diff options
Diffstat (limited to 'emcc')
| -rwxr-xr-x | emcc | 266 |
1 files changed, 202 insertions, 64 deletions
@@ -1,4 +1,5 @@ #!/usr/bin/env python2 +# -*- Mode: python -*- ''' emcc - compiler helper script @@ -74,7 +75,7 @@ emcc can be influenced by a few environment variables: EMMAKEN_COMPILER - The compiler to be used, if you don't want the default clang. ''' -import os, sys, shutil, tempfile, subprocess, shlex, time +import os, sys, shutil, tempfile, subprocess, shlex, time, re from subprocess import PIPE, STDOUT from tools import shared from tools.shared import Compression, execute, suffix, unsuffixed, unsuffixed_basename @@ -90,7 +91,10 @@ LLVM_OPT_LEVEL = { 3: 3, } -DEBUG = int(os.environ.get('EMCC_DEBUG') or 0) +DEBUG = os.environ.get('EMCC_DEBUG') +if DEBUG == "0": + DEBUG = None + TEMP_DIR = os.environ.get('EMCC_TEMP_DIR') LEAVE_INPUTS_RAW = os.environ.get('EMCC_LEAVE_INPUTS_RAW') # Do not compile .ll files into .bc, just compile them with emscripten directly # Not recommended, this is mainly for the test runner, or if you have some other @@ -118,12 +122,43 @@ if len(sys.argv) == 1: print 'emcc: no input files' exit(1) +# read response files very early on +response_file = True +while response_file: + response_file = None + for index in range(1, len(sys.argv)): + if sys.argv[index][0] == '@': + # found one, loop again next time + print >>sys.stderr, 'emcc: using response file: %s' % response_file + response_file = sys.argv[index][1:] + if not os.path.exists(response_file): + print >>sys.stderr, 'emcc: error: Response file not found: %s' % response_file + exit(1) + + response_fd = open(response_file, 'r') + extra_args = shlex.split(response_fd.read()) + response_fd.close() + + # slice in extra_args in place of the response file arg + sys.argv[index:index+1] = extra_args + #if DEBUG: print >>sys.stderr, "Expanded response file: " + " | ".join(sys.argv) + break + if sys.argv[1] == '--version': - print '''emcc (Emscripten GCC-like replacement) 2.0 -Copyright (C) 2012 the Emscripten authors. + revision = '(unknown revision)' + here = os.getcwd() + os.chdir(shared.path_from_root()) + try: + revision = execute(['git', 'show'], stdout=PIPE, stderr=PIPE)[0].split('\n')[0] + except: + pass + finally: + os.chdir(here) + print '''emcc (Emscripten GCC-like replacement) %s (%s) +Copyright (C) 2013 the Emscripten authors (see AUTHORS.txt) This is free and open source software under the MIT license. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. - ''' + ''' % (shared.EMSCRIPTEN_VERSION, revision) exit(0) elif sys.argv[1] == '--help': this = os.path.basename('em++' if os.environ.get('EMMAKEN_CXX') else 'emcc') @@ -146,16 +181,23 @@ Options that are modified or new in %s include: tools/shared.py and also src/settings.js.) Note: Optimizations are only done when compiling to JavaScript, not to intermediate - bitcode. + bitcode, *unless* you build with + EMCC_OPTIMIZE_NORMALLY=1 (not recommended + unless you know what you are doing!) -O2 As -O1, plus the relooper (loop recreation), - plus closure compiler advanced opts, plus - LLVM -O2 optimizations - Warning: Compiling with this takes a long time! + plus LLVM -O2 optimizations -O3 As -O2, plus dangerous optimizations that may - break the generated code! This is not - recommended at all, see the wiki for more - details (you can try -O2 and then add - dangerous optimizations one by one). + break the generated code! This adds + + -s DOUBLE_MODE=0 + -s PRECISE_I64_MATH=0 + --closure 1 + --llvm-lto 1 + + This is not recommended at all. A better idea + is to try each of these separately on top of + -O2 to see what works. See the wiki for more + information. -s OPTION=VALUE JavaScript code generation option passed into the emscripten compiler. For the @@ -177,6 +219,12 @@ Options that are modified or new in %s include: the last compilation phase from bitcode to JavaScript, or else we will remove it by default in -O1 and above. + In -O0, line numbers wil be shown in the + generated code. In -O1 and above, the optimizer + removes those comments. This flag does however + have the effect of disabling anything that + causes name mangling or minification (closure + or the registerize pass). --typed-arrays <mode> 0: No typed arrays 1: Parallel typed arrays @@ -187,14 +235,23 @@ Options that are modified or new in %s include: 2: -O2 LLVM optimizations 3: -O3 LLVM optimizations (default in -O2+) - --llvm-lto <level> 0: No LLVM LTO (default in -O0) - 1: LLVM LTO (default in -O1+) + --llvm-lto <level> 0: No LLVM LTO (default in -O2 and below) + 1: LLVM LTO (default in -O3) Note: If LLVM optimizations are not run (see --llvm-opts), setting this to 1 has no effect. - --closure <on> 0: No closure compiler (default in -O0, -O1) - 1: Run closure compiler (default in -O2, -O3) + --closure <on> 0: No closure compiler (default in -O2 and below) + 1: Run closure compiler. This greatly reduces + code size and may in some cases increase + runtime speed (although the opposite can also + occur). Note that it takes time to run, and + may require some changes to the code. This + is run by default in -O3. + + Note: If closure compiler hits an out-of-memory, + try adjusting JAVA_HEAP_SIZE in the environment + (for example, to 4096m for 4GB). --js-transform <cmd> <cmd> will be called on the generated code before it is optimized. This lets you modify @@ -345,6 +402,41 @@ Options that are modified or new in %s include: for a later incremental build (where you also enable it) to be sped up. + Caching works separately on 4 parts of compilation: + 'pre' which is types and global variables; that + information is then fed into 'funcs' which are + the functions (which we parallelize), and then + 'post' which adds final information based on + the functions (e.g., do we need long64 support + code). Finally, 'jsfuncs' are JavaScript-level + optimizations. Each of the 4 parts can be cached + separately, but note that they can affect each + other: If you recompile a single C++ file that + changes a global variable - e.g., adds, removes + or modifies a global variable, say by adding + a printf or by adding a compile-time timestamp, + then 'pre' cannot be loaded from the cache. And + since 'pre's output is sent to 'funcs' and 'post', + they will get invalidated as well, and only + 'jsfuncs' will be cached. So avoid modifying + globals to let caching work fully. + + To work around the problem mentioned in the + previous paragraph, you can use + + emscripten_jcache_printf + + when adding debug printfs to your code. That + function is specially preprocessed so that it + does not create a constant string global for + its first argument. See emscripten.h for more + details. Note in particular that you need to + already have a call to that function in your + code *before* you add one and do an incremental + build, so that adding an external reference + (also a global property) does not invalidate + everything. + --clear-cache Manually clears the cache of compiled emscripten system libraries (libc++, libc++abi, libc). This is normally @@ -600,7 +692,8 @@ try: ignore_dynamic_linking = False shell_path = shared.path_from_root('src', 'shell.html') js_libraries = [] - keep_debug = False + keep_llvm_debug = False + keep_js_debug = False bind = False jcache = False if use_cxx: @@ -613,18 +706,21 @@ try: absolute_warning_shown = False + settings_changes = [] + for i in range(len(newargs)): newargs[i] = newargs[i].strip() # On Windows Vista (and possibly others), excessive spaces in the command line leak into the items in this array, so trim e.g. 