Compiler output files

For a GCC or Clang command, there is typically one primary output file, specified by -o or the default (a.out or a.exe). There can also be temporary files and auxiliary files.

Primary output file

We can specify the primary output file with the -o option. When unspecified, a default output file name is inferred from the input files and the final phase. When the final phase is linking, the default output file name is a.out or a.exe.

gcc -S d/a.c      
gcc -c d/a.c      
gcc d/a.c         
gcc d/a.c -o e/a

For -S and -c, specifying -o in the presence of more than one input files leads to an error.

1
2
3

% gcc -c d/a.c d/b.c -o e/x
gcc: fatal error: cannot specify ‘-o’ with ‘-c’, ‘-S’ or ‘-E’ with multiple files
compilation terminated.

Temporary files

For compilation and linking in one command, the linker output file is the primary output file while the relocatable object files are temporary files.

For GCC, when generating a relocatable object file, it needs to generate a temporary assembly file, then feeds it to GNU assembler.

We can set the temporary directory with one of the environment variables TMPDIR, TMP, and TEMP. When none is specified, compilers have a fallback, /tmp on *NIX systems.

Auxiliary files

Beside primary output files and temporary output files, we have a third output file type called auxiliary output files. Some options such as -ftest-coverage (see below) and -gsplit-dwarf cause the compiler to generate auxiliary output files.

For compilation without linking (-c, -S, etc), the auxiliary output file names are derived from the primary output file name.

For compilation and linking in one command, the primary output file name affects the auxiliary output file names.

gcc -c -g -gsplit-dwarf d/a.c d/b.c      
gcc -c -g -gsplit-dwarf d/a.c -o e/a.o   
gcc -g -gsplit-dwarf d/a.c d/b.c -o e/x  
gcc -g -gsplit-dwarf d/a.c d/b.c         


gcc -g -gsplit-dwarf d/a.c.c -o e/x.out  
gcc -g -gsplit-dwarf d/a.c.c -o e/a.out

GCC provides some options that are primarily of interest to GCC developers. These dump output files are treated the same way as auxiliary output files.

-dumpdir and -dumpbase are provided to control the auxiliary output file names. The official documentation may be difficult to follow. Let's see some examples.

gcc -g -gsplit-dwarf -dumpdir f d/a.c -c 
gcc -g -gsplit-dwarf -dumpdir f d/a.c    
gcc -g -gsplit-dwarf -dumpdir f/ d/a.c   
gcc -g -gsplit-dwarf -dumpbase f d/a.c   
gcc -g -gsplit-dwarf -dumpbase f/ d/a.c  

gcc -g -gsplit-dwarf d/a.c -o e/x -dumpdir f/g  
gcc -c -g -gsplit-dwarf d/a.c -o e/xa.o -dumpdir f/g

In the absence of -dumpdir, -dumpbase appends a dash, which makes it inconvenient.

If we specify both -dumpdir and -dumpbase, we can avoid the influence of the source filename when there is one input file.

1 2	gcc -g -gsplit-dwarf -dumpdir f/ -dumpbase x d/a.c gcc -g -gsplit-dwarf -dumpdir f/ -dumpbase x d/a.c d/b.c

I suggest that you only use -dumpdir. When only -dumpdir is used, the behavior is still easy to explain.

I added -dumpdir to Clang 17 (D149193). The behavior is simplified from GCC's. In the common cases the behavior are identical.

Some GCC's strange rules are not ported. For example, clang -c -g -gsplit-dwarf d/a.c -o e/xa.o -dumpdir f/g creates f/ga.dwo instead of f/gxa.dwo. When -dumpdir is specified, -o is completely for deriving the auxiliary output file names.

`-save-temps`

-save-temps and -save-temps={cwd,obj} generate intermediate files, which are treated as auxiliary output files in GCC.

