aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rwxr-xr-xemcc21
-rwxr-xr-xemscripten.py10
-rw-r--r--src/jsifier.js4
-rw-r--r--src/parseTools.js4
-rw-r--r--tests/cases/zeroextarg.ll22
-rwxr-xr-xtests/runner.py160
-rw-r--r--tools/js-optimizer.js187
-rw-r--r--tools/shared.py6
-rw-r--r--tools/test-js-optimizer-asm-last-output.js1
-rw-r--r--tools/test-js-optimizer-asm-last.js1
10 files changed, 296 insertions, 120 deletions
diff --git a/emcc b/emcc
index 90a2ea71..bffe8d5d 100755
--- a/emcc
+++ b/emcc
@@ -522,14 +522,14 @@ if CONFIGURE_CONFIG or CMAKE_CONFIG:
open(tempout, 'w').write('//\n')
src = None
- for i in range(len(sys.argv)):
- if sys.argv[i].endswith('.c'):
+ for arg in sys.argv:
+ if arg.endswith('.c'):
try:
- src = open(sys.argv[i]).read()
- if debug_configure: open(tempout, 'a').write('============= ' + sys.argv[i] + '\n' + src + '\n=============\n\n')
+ src = open(arg).read()
+ if debug_configure: open(tempout, 'a').write('============= ' + arg + '\n' + src + '\n=============\n\n')
except:
pass
- if sys.argv[i].endswith('.s'):
+ if arg.endswith('.s'):
if debug_configure: open(tempout, 'a').write('(compiling .s assembly, must use clang\n')
use_js = 0
@@ -826,7 +826,7 @@ try:
newargs[i] = ''
newargs[i+1] = ''
elif newargs[i].startswith('--use-preload-cache'):
- use_preload_cache = True;
+ use_preload_cache = True
newargs[i] = ''
elif newargs[i] == '--ignore-dynamic-linking':
ignore_dynamic_linking = True
@@ -952,7 +952,7 @@ try:
if not prefix: continue
if l.startswith(prefix):
l = l[len(prefix):]
- break;
+ break
libs.append(l)
newargs[i] = ''
else:
@@ -1121,9 +1121,8 @@ try:
# Optimize source files
if llvm_opts > 0:
- for i in range(len(input_files)):
- input_file = input_files[i]
- if input_files[i].endswith(SOURCE_SUFFIXES):
+ for i, input_file in enumerate(input_files):
+ if input_file.endswith(SOURCE_SUFFIXES):
temp_file = temp_files[i]
logging.debug('optimizing %s with -O%d' % (input_file, llvm_opts))
shared.Building.llvm_opt(temp_file, llvm_opts)
@@ -1214,7 +1213,7 @@ try:
os.path.join('libc', 'gen', 'vwarn.c'),
os.path.join('libc', 'gen', 'vwarnx.c'),
os.path.join('libc', 'stdlib', 'strtod.c'),
- ];
+ ]
return build_libc('libc.bc', libc_files)
def apply_libc(need):
diff --git a/emscripten.py b/emscripten.py
index df0587f9..d9367566 100755
--- a/emscripten.py
+++ b/emscripten.py
@@ -163,10 +163,10 @@ def emscript(infile, settings, outfile, libraries=[], compiler_engine=None,
if DEBUG_CACHE and not out:
dfpath = os.path.join(get_configuration().TEMP_DIR, "ems_" + shortkey)
dfp = open(dfpath, 'w')
- dfp.write(pre_input);
- dfp.write("\n\n========================== settings_text\n\n");
- dfp.write(settings_text);
- dfp.write("\n\n========================== libraries\n\n");
+ dfp.write(pre_input)
+ dfp.write("\n\n========================== settings_text\n\n")
+ dfp.write(settings_text)
+ dfp.write("\n\n========================== libraries\n\n")
dfp.write("\n".join(libraries))
dfp.close()
print >>sys.stderr, ' cache miss, key data dumped to %s' % dfpath
@@ -485,7 +485,7 @@ def emscript(infile, settings, outfile, libraries=[], compiler_engine=None,
global_vars = map(lambda g: g['name'], filter(lambda g: settings['NAMED_GLOBALS'] or g.get('external') or g.get('unIndexable'), forwarded_json['Variables']['globals'].values()))