'foo.cpp ' -> 'foo.cpp' if newargs[i].startswith('-O'): - requested_level = newargs[i][2] + # Let -O default to -O2, which is what gcc does. + requested_level = newargs[i][2:] or '2' if requested_level == 's': - print >> sys.stderr, 'emcc: warning: -Os is ignored (use -O0, -O1, -O2)' - else: - try: - opt_level = int(requested_level) - assert 0 <= opt_level <= 3 - except: - raise Exception('Invalid optimization level: ' + newargs[i]) + requested_level = 2 + settings_changes.append('INLINING_LIMIT=50') + try: + opt_level = int(requested_level) + assert 0 <= opt_level <= 3 + except: + raise Exception('Invalid optimization level: ' + newargs[i]) newargs[i] = '' elif newargs[i].startswith('--llvm-opts'): check_bad_eq(newargs[i]) @@ -667,7 +763,8 @@ try: newargs[i] = '' newargs[i+1] = '' elif newargs[i] == '-g': - keep_debug = True + keep_llvm_debug = True + keep_js_debug = True elif newargs[i] == '--bind': bind = True newargs[i] = '' @@ -737,16 +834,16 @@ try: newargs = newargs + [default_cxx_std] if llvm_opts is None: llvm_opts = LLVM_OPT_LEVEL[opt_level] - if llvm_lto is None: llvm_lto = llvm_opts > 0 - if closure is None: closure = 1 if opt_level >= 2 else 0 - if opt_level <= 0: keep_debug = True # always keep debug in -O0 + if llvm_lto is None: llvm_lto = opt_level >= 3 + if opt_level <= 0: keep_llvm_debug = keep_js_debug = True # always keep debug in -O0 + if opt_level > 0: keep_llvm_debug = False # JS optimizer wipes out llvm debug info from being visible + if closure is None and opt_level == 3: closure = True if DEBUG: start_time = time.time() # done after parsing arguments, which might affect debug state if closure: assert os.path.exists(shared.CLOSURE_COMPILER), 'emcc: fatal: Closure compiler (%s) does not exist' % shared.CLOSURE_COMPILER - settings_changes = [] for i in range(len(newargs)): if newargs[i] == '-s': if is_minus_s_for_emcc(newargs, i): @@ -879,7 +976,11 @@ try: shared.Settings.CORRECT_OVERFLOWS = 1 if shared.Settings.CORRECT_SIGNS >= 2 or shared.Settings.CORRECT_OVERFLOWS >= 2 or shared.Settings.CORRECT_ROUNDINGS >= 2: - keep_debug = True # must keep debug info to do line-by-line operations + keep_llvm_debug = True # must keep debug info to do line-by-line operations + + if (keep_llvm_debug or keep_js_debug) and closure: + print >> sys.stderr, 'emcc: warning: disabling closure because debug info was requested' + closure = False if minify_whitespace is None: minify_whitespace = closure # if closure is run, minify whitespace @@ -894,6 +995,7 @@ try: for input_file in input_files: if input_file.endswith(SOURCE_SUFFIXES): if DEBUG: print >> sys.stderr, 'emcc: compiling source file: ', input_file + input_file = shared.Building.preprocess(input_file, in_temp(uniquename(input_file))) output_file = in_temp(unsuffixed(uniquename(input_file)) + '.o') temp_files.append(output_file) args = newargs + ['-emit-llvm', '-c', input_file, '-o', output_file] @@ -927,7 +1029,15 @@ try: # If we were just asked to generate bitcode, stop there if final_suffix not in JS_CONTAINING_SUFFIXES: if llvm_opts > 0: - print >> sys.stderr, 'emcc: warning: -Ox flags ignored, since not generating JavaScript' + if not os.environ.get('EMCC_OPTIMIZE_NORMALLY'): + print >> sys.stderr, 'emcc: warning: -Ox flags ignored, since not generating JavaScript' + else: + for input_file in input_files: + if input_file.endswith(SOURCE_SUFFIXES): + if DEBUG: print >> sys.stderr, 'emcc: optimizing %s with -O%d since EMCC_OPTIMIZE_NORMALLY defined' % (input_file, llvm_opts) + shared.Building.llvm_opt(in_temp(unsuffixed(uniquename(input_file)) + '.o'), llvm_opts) + else: + if DEBUG: print >> sys.stderr, 'emcc: not optimizing %s despite EMCC_OPTIMIZE_NORMALLY since not source code' % (input_file) if not specified_target: for input_file in input_files: shutil.move(in_temp(unsuffixed(uniquename(input_file)) + '.o'), unsuffixed_basename(input_file) + '.' + final_suffix) @@ -959,7 +1069,22 @@ try: def create_libc(): if DEBUG: print >> sys.stderr, 'emcc: building libc for cache' o_s = [] - for src in ['dlmalloc.c', os.path.join('libcxx', 'new.cpp')]: + libc_files = [ + 'dlmalloc.c', + os.path.join('libcxx', 'new.cpp'), + os.path.join('libc', 'stdlib', 'getopt_long.c'), + os.path.join('libc', 'gen', 'err.c'), + os.path.join('libc', 'gen', 'errx.c'), + os.path.join('libc', 'gen', 'warn.c'), + os.path.join('libc', 'gen', 'warnx.c'), + os.path.join('libc', 