isel has its own particular features that it wants in the CFG, in order to
reduce the number of times a constant is computed, etc.  Make sure that we
clean up the CFG before doing any other things for isel.  Doing so can
dramatically reduce the number of split edges and reduce the number of
places that constants get computed.  For example, this shrinks
CodeGen/Generic/phi-immediate-factoring.ll from 44 to 37 instructions on X86,
and from 21 to 17 MBB's in the output.  This is primarily a code size win,
not a performance win.

This implements CodeGen/Generic/phi-immediate-factoring.ll and PR1296.

llvm-svn: 35575

flags (ARMISD::CMPNZ) and comparisons that use all flags (ARMISD::CMP).
- Defines the instructions: TST, TEQ (ARM) and TST (Thumb).

llvm-svn: 35573

Make sure that the CTPOP result is casted to i32 as the bit counting
intrinsics all return i32 now (this affects CTLZ and CTTZ as well).

llvm-svn: 35567

llvm-svn: 35563

llvm-svn: 35561

llvm-svn: 35560

last night's regression in SingleSource/UnitTests/2005-05-11-Popcount-ffs-fls

llvm-svn: 35556

llvm-svn: 35555

llvm-svn: 35551

Support overloaded intrinsics bswap, ctpop, cttz, ctlz.

llvm-svn: 35547

Adjust for changes in the bit counting intrinsics. They all return i32
now so we have to trunc/zext the DAG node accordingly.

llvm-svn: 35546

Change getOperationName to return std::string instead of const char*

llvm-svn: 35545

Implement "actual" argument types for the Intrinsic member functions. This
involves changing the getName, getType, and getDeclaration methods to have
optional parameters for the actual types. These are necessary in order for
the type/name to be constructed properly for overloaded intrinsics. Only
the caller knows the actual argument types desired.

llvm-svn: 35541

1. Clear up confusion between "GotBits" and "ExpectBits". GotBits is the
   type actually provided. ExpectedBits is the type expected for the
   intrinsics. Before this patch, it was reversed!
2. Implement checks for overloaded intrinsics. This involves computing the
   suffix expected and making sure the suffix matches the function name. It
   also includes some intrinsic-specific checks such as ensuring that the
   bswap parameter and result are the same width and an even number of bytes.

llvm-svn: 35540

llvm-svn: 35536

llvm-svn: 35534

llvm-svn: 35533

llvm-svn: 35532

llvm-svn: 35531

llvm-svn: 35530

llvm-svn: 35529

llvm-svn: 35528

llvm-svn: 35527

llvm-svn: 35526

llvm-svn: 35525

llvm-svn: 35524

llvm-svn: 35523

llvm-svn: 35522

llvm-svn: 35521

llvm-svn: 35520

combination.

llvm-svn: 35517

register more than once.

llvm-svn: 35513

of dead def live interval at 1 to avoid multiple def's targeting the same
register. The previous patch missed a case where the source operand is live-in.
In that case, remove the whole interval.

llvm-svn: 35512

llvm-svn: 35510

* Rename the FunctionType* parameter from M to FT on all the functions.
* Implement a fix for PR1293 by just asserting that library functions that
  must return pointers should have pointer typed results. This just makes
  sure that we don't attempt to use an uninitialized integer or something
  later on.

llvm-svn: 35508

   amount is safe.
2. Use new method on ConstantInt instead of (? :) operator.
3. Use new method uge() on ConstantInt to simplify codes.

llvm-svn: 35505

llvm-svn: 35503

llvm-svn: 35500

Check at the end of the parse that there are no unresolved types and no
undefined values. Issue errors if there are. This gets rid of the need
for implementation or checkpoint by ensuring you can't finish a parse with
undefined things.

llvm-svn: 35499

llvm-svn: 35489