Targetable Compiler for Arbitrary Architecture

Question

Are there compilers for high-level languages (such as C) which can be targeted to various architectures by specifying the hardware resources of the target?

I'm wondering if there are compilers which can target an architecture by specifying features such as the available registers and instruction set of the processor (i.e. how each instruction changes the state of the processor), and memory layout of the architecture.

I am aware that compilers such as gcc can target for multiple archtectures, but I'd like to know if there are compilers that can compile for new architectures by swapping out, say, configuration files, in order to target a new architecture.

Answer 1

I was listening to a podcast a while ago which talked about a Java virtual machine written (mostly) in Java, which targets multiple different platforms with some kind of platform definition file. I think I was listening to Java Posse and I think it may have been the Maxine Research Virtual Machine.

Answer 2

AFAIK, gcc does work just like that. check "Machine Descriptions" in the gcc docs. Of course, creating those .md files is a major programming effort.

Answer 3

The fundamental issue is that the features of various processors are not particularly uniform. Look at bundling on IA64, dual instruction sets on ARM/Thumb, register windows on SPARC, the count register on PPC. Even features which may be on mutliple architectures (like predication on ARM and Itanium) may be sufficiently different that you would not share an implementation. These are all very unique features you need to understand in order to write a good compiler, they are not simply different configurations the designers chose. Given the fact that they tend to be unique to a given processor, any toggle you would have in a compiler to enable them would effectively be specific code in the compiler to support that processor, hiding it behind some config option to allow you to enable it is just an obfuscation of that.

Having said that, there are a number of common features that a lot of processors have to deal with. So you will find that most modern compilers define things like instructions encodings, register files (and aliasing), and various other tidbits in config files which they then use to generate part of the compiler source which integrates with the hand written bits.

Looking at the LLVM X86 backenda significant portion (~30%) of it are the .td (target definition) files

Phoenix:X86 louis$ ls -al | grep cpp 
-rw-r--r--   1 louis  louis    27627 Nov  1 03:32 X86CodeEmitter.cpp
-rw-r--r--   1 louis  louis      661 Oct 29 18:56 X86ELFWriterInfo.cpp
-rw-r--r--   1 louis  louis    46558 Oct 29 18:56 X86FastISel.cpp
-rw-r--r--   1 louis  louis    43660 Oct 29 18:56 X86FloatingPoint.cpp
-rw-r--r--   1 louis  louis    59915 Oct 29 18:56 X86ISelDAGToDAG.cpp
-rw-r--r--   1 louis  louis   312709 Nov  1 03:32 X86ISelLowering.cpp
-rw-r--r--   1 louis  louis   109229 Oct 29 18:56 X86InstrInfo.cpp
-rw-r--r--   1 louis  louis    17396 Oct 29 18:56 X86JITInfo.cpp
-rw-r--r--   1 louis  louis    44111 Nov  1 03:32 X86RegisterInfo.cpp
-rw-r--r--   1 louis  louis    10369 Oct 29 18:56 X86Subtarget.cpp
-rw-r--r--   1 louis  louis    15265 Oct 29 18:56 X86TargetAsmInfo.cpp
-rw-r--r--   1 louis  louis     9365 Oct 29 18:56 X86TargetMachine.cpp
Phoenix:X86 louis$ ls -al | grep td  
-rw-r--r--   1 louis  louis     6870 Oct 29 18:56 X86.td
-rw-r--r--   1 louis  louis    13480 Oct 29 18:56 X86CallingConv.td
-rw-r--r--   1 louis  louis    77361 Nov  1 03:32 X86Instr64bit.td
-rw-r--r--   1 louis  louis    31517 Oct 29 18:56 X86InstrFPStack.td
-rw-r--r--   1 louis  louis    11690 Oct 29 18:56 X86InstrFormats.td
-rw-r--r--   1 louis  louis   156188 Oct 29 18:56 X86InstrInfo.td
-rw-r--r--   1 louis  louis    32971 Oct 29 18:56 X86InstrMMX.td
-rw-r--r--   1 louis  louis   183502 Oct 29 18:56 X86InstrSSE.td
-rw-r--r--   1 louis  louis    23653 Nov  1 03:32 X86RegisterInfo.td