Test behavior of qNaN and sNaN

Test behavior of qNaN and sNaN


BUG=v8:2607
Committed: https://code.google.com/p/v8/source/detail?r=14133

[Dmitry Lomov (no reviews), 2013-04-02 09:31:28]

Add test that makes sure that s/qNaNs do not crash V8

[danno, 2013-04-02 10:31:04]

So, why don't we crash? Unless we mask the signal, I would assume we would.

Also, please add variants that test the runtime (first time called), ICs (2+
times), and Crankshaft (force the code to be optimized with
%OptimizeFunctionOnNextCall

[Dmitry Lomov (no reviews), 2013-04-02 11:12:08]

We do not crash because we somehow mask it:

Breakpoint 2, v8::Shell::Main (argc=2, argv=0xffffd104) at ../src/d8.cc:1937
1937	    if (options.stress_opt || options.stress_deopt) {
(gdb) p $mxcsr
$1 = [ IM DM ZM OM UM PM ]
(gdb) 

Good point about all other execution regimes, will add tests

On 2013/04/02 10:31:04, danno wrote:
> So, why don't we crash? Unless we mask the signal, I would assume we would.
> 
> Also, please add variants that test the runtime (first time called), ICs (2+
> times), and Crankshaft (force the code to be optimized with
> %OptimizeFunctionOnNextCall

[Dmitry Lomov (no reviews), 2013-04-02 11:21:03]

Indeed, the following prints "true\false" (only for sNaN):

var bytes = new Uint32Array([1, 0x7FF00000]);
var doubles = new Float64Array(bytes.buffer);
 
function f() {
  print(isNaN(doubles[0]));
}

f(); f()


On 2013/04/02 11:12:08, Dmitry Lomov (chromium) wrote:
> We do not crash because we somehow mask it:
> 
> Breakpoint 2, v8::Shell::Main (argc=2, argv=0xffffd104) at ../src/d8.cc:1937
> 1937	    if (options.stress_opt || options.stress_deopt) {
> (gdb) p $mxcsr
> $1 = [ IM DM ZM OM UM PM ]
> (gdb) 
> 
> Good point about all other execution regimes, will add tests
> 
> On 2013/04/02 10:31:04, danno wrote:
> > So, why don't we crash? Unless we mask the signal, I would assume we would.
> > 
> > Also, please add variants that test the runtime (first time called), ICs (2+
> > times), and Crankshaft (force the code to be optimized with
> > %OptimizeFunctionOnNextCall

[Dmitry Lomov (no reviews), 2013-04-02 11:24:09]

(tracked by https://code.google.com/p/v8/issues/detail?id=2607)
On 2013/04/02 11:21:03, Dmitry Lomov (chromium) wrote:
> Indeed, the following prints "true\false" (only for sNaN):
> 
> var bytes = new Uint32Array([1, 0x7FF00000]);
> var doubles = new Float64Array(bytes.buffer);
>  
> function f() {
>   print(isNaN(doubles[0]));
> }
> 
> f(); f()
> 
> 
> On 2013/04/02 11:12:08, Dmitry Lomov (chromium) wrote:
> > We do not crash because we somehow mask it:
> > 
> > Breakpoint 2, v8::Shell::Main (argc=2, argv=0xffffd104) at ../src/d8.cc:1937
> > 1937	    if (options.stress_opt || options.stress_deopt) {
> > (gdb) p $mxcsr
> > $1 = [ IM DM ZM OM UM PM ]
> > (gdb) 
> > 
> > Good point about all other execution regimes, will add tests
> > 
> > On 2013/04/02 10:31:04, danno wrote:
> > > So, why don't we crash? Unless we mask the signal, I would assume we
would.
> > > 
> > > Also, please add variants that test the runtime (first time called), ICs
(2+
> > > times), and Crankshaft (force the code to be optimized with
> > > %OptimizeFunctionOnNextCall

[Dmitry Lomov (no reviews), 2013-04-02 11:45:07]

Added tests for all modes, the test that uncovered a bug is commented out for
now. Will be interesting to run this on all architecures

[Dmitry Lomov (no reviews), 2013-04-03 14:38:20]

Fix 2607 by removing special handling of identical heap numbers in generic
compare IC.

[danno, 2013-04-04 09:19:43]

lgtm

[Dmitry Lomov (no reviews), 2013-04-04 09:31:34]

Committed patchset #5 manually as r14133 (presubmit successful).