Subzero: Fix a regalloc bug involving too-aggressive AllowRegisterOverlap.

Subzero: Fix a regalloc bug involving too-aggressive AllowRegisterOverlap.

See the BUG description for more details.  In short, the register allocator
was inappropriately honoring AllowRegisterOverlap even when the variable's
live range overlaps with an Unhandled variable precolored to the preferred
register.

Also changes legalize() logic to recognize when a variable is guaranteed
to ultimately have a physical register due to infinite weight, and not
create a new temporary in those cases.

Finally, dumps RegisterPreference and AllowRegisterOverlap info for
Variables for improved diagnostics.

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3897
R=jvoung@chromium.org

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a=commit;h=ca662e9

[jvoung (off chromium), 2014-07-10 21:55:32]

LGTM

https://codereview.chromium.org/380363002/diff/1/src/IceRegAlloc.cpp
File src/IceRegAlloc.cpp (right):

https://codereview.chromium.org/380363002/diff/1/src/IceRegAlloc.cpp#newcode224
src/IceRegAlloc.cpp:224: << "(U=" << RegUses[i] << ",F=" << Free[i] << ") ";
Maybe print the PrecoloredUnhandled[] status too.

[jvoung (off chromium), 2014-07-10 21:56:43]

https://codereview.chromium.org/380363002/diff/1/src/IceRegAlloc.cpp
File src/IceRegAlloc.cpp (right):

https://codereview.chromium.org/380363002/diff/1/src/IceRegAlloc.cpp#newcode224
src/IceRegAlloc.cpp:224: << "(U=" << RegUses[i] << ",F=" << Free[i] << ") ";
On 2014/07/10 21:55:32, jvoung wrote:
> Maybe print the PrecoloredUnhandled[] status too.

Hmm, or I guess you can look at the Unhandled list, and spot the pre-colored
registers -- so it's fine to leave out too.

[Jim Stichnoth, 2014-07-10 22:32:10]

https://codereview.chromium.org/380363002/diff/1/src/IceRegAlloc.cpp
File src/IceRegAlloc.cpp (right):

https://codereview.chromium.org/380363002/diff/1/src/IceRegAlloc.cpp#newcode224
src/IceRegAlloc.cpp:224: << "(U=" << RegUses[i] << ",F=" << Free[i] << ") ";
On 2014/07/10 21:56:42, jvoung wrote:
> On 2014/07/10 21:55:32, jvoung wrote:
> > Maybe print the PrecoloredUnhandled[] status too.
> 
> Hmm, or I guess you can look at the Unhandled list, and spot the pre-colored
> registers -- so it's fine to leave out too.

Good idea, done.  You're right that you can study the Unhandled list, but then
you also have to be careful to look at only the Unhandled live ranges that
overlap with the current live range.

[Jim Stichnoth, 2014-07-10 22:32:40]

Committed patchset #2 manually as rca662e9 (presubmit successful).