1. Decorate the list of live-in and live-out variables with register assignments in the dump() output.  This helps one to assess register pressure.

2. Fix a bug where the DisableInternal flag wasn't being honored for function definitions.

3. Add a -translate-only=<symbol> to limit translation to a single function or global variable.  This makes it easier to focus on debugging a single function.

4. Change the -no-phi-edge-split option to -phi-edge-split and invert the meaning, to better not avoid the non double negatives.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/673783002

Helps make it work with p2i instead of lc2i. This affected the
address mode optimizations, so some of the test expectations
have changed.

BUG=none
(happened to notice it while trying to test some things manually)

R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/671443003

The definition of ExternName4 in crosstest/test_global_main.cpp was
not defined as a pointer, but was in crosstest/test_global.cpp. As
a result, that name was not linked properly.

BUG=None
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/651673003

Fixes bug in the representation of relocation names (in either global
initializers or as constant expressions in code) so that they
understand when the name is externally defined. This allows us to test
this property using command line arguments, and fixes relocation tests
in cross compilations (where externnally referenced names shouldn't be
name mangled).

BUG=
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/667763002

Similar to https://codereview.chromium.org/656123003/, but cover some of
the assembler files which were avoided to avoid conflicts.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/643903006

We can't do direct calls via the .long sym hack, since that is
normally for an absolute relocation, but calls are expecting
relative relocations (except for reg/mem forms).

Nop-out the InstFake emitIAS methods. Remove the generic dispatcher
that redirects emitIAS() to emit(), since only branches and labels
are left.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/647193003

Force dest to be the full 32-bit reg instead of sometimes being
a 16-bit reg. This is to save on a operand size prefix (and
avoid passing the DestTy down to the dispatchers).

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/647223004

Currently, O2 calls VariablesMetadata::init() 4 times:

- Twice for liveness analysis, where only multi-block use information is needed for dealing with sparse bit vectors.

- Once for address mode inference, where single-definition information is needed.

- Once for register allocation, where all information is needed, including the set of all definitions which is needed for determining AllowOverlap.

So we limit the amount of data we gather based on the actual need.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/650613003

For consistency, put deleted ctors at the beginning of the class
definition.

If the default copy ctor or assignment operator is not deleted,
and the default implementation is used, leave it commented out to
indicate it is intentional.

Also, fixed one C++11 related TODO.

BUG= none
R=jvoung@chromium.org, kschimpf@google.com

Review URL: https://codereview.chromium.org/656123003

This removes the redundancy between live ranges stored in the Variable and those stored in Liveness, by removing the Liveness copy.  After liveness analysis, live ranges are constructed directly into the Variable.

Also, the LiveRangeWrapper is removed and Variable * is directly used instead.  The original thought behind LiveRangeWrapper was that it could be extended to include live range splitting.  However, when/if live range splitting is implemented, it will probably involve creating a new variable with its own live range, and carrying around some extra bookkeeping until the split is committed, so such a wrapper probably won't be needed.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/656023002

Give a different name to the crosstest .s and .o files depending on the
CPU features as well. That way the SSE2 and SSE4.1 .s and .o are separate.

The encodings for Pextrw and Pextrb/d... make me sad.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/656983002

This adds update counts to the output, e.g.:

Total across all functions - Flat times:
    0.262297 (13.0%): [ 1287] linearScan
    0.243965 (12.1%): [ 1287] emit
...

This is useful to know when some passes are called once per function and others are called several times per function.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/655563005

The key performance problem was that the per-block LiveBegin and LiveEnd vectors were dense with respect to the multi-block "global" variables, even though very few of the global variables are ever live within the block.  This led to large vectors needlessly initialized and iterated over.

The new approach is to accumulate two small vectors of <variable,instruction_number> tuples (LiveBegin and LiveEnd) as each block is processed, then sort the vectors and iterate over them in parallel to construct the live ranges.

Some of the anomalies in the original liveness analysis code have been straightened out:

1. Variables have an IgnoreLiveness attribute to suppress analysis.  This is currently used only on the esp register.

2. Instructions have a DestNonKillable attribute which causes the Dest variable not to be marked as starting a new live range at that instruction.  This is used when a variable is non-SSA and has more than one assignment within a block, but we want to treat it as a single live range.  This lets the variable have zero or one live range begins or ends within a block.  DestNonKillable is derived automatically for two-address instructions, and annotated manually in a few other cases.

