvm/prebuilt/${host}/bin/${triple}-as

r23 only has # 4, but I think we also need # 3 for -fno-integrated-as.

bog-dan-ro · 2021-08-25T07:22:10Z

Instead to force everyone to change their asm file to make them llvm compatible, isn't better to make llvm as compatible with gas?
This way you can nuke gas without no worries.

bog-dan-ro · 2021-08-25T07:26:09Z

Just for the record, compiling pixman with r21 & r22 with -fno-integrated-as spawns tones of Rd and Rm should be different in mul but at least it compiles ... :)

stephenhines · 2021-08-25T07:31:29Z

Instead to force everyone to change their asm file to make them llvm compatible, isn't better to make llvm as compatible with gas?
This way you can nuke gas without no worries.

There are inconsistencies and syntax relaxations for gas that ARM (among other vendors) have said that they won't support in any new compilers/assemblers. If the syntax being used is UAL (unified assembly language), then we can work with ARM to add the missing support. In that particular case, LLVM is correct in telling you that the instruction you're using isn't valid, and gas is merely transforming it into something equivalent. According to this link, "Rd cannot be the same as Rm."

bog-dan-ro · 2021-08-25T07:56:19Z

At first look it doesn't seem to be an ASM language problem, but llvm as is missing some features:

/path/to/android-sdk/ndk/21.1.6352462/toolchains/llvm/prebuilt/linux-x86_64/bin/clang -
Ipixman/libpixman-arm-simd.a.p -Ipixman -I../pixman -Xclang -fcolor-diagnostics -pipe -D_FILE_OFFSE
T_BITS=64 -Wall -Winvalid-pch -O2 -g -Wdeclaration-after-statement -fno-strict-aliasing -fvisibilit
y=hidden -Wundef -ftrapping-math -Wno-unused-local-typedefs -DHAVE_CONFIG_H --target=armv7a-linux-a
ndroideabi28 -fPIC -MD -MQ pixman/libpixman-arm-simd.a.p/pixman-arm-simd-asm.S.o -MF pixman/libpixm
an-arm-simd.a.p/pixman-arm-simd-asm.S.o.d -o pixman/libpixman-arm-simd.a.p/pixman-arm-simd-asm.S.o 
-c ../pixman/pixman-arm-simd-asm.S
<instantiation>:1:1: error: unknown directive
.func fname
^
../pixman/pixman-arm-simd-asm.h:599:5: note: while in macro instantiation
    pixman_asm_function fname
    ^
/tmp/pixman-arm-simd-asm-0ff417.s:913:1: note: while in macro instantiation
generate_composite_function pixman_composite_src_8888_8888_asm_armv6, 32, 0, 32, FLAG_DST_WRITEONLY
 | FLAG_COND_EXEC | FLAG_SPILL_LINE_VARS_WIDE | FLAG_PROCESS_PRESERVES_SCRATCH, 4, blit_init, nop_m
acro, nop_macro, blit_process_head, nop_macro, blit_inner_loop
^
../pixman/pixman-arm-simd-asm.h:614:6: error: expected absolute expression
 .if prefetch_distance == 0
     ^
/tmp/pixman-arm-simd-asm-0ff417.s:913:1: note: while in macro instantiation
generate_composite_function pixman_composite_src_8888_8888_asm_armv6, 32, 0, 32, FLAG_DST_WRITEONLY
 | FLAG_COND_EXEC | FLAG_SPILL_LINE_VARS_WIDE | FLAG_PROCESS_PRESERVES_SCRATCH, 4, blit_init, nop_m
acro, nop_macro, blit_process_head, nop_macro, blit_inner_loop
^
../pixman/pixman-arm-simd-asm.h:620:6: error: expected absolute expression
 .if src_bpp == 32
     ^

Attached the entire output in case you need it : out.txt

bog-dan-ro · 2021-08-25T08:04:32Z

According to this link, "Rd cannot be the same as Rm."

Just for the record, I'm seeing those messages when I'm compiling .c files with -fno-integrated-as not .s files.

rprichard · 2021-08-25T09:52:31Z

I think the problems with .if results from Clang's incomplete .altmacro support. pixman-arm-simd-asm.S uses .altmacro. Here's the non-altmacro behavior:

$ cat >test.S <<EOF
    .macro foo arg_
    .set arg, arg_
    .if arg != 42
    .error "arg must be 42"
    .endif
    .endm
    foo 42
EOF

$ gcc -c test.S
test.S: Assembler messages:
test.S:7: Error: arg must be 42

$ clang -c test.S
<instantiation>:2:9: error: expected absolute expression
    .if arg != 42
        ^
...

