SIMD plumbing

SIMD plumbing

Committed: https://code.google.com/p/dart/source/detail?r=20287

[Cutch, 2013-03-16 23:22:39]

Lays the ground work for supporting Float32x4 and Uint32x4 in optimizer. 

../tools/test.py -m all -a ia32,x64,simarm,simmips
Test configurations: none_vm_debug_ia32 none_vm_release_ia32 none_vm_debug_x64
none_vm_release_x64 none_vm_debug_simarm none_vm_release_simarm
none_vm_debug_simmips none_vm_release_simmips
[26:43 |  65% | +17612 | -    0]
FAILED: none-vm debug_x64 pkg/scheduled_test/test/scheduled_test_test
Expected: pass
Actual: timeout

stdout:
unittest-suite-wait-for-done


Command line: xcodebuild/DebugX64/dart "--ignore-unrecognized-flags"
"--package-root=xcodebuild/DebugX64/packages/"
"/Users/johnmccutchan/workspace/dart-repo/dart/pkg/scheduled_test/test/scheduled_test_test.dart"
[37:47 | 100% | +27020 | -    1]

[Vyacheslav Egorov (Google), 2013-03-17 18:17:10]

Some DBC.

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/flow_graph_comp...
File runtime/vm/flow_graph_compiler_ia32.cc (right):

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/flow_graph_comp...
runtime/vm/flow_graph_compiler_ia32.cc:1346: __ movups(Address(ESP, offset),
xmm_reg);
I wonder how much slower does this make runtime calls when there are live
doubles.

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/isolate.h
File runtime/vm/isolate.h (right):

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/isolate.h#newco...
runtime/vm/isolate.h:86: class DeferredFloat32x4 {
This is a lot of code duplication. 

Suggestion: move generic parts (e.g. next and slot_) into Base class, declare a
virtual function for materialization. 

class DeferredFloat32x4 : public DeferredSlot {
  public:
    virtual void Materialize() { /* Allocate and store into slot */ }
  private:
    const simd128_value_t value_
};

instead of 4 lists have a single list.

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/locations.h
File runtime/vm/locations.h (right):

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/locations.h#new...
runtime/vm/locations.h:62: kUint32x4StackSlot = 9,
It can't be 9 (see the comment above enum Kind).

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/locations.h#new...
runtime/vm/locations.h:274:
IndexField::encode(static_cast<uword>(kStackIndexBias + stack_index))
encoding of stack indexes repeats multiple times. consider introducing a helper.

[Cutch, 2013-03-17 21:52:37]

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/flow_graph_comp...
File runtime/vm/flow_graph_compiler_ia32.cc (right):

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/flow_graph_comp...
runtime/vm/flow_graph_compiler_ia32.cc:1346: __ movups(Address(ESP, offset),
xmm_reg);
On 2013/03/17 18:17:10, Vyacheslav Egorov (Google) wrote:
> I wonder how much slower does this make runtime calls when there are live
> doubles.

Can we add a tag in RegisterSet which indicates if an FPU register contains
64-bit or 128-bit data and issue the optimal move?

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/isolate.h
File runtime/vm/isolate.h (right):

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/isolate.h#newco...
runtime/vm/isolate.h:86: class DeferredFloat32x4 {
On 2013/03/17 18:17:10, Vyacheslav Egorov (Google) wrote:
> This is a lot of code duplication. 
> 
> Suggestion: move generic parts (e.g. next and slot_) into Base class, declare
a
> virtual function for materialization. 
> 
> class DeferredFloat32x4 : public DeferredSlot {
>   public:
>     virtual void Materialize() { /* Allocate and store into slot */ }
>   private:
>     const simd128_value_t value_
> };
> 
> instead of 4 lists have a single list.

Done.

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/locations.h
File runtime/vm/locations.h (right):

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/locations.h#new...
runtime/vm/locations.h:62: kUint32x4StackSlot = 9,
On 2013/03/17 18:17:10, Vyacheslav Egorov (Google) wrote:
> It can't be 9 (see the comment above enum Kind).

