1. Unconditional branch to the next basic block is removed.

2. For a conditional branch with a "false" edge to the next basic block, remove the unconditional branch to the fallthrough block.

3. For a conditional branch with a "true" edge to the next basic block, invert the condition and do like #2.

This is enabled only for O2, particularly because inverting the branch condition is a marginally risky operation.

This decreases the instruction count by about 5-6%.

Also, --stats prints a final tally to make it easier to post-process the output.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/580903005

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3892
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/576243002

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/559723003

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3892
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/576853002

This is needed since we are now using an absolute (and non-standard) path to clang++.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/567393007

The following are collected:
- Number of machine instructions emitted
- Number of registers saved/restored in prolog/epilog
- Number of stack frame bytes (non-alloca) allocated
- Number of "spills", or stores to stack slots
- Number of "fills", or loads/operations from stack slots
- Fill+Spill count (sum of above two)

These are somewhat reasonable approximations of code quality, and the primary intention is to compare before-and-after when trying out an optimization.

The statistics are dumped after translating each function.  Per-function and cumulative statistics are collected.  The output lines have a prefix that is easy to filter.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/580633002

In many cases, we expect a constant to be 32-bits or less.
This simplifies range checking for x86 memory operand displacements
(can only be 32-bit), or immediates in instructions (also 32-bit),
since we only store 32-bits (so it trivially fits in 32-bits).

Checks for whether a constant fits in 8-bits can be done on the 32-bit value
instead of the 64-bit value.

When TargetLowering sees a 64-bit immediate as an operand on a 64-bit
instruction, it should have split the 64-bit immediate into a 32-bit
loOperand(), and a 32-bit hiOperand(). So what's left for the Emit pass
should be 32-bit constants.

Other places which work with constants:
- intrinsic operands (the ABI only allows i32 params for atomic mem order,
  or atomic is lock free byte-size, or the longjmp param).
- addressing mode optimization (gep expansion should be working with
  i32 constants).
- insertelement, and extractelement constant indices (bitcode reader
  restricts the type of the index to be i32 also).

I guess now you may end up with multiple copies of what may be the
"same" constant (i64 0 vs i32 0).

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/569033002

Also fixes error messages on instruction operands, to print
out the operand (rather than pointer to it), since we can
now print out operands.

BUG= https://code.google.com/p/nativeclient/issues/detail?id=389
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/577703002

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/551723003

One wants to omit the --init option to make things go faster, but then there's the risk that the pexe or llvm2ice or llc has changed and --init is actually necessary.

This change automatically compares the modification timestamps of the pexe and the llvm2ice and llc binaries against the modification timestamps of the object files to determine whether re-translation is needed.

The --init option forces re-translation regardless of timestamps.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/551373008

A "standalone" version of dump() is provided, taking just an Ostream
argument and not requiring a Cfg or GlobalContext argument.

BUG= none
R=kschimpf@google.com

Review URL: https://codereview.chromium.org/570713006

LLVM objects/libraries are now built using clang from chrome. This
CL changes the compiler to chrome clang, and adds an appropriate
dynamic library path to the linked executable so that llvm2ice
can be run in any directory.

BUG=None
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/571973004

Also, refactor the key part of the address mode inference into separate functions, since it's getting unwieldy.

The main thing is that we mark phi temporaries as multi-definition, and disallow address mode inference transformations that involve such temporaries, because this is incorrect particular when there are backward branches involved.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/557953007

In x86-32, floating point values are returned to the caller on the top of the x87 floating point stack.  The caller is required to remove it from the x87 FP stack, e.g. via the fstp instruction.

This must be done even when the return value is not actually used anywhere else in the function, in which case O2 is likely to want to dead-code eliminate the fstp instruction.

We enforce this by adding a fake use of the fstp destination.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/563303003

ffs() vs findFirstSet() are slightly different, indexing is
0-based instead of 1-based.

Example mingw error:
http://build.chromium.org/p/tryserver.nacl/builders/nacl-toolchain-win7-pnacl-x86_64/builds/1920/steps/llvm_i686_w64_mingw32%20%28build%29/logs/stdio

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/563303002

This is a workaround for issue that Subzero currently assumes all
global addresses have a name, but finalized pexe files leave most
global addresses unnamed.

