|Version 7 (modified by 10 years ago) (diff),|
Native Code Generator
For other information related to this page, see:
- the Old GHC Commentary: Native Code Generator page (comments regarding Maximal Munch and register allocation optimisations are mostly still valid)
- BackEndNotes for optimisation ideas regarding the current NCG
Overview: Files, Parts
After GHC has produced Cmm (use -ddump-cmm or -ddump-opt-cmm to view), the Native Code Generator (NCG) transforms Cmm into architecture-specific assembly code. The NCG is located in compiler/nativeGen and is separated into eight modules:
top-level module for the NCG, imported by compiler/main/CodeOutput.lhs; also defines the Monad for optimising generic Cmm code,
generates architecture-specific instructions (a Haskell-representation of assembler) from Cmm code
contains data definitions and some functions (comparison, size, simple conversions) for machine instructions, mostly carried out through the
Instrdata type, defined here
defines the the main monad in the NCG: the Native code Machine instruction Monad,
NatM, and related functions. Note: the NCG switches between two monads at times, especially in
UniqSMMonad used throughout the compiler.
handles generation of position independent code and issues related to dynamic linking in the NCG; related to many other modules outside the NCG that handle symbol import, export and references, including
codeGenand the RTS, and the Mangler
Pretty prints machine instructions (
Instr) to assembler code (currently readable by GNU's
as), with some small modifications, especially for comparing and adding floating point numbers on x86 architectures
defines the main register information function,
regUsage, which takes a set of real and virtual registers and returns the actual registers used by a particular
Instr; register allocation is in AT&T syntax order (source, destination), in an internal function,
usage; defines the
one of the most complicated modules in the NCG,
RegisterAllocmanages the allocation of registers for each basic block of Haskell-abstracted assembler code: management involves liveness analysis, allocation or deletion of temporary registers, spilling temporary values to the spill stack (memory) and many optimisations. Note: much of this detail will be described later; basic block is defined below.
and one header file:
defines macros used to separate architecture-specific code in the Haskell NCG files; since GHC currently only generates machine code for the architecture on which it was compiled (GHC is not currently a cross-compiler), the Haskell NCG files become considerably smaller after preprocessing; ideally all architecture-specific code would reside in separate files and GHC would have them available to support cross-compiler capabilities.
The NCG has machine-independent and machine-dependent parts.
The machine-independent parts relate to generic operations, especially optimisations, on Cmm code. The main machine-independent parts begin with Cmm blocks. A Cmm block is roughly parallel to a Cmm function or procedure in the same way as a compiler may generate a C function into an assembler symbol used as a label, composed of smaller basic blocks (
BasicBlock) separated by branches (jumps)--every basic block ends in a branch instruction. Cmm blocks are held as lists of
Cmm statements (
[CmmStmt], defined in compiler/cmm/Cmm.hs, or the
CmmStmts, defined in compiler/cmm/CmmUtils.hs). A machine-specific (assembler) instruction is represented as a
Instr. The machine-independent NCG parts:
- optimise each Cmm block by reordering its basic blocks from the original order (the
Instrorder from the
Cmm) to minimise the number of branches between basic blocks, in other words, by maximising fallthrough of execution from one basic block to the next.
- lazily convert each Cmm block to abstract machine instructions (
Instr) operating on an infinite number of registers--since the NCG Haskell files only contain instructions for the host computer on which GHC was compiled, these
Instrare machine-specific; and,
- lazily allocate real registers for each basic block, based on the number of available registers on the target (currently, only the host) machine; for example, 32 integer and 32 floating-point registers on the PowerPC architecture. The NCG does not currently have support for SIMD registers such as the vector registers for Altivec or any variation of SSE.
Note: if a basic block simultaneously requires more registers than are available on the target machine and the temporary variable needs to be used (would sill be live) after the current instruction, it will be moved (spilled) into memory.
The machine-dependent parts:
- define the abstract (Haskell) assembler
Instrfor the target (host) machine and convert every Cmm block into it;
- define, manage and allocate the real registers available on the target system;
- pretty-print the Haskell-assembler to GNU AS (GAS) assembler code