Differential D63508
make -frewrite-includes also rewrite conditions in #if/#elif Authored by llunak on Jun 18 2019, 12:00 PM.
Details Qt5 has a wrapper macro that makes __has_include simpler to use: #ifdef __has_include # define QT_HAS_INCLUDE(x) __has_include(x) #else # define QT_HAS_INCLUDE(x) 0 #endif #if QT_HAS_INCLUDE(<chrono>) # include <chrono> #endif The code handling this so far ignores macros entirely. This patch makes -frewrite-includes rewrite all #if/#elif conditions, which is more reliable and simpler than trying to handle __has_include specifically.
Diff Detail
Event TimelineComment Actions Luboš, can you please clarify what is wrong with the current Clang behavior? I'm just thinking about what should we do with nested macros. Comment Actions The code uses a raw lexer, so it doesn't expand macros. Thus the piece of code quoted will not get the "#if QT_HAS_INCLUDE(<chrono>)" part rewritten. My specific use case is the Icecream distributed build tool which does remote builds in a chroot and so this #if will be false and the expanded contents of <chrono> will be ignored, leading to build errors (https://github.com/icecc/icecream/issues/471). I can change the code to use 'while' instead of 'if' to handle nested macros, but I think such an approach should go only so far. My patch is simple and handles a realistic scenario, but if more complex macro handling is needed (which is a question), then IMO the code should be changed to do macro expansion properly. It was a long time ago when I wrote the code, but seeing that usage of Preprocessor::SetMacroExpansionOnlyInDirectives() I think such an expansion could be done similarly how handling macro expansion is handled in other directives - which IIRC is getting that information from PPCallbacks. Comment Actions Just a thought: From the point of view of -frewrite-includes, it should technically do full/nested macro expansion right? But from the point of view of Icecream, that would defeat the purpose of using -frewrite-includes, since the goal is to embed all includes into a single source file, ship it to a remote node, and do the rest of the preprocessing there. Or is there a way to expand all the parts that may resolve to a __has_include but leave others as is (for later processing)? How does the code deal with #if FOO(123) && MACRO_HAS_INCLUDE(<foo.h>) eg? From the Qt side, this patch is good enough, since we only use one level of indirection (to handle compilers without this feature), so I'm just thinking out loud to understand the problem better 😃 Comment Actions Icecream's usage of -frewrite-includes is not special, the purpose is always to merge everything necessary for the compilation into one source file that can be later compiled on its own as if the original file was compiled. The #include expansion done during -frewrite-includes will evaluate all the #if conditions etc. the same way they will be evaluated again during the actual compilation. Comment Actions Perhaps we should rewrite all #if-like directives to #if 0 or #if 1? Either that or we could emit pragmas indicating the values of later __has_includes. Special-casing macros of this specific form is at best a partial solution. Comment Actions If I understand your suggestion correctly, wouldn't the latter break this case? #if QT_HAS_INCLUDE(<nope.h>)
# include <nope.h>
# define HAS_NOPE
#endif
#ifdef HAS_NOPE
void bar() { nope(); }
#endifComment Actions Patch generally looks good; just a minor concern about the output format.
Comment Actions
That was consistent with what the current code does, but I don't mind changing that.
That's ok. The code already inserts extra lines and always fixes them up using line markers. And I think that's unavoidable due to reasons below.
That's fragile. Do it with something like "#if value /*blah*/ " and you'll have a problem with comment nesting. That is why all the changes are inserted between the two extra #if 0 / #endif lines. IIRC I thought about this for a while when designing -frewrite-includes and this is the best I came up with.
It's strictly speaking not necessary, but it's there to make it easier to identify those extra #if lines added by -frewrite-includes. Comment Actions Updated the patch again, commenting out just a part of #if didn't work either, so it seems there's no good way to comment out unwanted #if. This patch surrounds rewritten #if conditions by extra #if 0 #endif block to disable them, moreover it makes the conditions be part of an empty #if #endif block, so they do not break nesting. I've also created https://reviews.llvm.org/D63979 to verify that the results in fact do compile.
Comment Actions I've noticed that the Developer Policy says that it's actually allowed to commit patches without approval for parts "that you have contributed or maintain". Given that I'm the author of -rewrite-includes I take it that it's ok if I commit this if there are no further comments. Comment Actions Just for understanding what's going on here, I'm going to ask some stupid questions 😄 So, this will make -frewrite-includes do more work, to ensure that it not only covers the "top level" #include or __has_include case, but also __has_include in one or more levels of macros? Does that effectively means it needs to do full preprocessing? Eg. to determine cases like this: #ifdef FOO #define BAR(x) __has_include(x) #else #define BAR(x) 0 #endif Or do we draw the line there and only support __has_include in macros if those are not conditional themselves? Does it depend on whether -E is passed with -frewrite-includes? Will this mean that cases like icecream and distcc that use -frewrite-includes will have fully preprocessed sources to send to the remote hosts (with correct line/column number when compilation fails), or should they still pass -D options during the remote compilation? I'm referring to this quote from the distcc manual:
How does ICECC_REMOTE_CPP work in this case? Comment Actions This patch does not make -frewrite-includes do any more work. It already does a partial preprocessing run which evaluates all preprocessor directives. The patch just make better use of that information.
You cannot pass -frewrite-includes without -E.
The patch does not change anything about the usage. -frewrite-includes still only expands all #include directives and that's it.
The same way as before. You preferably do nothing, which defaults to ICECC_REMOTE_CPP=1, which uses -frewrite-includes to send full source to the remote, which does full preprocessing and compilation and you get pretty much the same result as if compiling locally. Or you may go with ICECC_REMOTE_CPP=0, which will not do this and then you'll most likely get warnings caused by compiling already preprocessed source. As for ccache, I personally consider suggesting explicit passing of -frewrite-includes to the compilation as somewhat lame; if ccache requires it for the cpp run, it can silently add it under the hood the same way icecream does, or at least it can have it configurable the same as other options and not require messing with build options. I myself use ccache with CCACHE_DEPEND=1, which completely avoids any preprocessor usage. | ||||||||||||||||||||||||