[PATCH] Driver modifications for cross compilation


I noticed that the Clang driver requires the user to set “-ccc-host-triple” for cross-compilation (if not using the -arch behaviour on Darwin). I thought this was a little strange as the option is not documented in the --help, is nonstandard and nonobvious.

I delved a little deeper, and found a bunch of FIXMEs in the driver wanting to tablegen lots of if/else statements selecting specific architecture options (AddARMTargetArgs being the worst culprit). Adding to that, some of these monolithic functions were duplicated with static linkage in multiple files.

I’ve factored most of these if/else blocks into one tablegen file, which when #included forms an “architecture definition database” from which one can query default CPUs for given architecture/OS, which architecture a CPU belongs to, and any other target-specific feature of the CPU defined in the .td file.

The tablegen file is a list of known architecture variants and CPUs for different platforms; we are happy to maintain the ARM part of this.

The upshot of this is that several monolithic functions have disappeared:

  • Tool{s,Chain}.cpp:GetARMTargetCPU (replaced with Arch->DefaultCpu)

  • Tool{s,Chain}.cpp:getLLVMArchSuffixForARM (replaced with Arch->Properties[“LLVMArchSuffix”])

And “Tools.cpp:Add{ARM,Sparc,MIPS,X86}TargetArgs” have been simplified to contain fewer if/elses.

The above was all just general tidyup; after this refactor, a 3 line change in Driver.cpp now causes autodetection of the host triple from any -mcpu= or -march= options passed. For example:

clang -mcpu=cortex-a8 test.c -o test.o -c

will set the HostTriple to “armv7–”, which obviously then enables AddARMTargetArgs to be called instead of AddX86TargetArgs, and correct ARM code is produced.

I feel that this is a much more sane behaviour than mandating the use of -ccc-host-triple (although the behaviour of that option has not been altered and overrides any auto-detection); if a user specifies a CPU which is obviously part of an architecture, or a -march=something-that-isn’t-the-host, the HostTriple should be detected automatically.

As a result of this, taking the line of thought that -ccc-host-triple should be optional and should be able to be fully functionally replaced by -mcpu=/-march= and friends, I realised there was no way to dictate the host OS without use of -ccc-host-triple. So I added a -mos= option, which sets the OSAndEnvironment part of the llvm::Triple. This way, almost all functionality offered by overriding the triple (still wanted for power users) is available via more recognisable command line options.

The patch I have attached is the entire patch + testcases, so is rather large. If accepted, I intend to send to the commit list in several chunks:

  1. Tablegen file, new files ArchDefs.h and ArchDefs.cpp (+ new tablegen backend to llvm-commits)
  • This patch should not alter the behaviour of Clang at all, hence no tests being added.
  1. First usecase of the ArchDefs class: the ARM target. Removes GetARMTargetCPU and getLLVMArchSuffixForARM, and changes the default command line behaviour to infer target triple from -mcpu/-march if present.
  • As there is now a use case for the ArchDefs class, testcases will be added to this patch.
  1. Refactoring intel / mips target specifics to use ArchDefs. This is the most likely patch to cause problems with other uses that I haven’t forseen.

  2. Adding -mos= and making it have an effect on the host triple.

Patch 2. also contains several bugfixes to the ARM driver, namely:

  • ARM target driver would not detect -mabi=eabi/gnueabi correctly.

  • ARM target driver when setting -mabi= would not update the Triple and so the float-abi check later would be incorrect.

  • ARM target driver did not have knowledge of several ARM cores: Cortex-A{8,9,15}, Cortex-R{4,5,7}, Cortex-M{0,3,4}, ARM11MPCore, ARM1176.

Review of both the patch and the intended functional change would be excellent.



clang-all.patch (37.8 KB)

tablegen-clang-gen-arch-defs.patch (7.12 KB)