diff options
author | Alon Zakai <alonzakai@gmail.com> | 2013-06-27 12:04:36 -0700 |
---|---|---|
committer | Alon Zakai <alonzakai@gmail.com> | 2013-06-27 12:05:53 -0700 |
commit | 0432f1c7910cf12b51ea84a0a4cacb3465f40c63 (patch) | |
tree | bae9316b19e3c4ae12b197f75d8e554fea6c2894 | |
parent | 681ab3e9157f15c8276695413648b2ba9775894b (diff) |
add option for -O3 style LTO without normal LLVM LTO, because java-nbody hits an LTO bug but benefits from -O3 on all the code
-rwxr-xr-x | emcc | 16 | ||||
-rwxr-xr-x | tests/runner.py | 4 |
2 files changed, 11 insertions, 9 deletions
@@ -139,7 +139,7 @@ Options that are modified or new in %s include: -s DOUBLE_MODE=0 -s PRECISE_I64_MATH=0 --closure 1 - --llvm-lto 2 + --llvm-lto 3 This is not recommended at all. A better idea is to try each of these separately on top of @@ -217,12 +217,14 @@ Options that are modified or new in %s include: --llvm-lto <level> 0: No LLVM LTO (default) 1: LLVM LTO is performed + 2: We combine all the bitcode and run LLVM opt -O3 + on that (which optimizes across modules, but is + not the same as normal LTO), but do not do normal + LTO + 3: We do both 2 and then 1 Note: If LLVM optimizations are not run - (see --llvm-opts), setting this to 1 has no + (see --llvm-opts), setting this has no effect. - 2: LLVM LTO is performed, and we supplement it - with pre-running LLVM opt -O3 on the full - combined bitcode --closure <on> 0: No closure compiler (default in -O2 and below) 1: Run closure compiler. This greatly reduces @@ -882,7 +884,7 @@ try: newargs = newargs + [default_cxx_std] if llvm_opts is None: llvm_opts = LLVM_OPT_LEVEL[opt_level] - if llvm_lto is None and opt_level >= 3: llvm_lto = 2 + if llvm_lto is None and opt_level >= 3: llvm_lto = 3 if opt_level == 0: debug_level = 4 if closure is None and opt_level == 3: closure = True @@ -1433,7 +1435,7 @@ try: 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(): + if llvm_lto and llvm_lto != 2 and shared.Building.can_use_unsafe_opts(): if not shared.Building.can_inline(): link_opts.append('-disable-inlining') # do not internalize in std-link-opts - it ignores internalize-public-api-list - and add a manual internalize link_opts += ['-disable-internalize'] + shared.Building.get_safe_internalize() + ['-std-link-opts'] diff --git a/tests/runner.py b/tests/runner.py index 8fd14609..d20d7c2f 100755 --- a/tests/runner.py +++ b/tests/runner.py @@ -13419,7 +13419,7 @@ process(sys.argv[1]) try_delete(final_filename) output = Popen([PYTHON, EMCC, filename, #'-O3', '-O2', '-s', 'DOUBLE_MODE=0', '-s', 'PRECISE_I64_MATH=0', - '--llvm-lto', '2', '--memory-init-file', '0', '--js-transform', 'python hardcode.py', + '--llvm-lto', '3', '--memory-init-file', '0', '--js-transform', 'python hardcode.py', '-s', 'TOTAL_MEMORY=128*1024*1024', '--closure', '1', #'-g', @@ -13737,7 +13737,7 @@ process(sys.argv[1]) args = [path_from_root('tests', 'nbody-java', x) for x in os.listdir(path_from_root('tests', 'nbody-java')) if x.endswith('.c')] + \ ['-I' + path_from_root('tests', 'nbody-java')] self.do_benchmark('nbody_java', '', '''Time(s)''', - force_c=True, emcc_args=args + ['-s', 'PRECISE_I64_MATH=1', '--llvm-lto', '0'], native_args=args + ['-lgc', '-std=c99', '-target', 'x86_64-pc-linux-gnu', '-lm']) + force_c=True, emcc_args=args + ['-s', 'PRECISE_I64_MATH=1', '--llvm-lto', '2'], native_args=args + ['-lgc', '-std=c99', '-target', 'x86_64-pc-linux-gnu', '-lm']) def lua(self, benchmark, expected, output_parser=None, args_processor=None): shutil.copyfile(path_from_root('tests', 'lua', benchmark + '.lua'), benchmark + '.lua') |