1 2	gcc -save-temps d/a.c ls a.i a.s a.o

In the absence of -dumpdir/-dumpbase, -save-temps=cwd places intermediate files in the current directory while -save-temps=obj places intermediate files in the directory of the primary output file. -save-temps behaves like -save-temps=obj when -o is specified, and -save-temps=cwd otherwise.

When -dumpdir is specified, there is complex interaction between -dumpdir and -save-temps/-save-temps={cwd,obj}.

1
2
3

# The last of -dumpdir and -save-temps wins.
gcc -g -gsplit-dwarf d/a.c -o e/x -dumpdir f/ -save-temps=obj # e/a.{i,s,o,dwo}
gcc -g -gsplit-dwarf d/a.c -o e/x -save-temps=obj -dumpdir f/ # f/a.{i,s,o,dwo}

For Clang, I think we should abandon the idea to treat these intermdiate files (*.i, *.bc, *.s, etc) as auxiliary output files. Just make -dumpdir and -save-temps/-save-temps={cwd,obj} orthogonal.

1	clang -g -gsplit-dwarf d/a.c -o e/x -save-temps=obj -dumpdir f/

Other auxiliary files

Clang supports a few options (e.g. -ftime-trace) to generate other auxiliary output files. I plan to change their file names to be controlled by -dumpdir.

`-ftime-trace`

This option instructs Clang to print time summaries for stages of the compilation process. The output is a time trace file in JSON that can be loaded into https://ui.perfetto.dev or the legacy chrome://tracing UI.

There is a variant -ftime-trace= that specifies the output JSON file or a directory to place the trace file.

After my https://reviews.llvm.org/D150282, Clang driver derives the trace file from -o or -dumpdir using a -gsplit-dwarf style approach.

#!/bin/sh -e
PATH=/tmp/Rel/bin:$PATH
mkdir -p d e f
echo 'int main() {}' > d/a.c
echo > d/b.c

a() { rm $1 || exit 1; }

clang -ftime-trace d/a.c d/b.c         
a a-a.json; a a-b.json
clang -ftime-trace d/a.c d/b.c -o e/x  
a e/x-a.json; a e/x-b.json
clang -ftime-trace d/a.c d/b.c -o e/x -dumpdir f/
a f/a.json; a f/b.json
clang -ftime-trace=f d/a.c d/b.c -o e/x
a f/a-*.json; a f/b-*.json

clang -c -ftime-trace d/a.c d/b.c
a a.json b.json
clang -c -ftime-trace=f d/a.c d/b.c
a f/a.json f/b.json

clang -c -ftime-trace d/a.c -o e/xa.o
a e/xa.json
clang -c -ftime-trace d/a.c -o e/xa.o -dumpdir f/g
a f/ga.json

Feature request to support device compilation for offloading targets: https://github.com/llvm/llvm-project/issues/55455. Feature request to support clang++ -c -std=c++20 -ftime-trace a.cppm -o a.o: https://github.com/llvm/llvm-project/issues/60555.

GCC feature request to implement the option.

`-fproc-stat-report`

This option (D78903) dumps the primary output filename and memory usage to stdout.

There is a variant -fproc-stat-report=a.txt that dumps the information to a text file. The file will be opened in the append mode.

The output behavior is quite unusual.

gcov

GCC implements a code coverage feature called gcov. The driver provides -ftest-coverage for generating notes files (.gcno) and -fprofile-arcs for instrumenting code to produce data files (.gcda) during execution. For convenience, --coverage encompasses both options.

Previously, when Clang performs compilation and linking in one command, it places .gcno files in the current directory. The generated .gcda files during program execution go to the current directory as well. I have changed Clang to consult the primary output file for .gcno and .gcda files.