'gen', 'verr.c'), + os.path.join('libc', 'gen', 'verrx.c'), + os.path.join('libc', 'gen', 'vwarn.c'), + os.path.join('libc', 'gen', 'vwarnx.c'), + os.path.join('libc', 'stdlib', 'strtod.c'), + ]; + + for src in libc_files: o = in_temp(os.path.basename(src) + '.o') execute([shared.PYTHON, shared.EMCC, shared.path_from_root('system', 'lib', src), '-o', o], stdout=stdout, stderr=stderr) o_s.append(o) @@ -984,7 +1109,7 @@ try: def create_libcxx(): if DEBUG: print >> sys.stderr, 'emcc: building libcxx for cache' os = [] - for src in ['algorithm.cpp', 'condition_variable.cpp', 'future.cpp', 'iostream.cpp', 'memory.cpp', 'random.cpp', 'stdexcept.cpp', 'system_error.cpp', 'utility.cpp', 'bind.cpp', 'debug.cpp', 'hash.cpp', 'mutex.cpp', 'string.cpp', 'thread.cpp', 'valarray.cpp', 'chrono.cpp', 'exception.cpp', 'ios.cpp', 'locale.cpp', 'regex.cpp', 'strstream.cpp', 'typeinfo.cpp']: + for src in ['algorithm.cpp', 'condition_variable.cpp', 'future.cpp', 'iostream.cpp', 'memory.cpp', 'random.cpp', 'stdexcept.cpp', 'system_error.cpp', 'utility.cpp', 'bind.cpp', 'debug.cpp', 'hash.cpp', 'mutex.cpp', 'string.cpp', 'thread.cpp', 'valarray.cpp', 'chrono.cpp', 'exception.cpp', 'ios.cpp', 'locale.cpp', 'regex.cpp', 'strstream.cpp']: o = in_temp(src + '.o') execute([shared.PYTHON, shared.EMXX, shared.path_from_root('system', 'lib', 'libcxx', src), '-o', o], stdout=stdout, stderr=stderr) os.append(o) @@ -1003,7 +1128,7 @@ try: def create_libcxxabi(): if DEBUG: print >> sys.stderr, 'emcc: building libcxxabi for cache' os = [] - for src in ['private_typeinfo.cpp']: + for src in ['private_typeinfo.cpp', 'typeinfo.cpp']: o = in_temp(src + '.o') execute([shared.PYTHON, shared.EMXX, shared.path_from_root('system', 'lib', 'libcxxabi', 'src', src), '-o', o], stdout=stdout, stderr=stderr) os.append(o) @@ -1017,29 +1142,32 @@ try: libcxxabi_symbols = filter(lambda symbol: symbol not in libc_symbols, libcxxabi_symbols) libcxxabi_symbols = set(libcxxabi_symbols) - force = False # If we have libcxx, we must force inclusion of libc, since libcxx uses new internally. Note: this is kind of hacky - + # If we have libcxx, we must force inclusion of libc, since libcxx uses new internally. Note: this is kind of hacky + # Settings this in the environment will avoid checking dependencies and make building big projects a little faster + force = os.environ.get('EMCC_FORCE_STDLIBS') + has = need = None for name, create, fix, library_symbols in [('libcxx', create_libcxx, fix_libcxx, libcxx_symbols), ('libcxxabi', create_libcxxabi, fix_libcxxabi, libcxxabi_symbols), ('libc', create_libc, fix_libc, libc_symbols)]: - need = set() - has = set() - for temp_file in temp_files: - symbols = shared.Building.llvm_nm(temp_file) - for library_symbol in library_symbols: - if library_symbol in symbols.undefs: - need.add(library_symbol) - if library_symbol in symbols.defs: - has.add(library_symbol) - for haz in has: # remove symbols that are supplied by another of the inputs - if haz in need: - need.remove(haz) - if DEBUG: print >> sys.stderr, 'emcc: considering including %s: we need %s and have %s' % (name, str(need), str(has)) + if not force: + need = set() + has = set() + for temp_file in temp_files: + symbols = shared.Building.llvm_nm(temp_file) + for library_symbol in library_symbols: + if library_symbol in symbols.undefs: + need.add(library_symbol) + if library_symbol in symbols.defs: + has.add(library_symbol) + for haz in has: # remove symbols that are supplied by another of the inputs + if haz in need: + need.remove(haz) + if DEBUG: print >> sys.stderr, 'emcc: considering including %s: we need %s and have %s' % (name, str(need), str(has)) if force or len(need) > 0: # We need to build and link the library in if DEBUG: print >> sys.stderr, 'emcc: including %s' % name libfile = shared.Cache.get(name, create) - if len(has) > 0: + if has and len(has) > 0: # remove the symbols we do not need fixed = in_temp(uniquename(libfile)) + '.bc' shutil.copyfile(libfile, fixed) @@ -1049,7 +1177,7 @@ try: libfile = fixed extra_files_to_link.append(libfile) force = True - if fix: + if fix and need: fix(need) # First, combine the bitcode files if there are several. We must also link if we have a singleton .a @@ -1057,7 +1185,10 @@ try: (not LEAVE_INPUTS_RAW and not (suffix(temp_files[0]) in BITCODE_SUFFIXES or suffix(temp_files[0]) in DYNAMICLIB_SUFFIXES) and shared.Building.is_ar(temp_files[0])): linker_inputs = temp_files + extra_files_to_link if DEBUG: print >> sys.stderr, 'emcc: linking: ', linker_inputs + t0 = time.time() shared.Building.link(linker_inputs, in_temp(target_basename + '.bc')) + t1 = time.time() + if DEBUG: print >> sys.stderr, 'emcc: linking took %.2f seconds' % (t1 - t0) final = in_temp(target_basename + '.bc') else: if not LEAVE_INPUTS_RAW: @@ -1084,11 +1215,14 @@ try: # Optimize, if asked to if not LEAVE_INPUTS_RAW: - link_opts = [] if keep_debug else ['-strip-debug'] + link_opts = [] if keep_llvm_debug else ['-strip-debug'] # remove LLVM debug info in -O1+, since the optimizer removes it anyhow if llvm_opts > 0: - shared.Building.llvm_opt(in_temp(target_basename + '.bc'), llvm_opts) - if DEBUG: save_intermediate('opt', 'bc') - # Do LTO in a separate pass to work around LLVM bug XXX (see failure e.g. in cubescript) + if not os.environ.get('EMCC_OPTIMIZE_NORMALLY'): + shared.Building.llvm_opt(in_temp(target_basename + '.bc'), llvm_opts) + if DEBUG: save_intermediate('opt', 'bc') + # Do LTO in a separate pass to work around LLVM bug XXX (see failure e.g. in cubescript) + else: + if DEBUG: print >> sys.stderr, 'emcc: not running opt because EMCC_OPTIMIZE_NORMALLY was specified, opt should have been run before' if shared.Building.can_build_standalone(): # If we can LTO, do it before dce, since it opens up dce opportunities if llvm_lto and shared.Building.can_use_unsafe_opts(): @@ -1167,15 +1301,12 @@ try: execute(shlex.split(js_transform, posix=posix) + [os.path.abspath(final)]) if DEBUG: save_intermediate('transformed') - if shared.Settings.ASM_JS: # XXX temporary wrapping for testing purposes - print >> sys.stderr, 'emcc: ASM_JS mode is highly experimental, and will not work on most codebases yet. It is NOT recommended that you try this yet.' - # It is useful to run several js optimizer passes together, to save on unneeded unparsing/reparsing js_optimizer_queue = [] def flush_js_optimizer_queue(): global final, js_optimizer_queue if len(js_optimizer_queue) > 0 and not(len(js_optimizer_queue) == 1 and js_optimizer_queue[0] == 'last'): - if DEBUG < 2: + if DEBUG != '2': if shared.Settings.ASM_JS: js_optimizer_queue = ['asm'] + js_optimizer_queue if DEBUG: print >> sys.stderr, 'emcc: applying js optimization passes:', js_optimizer_queue @@ -1194,7 +1325,7 @@ try: if opt_level >= 1: if DEBUG: print >> sys.stderr, 'emcc: running pre-closure post-opts' - if DEBUG >= 2: + if DEBUG == '2': # Clean up the syntax a bit final = shared.Building.js_optimizer(final, [], jcache) if DEBUG: save_intermediate('pretty') @@ -1216,8 +1347,9 @@ try: if DEBUG: print >> sys.stderr, 'emcc: running closure' final = shared.Building.closure_compiler(final) if DEBUG: save_intermediate('closure') - elif shared.Settings.ASM_JS and shared.Settings.RELOOP: - js_optimizer_queue += ['registerize'] # we can't use closure in asm, but this does much of the same + elif shared.Settings.RELOOP and not closure and not keep_js_debug: + # do this if closure is not enabled (it gives similar speedups), and we do not need to keep debug info around + js_optimizer_queue += ['registerize'] if opt_level >= 1: if DEBUG: print >> sys.stderr, 'emcc: running post-closure post-opts' @@ -1230,6 +1362,12 @@ try: flush_js_optimizer_queue() + if not minify_whitespace: + # Remove some trivial whitespace + src = open(final).read() + src = re.sub(r'\n+[ \n]*\n+', '\n', src) + open(final, 'w').write(src) + # If we were asked to also generate HTML, do that if final_suffix == 'html': if DEBUG: print >> sys.stderr, 'emcc: generating HTML' |