# Changes between Version 34 and Version 35 of Commentary/CodingStyle

Ignore:
Timestamp:
Apr 24, 2011 10:05:04 PM (5 years ago)
Comment:

yet more restructuring

### Legend:

Unmodified
 v34 If the module can be compiled multiple ways (eg. GHCI vs. non-GHCI), make sure the imports are properly #ifdefed too, so as to avoid spurious unused import warnings. == Literate Haskell == In GHC we use a mixture of literate ({{{.lhs}}}) and non-literate ({{{.hs}}}) source. I (Simon M.) prefer to use non-literate style, because I think the {{{\begin{code}..\end{code}}}} clutter up the source too much, and I like to use Haddock-style comments (we haven't tried processing the whole of GHC with Haddock yet, though). == The C Preprocessor (CPP) == Currently we pass all the compiler sources through CPP. The -cpp flag is always added by the build system. However, whenever possible we try to avoid using CPP, as it can hide code from the compiler (which means changes that work on one platform can break the build on another) and code using CPP can be harder to understand. The following CPP symbols are used throughout the compiler: '''DEBUG''':: Used to enables extra checks and debugging output in the compiler. The ASSERT macro (see {{{HsVersions.h}}}) provides assertions which disappear when DEBUG is not defined. However, whenever possible, it is better to us debugIsOn from the Util module, which is defined to be True when DEBUG is defined and False otherwise.  The ideal way to provide debugging output is to use a Haskell expression "when debugIsOn $..." to arrange that the compiler will be silent when DEBUG is off (unless of course something goes wrong or the verbosity level is nonzero). When option -O is used, GHC will easily sweep away the unreachable code. As a last resort, debugging code can be placed inside #ifdef DEBUG, but since this strategy guarantees that only a fraction of the code is seen be the compiler on any one compilation, it is to be avoided when possible. Regarding performance, a good rule of thumb is that DEBUG shouldn't add more than about 10-20% to the compilation time. This is the case at the moment. If it gets too expensive, we won't use it. For more expensive runtime checks, consider adding a flag - see for example -dcore-lint. '''Trap, pitfall for using the ASSERT macro''': The ASSERT macro uses CPP, and if you are unwise enough to try to write assertions using primed variables ({{{ASSERT (not$ intersectsBlockEnv b b')}}}), one possible outcome is that CPP silently fails to expand the ASSERT, and you get this very baffling error message: {{{ Not in scope: data constructor ASSERT' }}} Now you can Google for this error message :-) '''GHCI''':: Enables GHCi support, including the byte code generator and interactive user interface. This isn't the default, because the compiler needs to be bootstrapped with itself in order for GHCi to work properly. The reason is that the byte-code compiler and linker are quite closely tied to the runtime system, so it is essential that GHCi is linked with the most up-to-date RTS. Another reason is that the representation of certain datatypes must be consistent between GHCi and its libraries, and if these were inconsistent then disaster could follow. == Compiler versions and language extensions == }}} === Literate Haskell === In GHC we use a mixture of literate ({{{.lhs}}}) and non-literate ({{{.hs}}}) source. I (Simon M.) prefer to use non-literate style, because I think the {{{\begin{code}..\end{code}}}} clutter up the source too much, and I like to use Haddock-style comments (we haven't tried processing the whole of GHC with Haddock yet, though). === The C Preprocessor (CPP) === Currently we pass all the compiler sources through CPP. The -cpp flag is always added by the build system. However, whenever possible we try to avoid using CPP, as it can hide code from the compiler (which means changes that work on one platform can break the build on another) and code using CPP can be harder to understand. The following CPP symbols are used throughout the compiler: '''DEBUG''':: Used to enables extra checks and debugging output in the compiler. The ASSERT macro (see {{{HsVersions.h}}}) provides assertions which disappear when DEBUG is not defined. However, whenever possible, it is better to us debugIsOn from the Util module, which is defined to be True when DEBUG is defined and False otherwise.  The ideal way to provide debugging output is to use a Haskell expression "when debugIsOn $..." to arrange that the compiler will be silent when DEBUG is off (unless of course something goes wrong or the verbosity level is nonzero). When option -O is used, GHC will easily sweep away the unreachable code. As a last resort, debugging code can be placed inside #ifdef DEBUG, but since this strategy guarantees that only a fraction of the code is seen be the compiler on any one compilation, it is to be avoided when possible. Regarding performance, a good rule of thumb is that DEBUG shouldn't add more than about 10-20% to the compilation time. This is the case at the moment. If it gets too expensive, we won't use it. For more expensive runtime checks, consider adding a flag - see for example -dcore-lint. '''Trap, pitfall for using the ASSERT macro''': The ASSERT macro uses CPP, and if you are unwise enough to try to write assertions using primed variables ({{{ASSERT (not$ intersectsBlockEnv b b')}}}), one possible outcome is that CPP silently fails to expand the ASSERT, and you get this very baffling error message: {{{ Not in scope: data constructor ASSERT' }}} Now you can Google for this error message :-) '''GHCI''':: Enables GHCi support, including the byte code generator and interactive user interface. This isn't the default, because the compiler needs to be bootstrapped with itself in order for GHCi to work properly. The reason is that the byte-code compiler and linker are quite closely tied to the runtime system, so it is essential that GHCi is linked with the most up-to-date RTS. Another reason is that the representation of certain datatypes must be consistent between GHCi and its libraries, and if these were inconsistent then disaster could follow.