#!/bin/zsh -e
PATH=/tmp/Rel/bin:$PATH                
mkdir -p d e f
echo 'int main() {}' > d/a.c
echo > d/b.c

a() { rm [email protected] || exit 1; }

clang --coverage d/a.c d/b.c && ./a.out
a a-[ab].gc{no,da}
clang --coverage d/a.c d/b.c -o e/x && e/x
a e/x-[ab].gc{no,da}
clang --coverage d/a.c d/b.c -o e/x -dumpdir f/ && e/x
a f/[ab].gc{no,da}
clang --coverage -fprofile-dir=f d/a.c d/b.c -o e/x && e/x
a e/x-[ab].gcno f$PWD/e/x-[ab].gcda


clang -c --coverage d/a.c d/b.c && clang --coverage a.o b.o && ./a.out
a [ab].gc{no,da}
clang -c --coverage -fprofile-dir=f d/a.c d/b.c && clang --coverage a.o b.o && ./a.out
a [ab].gcno f$PWD/[ab].gcda


clang -c --coverage d/a.c -o e/xa.o && clang --coverage e/xa.o && ./a.out
a e/xa.gc{no,da}
clang -c --coverage d/a.c -o e/xa.o -dumpdir f/g && clang --coverage e/xa.o && ./a.out
a f/ga.gc{no,da}

GCC provides -fprofile-dir= to set the location for generating .gcda files. When the primary output path is absolute, GCC mangles the absolute output file path to create a file within the -fprofile-dir= directory.

I find this behavior to be peculiar. Clang, on the other hand, always creates hierarchies within the -fprofile-dir= directory.

% c() { for i in **/*.gcno **/*.gcda; do echo $i; 'rm' -f $i; done; }
% gcc --coverage -fprofile-dir=f d/a.c -o e/x && e/x && c
./f/#tmp#c#e#x-a.gcda
./e/x-a.gcno
% gcc --coverage -fprofile-dir=f d/a.c -o $PWD/e/x && e/x && c
e/x-a.gcno
f/tmp/c/e/x-a.gcda
% clang --coverage -fprofile-dir=f d/a.c -o e/x && e/x && c
e/x-a.gcno
f/tmp/c/e/x-a.gcda
% clang --coverage -fprofile-dir=f d/a.c -o $PWD/e/x && e/x && c
e/x-a.gcno
f/tmp/c/e/x-a.gcda

The compiled object files encode the paths of the emitted .gcda files, so the directory of the final executable doesn't matter. The absolute .gcda paths have a benefit. For a recursive Make style build systems, we will get different *.gcda files even if two directories contain files of the same name.

1
2
3

(cd d0; gcc -c --coverage a.c)
(cd d1; gcc -c --coverage a.c)
gcc --coverage d0/a.o d1/a.o -o e/x

However, the absolute .gcda paths are not friendly to build determinism. To achieve local determinism (make builds independent of the build directory's name), we need to set the PWD environment variable.

% gcc --coverage d/a.c -c -o e/x.o && strings e/x.o | grep gcda
     b0 /tmp/c/e/x.gcda
% PWD=/proc/self/cwd gcc --coverage d/a.c -c -o e/x.o && strings e/x.o | grep gcda
     b0 /proc/self/cwd/e/x.gcda

Unfortunately, PWD=/proc/self/cwd/d doesn't work (say, d is a directory within the current directory). If we have a recursive Make style build system and files of the same name within differect directories, it's very tricky to avoid duplicate .gcda paths while achieving local determinism. In my experiments, -fprofile-prefix-map= and -ffile-prefix-map= don't affect the emitted .gcda path.

Offloading

Clang supports offloading to various architectures using programming models like CUDA, HIP, and OpenMP. Using offloading options may generate multiple output files.

For example, the following command spawns a host compile action and two device compile actions. Many features with auxiliary output files don't support offloading. Uses are not rejected, but the behavior may be undesired (e.g. an auxiliary ouput file gets overwritten by multiple actions).

1	clang -c --target=x86_64-unknown-linux-gnu --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 -fgpu-rdc -nogpulib -nogpuinc -ftime-trace a.hip -ccc-print-phases

TODO