Completely stabilize NVCC_DEVICE_LINK. CONTRIB-150

Fix symbols that depend on mtime and cwd.

This ensures that nm contrib/libs/nvidia/nccl/src/collectives/device/devlink.o does not change after --rebuild.

ref:810eaf8dbc93b225494ac1d081f5021f065695d4

build/scripts/compile_cuda.py

@@ -22,7 +22,7 @@ def main():
         skip_nocxxinc = False
 
     spl = sys.argv.index('--cflags')
-    getpid1 = sys.argv[1]
+    mtime0 = sys.argv[1]
     command = sys.argv[2: spl]
     cflags = sys.argv[spl + 1:]
 
@@ -124,19 +124,23 @@ def good(arg):
     if compiler_args:
         command += ['--compiler-options', ','.join(compiler_args)]
 
-    # nvcc generates symbols like this:
-    # __cudaRegisterLinkedBinary_{len}_tmpxft_{pid}_00000000_6_{src}1_ii_{hash}
-    # They embed nvcc pid. This is the only unstable part. We stabilize it by
-    # preloading getpid() that always returns 1.
-    os.environ['LD_PRELOAD'] = getpid1
-    # nvcc deletes all files in TMPDIR that match tmpxft_{pid}*, even not
-    # created by it, so we provide a fresh TMPDIR.
-    os.environ['TMPDIR'] = tempfile.mkdtemp(prefix='compile_cuda.py.')
+    # --keep is necessary to prevent nvcc from embedding nvcc pid in generated
+    # symbols.  It makes nvcc use the original file name as the prefix in the
+    # generated files (otherwise it also prepends tmpxft_{pid}_00000000-5), and
+    # cicc derives the module name from its {input}.cpp1.ii file name.
+    command += ['--keep', '--keep-dir', tempfile.mkdtemp(prefix='compile_cuda.py.')]
+    # nvcc generates symbols like __fatbinwrap_{len}_{basename}_{hash} where
+    # {basename} is {input}.cpp1.ii with non-C chars translated to _, {len} is
+    # {basename} length, and {hash} is the hash of first exported symbol in
+    # {input}.cpp1.ii if there is one, otherwise it is based on its modification
+    # time and the current working directory.  To stabilize the names of these
+    # symbols we need to fix mtime and cwd.
+    os.environ['LD_PRELOAD'] = mtime0
 
     if dump_args:
         sys.stdout.write('\n'.join(command))
     else:
-        sys.exit(subprocess.Popen(command, stdout=sys.stderr, stderr=sys.stderr).wait())
+        sys.exit(subprocess.Popen(command, stdout=sys.stderr, stderr=sys.stderr, cwd='/').wait())
 
 
 if __name__ == '__main__':

build/ymake_conf.py

@@ -2414,7 +2414,7 @@ def print_macros(self):
         }
 
         if not self.cuda_use_clang.value:
-            cmd = '$YMAKE_PYTHON ${input:"build/scripts/compile_cuda.py"} ${tool:"tools/getpid1"} $NVCC $NVCC_FLAGS -c ${input:SRC} -o ${output;suf=${OBJ_SUF}${NVCC_OBJ_EXT}:SRC} %(skip_nocxxinc)s %(includes)s --cflags $C_FLAGS_PLATFORM $CFLAGS $SRCFLAGS $CUDA_HOST_COMPILER_ENV ${kv;hide:"p CC"} ${kv;hide:"pc light-green"}'
+            cmd = '$YMAKE_PYTHON ${input:"build/scripts/compile_cuda.py"} ${tool:"tools/mtime0"} $NVCC $NVCC_FLAGS -c ${input:SRC} -o ${output;suf=${OBJ_SUF}${NVCC_OBJ_EXT}:SRC} %(skip_nocxxinc)s %(includes)s --cflags $C_FLAGS_PLATFORM $CFLAGS $SRCFLAGS $CUDA_HOST_COMPILER_ENV ${kv;hide:"p CC"} ${kv;hide:"pc light-green"}'
         else:
             cmd = '$CXX_COMPILER --cuda-path=$CUDA_ROOT $C_FLAGS_PLATFORM -c ${input:SRC} -o ${output;suf=${OBJ_SUF}${NVCC_OBJ_EXT}:SRC} %(includes)s $CXXFLAGS $SRCFLAGS $TOOLCHAIN_ENV ${kv;hide:"p CU"} ${kv;hide:"pc green"}'
 

tools/mtime0/mtime0.c

@@ -0,0 +1,17 @@
+// Do not alias __xstat to __xstat64.
+#undef _FILE_OFFSET_BITS
+
+#include <dlfcn.h>
+#include <sys/stat.h>
+
+int __xstat(int ver, const char* path, struct stat* buf) {
+    static int (*xstat)(int, const char*, struct stat*) = 0;
+    if (!xstat) {
+        xstat = dlsym(RTLD_NEXT, "__xstat");
+    }
+    int rc = xstat(ver, path, buf);
+    if (rc == 0 && ver == _STAT_VER) {
+        buf->st_mtime = 0;
+    }
+    return rc;
+}

tools/mtime0/mtime0.exports

@@ -0,0 +1 @@
+C __xstat

tools/mtime0/ya.make

@@ -0,0 +1,19 @@
+DLL_TOOL(mtime0 PREFIX "")
+
+
+
+EXPORTS_SCRIPT(mtime0.exports)
+
+NO_RUNTIME()
+
+CFLAGS(
+    -fPIC
+)
+
+IF (OS_LINUX)
+    SRCS(
+        mtime0.c
+    )
+ENDIF()
+
+END()

tools/ya.make

@@ -1,7 +1,7 @@
 RECURSE(
     enum_parser
     fix_elf
-    getpid1
+    mtime0
     rescompiler
     rescompressor
     rorescompiler