This patch implements lowerLoad and extends existing lowerStore for byte, short and floating-point types.
The patch also modifies PostLoweringLegalizer for conversion of mov to load or store.

R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2301303003 .

Patch from Jaydeep Patil <jaydeep.patil@imgtec.com>.

Visual Studio 2015 does not support the use of typename outside of a
template definition. Instead alias the flag's storage type. This
even avoids some duplication.

BUG=swiftshader:7

Change-Id: I787701f6bfe8e36169e6ac5d63fcb6a1a114bdb0

There are two problems to address:

1. Google developers lack some tooling to test MIPS changes.

2. MIPS developers lack some tooling/packages to allow full "make presubmit" testing.

Until all this gets sorted out, we add the "make -f Makefile.standalone MIPS=1" option to control some of the testing targets:

check-lit: No changes, as all these lit tests should be runnable in both environments.

check-xtest: MIPS=1 runs just the mips32 tests, plus the x8664 native tests as a sanity-check.  Non-MIPS runs everything except the mips32 tests.

check-spec: MIPS=1 completely disables spec2k testing.

presubmit: MIPS=1 greatly reduces the number of tests (especially since so many are spec variants).

With this change, mips32 CLs should pass "make -f Makefile.standalone MIPS=1 presubmit" before upload, and other CLs should pass "make -f Makefile.standalone presubmit".

BUG= none
R=kschimpf@google.com, obucinac@gmail.com

Review URL: https://codereview.chromium.org/2271053006 .

BUG=swiftshader:7

Change-Id: I1c40c10b3c3d032d3f7e0f8634c7c9abcc21d130

Currently, TargetLowering::sortVarsByAlignment() uses the variable's required alignment as the primary key, but then std::sort() breaks ties arbitrarily.  This can give different results across different STL implementations, for example when building against LLVM trunk versus building in the PNaCl environment.

The fix is to use Variable::Number as the secondary key.

BUG= none
R=kschimpf@google.com

Review URL: https://codereview.chromium.org/2295393002 .

LLVM commit 1d79fff9e65e77f84bf80c2cf4f0155bd167c90d changed the way ilist::reverse_iterator is converted to ilist::iterator.

See https://github.com/llvm-mirror/llvm/commit/1d79fff9e65e77f84bf80c2cf4f0155bd167c90d

BUG= none
R=kschimpf@google.com

Review URL: https://codereview.chromium.org/2297363002 .

The patch implements lowerSwitch for i32 and i64 types.

R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2289043002 .

Patch from Jaydeep Patil <jaydeep.patil@imgtec.com>.

The original Ice::Inst needed a bit of template specialization so that llvm::ilist<Ice::Inst> methods wouldn't try to invoke private/deleted ctor/dtor methods.  The Subzero code copied a pattern fairly widely used in LLVM.

Recently, LLVM improved ilist<> so that this specialization is no longer needed.

BUG= none
R=jpp@chromium.org

Review URL: https://codereview.chromium.org/2287363002 .

This patch adds encoding of few arithmetic and branch instructions.

R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2259983004 .

Patch from Jaydeep Patil <jaydeep.patil@imgtec.com>.

R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2275883002 .

Patch from Jaydeep Patil <jaydeep.patil@imgtec.com>.

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2266593002 .

Patch from Thomas Lively <tlively@google.com>.

R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2257043003 .

Patch from Jaydeep Patil <jaydeep.patil@imgtec.com>.

Both the operands in InstMIPS32Memory are source. However in load instructions, first operand is a destination. The InstMIPS32Load treats first operand as destination and second operand as source.

R=jpp@chromium.org, stichnot@chromium.org

Review URL: https://codereview.chromium.org/2250203005 .

Patch from Jaydeep Patil <jaydeep.patil@imgtec.com>.

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2256673004 .

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=kschimpf@google.com, stichnot@chromium.org

Review URL: https://codereview.chromium.org/2256903003 .

Instead of the version string being like this:
  Subzero_revision_efdf4120
Make it more like this:
  Subzero_revision_1089_efdf4120
I.e., provide both git hash and commit count.

That way, it's easier to compare revision strings and determine which one is newer.

BUG= none
R=eholk@chromium.org

Review URL: https://codereview.chromium.org/2251153002 .

1. Some explicit conversions between Inst* and InstList::iterator are needed.

2. llvm::sys::PrintStackTraceOnErrorSignal() has a new argument.

3. A previous CL neglected to plumb in SUBZERO_REVISION for the cmake build.

BUG= none
R=eholk@chromium.org

Review URL: https://codereview.chromium.org/2247253005 .

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=kschimpf@google.com, stichnot@chromium.org

Review URL: https://codereview.chromium.org/2241383006 .

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=kschimpf@google.com, stichnot@chromium.org

Review URL: https://codereview.chromium.org/2242243003 .

R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2223783002 .

Patch from Mohit Bhakkad <mohit.bhakkad@imgtec.com>.

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=kschimpf@google.com

Review URL: https://codereview.chromium.org/2235023002 .

