This caused miscompilations on out-of-tree targets, and possibly i386 as
well.

I'll find some other way of hoisting %rip-relative loads from loops
containing calls.

llvm-svn: 150816

Don't lazily allocate eh_frame.  We're not lazily allocating things like the LSDA, which are only used when the eh frame is used, so this lazy allocation doesn't really make sense.

Fix the type of eh_frame on Solaris so that Sun ld doesn't fail to combine them (thus making it impossible for the unwind library to find them and breaking exceptions).

llvm-svn: 150814

... and it's probably best to use the correct alignment, rather than just guessing that it's the same as the size.

llvm-svn: 150813

It turns out that putting an 8-byte symbol in a 4-byte section makes Solaris ld sulk.  GNU ld is perfectly happy with it, which is worrying for a whole other set of reasons...

Thanks to Anton, Duncan and Rafael for helping me track this down.
Pointy hat to Rafael for introducing the bug in the first place.

llvm-svn: 150811

llvm-svn: 150805

llvm-svn: 150796

llvm-svn: 150795

useful to represent a variable that is const in the source but can't be constant
in the IR because of a non-trivial constructor. If globalopt evaluates the
constructor, and there was an invariant.start with no matching invariant.end
possible, it will mark the global constant afterwards.

llvm-svn: 150794

instructions to be emitted.

llvm-svn: 150782

llvm-svn: 150780

llvm-svn: 150779

llvm-svn: 150778

llvm-svn: 150775

entend flag.

llvm-svn: 150774

llvm-svn: 150773

llvm-svn: 150771

Re-enable 150652 and 150654 - Make FPSCR non-reserved, and make MachineCSE bail on reserved registers. This *should* be safe as of r150786.

llvm-svn: 150769

llvm-svn: 150768

llvm-svn: 150739

llvm-svn: 150737

Call clobbers are now represented with register mask operands.  The
regmask can easily represent the fact that xmm6 is call-preserved while
ymm6 isn't.  This is automatically computed by TableGen from the
CalleeSavedRegs containing xmm6.

llvm-svn: 150709

The different calling conventions and call-preserved registers are
represented with regmask operands that are added dynamically.

llvm-svn: 150708

llvm-svn: 150706

Disable machine copy propagation for now. It's known to be buggy (PR11940) and introduces subtle miscompiles in many places.

llvm-svn: 150703

Turns out this isn't safe, because the code below depends on LHS and RHS having
the same type.

llvm-svn: 150695

llvm-svn: 150694

llvm-svn: 150690

llvm-svn: 150679

llvm-svn: 150670

Modify the algorithm when traversing the DAGCombiner's worklist to be O(log N) for all operations. This fixes a horrible worst case with lots of nodes where 99% of the time was being spent in std::remove.

llvm-svn: 150669

With MSVCRT, prior checker missed emission of #INF and #NAN.

FIXME: Checking should be simpler.
llvm-svn: 150667

atof(3) might behave differently among platforms.

llvm-svn: 150661

llvm-svn: 150655

llvm-svn: 150654

llvm-svn: 150653

loop-rotate shouldn't hoist alloca instructions out of a loop.  Patch by Patrik Hägglund, with slightly modified test.  Issue reported by Patrik Hägglund on llvmdev.

llvm-svn: 150642

Call instructions no longer have a list of 43 call-clobbered registers.
Instead, they get a single register mask operand with a bit vector of
call-preserved registers.

This saves a lot of memory, 42 x 32 bytes = 1344 bytes per call
instruction, and it speeds up building call instructions because those
43 imp-def operands no longer need to be added to use-def lists. (And
removed and shifted and re-added for every explicit call operand).

Passes like LiveVariables, LiveIntervals, RAGreedy, PEI, and
BranchFolding are significantly faster because they can deal with call
clobbers in bulk.

Overall, clang -O2 is between 0% and 8% faster, uniformly distributed
depending on call density in the compiled code.  Debug builds using
clang -O0 are 0% - 3% faster.

I have verified that this patch doesn't change the assembly generated
for the LLVM nightly test suite when building with -disable-copyprop
and -disable-branch-fold.

Branch folding behaves slightly differently in a few cases because call
instructions have different hash values now.

Copy propagation flushes its data structures when it crosses a register
mask operand. This causes it to leave a few dead copies behind, on the
order of 20 instruction across the entire nightly test suite, including
SPEC. Fixing this properly would require the pass to use different data
structures.

llvm-svn: 150638

Don't attempt to move instructions with regmask operands. They are most
likely calls anyway.

llvm-svn: 150634

The existing framework for postra scheduling is library local. We want to keep it that way. Soon we will have a more general MachineScheduler interface. At that time, various bits will be exposed to targets. In the meantime, the VLIWPacketizer wants to use ScheduleDAGInstrs directly, so it needs to wrapped in a PIMPL to avoid exposing it to the target interface.

llvm-svn: 150633

llvm-svn: 150632