It can be fixed using \arg_:

$ cat >test.S <<EOF
    .macro foo arg_
    .set arg, \arg_
    .if arg != 42
    .error "arg must be 42"
    .endif
    .endm
    foo 42
EOF

$ gcc -c test.S
<success>

$ clang -c test.S
<success>

But .altmacro changes a bunch of stuff, and it can also be used instead of escaping the macro argument:

$ cat >test.S <<EOF
    .altmacro
    .macro foo arg_
    .set arg, arg_
    .if arg != 42
    .error "arg must be 42"
    .endif
    .endm
    foo 42
EOF

$ gcc -c test.S
<success>

$ clang -c test.S
<instantiation>:2:9: error: expected absolute expression
    .if arg != 42
        ^

There are a few open .altmacro bugs already, but I don't think this one is open yet:
https://bugs.llvm.org/buglist.cgi?quicksearch=altmacro

rprichard · 2021-08-25T23:05:05Z

FWIW: ART uses .altmacro but only with GNU as:
https://android.googlesource.com/platform/art/+/refs/tags/android-11.0.0_r40/runtime/arch/x86_64/asm_support_x86_64.S#30

ART uses .altmacro to work around this GCC-vs-Clang preprocessor difference. My guess, though, is that we could rewrite that code to work without .altmacro (and use the same code for GCC and Clang).

enh-google · 2021-08-25T23:28:43Z

FWIW: ART uses .altmacro but only with GNU as:
https://android.googlesource.com/platform/art/+/refs/tags/android-11.0.0_r40/runtime/arch/x86_64/asm_support_x86_64.S#30

ART uses .altmacro to work around this GCC-vs-Clang preprocessor difference. My guess, though, is that we could rewrite that code to work without .altmacro (and use the same code for GCC and Clang).

well, given that ART only builds with clang, we could just delete that code :-)

DanAlbert · 2021-09-01T00:27:33Z

Just putting back toolchains/llvm/prebuilt/${host}/${triple}/bin/as doesn't seem to be enough. I wonder if we need to add back all the no-op crt objects to get Clang to recognize the "GCC" directory...

DanAlbert · 2021-09-01T19:59:09Z

Yeah, I think that is it. The actual path that gets used for as in r22:

"/work/src/android-ndk-r22b/toolchains/llvm/prebuilt/linux-x86_64/bin/../lib/gcc/arm-linux-androideabi/4.9.x/../../../../arm-linux-androideabi/bin/as"

🤮

DanAlbert · 2021-09-01T21:27:13Z

https://android-review.googlesource.com/c/platform/ndk/+/1817218 fixes the problem. Sorry about that. In our effort to make sure we were -fno-integrated-as clean before asking users to do so we of course removed all our own uses of it, but neglected to add a test to cover that case. I've added an explicit test of -fno-integrated-as to go with the fix :)

I forked the missing altmacro support discussion into a new bug.

DanAlbert · 2021-09-01T23:03:59Z

The changelog update so we don't forget to cherry-pick it: https://android-review.googlesource.com/c/platform/ndk/+/1817316

In particular, link to our migration advice. This commit is based on 3e47e46, with the Changelog-r24.md part removed so it can be cherry-picked to aosp/ndk-release-r23. Bug: http://b/200995199 Bug: android/ndk#1569 Test: N/A Change-Id: I61c3107bd389c2da1a5641babdef6f56ca487b1f (cherry picked from commit 3236d6773fe6ed59f4b44c58c0a5a2fd0e28f86a)

The change can't be done in this branch since GAS is fully gone in r24, but the changelog is being updated here to reduce merge conflicts. Test: None Bug: android/ndk#1569 Change-Id: I7de8ffe8880ed05f89d46302f253b5d1d89677c6 (cherry picked from commit 175b1c2)

Grimler91 · 2021-10-23T13:48:27Z

The as symlinks that were added in https://android-review.googlesource.com/c/platform/ndk/+/1817218 seem to be broken in r23b:

