llvm-svn: 75909

llvm-svn: 75908

inline in the XCore.

llvm-svn: 75906

llvm-svn: 75903

offset from a 32 bit aligned base as follows:

  ldw low, base[offset >> 2]
  ldw high, base[(offset >> 2) + 1]
  shr low_shifted, low, (offset & 0x3) * 8
  shl high_shifted, high, 32 - (offset & 0x3) * 8
  or result, low_shifted, high_shifted

Expand 32 bit loads / stores with 16 bit alignment into two 16 bit
loads / stores.

llvm-svn: 75902

a big code size win since before they were expanding to upto 16
instructions.

llvm-svn: 75901

Avoid remat'ing instructions whose def have sub-register indices for now. It's just really really hard to get all the cases right.

llvm-svn: 75900

generator can know the top bits are zero, not undefined.
Thanks to Dan for pointing this out.

llvm-svn: 75899

llvm-svn: 75898

Varga!

llvm-svn: 75897

llvm-svn: 75896

llvm-svn: 75894

llvm-svn: 75893

llvm-svn: 75892

llvm-svn: 75891

llvm-svn: 75890

llvm-svn: 75889

processing.

llvm-svn: 75888

variables.
 - Module initialization functions supplanted the need for these.

llvm-svn: 75886

means that two separate raw_ostreams are doing buffering before data is
sent to the underlying stream. Besides the inefficiency of redundant
buffering, the second level of buffering doesn't recieve flush()
requests.

Fix this by having formatted_raw_ostream set the underlying raw_ostream
to be unbuffered. This eliminates inefficiency due to redundant buffering,
and it makes the flush() disconnect harmless.

This fixes PR4559.

llvm-svn: 75883

section on ppc.

llvm-svn: 75881

multiple times.

llvm-svn: 75880

This prefix is used by LLVM to inform the asm printer to not emit usual global symbol prefix before the symbol name.

llvm-svn: 75875

 - Added a sed script to do this in the comments.

llvm-svn: 75872

 - This isn't needed anymore, we require clients to call the appropriate
   initialization routines using TargetSelect.h.

llvm-svn: 75871

llvm-svn: 75867

llvm-svn: 75866

(although we don't get a very good error message).

llvm-svn: 75864

llvm-svn: 75863

llvm-svn: 75862

This extra check is not trigged when runnning "make check" on top-of-tree.

Change error message to better match llvm_unreachable() grammar.

Don't call llvm_unreachable() when writing error messages to a file, but keep going.

llvm-svn: 75860

 - No functionality change.

llvm-svn: 75859

llvm-svn: 75858

immediately on every output error. Instead, add a flag to
raw_ostream, and set the flag whenever an error is detected.

The flag can be queried and cleared from the public API. This
gives applications more flexibility to handling errors in
application-specific ways.

If the flag is not cleared when the raw_ostream is destructed,
llvm_report_error is called from the destructor. This ensures
that errors are not implicitly silenced, and provides
convenient default behavior for tools like llc and opt.
Clients wishing to avoid llvm_report_error calls from
raw_ostream should check for errors and clear the error flag.

llvm-svn: 75857

 - Switch to standard addAssemblyEmitter logic.

llvm-svn: 75854

Add an LLVMContext to Type, hardwired to the global context until Type uniquing is moved on the contexts themselves.

llvm-svn: 75853

using it.

llvm-svn: 75852

llvm-svn: 75851

 - No intended functionality change.

llvm-svn: 75848

Reserved registers are not candidates for scavenging, and they were removed
from the candidate list like this:

CreateRegClassMask(RC, Candidates);
Candidates ^= ReservedRegs;

However, when there are reserved registers outside RC, this causes invalid
bits to be set in Candidates.

llvm-svn: 75847