global_funcs = ['_' + key for key, value in forwarded_json['Functions']['libraryFunctions'].iteritems() if value != 2]
def math_fix(g):
- return g if not g.startswith('Math_') else g.split('_')[1];
+ return g if not g.startswith('Math_') else g.split('_')[1]
asm_global_funcs = ''.join([' var ' + g.replace('.', '_') + '=global.' + g + ';\n' for g in maths]) + \
''.join([' var ' + g + '=env.' + math_fix(g) + ';\n' for g in basic_funcs + global_funcs])
asm_global_vars = ''.join([' var ' + g + '=env.' + g + '|0;\n' for g in basic_vars + global_vars]) + \
diff --git a/src/jsifier.js b/src/jsifier.js
index 86931f3e..82b78d0a 100644
--- a/src/jsifier.js
+++ b/src/jsifier.js
@@ -1479,8 +1479,8 @@ function JSify(data, functionsOnly, givenFunctions) {
return inline.apply(null, args); // Warning: inlining does not prevent recalculation of the arguments. They should be simple identifiers
}
- if (ASM_JS) {
- // remove unneeded arguments, which the asm sig can show us. this lets us alias memset with llvm.memset, we just
+ if (ASM_JS && ident.indexOf('llvm_') >= 0) {
+ // remove unneeded arguments in llvm intrinsic functions, which the asm sig can show us. this lets us alias memset with llvm.memset, we just
// drop the final 2 args so things validate properly in asm
var libsig = LibraryManager.library[simpleIdent + '__sig'];
if (libsig) {
diff --git a/src/parseTools.js b/src/parseTools.js
index dfd4b7ed..eb200c65 100644
--- a/src/parseTools.js
+++ b/src/parseTools.js
@@ -2222,9 +2222,9 @@ function processMathop(item) {
// basic integer ops
case 'add': return handleOverflow(getFastValue(idents[0], '+', idents[1], item.type), bits);
case 'sub': return handleOverflow(getFastValue(idents[0], '-', idents[1], item.type), bits);
- case 'sdiv': case 'udiv': return makeRounding(getFastValue(idents[0], '/', idents[1], item.type), bits, op[0] === 's');
+ case 'sdiv': case 'udiv': return makeRounding(getFastValue(idents[0], '/', idents[1], item.type), bits, true);
case 'mul': return getFastValue(idents[0], '*', idents[1], item.type); // overflow handling is already done in getFastValue for '*'
- case 'urem': case 'srem': return makeRounding(getFastValue(idents[0], '%', idents[1], item.type), bits, op[0] === 's');
+ case 'urem': case 'srem': return makeRounding(getFastValue(idents[0], '%', idents[1], item.type), bits, true);
case 'or': {
if (bits > 32) {
assert(bits === 64, 'Too many bits for or: ' + bits);
diff --git a/tests/cases/zeroextarg.ll b/tests/cases/zeroextarg.ll
new file mode 100644
index 00000000..25efb7ec
--- /dev/null
+++ b/tests/cases/zeroextarg.ll
@@ -0,0 +1,22 @@
+; ModuleID = 'tests/hello_world.bc'
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32-n8:16:32-S128"
+target triple = "i386-pc-linux-gnu"
+
+@.str = private unnamed_addr constant [15 x i8] c"hello, world!\0A\00", align 1 ; [#uses=1 type=[15 x i8]*]
+
+define void @glSampleCoverage(float %18, i8 zeroext %invert) {
+entry:
+ %call = call i32 (i8*, ...)* @printf(i8* getelementptr inbounds ([15 x i8]* @.str, i32 0, i32 0)) ; [#uses=0 type=i32]
+ ret void
+}
+
+; [#uses=0]
+define i32 @main() {
+entry:
+ tail call void @glSampleCoverage(float 3.5, i8 zeroext 12)
+ ret i32 1
+}
+
+; [#uses=1]
+declare i32 @printf(i8*, ...)
