Commit graph

44 commits

Author SHA1 Message Date
Alexey Prokhin
8c5cf40bb5 Fixed compilation of ldc1 2012-04-22 12:26:25 +04:00
Alexey Prokhin
bca73ec8b9 Merged dmd 2.059beta 2012-04-05 15:30:00 +04:00
Alexey Prokhin
629f13929e WIP: port to llvm 3.0 2011-10-25 15:43:39 +04:00
David Nadlinger
3e28eecfea Quick workaround for a LLVM 2.9 crash with higher optimization levels.
createFunctionAttrsPass() is a LLVM pass and the crash is deep in LLVM code, maybe something is wrong with the pass manager state.
2011-04-20 20:44:49 +02:00
Kelly Wilson
72ea043124 Recommit klickverbot's patch for LLVM-2.9. 2011-04-20 09:28:37 -06:00
Kelly Wilson
7c81278b9c Backed out changeset 5253bfbb3890 2011-04-20 00:28:23 -06:00
Kelly Wilson
b26b0f4196 Upgrade to LLVM-2.9. Thanks to David Nadlinger for the patch. 2011-04-19 21:57:15 -06:00
Tomas Lindquist Olsen
1fac40d2bd Update to work with LLVM 2.7.
Removed use of dyn_cast, llvm no compiles
without exceptions and rtti by
default. We do need exceptions for the libconfig stuff, but rtti isn't
necessary (anymore).

Debug info needs to be rewritten, as in LLVM 2.7 the format has
completely changed. To have something to look at while rewriting, the
old code has been wrapped inside #ifndef DISABLE_DEBUG_INFO , this means
that you have to define this to compile at the moment.

Updated tango 0.99.9 patch to include updated EH runtime code, which is
needed for LLVM 2.7 as well.
2010-05-19 12:42:32 +02:00
Benjamin Kramer
8b6f11938a Kill off StripMetaData. 2009-08-25 21:35:43 +02:00
Tomas Lindquist Olsen
d5d3ecdb05 Added back option to disable metadata generation and users. Set USE_METADATA to OFF in ccmake. 2009-07-31 11:28:31 +02:00
Benjamin Kramer
51c02d192a Some minor cleanups
* remove an #ifdef USE_METADATA I accidently left in
* remove now unneeded llvm-version includes
* fix indentation in metadata.h
* prevent the "Found native target" message from interrupting ccmake
2009-07-20 18:16:11 +02:00
Benjamin Kramer
8576788245 Get rid of USE_METADATA 2009-07-11 14:19:21 +02:00
Benjamin Kramer
45f7401a5b Remove all of the now unecessary #if LLVM_REV 2009-07-11 14:09:33 +02:00
Frits van Bommel
5ecdf1a70a Eliminate comparisons between GC allocations and constants. This removes some
`assert(this !is null)`s when member functions get inlined.

Also tweak pass order a bit.
2009-06-20 14:28:59 +02:00
Frits van Bommel
465f15eda0 Return void* from _d_allocclass so LLVM doesn't do weird things with it...
This allows `-instcombine` followed by `-gvn` to do devirtualization, so add
`-gvn` in strategic places in the default pass order.
2009-06-20 11:39:13 +02:00
Frits van Bommel
bc7737a03b Add some passes that opt runs to the default pass list. 2009-06-11 16:57:45 +02:00
Frits van Bommel
b999f679a7 Teach -dgc2stack to promote GC allocations in simple loops to stack
allocations too. (A "simple" loop is one where the allocation isn't used in a
subsequent iteration)

This also means it's no longer necessary to run this pass multiple times.
Running it once after inlining should now catch all cases.
2009-06-08 12:35:55 +02:00
Frits van Bommel
96650855ed Add a pass to strip the bodies of available_externally functions so string
literals and `TypeInfo`s only referenced by them can be deleted by `-globaldce`.
2009-06-07 16:00:13 +02:00
Frits van Bommel
96fdcd6347 Run semantic3 on imported modules, and emit new symbols with
`available_externally` linkage. This allows the inliner to inline functions from
other modules while telling the code generator to ignore those functions (treat
them as declarations)
Still generates a few extra `TypeInfo`s and strings...
Disabled when generating debug info because I don't really understand it, and it
doesn't like this.
2009-06-07 16:00:13 +02:00
Frits van Bommel
b972dcb92d Add -memcpyopt to the optimization pass list. 2009-06-07 15:07:29 +02:00
Frits van Bommel
5edb7545f2 Move -tailcallelim to a place in the optimization sequence where it seems to
be more useful.
2009-06-04 15:59:51 +02:00
Frits van Bommel
6619190ba5 * Add -functionattrs to default pass list so -dgc2stack is more effective
when callees aren't inlined. This should also improve various standard LLVM
  optimizations.
* Comment out some verbose logging.
2009-05-28 00:07:21 +02:00
Frits van Bommel
a2eee70761 Add -verify-each option to ease debugging 2009-05-03 21:58:28 +02:00
Frits van Bommel
865ae79e27 Strip metadata even when addPassesForOptLevel() isn't called, for instance when
only inlining or when manually specifying passes.
2009-05-03 20:16:15 +02:00
Matti Niemenmaa
87d960c29a Fix build when USE_METADATA is off. 2009-05-02 19:03:33 +03:00
Frits van Bommel
262ef97945 There's an issue with LLVM metadata support; it triggers an assert when trying
to generate asm for code with metadata globals.
This new pass is used as a workaround: it strips metadata from the module before
it reaches the code generator.
Obviously, this is disabled if LLVM doesn't support metadata.
2009-05-02 12:19:43 +02:00
Frits van Bommel
cff0f1f446 Implement another D-specific pass: -dgc2stack
This one promotes GC allocations to stack memory when it can determine it's safe
to do so.
Not all GC calls are recognized yet (in fact only one *is* recognized for now).
Needs metadata, so disabled for LLVM versions that don't support it.
2009-05-02 11:58:50 +02:00
Frits van Bommel
da4450d2f5 Make sure this still compiles after LLVM r70437, which introduces a
specialization for command-line option template parser<char> which does the
wrong thing for us...
2009-04-30 12:25:04 +02:00
Frits van Bommel
0df1e34eb2 Implement first D-specific optimization pass: -simplify-drtcalls.
It uses the machinery of the standard -simplify-libcalls pass, but optimizes
calls to the D runtime instead of calls to C libraries.