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=kschimpf@google.com

Review URL: https://codereview.chromium.org/2227353002 .

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=kschimpf@google.com, stichnot@chromium.org

Review URL: https://codereview.chromium.org/2194853003 .

BUG= none

Review URL: https://codereview.chromium.org/2215623002 .

Modify the Makefiles to pass in the current git hash, which is embedded into the translated output.  As a side effect, it is also embedded into the Subzero translator binary.  This is useful for two reasons:

1. The PNaCl component update process is somewhat manual, making it tricky long after the fact to know exactly which revision was pushed, e.g. when trying to reproduce a bug or crash.

2. A translated binary can be inspected to make sure Chrome used the expected revision of Subzero.  (And also to verify that pnacl-sz was used rather than pnacl-llc.)

The revision string is suppressed for lit tests, because a number of tests seem overly strict about global initializer expectations.

BUG= none
R=jpp@chromium.org

Review URL: https://codereview.chromium.org/2218363002 .

The opt level (O2 versus Om1) should be tested using Cfg::getOptLevel() instead of getFlags().getOptLevel() whenever possible.

This is because if you run "-Om1 -force-O2=foo", and you're compiling foo, the first form tells you O2 while the second form tells you Om1.

BUG= none
R=eholk@chromium.org

Review URL: https://codereview.chromium.org/2210773002 .

Memset lowering was using the memcpy inline threshold instead of the memset threshold.

Using the memset threshold as specified (16) seems to make spec2k performance slightly worse, so change it to the original value (8).

BUG= none
R=eholk@chromium.org

Review URL: https://codereview.chromium.org/2217983003 .

LCSE is local common sub-expression elimination.
LICM is loop invariant code motion.
Short circuit splits basic blocks and introduces early jumps.
Global Splitting is a post regalloc live range splitting pass.

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2217773003 .

Convert adds with a constant operand to lea on -aggressive-lea

BUG=none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2135403002 .

Load multiple uses of a floating point constant (between two call
instructions or block start/end) into a variable before its first use.
  t1 = b + 1.0
  t2 = c + 1.0
Gets transformed to:
  t0 = 1.0
  t0_1 = t0
  t1 = b + t0_1
  t2 = c + t0_1
Call instructions reset the procedure, but uses the same variable, just
in case it got a register. We are assuming floating point registers are
not calee saved in general. Example, continuing from before:
  result = call <some function>
  t3 = d + 1.0
Gets transformed to:
  result = call <some function>
  t0_2 = t0
  t3 = d + t0_2

BUG= none
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2208523002 .

After register allocation is done once, this pass targets
the variables that do not get registers, break them into
multiple variables with shorter (at most spanning a basic
block) live ranges. After discarding the new variables with
too few uses, the register allocator is run again and
the new variables that manage to get registers are inserted.

BUG=None
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2172313002 .

Added load/store and stack/heap/global information.

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=kschimpf@google.com

Review URL: https://codereview.chromium.org/2211733002 .

For certain cases of inlined memset lowering, the 8-bit value wasn't being properly spread/replicated into the 32-bit immediate to be stored.

Specifically, if the 8-bit value is between -128 and -1 (i.e. 0x80 to 0xff), the spread value would be something like 0xffffff80 instead of 0x80808080.

BUG= b/30502279
R=jpp@chromium.org

Review URL: https://codereview.chromium.org/2215553002 .

Previous CL forgot to "make format".

BUG= none

Review URL: https://codereview.chromium.org/2206743003 .

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=kschimpf@google.com

Review URL: https://codereview.chromium.org/2209563002 .

R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2197233002 .

Patch from Mohit Bhakkad <mohit.bhakkad@imgtec.com>.

Reduce the default number of iterations to 1
Put the optional code behind the -lcse-no-ssa flag, which is disabled by
default. This brings down the overhead of enabling this to about 2%.

BUG=
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2185193002 .

The linear-scan register allocator takes an all-or-nothing approach -- either the variable's entire live range gets a register, or none of it does.

To help with this, we add a pass that splits successive uses of a variable within a basic block into a chain of linked variables.  This gives the register allocator the chance to allocate registers to subsets of the original live range.

The split variables are linked to each other so that if they don't get a register, they share a stack slot with the original variable, and redundant writes to that stack slot are recognized and elided.

This pass is executed after target lowering and right before register allocation.  As such, it has to deal with some idiosyncrasies of target lowering, specifically the possibility of intra-block control flow.  We experimented with doing this as a pre-lowering pass.  However, the transformations interfered with some of the target lowering's pattern matching, such as bool folding, so we concluded that post-lowering was a better place for it.

Note: Some of the lit tests are overly specific about registers, and in these cases it was the path of least resistance to just disable local variable splitting.

BUG= none
R=eholk@chromium.org, jpp@chromium.org

Review URL: https://codereview.chromium.org/2177033002 .

BUG=https://bugs.chromium.org/p/nativeclient/issues/detail?id=4374
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2183683003 .