This is tested by comparing the O2 asm output in each Spec2K component.  In theory, the output should be the same except for some differences in pseudo-instructions output as comments.  However, some actual differences showed up, related to the i64 shl instruction followed by trunc to i32.  This turned out to be a liveness bug that was accidentally fixed.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/652633002

1. Use a sorted std::vector instead of std::set to improve management
of the Unhandled sets.  This is the main performance gain.

2. Use std::list.splice() to move items between lists, instead of
erase()+push_back().  This doesn't really save much, but the intention
is somewhat clearer.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/642603005

Currently, this only checks and emits the segment
override only for GPR instructions, assuming it's mostly
only used for nacl.read.tp. The code will assert when used
in other situations.

The lea hack is still tested in some files, but it's not
emitted with emitIAS, and instead the "immediate" operand
now has a fixup.

There is a more compact encoding for "mov eax, moffs32",
etc., but that isn't used right now.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/649463002

Introduces the notion of a function address, to replace using LLVM
IR's Function class. Modifies Ice converter, and Subzero's bitcode
reader, to build function addresses.

BUG=None
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/641193002

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/650573002

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/634333002

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/640603002

Also changes the szbuild.py script to add -fdata-sections, and entirely removes the -disable-globals option.

The global initializer emission basically copies what llc does, based on 3 properties of the global: constant vs non-constant, internal vs external, and full zero-initializer vs non-trivial initializer.

BUG= none
R=jvoung@chromium.org, kschimpf@google.com

Review URL: https://codereview.chromium.org/631383003

Changes Subzero's bitcode reader to build and store ICE types,
instead of using LLVM's types. Note: This code doesn't remove
all uses of LLVM types. They are still used to check types for
instructions and to generate function addresses.

BUG=None
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/625243002

Makes sure the percentages represent only the function(s) focused on, and not with respect to the total translation time across all functions.

Reset the timings between functions so that --timing-focus=* gives reasonable numbers.

Also, adds a timer for the live range construction phase.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/640713003

Add the SZ runtime functions for unsigned conversion.
Add some more cast tests before doing emitIAS for cvt.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/639543002

Since push isn't used for args passing anymore, the cases of handling
push for vectors and floats/doubles isn't needed anymore. Passing
vectors requires a bit more care of alignment, so that was changed.

I can imagine push needing to handle addresses later (at least on
x86-64 to push the lower 32-bits of return address), but for now,
this means only handling GPRs. The XMM registers are not callee saved.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/633553003

The main optimization is for the repeated overlaps() calls against the Inactive set, by iteratively trimming away the early sections of the Inactive live ranges that can no longer overlap with Cur.

A more minor optimization doesn't bother checking pure point-valued Inactive ranges for expiring or reactivating.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/627203002

Modifies both LLVM to ICE converter, and Subzero's bitcode reader,
to build Subzero's global initializers. Modifies target lowering
routines for global initializers to use this new model.

Also modifies both to now handle relocations in global variable
initializers.

BUG=None
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/624663002

--timing-funcs - Produces a sorted list of total time spent translating each function.

--timing-focus=<F> - Turns on the --timing equivalent just for one function.  Use '*' to do this for all functions, i.e. get complete timing breakdowns across all functions.

--verbose-focus=<F> - Temporarily turns on --verbose=all for one function.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/620373004

The "test" instruction is used in very limited situations. I've made a best effort
to fill in the possible forms (address for the first operand), but it's not tested,
so I put the *untested* parts behind an assert. Otherwise it's very similar to
icmp, so if it starts to be used and tested then the asserts can be taken out,
and the code shared with icmp.

Tighten some of the XMM dispatch/emitters. Most of those XMM instructions
can only encode the variant where dest is a register. Rather than waste
a slot for a NULL method pointer, just make the struct type have two variants
instead of three.

Fill out a couple of XMM instructions which *do* allow mem-ops as dest
(mov instructions).

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/624263002

Call instruction lowering includes the FakeKill instruction, which creates several precolored variables, one for each scratch register.  The live range for each of these variables consists of a set of "point" ranges, one point for every FakeKill instruction.  The overlaps() logic is such that a point range never overlaps with an individual instruction, but it can overlap with a normal non-point range.