+
diff --git a/tests/runner.py b/tests/runner.py
index 7b894110..b21eee08 100755
--- a/tests/runner.py
+++ b/tests/runner.py
@@ -10020,97 +10020,97 @@ finalizing 3 (global == 0)
''')
# Generate tests for everything
- def make_run(fullname, name=-1, compiler=-1, embetter=0, quantum_size=0, typed_arrays=0, emcc_args=None, env='{}'):
- exec('''
-class %s(T):
- run_name = '%s'
- env = %s
+ def make_run(fullname, name=-1, compiler=-1, embetter=0, quantum_size=0,
+ typed_arrays=0, emcc_args=None, env=None):
- def tearDown(self):
- super(%s, self).tearDown()
+ if env is None: env = {}
- for k, v in self.env.iteritems():
- del os.environ[k]
+ TT = type(fullname, (T,), dict(run_name = fullname, env = env))
+
+ def tearDown(self):
+ super(TT, self).tearDown()
+
+ for k, v in self.env.iteritems():
+ del os.environ[k]
+
+ TT.tearDown = tearDown
+
+ def setUp(self):
+ super(TT, self).setUp()
+ for k, v in self.env.iteritems():
+ assert k not in os.environ, k + ' should not be in environment'
+ os.environ[k] = v
+
+ global checked_sanity
+ if not checked_sanity:
+ print '(checking sanity from test runner)' # do this after we set env stuff
+ check_sanity(force=True)
+ checked_sanity = True
+
+ Building.COMPILER_TEST_OPTS = ['-g']
+ os.chdir(self.get_dir()) # Ensure the directory exists and go there
+ Building.COMPILER = compiler
+
+ self.emcc_args = None if emcc_args is None else emcc_args[:]
+ if self.emcc_args is not None:
+ Settings.load(self.emcc_args)
+ Building.LLVM_OPTS = 0
+ if '-O2' in self.emcc_args:
+ Building.COMPILER_TEST_OPTS = [] # remove -g in -O2 tests, for more coverage
+ #Building.COMPILER_TEST_OPTS += self.emcc_args
+ for arg in self.emcc_args:
+ if arg.startswith('-O'):
+ Building.COMPILER_TEST_OPTS.append(arg) # so bitcode is optimized too, this is for cpp to ll
+ else:
+ try:
+ key, value = arg.split('=')
+ Settings[key] = value # forward -s K=V
+ except:
+ pass
+ return
+
+ # TODO: Move much of these to a init() function in shared.py, and reuse that
+ Settings.USE_TYPED_ARRAYS = typed_arrays
+ Settings.INVOKE_RUN = 1
+ Settings.RELOOP = 0 # we only do them in the "o2" pass
+ Settings.MICRO_OPTS = embetter
+ Settings.QUANTUM_SIZE = quantum_size
+ Settings.ASSERTIONS = 1-embetter
+ Settings.SAFE_HEAP = 1-embetter
+ Settings.CHECK_OVERFLOWS = 1-embetter
+ Settings.CORRECT_OVERFLOWS = 1-embetter
+ Settings.CORRECT_SIGNS = 0
+ Settings.CORRECT_ROUNDINGS = 0
+ Settings.CORRECT_OVERFLOWS_LINES = CORRECT_SIGNS_LINES = CORRECT_ROUNDINGS_LINES = SAFE_HEAP_LINES = []
+ Settings.CHECK_SIGNS = 0 #1-embetter
+ Settings.RUNTIME_TYPE_INFO = 0
+ Settings.DISABLE_EXCEPTION_CATCHING = 0
+ Settings.INCLUDE_FULL_LIBRARY = 0
+ Settings.BUILD_AS_SHARED_LIB = 0
+ Settings.RUNTIME_LINKED_LIBS = []
+ Settings.EMULATE_UNALIGNED_ACCESSES = int(Settings.USE_TYPED_ARRAYS == 2 and Building.LLVM_OPTS == 2)
+ Settings.DOUBLE_MODE = 1 if Settings.USE_TYPED_ARRAYS and Building.LLVM_OPTS == 0 else 0
+ Settings.PRECISE_I64_MATH = 0
+ Settings.NAMED_GLOBALS = 0 if not embetter else 1
+
+ TT.setUp = setUp
- def setUp(self):
- super(%s, self).setUp()
-
- for k, v in self.env.iteritems():
- assert k not in os.environ, k + ' should not be in environment'
- os.environ[k] = v
-
- global checked_sanity
- if not checked_sanity:
- print '(checking sanity from test runner)' # do this after we set env stuff
- check_sanity(force=True)
- checked_sanity = True
-
- Building.COMPILER_TEST_OPTS = ['-g']
- os.chdir(self.get_dir()) # Ensure the directory exists and go there
- Building.COMPILER = %r
-
- self.emcc_args = %s
- if self.emcc_args is not None:
- Settings.load(self.emcc_args)
- Building.LLVM_OPTS = 0
- if '-O2' in self.emcc_args:
- Building.COMPILER_TEST_OPTS = [] # remove -g in -O2 tests, for more coverage
- #Building.COMPILER_TEST_OPTS += self.emcc_args
