Do programming language compilers first translate to assembly or directly to machine code?
I'm primarily interested in popular and widely used compilers, such as gcc. But if things are done differently with different compilers, I'd like to know that, too.
Taking gcc as an example, does it compile a short program written in C directly to machine code, or does it first translate it to human-readable assembly, and only then uses an (in-built?) assembler to translate the assembly program into binary, machine code -- a series of instructions to the CPU?
Is using assembly code to create a binary executable a significantly expensive operation? Or is it a relatively simple and quick thing to do?
(Let's assume we're dealing with only the x86 family of processors, and all programs are written for Linux.)
gcc actually produces assembler and assembles it using the as assembler. Not all compilers do this - the MS compilers produce object code directly, though you can make them generate assembler output. Translating assembler to object code is a pretty simple process, at least compared with compilation.
Some compilers produce other high-level language code as their output - for example, cfront, the first C++ compiler produced C as its output which was then compiled by a C compiler.
Note that neither direct compilation or assembly actually produce an executable. That is done by the linker, which takes the various object code files produced by compilation/assembly, resolves all the names they contain and produces the final executable binary.
Almost all compilers, including gcc, produce assembly code because it's easier---both to produce and to debug the compiler. The major exceptions are usually just-in-time compilers or interactive compilers, whose authors don't want the performance overhead or the hassle of forking a whole process to run the assembler. Some interesting examples include
Standard ML of New Jersey, which runs interactively and compiles every expression on the fly.
The tinycc compiler, which is designed to be fast enough to compile, load, and run a C script in well under 100 milliseconds, and therefore doesn't want the overhead of calling the assembler and linker.
What these cases have in common is a desire for "instantaneous" response. Assemblers and linkers are plenty fast, but not quite good enough for interactive response. Yet.
There are also a large family of languages, such as Smalltalk, Java, and Lua, which compile to bytecode, not assembly code, but whose implementations may later translate that bytecode directly to machine code without benefit of an assembler.
(Footnote: in the early 1990s, Mary Fernandez and I wrote the New Jersey Machine Code Toolkit, for which the code is online, which generates C libraries that compiler writers can use to bypass the standard assembler and linker. Mary used it to roughly double the speed of her optimizing linker when generating a.out. If you don't write to disk, speedups are even greater...)
GCC compiles to assembler. Some other compilers don't. For example, LLVM-GCC compiles to LLVM-assembly or LLVM-bytecode, which is then compiled to machine code. Almost all compilers have some sort of internal representation, LLVM-GCC use LLVM, and, IIRC, GCC uses something called GIMPLE.
According to chapter 2 of Introduction to Reverse Engineering Software (by Mike Perry and Nasko Oskov), both gcc and cl.exe (the back end compiler for MSVC++) have the -S switch you can use to output the assembly that each compiler produces.
You can also run gcc in verbose mode (gcc -v) to get a list of commands that it executes to see what it's doing behind the scenes.