It turns out that during register allocation, usually most of the variables on the Inactive list are these FakeKill instructions.  The live range representation can be quite large if there are many calls in the function.  In the "Check for inactive ranges that have expired or reactivated" section, a lot of time was spent on overlapsStart() calls that were doomed to return false.

This change lets the live range keep track of whether it contains non-point segments, and if not, optimize the overlaps(InstNumberT) method.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/631483002

For the integer shift ops, since the Src1 operand is forced
to be an immediate or register (cl), it should be legal to
have Dest+Src0 be either register or memory. However, we
are currently only using the register form. It might be the
case that shift w/ Dest+Src0 as mem are less optimized
on some micro-architectures though, since it has to load,
shift, and store all in one operation, but I'm not sure.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/622113002

A lot of time was being spent in the two loops that check precolored ranges in the Unhandled set, specifically in the endsBefore() check.

Solve this by keeping a shadow copy of Unhandled, restricted to the ranges that are precolored.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/622553003

sed -i 's/LLVM_DELETED_FUNCTION/= delete/' src/*.{h,cpp}

BUG= https://codereview.chromium.org/512933006/
R=jfb@chromium.org

Review URL: https://codereview.chromium.org/619983002

Use C++11 'auto' where practical to make iteration more concise.
Use C++11 range-based for loops where possible.

BUG= none
R=jfb@chromium.org, kschimpf@google.com

Review URL: https://codereview.chromium.org/619893002

BUG= none
R=jvoung@chromium.org, kschimpf@google.com

Review URL: https://codereview.chromium.org/618313003

1. Setting command-line make variable NOASSERT=1 adds -DNDEBUG and builds in a separate directory.  By default, we still get Release+Asserts.

2. Add "(void)foo;" as necessary when foo is only used in an assert(), to remove warnings.

3. Minimize inclusion of llvm/Support/Timer.h because it adds warnings.

4. Call validateLiveness() only when asserts are enabled, because it's relatively expensive.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/623493002

While I'm at it, normalize the #include order:

1. C++ library headers
2. LLVM headers
3. Subzero headers

A blank line between each group.  Each group sorted alphabetically, case-insensitive.

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3930
R=jfb@chromium.org, jvoung@chromium.org

Review URL: https://codereview.chromium.org/622443002

This makes it much more useful for individual analysis and long-term translation performance tracking.

1. Collect and report aggregated across the entire translation, instead of function-by-function.  If you really care about a single function, just extract it and translate it separately for analysis.

2. Remove "-verbose time" and just use -timing.

3. Collects two kinds of timings: cumulative and flat.  Cumulative measures the total time, even if a callee also times itself.  Flat only measures the currently active timer at the top of the stack.  The flat times should add up to 100%, but cumulative will usually add up to much more than 100%.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/610813002

This makes it much easier to copy/paste the output.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/611983003

Be sure to legalize 8-bit imul immediates (there is only the r/m form).
Add a test for that, and cover a couple of other ops too...

There is a one-byte-shorter form when Dest/Src0 == EAX and Src1 is not
an immediate, but that isn't taken advantage of.

Go ahead and add the optimization for 8-bit immediates for i16/i32
(not allowed for i8). It shows up sometimes in spec, e.g., to multiply by 10.
There is a lot of multiply by 4 as well, that we could strength-reduce.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/617593002

For some arithmetic assembler methods, instead of checking
IceType_i8 || IceType_i1, only allow IceType_i8 and assert if
an i1 leaked to that stage (should have been vetted earlier
by the bitcode reader / ABI checks). Could have looked up the
type width and isIntegerArithmeticType, etc. in the property table,
but that seemed a bit heavy for just checking one type
(or one of two types).

Also changed some f32 || f64 checks into just using
isScalarFloatingType() which looks things up in a property table.
Could alternatively just keep it as an simple f32 || f64 check,
and I could change isScalarFloatingType()'s implementation.

In some places where we assume something is either i32 or i64
and do a select, change that into using a helper function
so that we can do one compare, and then assert. Some of the
asserts are really redundant (already within a branch which
already checked that), but hopefully that disappears if
we compile in release mode.
Similar for f32 or f64 (which happened a lot in the assembler).

BUG=none
R=kschimpf@google.com, stichnot@chromium.org

Review URL: https://codereview.chromium.org/613483002