Convert runtime atomics functions to inline asm

Convert runtime atomics functions to inline asm

This is the first step to optimizing Atomics accesses in TF.

The instruction sequences must be identical for all compilers. Since
runtime functions are shared by all compilers by default, it's a
convenient way to share the instruction sequences as well.

This CL only implements the inline assembly for ia32 and x64 on
gcc/clang and for ia32 on MSVC. MSVC does not support inline assembly
for x64, so in that case we use a separate assembly source file.

For all other architectures, we use the gcc/clang intrinsics for now.

BUG=v8:4614

[binji, 2016-01-09 02:50:51]

Description was changed from

==========
WIP inline assembly

BUG=
==========

to

==========
Convert runtime atomics functions to inline asm

This is the first step to optimizing Atomics accesses in TF.

The instruction sequences must be identical for all compilers. Since
runtime functions are shared by all compilers by default, it's a
convenient way to share the instruction sequences as well.

This CL only implements the inline assembly for ia32 and x64 on
gcc/clang and for ia32 on MSVC. MSVC does not support inline assembly
for x64, so in that case we use a separate assembly source file.

For all other architectures, we use the gcc/clang intrinsics for now.

BUG=https://bugs.chromium.org/p/v8/issues/detail?id=4614
==========

[binji, 2016-01-09 02:51:45]

Description was changed from

==========
Convert runtime atomics functions to inline asm

This is the first step to optimizing Atomics accesses in TF.

The instruction sequences must be identical for all compilers. Since
runtime functions are shared by all compilers by default, it's a
convenient way to share the instruction sequences as well.

This CL only implements the inline assembly for ia32 and x64 on
gcc/clang and for ia32 on MSVC. MSVC does not support inline assembly
for x64, so in that case we use a separate assembly source file.

For all other architectures, we use the gcc/clang intrinsics for now.

BUG=https://bugs.chromium.org/p/v8/issues/detail?id=4614
==========

to

==========
Convert runtime atomics functions to inline asm

This is the first step to optimizing Atomics accesses in TF.

The instruction sequences must be identical for all compilers. Since
runtime functions are shared by all compilers by default, it's a
convenient way to share the instruction sequences as well.

This CL only implements the inline assembly for ia32 and x64 on
gcc/clang and for ia32 on MSVC. MSVC does not support inline assembly
for x64, so in that case we use a separate assembly source file.

For all other architectures, we use the gcc/clang intrinsics for now.

BUG=v8:4614
==========

[binji, 2016-01-09 02:52:52]

binji@chromium.org changed reviewers:
+ bmeurer@chromium.org, jarin@chromium.org

[binji, 2016-01-09 02:53:05]

PTAL

[binji, 2016-01-12 21:14:15]

On 2016/01/09 02:53:05, binji wrote:
> PTAL

ping