- for arg in self.emcc_args:
- if arg.startswith('-O'):
- Building.COMPILER_TEST_OPTS.append(arg) # so bitcode is optimized too, this is for cpp to ll
- else:
- try:
- key, value = arg.split('=')
- Settings[key] = value # forward -s K=V
- except:
- pass
- return
-
- embetter = %d
- quantum_size = %d
- # TODO: Move much of these to a init() function in shared.py, and reuse that
- Settings.USE_TYPED_ARRAYS = %d
- Settings.INVOKE_RUN = 1
- Settings.RELOOP = 0 # we only do them in the "o2" pass
- Settings.MICRO_OPTS = embetter
- Settings.QUANTUM_SIZE = quantum_size
- Settings.ASSERTIONS = 1-embetter
- Settings.SAFE_HEAP = 1-embetter
- Settings.CHECK_OVERFLOWS = 1-embetter
- Settings.CORRECT_OVERFLOWS = 1-embetter
- Settings.CORRECT_SIGNS = 0
- Settings.CORRECT_ROUNDINGS = 0
- Settings.CORRECT_OVERFLOWS_LINES = CORRECT_SIGNS_LINES = CORRECT_ROUNDINGS_LINES = SAFE_HEAP_LINES = []
- Settings.CHECK_SIGNS = 0 #1-embetter
- Settings.RUNTIME_TYPE_INFO = 0
- Settings.DISABLE_EXCEPTION_CATCHING = 0
- Settings.INCLUDE_FULL_LIBRARY = 0
- Settings.BUILD_AS_SHARED_LIB = 0
- Settings.RUNTIME_LINKED_LIBS = []
- Settings.EMULATE_UNALIGNED_ACCESSES = int(Settings.USE_TYPED_ARRAYS == 2 and Building.LLVM_OPTS == 2)
- Settings.DOUBLE_MODE = 1 if Settings.USE_TYPED_ARRAYS and Building.LLVM_OPTS == 0 else 0
- Settings.PRECISE_I64_MATH = 0
- Settings.NAMED_GLOBALS = 0 if not embetter else 1
-
-TT = %s
-''' % (fullname, fullname, env, fullname, fullname, compiler, str(emcc_args), embetter, quantum_size, typed_arrays, fullname))
return TT
# Make one run with the defaults
- exec('default = make_run("default", compiler=CLANG, emcc_args=[])')
+ default = make_run("default", compiler=CLANG, emcc_args=[])
# Make one run with -O1, with safe heap
- exec('o1 = make_run("o1", compiler=CLANG, emcc_args=["-O1", "-s", "ASM_JS=0", "-s", "SAFE_HEAP=1"])')
+ o1 = make_run("o1", compiler=CLANG, emcc_args=["-O1", "-s", "ASM_JS=0", "-s", "SAFE_HEAP=1"])
# Make one run with -O2, but without closure (we enable closure in specific tests, otherwise on everything it is too slow)
- exec('o2 = make_run("o2", compiler=CLANG, emcc_args=["-O2", "-s", "ASM_JS=0", "-s", "JS_CHUNK_SIZE=1024"])')
+ o2 = make_run("o2", compiler=CLANG, emcc_args=["-O2", "-s", "ASM_JS=0", "-s", "JS_CHUNK_SIZE=1024"])
# asm.js
- exec('asm1 = make_run("asm1", compiler=CLANG, emcc_args=["-O1", "-s", "CHECK_HEAP_ALIGN=1"])')
- exec('asm2 = make_run("asm2", compiler=CLANG, emcc_args=["-O2"])')
- exec('asm2g = make_run("asm2g", compiler=CLANG, emcc_args=["-O2", "-g", "-s", "ASSERTIONS=1", "--memory-init-file", "1"])')
- exec('''asm2x86 = make_run("asm2x86", compiler=CLANG, emcc_args=["-O2", "-g", "-s", "CHECK_HEAP_ALIGN=1"], env='{"EMCC_LLVM_TARGET": "i386-pc-linux-gnu"}')''')
+ asm1 = make_run("asm1", compiler=CLANG, emcc_args=["-O1", "-s", "CHECK_HEAP_ALIGN=1"])
+ asm2 = make_run("asm2", compiler=CLANG, emcc_args=["-O2"])
+ asm2g = make_run("asm2g", compiler=CLANG, emcc_args=["-O2", "-g", "-s", "ASSERTIONS=1", "--memory-init-file", "1"])
+ asm2x86 = make_run("asm2x86", compiler=CLANG, emcc_args=["-O2", "-g", "-s", "CHECK_HEAP_ALIGN=1"], env={"EMCC_LLVM_TARGET": "i386-pc-linux-gnu"})
# Make custom runs with various options
for compiler, quantum, embetter, typed_arrays in [
@@ -10120,7 +10120,7 @@ TT = %s
fullname = 's_0_%d%s%s' % (
embetter, '' if quantum == 4 else '_q' + str(quantum), '' if typed_arrays in [0, 1] else '_t' + str(typed_arrays)
)
- exec('%s = make_run(fullname, %r,%r,%d,%d,%d)' % (fullname, fullname, compiler, embetter, quantum, typed_arrays))
+ locals()[fullname] = make_run(fullname, fullname, compiler, embetter, quantum, typed_arrays)
del T # T is just a shape for the specific subclasses, we don't test it itself
diff --git a/tools/js-optimizer.js b/tools/js-optimizer.js