At the moment, these optimizations are implemented by this pass:
 - Avoid the runtime call for `arr.length = newlen` if it can determine that
   the new length isn't longer than the old one.
 - Ditto for `cast(T[]) arr` if it will clearly always succeed.
   (e.g. if the length of the original array is zero, or if the old element
   size is a multiple of the new element size)
2009-04-28 21:58:06 +02:00
Frits van Bommel
664c2c333e Remove a bit of code duplication. 2009-04-26 20:36:53 +02:00
Frits van Bommel
5d63aeaed9 Fix a logic bug. 2009-04-26 20:31:48 +02:00
Frits van Bommel
7782d80919 Fix typo in comment 2009-04-16 12:01:28 +02:00
Frits van Bommel
5ea141e919 Tweak some optimizations.
Delegates passed to inlined functions now also stand a chance of being inlined.
This should make opApply as efficient as a regular loop, as long as both opApply
and the foreachbody are eligible for inlining; which is to say most non-virtual
opApply invocations will likely get fully inlined now.
(Note: above requires -O2 -enable-inlining or -O3)
2009-04-16 11:58:43 +02:00
Frits van Bommel
424bb88da0 Enable inlining by default for -O3+. 2009-03-29 19:38:59 +02:00
Frits van Bommel
6135e4201a Small cleanup: remove mars.h #include 2009-03-29 16:47:19 +02:00
Frits van Bommel
257d305032 Change meaning of optimization levels: -O0 now means 'no optimization' like with
other compilers.
2009-03-29 15:51:14 +02:00
Frits van Bommel
e129494812 Allow specific optimization passes to be requested from the command line.
Now you can run "`ldc test.d -c -mem2reg -simplifycfg`" if you feel the urge.
The -O<N> options are still supported, and are inserted in the passes list in
the position where they appear on the command line.
(so -simplifycfg -O1 -instcombine does the "right thing")

One small change: -inline is renamed to -enable-inlining due to a naming
conflict with the option to add the -inline pass. -inline now inserts the
inlining pass in the position specified, not in the middle of -O<N>.
(ldmd has been updated to translate -inline to -enable-inlining)
2009-03-29 15:46:55 +02:00
Christian Kamm
6eeeafdba6 First part of rename to LDC. 2008-10-06 22:46:55 +02:00
Tomas Lindquist Olsen
6c5c7ee1c1 Fixed the optimizer thing, since llvm PR 2800 is already fixed, users need to upgrade LLVM to latest svn. 2008-09-18 21:07:35 +02:00
Tomas Lindquist Olsen
1daa67ba50 Removed useless assert in ArrayLiteralExp::toConstElem
Added second verification pass after optimization
Commented an optimization pass out from lvl2, it turns the IR invalid, see llvm PR 2800
2008-09-16 16:06:39 +02:00
Tomas Lindquist Olsen
928f7d4de5 [svn r296] Removed: the 'suite' dir, it never took off!
Fixed: foreach statement, key-type checks were buggy.
Fixed: setting LLVMDC versions on the command line is now an error.
Fixed: array compare runtime had incorrect param attrs on call.
Fixed: index expressions on dynamic array slices w/o storage was broken.
Fixed: scope classes had incorrect finalization in some cases.
Fixed: when outputting !ClassInfoS !OffsetTypeInfoS, static class members were trying to be included, crashing the compiler.
Fixed: calling LLVMDC with -inline but not any -O option caused assertion failure.
Changed: the runtime now uses a single interface to "get" to !TypeInfoS, part of eliminating duplicate !TypeInfo codegen.
2008-06-19 17:30:32 +02:00
Tomas Lindquist Olsen
bc08c6fcb1 [svn r135] * Merged DMD 1.025 *
* Fixed a minor linking order mishap *
* Added an command line option -annotate *
* Fixed some problems with running optimizations *
* Added std.stdio and dependencies to lphobos (still not 100% working, but compiles and links) *
* Fixed problems with passing aggregate types to variadic functions *
* Added initial code towards full GC support, currently based on malloc and friends, not all the runtime calls the GC yet for memory *
* Fixed problems with resolving nested function context pointers for some heavily nested cases *
* Redid function argument passing + other minor code cleanups, still lots to do on this end... *
2008-01-04 01:38:42 +01:00
Tomas Lindquist Olsen
5eb88f9e80 [svn r133] Fixed some problems with inlining not happening :P
Fixed problems with certain cases of deeply nested classes/functions.
2007-12-28 22:55:24 +01:00
Tomas Lindquist Olsen
5ab2166b05 [svn r129] Started AA literals.
Fixed #15, passing -O will now invoke the optimizer before writing bitcode.
2007-11-28 04:52:35 +01:00