Merge branch 'incoming' into merge-mar-13-2015

Author: kripken

emcc

@@ -1504,6 +1504,11 @@ try:
         # add explicit label setting, as we will run aggressiveVariableElimination late, *after* 'label' is no longer notable by name
         js_optimizer_queue += ['safeLabelSetting']
 
+    if shared.Settings.EMTERPRETIFY:
+      # add explicit label setting, as we will run aggressiveVariableElimination late, *after* 'label' is no longer notable by name
+      js_optimizer_queue += ['safeLabelSetting']
+
+    if opt_level >= 2:
       if shared.Settings.RELOOP and not shared.Settings.ASM_JS:
         js_optimizer_queue += ['optimizeShiftsAggressive', get_eliminate()] # aggressive shifts optimization requires loops, it breaks on switches
 

tests/test_other.py

@@ -4588,7 +4588,7 @@ def do_log_test(source, expected, func):
       print '  seen', seen, ', expected ', expected, type(seen), type(expected)
       assert expected == seen or (seen in expected if type(expected) in [list, tuple] else False), ['expect', expected, 'but see', seen]
 
-    do_log_test(path_from_root('tests', 'primes.cpp'), 86, 'main')
+    do_log_test(path_from_root('tests', 'primes.cpp'), 87, 'main')
     do_log_test(path_from_root('tests', 'fannkuch.cpp'), 230, 'fannkuch_worker')
 
     # test non-native as well, registerizeHarder can be a little more efficient here

tools/emterpretify.py

@@ -221,7 +221,7 @@ def handle_arg(arg):
 
   'SWITCH',  # [lx, ly, lz]         switch (lx) { .. }. followed by a jump table for values in range [ly..ly+lz), after which is the default (which might be empty)
   'RET',     # [l, 0, 0]            return l (depending on which emterpreter_x we are in, has the right type)
-  'FUNC',    # [num params, total locals (low 8 bits), total locals (high 8 bits)] [which emterpreter (0 = normal, 1 = zero), 0, last zeroinit = num params + num zero-inits (low 8), (high 8)]           function with n locals (each taking 64 bits), of which the first are params
+  'FUNC',    # [num params, total locals (low 8 bits), total locals (high 8 bits)] [which emterpreter (0 = normal, 1 = zero), 0, 0, 0]           function with n locals (each taking 64 bits), of which the first are params
              # this is read in the emterpreter prelude, and also in intcalls
 
   # slow locals support - copying from/to slow locals
@@ -679,12 +679,6 @@ def process(code):
  assert(((HEAPU8[pc>>0]>>>0) == %d)|0);
  lx = HEAPU16[pc + 2 >> 1] | 0; // num locals
 %s
- ly = HEAPU8[pc + 1 >> 0] | 0; // first zeroinit (after params)
- lz = HEAPU16[pc + 6 >> 1] | 0; // offset of last zeroinit
- while ((ly | 0) < (lz | 0)) { // clear the zeroinits
-  %s = +0;
-  ly = ly + 1 | 0;
- }
 %s
  //print('enter func ' + [pc, HEAPU8[pc + 0],HEAPU8[pc + 1],HEAPU8[pc + 2],HEAPU8[pc + 3],HEAPU8[pc + 4],HEAPU8[pc + 5],HEAPU8[pc + 6],HEAPU8[pc + 7]].join(', '));
  //var first = true;
@@ -702,7 +696,6 @@ def process(code):
   ROPCODES['FUNC'],
   (''' EMTSTACKTOP = EMTSTACKTOP + (lx ''' + (' + 1 ' if ASYNC else '') + '''<< 3) | 0;
  assert(((EMTSTACKTOP|0) <= (EMT_STACK_MAX|0))|0);\n''' + (' if ((asyncState|0) != 2) {' if ASYNC else '')) if not zero else '',
-  get_access('ly', s='d'),
   ' } else { pc = (HEAP32[sp - 4 >> 2] | 0) - 8 | 0; }' if ASYNC else '',
   main_loop,
 ))

tools/js-optimizer.js

@@ -7610,23 +7610,9 @@ function emterpretify(ast) {
     }
     assert(numLocals < FAST_LOCALS, 'way too many params!');
     assert(FAST_LOCALS < 256);
-    var zeroInits;
-    if (numVars < FAST_LOCALS*2) {
-      zeroInits = findUninitializedVars(func, asmData);
-    } else {
-      zeroInits = copy(asmData.vars); // give up - tons of vars, this will be slow anyhow, and findUninitializedVars is non-linear so can be very slow on massive function
-    }
-    for (var zero in zeroInits) {
-      locals[zero] = numLocals++;
-      if (numLocals === FAST_LOCALS) numLocals = 256; // jump over the temps, remaining locals are slow locals
-    }
-    var lastZeroInit = func[2].length + numLocals; // may be higher than func[2].length + numZeroInits, due to jumping over temps
-                                                   // this means if we have very many zeroinits, we end up zeroing out the temps, oh well
     for (var i in asmData.vars) {
-      if (!(i in zeroInits)) {
-        locals[i] = numLocals++; // TODO: sort by frequency of appearance, so common ones are fast, rare are slow
-        if (numLocals === FAST_LOCALS) numLocals = 256; // jump over the temps, remaining locals are slow locals
-      }
+      locals[i] = numLocals++; // TODO: sort by frequency of appearance, so common ones are fast, rare are slow
+      if (numLocals === FAST_LOCALS) numLocals = 256; // jump over the temps, remaining locals are slow locals
     }
     var withSlowLocals = numLocals;
     numLocals = Math.min(numLocals, FAST_LOCALS); // ignore the slow locals
@@ -7650,7 +7636,7 @@ function emterpretify(ast) {
     // calculate final count of local variables, and emit func header
     var finalLocals = Math.max(numLocals, maxLocal+1, withSlowLocals);
     assert(finalLocals < 65535, 'too many locals ' + [maxLocal, numLocals, withSlowLocals]);
-    code = ['FUNC', func[2].length, finalLocals & 255, finalLocals >>> 8, 0, 0, lastZeroInit & 255, lastZeroInit >>> 8].concat(constants).concat(code);
+    code = ['FUNC', func[2].length, finalLocals & 255, finalLocals >>> 8, 0, 0, 0, 0].concat(constants).concat(code);
     verifyCode(code);
 
     finalizeJumps(code);