[BUG] Clang needs gold workaround reapplied for upcoming NDK release

[rprichard]

LLVM has a workaround for a bug in ld.gold (https://sourceware.org/bugzilla/show_bug.cgi?id=16794). Unfortunately, the SVN revision used for the upcoming NDK release (365631) fell during a brief window where the gold workaround had been reverted.

• LLVM review: https://reviews.llvm.org/D64327
• remove hack: https://reviews.llvm.org/rL365618
• restore hack: https://reviews.llvm.org/rL365685
• remove hack again: https://reviews.llvm.org/rL366686
• restore hack again: https://reviews.llvm.org/rL366708 (This link has some discussion on it.)

$ /x/ndk/out/android-ndk-r21-canary/toolchains/llvm/prebuilt/linux-x86_64/bin/clang --version
Android (5821526 based on r365631) clang version 9.0.6 (https://android.googlesource.com/toolchain/llvm-project 85305eaf1e90ff529d304abac8a979e1d967f0a2) (based on LLVM 9.0.6svn)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /x/ndk/out/android-ndk-r21-canary/toolchains/llvm/prebuilt/linux-x86_64/bin

I uploaded an x86-32 Linux test case to the gold bugtracker. Here's a test case more closely resembling what I reduced today:

test.cpp:

#include <stdio.h>

void call_printf(double x) {
  printf("%f\n", x);
}

void test1_func() {
  double tmp = 10;
  call_printf(tmp);
}

void test2_func() {
  double tmp = 10;
  call_printf(tmp);
}

void test3_func() {
  double tmp = 20;
  call_printf(tmp);
}

void test4_func() {
  double tmp = 20;
  call_printf(tmp);
}

int main(int, char**) {
  printf("10: "); test1_func();
  printf("10: "); test2_func();
  printf("20: "); test3_func();
  printf("20: "); test4_func();
  return 0;
}

$ /x/ndk/out/android-ndk-r21-canary/toolchains/llvm/prebuilt/linux-x86_64/bin/i686-linux-android16-clang++ \
    test.cpp -static-libstdc++ -o test -fuse-ld=gold
$ export ANDROID_SERIAL=emulator-5554
$ adb push test /data/local/tmp/test && adb shell /data/local/tmp/test
test: 1 file pushed. 2.3 MB/s (6620 bytes in 0.003s)
10: 10.000000
10: 20.000000
20: 0.000000
20: 0.000000
$ /x/ndk/out/android-ndk-r21-canary/toolchains/llvm/prebuilt/linux-x86_64/bin/i686-linux-android16-clang++ \
    test.cpp -static-libstdc++ -o test -fuse-ld=bfd
$ adb push test /data/local/tmp/test && adb shell /data/local/tmp/test
test: 1 file pushed. 3.2 MB/s (7092 bytes in 0.002s)
10: 10.000000
10: 10.000000
20: 20.000000
20: 20.000000

With the workaround, Clang outputs relocations to local symbols within the mergeable section:

$ readelf -r test.o | grep GOTOFF
0000001d  00000209 R_386_GOTOFF           00000000   .L.str
0000006a  00000509 R_386_GOTOFF           00000000   .LCPI1_0
000000aa  00000609 R_386_GOTOFF           00000008   .LCPI2_0
000000ea  00000709 R_386_GOTOFF           00000010   .LCPI3_0
0000012a  00000809 R_386_GOTOFF           00000018   .LCPI4_0
0000016f  00000309 R_386_GOTOFF           00000004   .L.str.1
000001a4  00000309 R_386_GOTOFF           00000004   .L.str.1
000001c5  00000409 R_386_GOTOFF           00000009   .L.str.2
000001e6  00000409 R_386_GOTOFF           00000009   .L.str.2

Without the workaround, it references the sections directly (and uses addends to identify pieces within the mergeable section):

$ readelf -r test.o | grep GOTOFF
0000001d  00000409 R_386_GOTOFF           00000000   .rodata.str1.1
0000006a  00000309 R_386_GOTOFF           00000000   .rodata.cst8
000000aa  00000309 R_386_GOTOFF           00000000   .rodata.cst8
000000ea  00000309 R_386_GOTOFF           00000000   .rodata.cst8
0000012a  00000309 R_386_GOTOFF           00000000   .rodata.cst8
0000016f  00000409 R_386_GOTOFF           00000000   .rodata.str1.1
000001a4  00000409 R_386_GOTOFF           00000000   .rodata.str1.1
000001c5  00000409 R_386_GOTOFF           00000000   .rodata.str1.1
000001e6  00000409 R_386_GOTOFF           00000000   .rodata.str1.1

[rprichard]

@stephenhines

[stephenhines]

https://android-review.googlesource.com/c/toolchain/llvm-project/+/1118093 will be the cherry-pick to our current toolchain sources.