Subzero: Improve effectiveness of local register availability peephole.

Subzero: Improve effectiveness of local register availability peephole.

X86 only.  The register availability peephole optimization during lowering disallows available register substitution when the variable is pre-colored.  This is for good reasons (too complex to be discussed here).  However, that leaves some potential substitutions on the table.

Specifically, this happens a lot around register arguments to function calls, both at the call site and in the prolog.

The simplest solution seems to be to launder the pre-colored variable through a separate infinite-weight variable, as implemented in this CL through a combination of such copies and extra legalize() calls.

There are other situations where this technique can also work, which may be handled in a separate CL.

This CL also fixes a problem where the stack pointer adjustment in the prolog is subject to dead-code elimination if the function has no epilog.  This would only happen in asm-verbose mode, in the final liveness analysis pass prior to code emission.

BUG= none
R=eholk@chromium.org

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a=commit;h=f531931fb45172157bc5816a8f55309556996dc7

[Jim Stichnoth, 2016-06-09 20:49:08]

Description was changed from

==========
Subzero: Improve effectiveness of local register availability peephole.

BUG= none
==========

to

==========
Subzero: Improve effectiveness of local register availability peephole.

X86 only.  The register availability peephole optimization during lowering
disallows available register substitution when the variable is pre-colored. 
This is for good reasons (too complex to be discussed here).  However, that
leaves some potential substitutions on the table.

Specifically, this happens a lot around register arguments to function calls,
both at the call site and in the prolog.

The simplest solution seems to be to launder the pre-colored variable through a
separate infinite-weight variable, as implemented in this CL through a
combination of such copies and extra legalize() calls.

There are other situations where this technique can also work, which may be
handled in a separate CL.

This CL also fixes a problem where the stack pointer adjustment in the prolog is
subject to dead-code elimination if the function has no epilog.  This would only
happen in asm-verbose mode, in the final liveness analysis pass prior to code
emission.

BUG= none
==========

[Jim Stichnoth, 2016-06-09 20:50:02]

stichnot@chromium.org changed reviewers:
+ eholk@chromium.org, jpp@chromium.org, kschimpf@google.com, manasijm@google.com

[Jim Stichnoth, 2016-06-09 20:50:37]

This seems to give about 0.6% geomean improvement on spec2k/x86-64.

[Eric Holk, 2016-06-09 23:12:08]

lgtm

[Jim Stichnoth, 2016-06-10 19:21:21]

Description was changed from

==========
Subzero: Improve effectiveness of local register availability peephole.

X86 only.  The register availability peephole optimization during lowering
disallows available register substitution when the variable is pre-colored. 
This is for good reasons (too complex to be discussed here).  However, that
leaves some potential substitutions on the table.

Specifically, this happens a lot around register arguments to function calls,
both at the call site and in the prolog.

The simplest solution seems to be to launder the pre-colored variable through a
separate infinite-weight variable, as implemented in this CL through a
combination of such copies and extra legalize() calls.

There are other situations where this technique can also work, which may be
handled in a separate CL.

This CL also fixes a problem where the stack pointer adjustment in the prolog is
subject to dead-code elimination if the function has no epilog.  This would only
happen in asm-verbose mode, in the final liveness analysis pass prior to code
emission.

BUG= none
==========

to

==========
Subzero: Improve effectiveness of local register availability peephole.

X86 only.  The register availability peephole optimization during lowering
disallows available register substitution when the variable is pre-colored. 
This is for good reasons (too complex to be discussed here).  However, that
leaves some potential substitutions on the table.

Specifically, this happens a lot around register arguments to function calls,
both at the call site and in the prolog.

The simplest solution seems to be to launder the pre-colored variable through a
separate infinite-weight variable, as implemented in this CL through a
combination of such copies and extra legalize() calls.

There are other situations where this technique can also work, which may be
handled in a separate CL.

This CL also fixes a problem where the stack pointer adjustment in the prolog is
subject to dead-code elimination if the function has no epilog.  This would only
happen in asm-verbose mode, in the final liveness analysis pass prior to code
emission.

BUG= none
R=eholk@chromium.org

Committed:
https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a...
==========

[Jim Stichnoth, 2016-06-10 19:21:22]

Committed patchset #3 (id:40001) manually as
f531931fb45172157bc5816a8f55309556996dc7 (presubmit successful).