index 06d82752..6987511c 100644
--- a/tools/js-optimizer.js
+++ b/tools/js-optimizer.js
@@ -136,6 +136,9 @@ var CONTROL_FLOW = set('do', 'while', 'for', 'if', 'switch');
var NAME_OR_NUM = set('name', 'num');
var ASSOCIATIVE_BINARIES = set('+', '*', '|', '&', '^');
+var BREAK_CAPTURERS = set('do', 'while', 'for', 'switch');
+var CONTINUE_CAPTURERS = LOOP;
+
var NULL_NODE = ['name', 'null'];
var UNDEFINED_NODE = ['unary-prefix', 'void', ['num', 0]];
var TRUE_NODE = ['unary-prefix', '!', ['num', 0]];
@@ -1531,10 +1534,15 @@ function unVarify(vars, ret) { // transform var x=1, y=2 etc. into (x=1, y=2), i
var ASM_INT = 0;
var ASM_DOUBLE = 1;
-function detectAsmCoercion(node) {
+function detectAsmCoercion(node, asmInfo) {
// for params, +x vs x|0, for vars, 0.0 vs 0
if (node[0] === 'num' && node[1].toString().indexOf('.') >= 0) return ASM_DOUBLE;
- return node[0] === 'unary-prefix' ? ASM_DOUBLE : ASM_INT;
+ if (node[0] === 'unary-prefix') return ASM_DOUBLE;
+ if (asmInfo && node[0] == 'name') {
+ if (node[1] in asmInfo.vars) return asmInfo.vars[node[1]];
+ if (node[1] in asmInfo.params) return asmInfo.params[node[1]];
+ }
+ return ASM_INT;
}
function makeAsmParamCoercion(param, type) {
@@ -2974,17 +2982,31 @@ function outline(ast) {
for (var i = 0; i < stack.length; i++) {
asmData.stackPos[stack[i]] = i*8;
}
-
+ // Reserve an extra two spots: one for control flow var, the other for control flow data
+ asmData.stackSize = (stack.length + 2)*8;
+ asmData.controlStackPos = asmData.stackSize - 16;
+ asmData.controlDataStackPos = asmData.stackSize - 8;
asmData.splitCounter = 0;
}
// Analyze uses - reads and writes - of variables in part of the AST of a function
function analyzeCode(func, asmData, ast) {
+ var labels = {};
+ var labelCounter = 1; // 0 means no label
+
+ traverse(ast, function(node, type) {
+ if ((type == 'label' || type in LOOP_FLOW) && node[1] && !(node[1] in labels)) {
+ labels[node[1]] = labelCounter++;
+ }
+ });
+
var writes = {};
var appearances = {};
var hasReturn = false, hasBreak = false, hasContinue = false;
var breaks = {}; // set of labels we break or continue
var continues = {}; // to. '0' is an unlabeled one
+ var breakCapturers = 0;
+ var continueCapturers = 0;
traverse(ast, function(node, type) {
if (type == 'assign' && node[2][0] == 'name') {
@@ -3001,11 +3023,31 @@ function outline(ast) {
} else if (type == 'return') {
hasReturn = true;
} else if (type == 'break') {
- breaks[node[1] || 0] = 0;
+ var label = node[1] || 0;
+ if (!label && breakCapturers > 0) return; // no label, and captured
+ if (label && (label in labels)) return; // label, and defined in this code, so captured
+ breaks[label || 0] = 0;
hasBreak = true;
} else if (type == 'continue') {
- continues[node[1] || 0] = 0;
+ var label = node[1] || 0;
+ if (!label && continueCapturers > 0) return; // no label, and captured
+ if (label && (label in labels)) return; // label, and defined in this code, so captured
+ continues[label || 0] = 0;
hasContinue = true;
+ } else {
+ if (type in BREAK_CAPTURERS) {
+ breakCapturers++;
+ }
+ if (type in CONTINUE_CAPTURERS) {
+ continueCapturers++;
+ }
+ }
+ }, function(node, type) {
+ if (type in BREAK_CAPTURERS) {
+ breakCapturers--;
+ }
+ if (type in CONTINUE_CAPTURERS) {
+ continueCapturers--;
}
});
@@ -3015,9 +3057,26 @@ function outline(ast) {
if (appearances[name] > 0) reads[name] = 0;
}
- return { writes: writes, reads: reads, hasReturn: hasReturn, breaks: breaks, continues: continues };
+ return { writes: writes, reads: reads, hasReturn: hasReturn, breaks: breaks, continues: continues, labels: labels };
+ }
+
+ function makeAssign(dst, src) {
+ return ['assign', true, dst, src];
+ }
+ function makeStackAccess(type, pos) { // TODO: float64, not 32
