Reimplement dynamic frame alignment for frames that are compiled via OSR or have more than 2 double…

Reimplement dynamic frame alignment for frames that are compiled via OSR or have more than 2 double spill slots.

The first spill slot is now reserved on all optimized frames to distinguish frames that were aligned.

Committed: https://code.google.com/p/v8/source/detail?r=11772

[Vyacheslav Egorov (Google), 2012-06-08 14:09:34]

I have re-implemented dynamic stack alignment in the correct way while
researching some performance flakiness on the older hardware.

Uploading this CL so that this work will not be lost.

Leaving it for later to decide whether we want this or not.

[Massi, 2012-06-11 10:03:32]

http://codereview.chromium.org/10532066/diff/1/src/ia32/deoptimizer-ia32.cc
File src/ia32/deoptimizer-ia32.cc (right):

http://codereview.chromium.org/10532066/diff/1/src/ia32/deoptimizer-ia32.cc#n...
src/ia32/deoptimizer-ia32.cc:418: // All OSR stack frames are dynamically
aligned to an 8-byte boundary.
Why don't we need to (eventually) move the incoming arguments like we do in the
prologue in the non-OSR case?

[danno, 2012-06-11 10:53:14]

LGTM with comments.

http://codereview.chromium.org/10532066/diff/4001/src/ia32/lithium-codegen-ia...
File src/ia32/lithium-codegen-ia32.cc (right):

http://codereview.chromium.org/10532066/diff/4001/src/ia32/lithium-codegen-ia...
src/ia32/lithium-codegen-ia32.cc:215: __ mov(Operand(ebp,
JavaScriptFrameConstants::kLocal0Offset), edx);
To reduce the theoretical expense of the tag in the non-dynamic_frame_alignment_
case, can you do the immediate store in a single instruction?

http://codereview.chromium.org/10532066/diff/4001/src/ia32/lithium-codegen-ia...
src/ia32/lithium-codegen-ia32.cc:2133: __ mov(edx, Operand(ebp,
JavaScriptFrameConstants::kLocal0Offset));
Can you alias something like kFrameAlignmentCookieOffset to kLocalOffset and use
it instead?

[Vyacheslav Egorov (Google), 2012-06-11 11:02:43]

http://codereview.chromium.org/10532066/diff/1/src/ia32/deoptimizer-ia32.cc
File src/ia32/deoptimizer-ia32.cc (right):

http://codereview.chromium.org/10532066/diff/1/src/ia32/deoptimizer-ia32.cc#n...
src/ia32/deoptimizer-ia32.cc:418: // All OSR stack frames are dynamically
aligned to an 8-byte boundary.
On 2012/06/11 10:03:32, Massi wrote:
> Why don't we need to (eventually) move the incoming arguments like we do in
the
> prologue in the non-OSR case?

Incomming arguments are part of the frame translation so they will get moved
automatically by the translation itself.

See line 376 above.

[Vyacheslav Egorov (Google), 2012-06-11 14:10:00]

PTAL

I fixed problems with stepping in and stepping out in the debugger by adding
UnpaddedFP accessor to the frame and using that in debugger.

I also added alignment checks.

Additionally I altered alignment and numbering scheme for double spill slots so
that oddly numbered slots are aligned while evenly numbered are misaligned.
Previously spill-slot area itself was misaligned, now it is aligned.

[danno, 2012-06-12 08:52:05]

Still LGTM