It does this by allowing two optional command-line flag name prefixes
that are used to generate names for unnamed global addresses.

BUG= None
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/567703003

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/567553003

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3892
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/568473002

BUG= none
R=jfb@chromium.org

Review URL: https://codereview.chromium.org/565553002

Also fixes minor issues with branches in instructions (i.e. defining
entry node and computing predecessors).

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3892
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/561823002

Fixes call to extractElement.

BUG= None
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/562783002

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3892
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/561883002

This addresses a TODO and fixes a code generation bug that was causing sprintf and _vfprintf_r to execute incorrectly.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/560773002

Ended up needing to fix the InstX8632Lockable error that JF ran into
now that -O0 is really -O0:
https://codereview.chromium.org/512933006/

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/557933002

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/560493002

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3894
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/545623005

The script translates a pexe using both Subzero and llc, and then uses linker tricks to resolve each symbol into either its Subzero or llc version.  This enables quick bisection-based debugging of Subzero code generation.

BUG= none
R=jfb@chromium.org, jvoung@chromium.org

Review URL: https://codereview.chromium.org/551953002

1. Modify dump() to match LLVM.

2. If it weren't for minimum stack alignment, the alignment code would be broken, so add a test in case the alignment code changes.

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

Review URL: https://codereview.chromium.org/557533003

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3894
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/548553002

This makes it much easier to run scripts from different working directories, as long as they are run somewhere under the native_client directory.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/554013002

One issue is that the test_arith cross test defined functions on i1 but never actually invoked them.

Another issue is that the lowering was using 8-bit registers for i1 values, but was being sloppy about leaving stuff in the upper 7 bits, and then using all 8 bits for tests.

This takes the approach of explicitly masking the result whenever it's possible for the result to exceed one bit, such as trunc, fptosi, fptoui.

Another possibility might be to allow the upper 7 bits to stay sloppy, and explicitly only test the lower bit.

Additionally, some "CHECK: ret" lines were removed, since they aren't actually needed after the change to use CHECK-LABEL, and they are affected by an llvm-dump bug (which is fixed in LLVM 3.6).

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/547033002

This is the truncating cvt instruction instead of rounding.

A few interesting floating point inputs are added to the cross tests.

Also, the cross test error output is modified to be more clear.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/550723002

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3894
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/545603003

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3894
R=jvoung@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/543793003

Also declare a few variables as Constant* instead of Operand* when they hold the result of Ctx->getConstantInt(), to be consistent with the rest of the code.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/541093002

BUG=https: //code.google.com/p/nativeclient/issues/detail?id=3894
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/531123002

There's already a hack that emits asm like:
  lea eax, myglobal
instead of:
  mov eax, [myglobal]
because of an llvm-mc parser bug.  However, the lea hack still doesn't work if the symbol is a reserved word, e.g.:
  lea eax, flags

The extra hack is to drop into AT&T syntax temporarily:
.att_syntax
  leal flags, %eax
.intel_syntax

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/543803002

The bug was first spotted in the optimized gl_Color4ub() from spec2k's mesa.  The lowering sequences for fptosi and fptoui with i8 or i16 include "mov T_2, T_1" where T_1 and T_2 may have different integer types, and the statement:
      T_2->setPreferredRegister(T_1, true);

If T_2's type is i8 and T_1 is assigned a register that has no 8-bit version, then T_2 gets an unsuitable register.

The fix is to honor RegisterOverlap only when RegMask allows.

It's hard to construct a good test for this, since it depends heavily on register allocation decisions, which will change over time.

BUG= none
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/544713002

Integers are generally dumped as signed instead of unsigned values.
Integers of i1 type are dumped as 'false' and 'true'.  Floating point
values still don't match LLVM.

BUG= none
R=kschimpf@google.com

Review URL: https://codereview.chromium.org/539743002

BUG= //code.google.com/p/nativeclient/issues/detail?id=3894
R=jvoung@chromium.org

Review URL: https://codereview.chromium.org/529113002