+ return ['sub', ['name', type == ASM_INT ? 'HEAP32' : 'HEAPF32'], ['binary', '>>', ['binary', '+', ['name', 'sp'], ['num', pos]], ['num', '2']]];
+ }
+ function makeIf(cond, then, else_) {
+ var ret = ['if', cond, ['block', then]];
+ if (else_) ret.push(['block', else_]);
+ return ret;
+ }
+ function makeComparison(left, comp, right) {
+ return ['binary', comp, left, right];
}
+ var CONTROL_BREAK = 1, CONTROL_BREAK_LABEL = 2, CONTROL_CONTINUE = 3, CONTROL_CONTINUE_LABEL = 4, CONTROL_RETURN_VOID = 5, CONTROL_RETURN_INT = 6, CONTROL_RETURN_DOUBLE = 7;
+
var sizeToOutline = extraInfo.sizeToOutline;
var level = 0;
@@ -3025,7 +3084,7 @@ function outline(ast) {
printErr(' do outline ' + [func[1], level, 'range:', start, end, 'of', stats.length]);
var code = stats.slice(start, end+1);
var newIdent = func[1] + '$' + (asmData.splitCounter++);
- // add spills and reads before and after the call to the outlined code
+ // add spills and reads before and after the call to the outlined code, and in the outlined code itself
var codeInfo = analyzeCode(func, asmData, code);
var reps = [];
for (var v in codeInfo.reads) {
@@ -3039,21 +3098,109 @@ function outline(ast) {
reps.push(['stat', ['assign', true, ['name', v], ['sub', ['name', getAsmType(asmData, v) == ASM_INT ? 'HEAP32' : 'HEAPF32'], ['binary', '>>', ['binary', '+', ['name', 'sp'], ['num', asmData.stackPos[v]]], ['num', '2']]]]]);
code.push(['stat', ['assign', true, ['sub', ['name', getAsmType(asmData, v) == ASM_INT ? 'HEAP32' : 'HEAPF32'], ['binary', '>>', ['binary', '+', ['name', 'sp'], ['num', asmData.stackPos[v]]], ['num', '2']]], ['name', v]]]);
}
- stats.splice.apply(stats, [start, end-start+1].concat(reps));
// Generate new function
if (codeInfo.hasReturn || codeInfo.hasBreak || codeInfo.hasContinue) {
// we need to capture all control flow using a top-level labeled one-time loop in the outlined function
code = [['label', 'OL', ['do', ['num', 0], ['block', code]]]];
+ var breakCapturers = 0;
+ var continueCapturers = 0;
+ traverse(code, function(node, type) {
+ // replace all break/continue/returns with code to break out of the main one-time loop, and set the control data
+ if (type == 'return') {
+ var ret = ['break', 'OL'];
+ if (!node[1]) {
+ ret = ['seq', makeAssign(makeStackAccess(ASM_INT, asmData.controlStackPos), ['num', CONTROL_RETURN_VOID]), ret];
+ } else {
+ var type = detectAsmCoercion(node[1], asmData);
+ ret = ['seq', makeAssign(makeStackAccess(ASM_INT, asmData.controlStackPos), ['num', type == ASM_INT ? CONTROL_RETURN_INT : CONTROL_RETURN_DOUBLE]), ret];
+ ret = ['seq', makeAssign(makeStackAccess(type, asmData.controlDataStackPos), node[1]), ret];
+ }
+ return ret;
+ } else if (type == 'break') {
+ var label = node[1] || 0;
+ if (label == 'OL') return; // this was just added before us, it is new replacement code
+ if (!label && breakCapturers > 0) return; // no label, and captured
+ if (label && (label in codeInfo.labels)) return; // label, and defined in this code, so captured
+ var ret = ['break', 'OL'];
+ ret = ['seq', makeAssign(makeStackAccess(ASM_INT, asmData.controlStackPos), ['num', label ? CONTROL_BREAK_LABEL : CONTROL_BREAK]), ret];
+ if (label) {
+ assert(label in codeInfo.labels, label + ' in ' + keys(codeInfo.labels));
+ ret = ['seq', makeAssign(makeStackAccess(ASM_INT, asmData.controlDataStackPos), ['num', codeInfo.labels[label]]), ret];
+ }
+ return ret;
+ } else if (type == 'continue') {
+ var label = node[1] || 0;
+ if (!label && continueCapturers > 0) return; // no label, and captured
+ if (label && (label in codeInfo.labels)) return; // label, and defined in this code, so captured
+ var ret = ['break', 'OL'];
+ ret = ['seq', makeAssign(makeStackAccess(ASM_INT, asmData.controlStackPos), ['num', label ? CONTROL_CONTINUE_LABEL : CONTROL_CONTINUE]), ret];
+ if (label) {