$ ls -l toolchains/llvm/prebuilt/linux-x86_64/*android*/bin/
toolchains/llvm/prebuilt/linux-x86_64/aarch64-linux-android/bin/:
total 4
lrwxrwxrwx 1 root root 131 Oct 23 15:10 as -> /buildbot/src/android/ndk-release-r23/out/linux/android-ndk-r23b/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android-as

toolchains/llvm/prebuilt/linux-x86_64/arm-linux-androideabi/bin/:
total 4
lrwxrwxrwx 1 root root 131 Oct 23 15:10 as -> /buildbot/src/android/ndk-release-r23/out/linux/android-ndk-r23b/toolchains/llvm/prebuilt/linux-x86_64/bin/arm-linux-androideabi-as

toolchains/llvm/prebuilt/linux-x86_64/i686-linux-android/bin/:
total 4
lrwxrwxrwx 1 root root 128 Oct 23 15:10 as -> /buildbot/src/android/ndk-release-r23/out/linux/android-ndk-r23b/toolchains/llvm/prebuilt/linux-x86_64/bin/i686-linux-android-as

toolchains/llvm/prebuilt/linux-x86_64/x86_64-linux-android/bin/:
total 4
lrwxrwxrwx 1 root root 130 Oct 23 15:10 as -> /buildbot/src/android/ndk-release-r23/out/linux/android-ndk-r23b/toolchains/llvm/prebuilt/linux-x86_64/bin/x86_64-linux-android-as

Fixing so they instead point towards ../../bin/{triple}-as makes -fno-integrated-as work as intended.

rprichard · 2021-10-29T23:59:56Z

I uploaded a fix making the as symlink target relative: https://android-review.googlesource.com/c/platform/ndk/+/1875878

DanAlbert · 2021-11-01T18:51:52Z

And of course the test passes because it's being run on the same machine (－‸ლ)

rprichard · 2021-11-03T00:37:16Z

And of course the test passes because it's being run on the same machine (－‸ლ)

I think detecting this issue would have required building the NDK and the tests on different machines.

Alternately, maybe we could have a test that walks the NDK install directory and verifies that all symlinks are relative.

enh-google · 2021-11-03T01:14:37Z

gah, i make this mistake all the time when a configure script or whatever creates an absolute symlink. why do they even do that?! looking at you, pcre2... :-)

Bug: android/ndk#1569 Test: checkbuild.py Test: verify that all symlinks are relative Change-Id: I65073f8043029e39ee0667cf880375a3869ae1e0 (cherry picked from commit 15fd1e3)

Bug: android/ndk#1569 Test: checkbuild.py Test: verify that all symlinks are relative Test: Use -fno-integrated-as manually Change-Id: I977734b3681b341f86cf72a95a19a798682fb93c

orifmilod · 2022-03-29T16:23:41Z

For those who are facing the same issue as @Grimler91 - (seeing /buildbot/...). Please run the command below, it should remove the old broken sym-link files and create new ones.

rm ./toolchains/llvm/prebuilt/linux-x86_64/aarch64-linux-android/bin/as && \
ln -sv ./toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android-as ./toolchains/llvm/prebuilt/linux-x86_64/aarch64-linux-android/bin/as && \
rm ./toolchains/llvm/prebuilt/linux-x86_64/arm-linux-androideabi/bin/as && \
ln -sv ./toolchains/llvm/prebuilt/linux-x86_64/bin/arm-linux-androideabi-as ./toolchains/llvm/prebuilt/linux-x86_64/arm-linux-androideabi/bin/as && \
rm ./toolchains/llvm/prebuilt/linux-x86_64/i686-linux-android/bin/as && \
ln -sv ./toolchains/llvm/prebuilt/linux-x86_64/bin/i686-linux-android-as ./toolchains/llvm/prebuilt/linux-x86_64/i686-linux-android/bin/as && \
rm ./toolchains/llvm/prebuilt/linux-x86_64/x86_64-linux-android/bin/as && \
ln -sv ./toolchains/llvm/prebuilt/linux-x86_64/bin/x86_64-linux-android-as ./toolchains/llvm/prebuilt/linux-x86_64/x86_64-linux-android/bin/as

DanAlbert · 2022-04-22T20:29:08Z

Should be fixed in r23 build 8486889.