Done.

https://codereview.chromium.org/12871015/diff/2001/runtime/vm/locations.h#new...
runtime/vm/locations.h:274:
IndexField::encode(static_cast<uword>(kStackIndexBias + stack_index))
On 2013/03/17 18:17:10, Vyacheslav Egorov (Google) wrote:
> encoding of stack indexes repeats multiple times. consider introducing a
helper.

Done.

[Cutch, 2013-03-17 23:35:34]

$ ./tools/test.py -a ia32,x64,simarm,simmips -m all
Test configurations: none_vm_debug_ia32 none_vm_release_ia32 none_vm_debug_x64
none_vm_release_x64 none_vm_debug_simarm none_vm_release_simarm
none_vm_debug_simmips none_vm_release_simmips
[00:00 |  --% | +    0 | -    0]
[26:58 |  65% | +17612 | -    0]
FAILED: none-vm debug_x64 pkg/scheduled_test/test/scheduled_test_test
Expected: pass
Actual: timeout

stdout:
unittest-suite-wait-for-done


Command line: xcodebuild/DebugX64/dart "--ignore-unrecognized-flags"
"--package-root=xcodebuild/DebugX64/packages/"
"/Users/johnmccutchan/workspace/dart-repo/dart/pkg/scheduled_test/test/scheduled_test_test.dart"

[37:47 | 100% | +27020 | -    1]

[srdjan, 2013-03-19 20:48:45]

Some initial comments before moving to the new CL.

https://codereview.chromium.org/12871015/diff/7001/runtime/vm/deopt_instructi...
File runtime/vm/deopt_instructions.h (right):

https://codereview.chromium.org/12871015/diff/7001/runtime/vm/deopt_instructi...
runtime/vm/deopt_instructions.h:56: return
*reinterpret_cast<double*>(&fpu_registers_copy_[reg]);
Is this machine specific (where the double is stored with the 128 bit word) or
at last little/big endian specific?

https://codereview.chromium.org/12871015/diff/10001/runtime/vm/isolate.h
File runtime/vm/isolate.h (right):

https://codereview.chromium.org/12871015/diff/10001/runtime/vm/isolate.h#newc...
runtime/vm/isolate.h:60: virtual void materialize() = 0;
s/materialize/Materialize/

[Cutch, 2013-03-19 21:48:33]

https://codereview.chromium.org/12871015/diff/7001/runtime/vm/deopt_instructi...
File runtime/vm/deopt_instructions.h (right):

https://codereview.chromium.org/12871015/diff/7001/runtime/vm/deopt_instructi...
runtime/vm/deopt_instructions.h:56: return
*reinterpret_cast<double*>(&fpu_registers_copy_[reg]);
On 2013/03/19 20:48:45, srdjan wrote:
> Is this machine specific (where the double is stored with the 128 bit word) or
> at last little/big endian specific?

This is machine specific where a FPU register is wider than a double.  On
platforms where the floating point register is the width of a double, there
should be no change. There could be an endian issue if the double is at the top
of a 128-bit register.

https://codereview.chromium.org/12871015/diff/10001/runtime/vm/isolate.h
File runtime/vm/isolate.h (right):

https://codereview.chromium.org/12871015/diff/10001/runtime/vm/isolate.h#newc...
runtime/vm/isolate.h:60: virtual void materialize() = 0;
On 2013/03/19 20:48:45, srdjan wrote:
> s/materialize/Materialize/

Done.

[srdjan, 2013-03-19 21:58:25]

lgtm

[Cutch, 2013-03-20 19:48:47]

Test suite from latest:

../tools/test.py -m all -a ia32,x64,simmips,simarm
Test configurations: none_vm_debug_ia32 none_vm_release_ia32 none_vm_debug_x64
none_vm_release_x64 none_vm_debug_simmips none_vm_release_simmips
none_vm_debug_simarm none_vm_release_simarm
[37:53 | 100% | +27113 | -    0]

[Cutch, 2013-03-20 20:29:14]

Committed patchset #5 manually as r20287 (presubmit successful).