+ ret = ['seq', makeAssign(makeStackAccess(ASM_INT, asmData.controlDataStackPos), ['num', codeInfo.labels[label]]), ret];
+ }
+ return ret;
+ } else {
+ if (type in BREAK_CAPTURERS) {
+ breakCapturers++;
+ }
+ if (type in CONTINUE_CAPTURERS) {
+ continueCapturers++;
+ }
+ }
+ }, function(node, type) {
+ if (type in BREAK_CAPTURERS) {
+ breakCapturers--;
+ }
+ if (type in CONTINUE_CAPTURERS) {
+ continueCapturers--;
+ }
+ });
+ // read the control data at the callsite to the outlined function
+ if (codeInfo.hasReturn) {
+ reps.push(makeIf(
+ makeComparison(makeStackAccess(ASM_INT, asmData.controlStackPos), '==', ['num', CONTROL_RETURN_VOID]),
+ [['stat', ['return']]]
+ ));
+ reps.push(makeIf(
+ makeComparison(makeStackAccess(ASM_INT, asmData.controlStackPos), '==', ['num', CONTROL_RETURN_INT]),
+ [['stat', ['return', makeStackAccess(ASM_INT, asmData.controlDataStackPos)]]]
+ ));
+ reps.push(makeIf(
+ makeComparison(makeStackAccess(ASM_INT, asmData.controlStackPos), '==', ['num', CONTROL_RETURN_DOUBLE]),
+ [['stat', ['return', makeStackAccess(ASM_DOUBLE, asmData.controlDataStackPos)]]]
+ ));
+ }
+ if (codeInfo.hasBreak) {
+ reps.push(makeIf(
+ makeComparison(makeStackAccess(ASM_INT, asmData.controlStackPos), '==', ['num', CONTROL_BREAK]),
+ ['stat', ['break']]
+ ));
+ reps.push(makeIf(
+ makeComparison(makeStackAccess(ASM_INT, asmData.controlStackPos), '==', ['num', CONTROL_BREAK_LABEL]),
+ ['stat', ['break', makeStackAccess(ASM_INT, asmData.controlDataStackPos)]] // XXX here and below, need a switch overall possible labels
+ ));
+ }
+ if (codeInfo.hasContinue) {
+ reps.push(makeIf(
+ makeComparison(makeStackAccess(ASM_INT, asmData.controlStackPos), '==', ['num', CONTROL_CONTINUE]),
+ ['stat', ['break']]
+ ));
+ reps.push(makeIf(
+ makeComparison(makeStackAccess(ASM_INT, asmData.controlStackPos), '==', ['num', CONTROL_CONTINUE_LABEL]),
+ ['stat', ['continue', makeStackAccess(ASM_INT, asmData.controlDataStackPos)]] // XXX
+ ));
+ }
}
var newFunc = ['defun', newIdent, ['sp'], code];
- var newAsmInfo = { params: { sp: ASM_INT }, vars: {} };
+ var newAsmData = { params: { sp: ASM_INT }, vars: {} };
for (var v in codeInfo.reads) {
- newAsmInfo.vars[v] = getAsmType(asmData, v);
+ newAsmData.vars[v] = getAsmType(asmData, v);
}
for (var v in codeInfo.writes) {
- newAsmInfo.vars[v] = getAsmType(asmData, v);
+ newAsmData.vars[v] = getAsmType(asmData, v);
}
- denormalizeAsm(newFunc, newAsmInfo);
+ denormalizeAsm(newFunc, newAsmData);
+ // replace in stats
+ stats.splice.apply(stats, [start, end-start+1].concat(reps));
return [newFunc];
}
@@ -3138,7 +3285,7 @@ function outline(ast) {
if (newFuncs.length > 0) {
// We have outlined. Add stack support: header in which we allocate enough stack space TODO
- // If sp was not present before, add it and before each return, pop the stack TODO
+ // If sp was not present before, add it and before each return, pop the stack. also a final pop if not ending with a return TODO
// (none of this should be done in inner functions, of course, just the original)
// add new functions to the toplevel, or create a toplevel if there isn't one
@@ -3175,12 +3322,13 @@ function fixDotZero(js) {
});
}
-function asmLoopOptimizer(ast) {
+function asmLastOpts(ast) {
traverseGeneratedFunctions(ast, function(fun) {
- // This is at the end of the pipeline, we can assume all other optimizations are done, and we modify loops
- // into shapes that might confuse other passes
traverse(fun, function(node, type) {
if (type === 'while' && node[1][0] === 'num' && node[1][1] === 1 && node[2][0] === 'block') {
+ // This is at the end of the pipeline, we can assume all other optimizations are done, and we modify loops
+ // into shapes that might confuse other passes
+
// while (1) { .. if (..) { break } } ==> do { .. } while(..)