Bug: android/ndk#1569 Test: n/a Change-Id: I633f07d281707997749ebbd99885dbc5026c293a (cherry picked from commit 3aa8b12e5ec47d5958b62ae05522057d8ed3e680) Merged-In: I633f07d281707997749ebbd99885dbc5026c293a

The validation happens immediately after the NDK is built. When the Windows NDK is cross-compiled from Linux, the scan verifies that there are no symlinks. Bug: android/ndk#1569 Test: checkbuild.py Change-Id: Idea7ef7b1652e4bef16482f4886332335f8fb758 (cherry picked from commit 8076ea3e61531961ced9a43298be446f779c7340) Merged-In: Idea7ef7b1652e4bef16482f4886332335f8fb758

bog-dan-ro added the bug label Aug 24, 2021

enh-google changed the title ~~[BUG] NDK r23 can't compile a simple file using -fno-integrated-as~~ [BUG] NDK r23 gives confusing error message for -fno-integrated-as Aug 24, 2021

DanAlbert mentioned this issue Sep 1, 2021

incomplete altmacro support in clang #1576

Open

DanAlbert self-assigned this Sep 1, 2021

SSE4 mentioned this issue Nov 24, 2021

android-ndk: Various Fixes and Version Upgrades conan-io/conan-center-index#8168

Merged

4 tasks

DanAlbert mentioned this issue Feb 16, 2022

[BUG] Broken symlinks in android-ndk - r23b #1671

Closed

DanAlbert closed this as completed Apr 22, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[BUG] NDK r23 gives confusing error message for `-fno-integrated-as` #1569

[BUG] NDK r23 gives confusing error message for `-fno-integrated-as` #1569

bog-dan-ro commented Aug 24, 2021

enh-google commented Aug 24, 2021

enh-google commented Aug 24, 2021

rprichard commented Aug 24, 2021

bog-dan-ro commented Aug 25, 2021

bog-dan-ro commented Aug 25, 2021

stephenhines commented Aug 25, 2021

bog-dan-ro commented Aug 25, 2021 •

edited

Loading

bog-dan-ro commented Aug 25, 2021

rprichard commented Aug 25, 2021

rprichard commented Aug 25, 2021

enh-google commented Aug 25, 2021

DanAlbert commented Sep 1, 2021

DanAlbert commented Sep 1, 2021

DanAlbert commented Sep 1, 2021 •

edited

Loading

DanAlbert commented Sep 1, 2021

Grimler91 commented Oct 23, 2021

rprichard commented Oct 29, 2021

DanAlbert commented Nov 1, 2021

rprichard commented Nov 3, 2021

enh-google commented Nov 3, 2021

orifmilod commented Mar 29, 2022 •

edited

Loading

DanAlbert commented Apr 22, 2022

[BUG] NDK r23 gives confusing error message for -fno-integrated-as #1569

[BUG] NDK r23 gives confusing error message for -fno-integrated-as #1569

Comments

bog-dan-ro commented Aug 24, 2021

Description

enh-google commented Aug 24, 2021

enh-google commented Aug 24, 2021

rprichard commented Aug 24, 2021

bog-dan-ro commented Aug 25, 2021

bog-dan-ro commented Aug 25, 2021

stephenhines commented Aug 25, 2021

bog-dan-ro commented Aug 25, 2021 • edited Loading

bog-dan-ro commented Aug 25, 2021

rprichard commented Aug 25, 2021

rprichard commented Aug 25, 2021

enh-google commented Aug 25, 2021

DanAlbert commented Sep 1, 2021

DanAlbert commented Sep 1, 2021

DanAlbert commented Sep 1, 2021 • edited Loading

DanAlbert commented Sep 1, 2021

Grimler91 commented Oct 23, 2021

rprichard commented Oct 29, 2021

DanAlbert commented Nov 1, 2021

rprichard commented Nov 3, 2021

enh-google commented Nov 3, 2021

orifmilod commented Mar 29, 2022 • edited Loading

DanAlbert commented Apr 22, 2022

[BUG] NDK r23 gives confusing error message for `-fno-integrated-as` #1569

[BUG] NDK r23 gives confusing error message for `-fno-integrated-as` #1569

bog-dan-ro commented Aug 25, 2021 •

edited

Loading

DanAlbert commented Sep 1, 2021 •

edited

Loading

orifmilod commented Mar 29, 2022 •

edited

Loading