var stats = node[2][1];
var last = stats[stats.length-1];
@@ -3208,6 +3356,11 @@ function asmLoopOptimizer(ast) {
node[1] = simplifyNotCompsDirect(['unary-prefix', '!', conditionToBreak]);
return node;
}
+ } else if (type == 'binary' && node[1] == '&' && node[3][0] == 'unary-prefix' && node[3][1] == '-' && node[3][2][0] == 'num' && node[3][2][1] == 1) {
+ // Change &-1 into |0, at this point the hint is no longer needed
+ node[1] = '|';
+ node[3] = node[3][2];
+ node[3][1] = 0;
}
});
});
@@ -3266,7 +3419,7 @@ arguments_.slice(1).forEach(function(arg) {
passes[arg](ast);
});
if (asm && last) {
- asmLoopOptimizer(ast); // TODO: move out of last, to make last faster when done later (as in side modules)
+ asmLastOpts(ast); // TODO: move out of last, to make last faster when done later (as in side modules)
prepDotZero(ast);
}
var js = astToSrc(ast, minifyWhitespace), old;
diff --git a/tools/shared.py b/tools/shared.py
index 46245fd1..0351a736 100644
--- a/tools/shared.py
+++ b/tools/shared.py
@@ -257,7 +257,7 @@ def check_clang_version():
def check_llvm_version():
try:
- check_clang_version();
+ check_clang_version()
except Exception, e:
logging.warning('Could not verify LLVM version: %s' % str(e))
@@ -1196,7 +1196,7 @@ set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)''' % { 'winfix': '' if not WINDOWS e
opts.append('-jump-threading')
opts.append('-correlated-propagation')
opts.append('-dse')
- #addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
+ #addExtensionsToPM(EP_ScalarOptimizerLate, MPM)
opts.append('-adce')
opts.append('-simplifycfg')
@@ -1300,7 +1300,7 @@ set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)''' % { 'winfix': '' if not WINDOWS e
'-O' + str(opt_level), '--closure', '0'], raw)
f = open(relooper, 'w')
f.write("// Relooper, (C) 2012 Alon Zakai, MIT license, https://github.com/kripken/Relooper\n")
- f.write("var Relooper = (function() {\n");
+ f.write("var Relooper = (function() {\n")
f.write(open(raw).read())
f.write('\n return Module.Relooper;\n')
f.write('})();\n')
diff --git a/tools/test-js-optimizer-asm-last-output.js b/tools/test-js-optimizer-asm-last-output.js
index 0f95d544..f850b18f 100644
--- a/tools/test-js-optimizer-asm-last-output.js
+++ b/tools/test-js-optimizer-asm-last-output.js
@@ -30,6 +30,7 @@ function finall(x) {
a = -999999984306749400.0;
a = -999999984306749400.0;
a = -0xde0b6b000000000;
+ f(g() | 0);
return 12.0e10;
}
function looop() {
diff --git a/tools/test-js-optimizer-asm-last.js b/tools/test-js-optimizer-asm-last.js
index 05e1049e..1d39b1a6 100644
--- a/tools/test-js-optimizer-asm-last.js
+++ b/tools/test-js-optimizer-asm-last.js
@@ -30,6 +30,7 @@ function finall(x) {
a = +-0xde0b6b000000000;
a = -+0xde0b6b000000000;
a = -0xde0b6b000000000;
+ f(g() & -1);
return +12e10;
}
function looop() {
generated by cgit v1.2.3-18-g5258 (git 2.32.0) at 2025-08-01 00:52:11 +0000