[DRAFT] Announcement for LLVM 1.6 [DRAFT]

Hi All,

I'm putting together the announcement for the LLVM 1.6 release. Here is what I have so far. Because so much has been done, it is very likely that I have forgotten something. If you have done something that is not on the list, please send me a private email so I can add it (also, please tell me if I've made a mistake or miscredited something)!

My next project is to start hacking on the release notes. I would appreciate it if people could look into the documentation and let us know about (or fix!) anything that is out of date, misleading, or could be improved.

I believe we are still on track to start rolling the release next monday (Oct 31), which makes the likely 'official' release to be about a week from then. Once the release is officially out, it would be great to get help making binary distros for additional platforms (e.g. Cygwin) and for packages to be built (e.g. for the various Linux distros).

Thanks all!

-Chris

----------------- 8< ----------------- 8< --------------------

Insert high-level description/overview blurb here :slight_smile:
Themes: Major code generator work. New excellent performance/stability on PPC/darwin. Going back to 3-month release cycle. LLVM being used by many different people projects, e.g. papers being published.

Major New Features:

X. Reid added guards to the JIT, enabling it to JIT multithreaded code
    (on systems with pthreads). Threads support can be disabled by
    configuring LLVM with the --disable-threads swith.
X. LLVM includes a new optimization to statically evaluate C++ static
    variable constructors when they are simple enough. For example, it can
    now statically initialize "struct X { int a; X() : a(4) {} } g;".
X. The Loop Strength Reduction pass is completely rewritten, far more
    aggressive and enabled by default on most RISC targets. On PPC, we
    find that it often speeds up programs from 10-40% depending on the
    program.
X. LLVM now includes support for auto-generating large portions of the
    instruction selectors from target descriptions. This allows us to
    write patterns in the target .td file, instead of writing lots of
    nasty C++ code. Most of the PowerPC instruction selector is now
    generated from the PowerPC target description files and other targets
    are adding support that will be live for LLVM 1.7.
X. Andrew has dramatically improved the LLVM Alpha backend, to the point
    where it is now considered fully functional and off the 'beta' list.

Major Code Generator Changes:

X. Nate wrote a new component for the backend, a DAG Combiner. This
    allows the backend to take advantage of identities and do low-level
    peephole-style optimizations on the DAG.
X. Nate added support for a new TargetSubtarget interface, which
    determines which parts of the target to enable based on the
    target-triple (e.g., whether to use GAS or Intel asm printers on X86).
X. Jim Laskey extended sub-target support to include -mcpu and -mattr,
    allowing the target to think about what to do when particular features
    are enabled, but allowing the end-user to think about what CPU they
    have.
X. Jim also contributed a new light-weight instruction scheduler,
    available to targets that use DAG-to-DAG instruction selectors. In
    this release, the scheduler is fully operational but needs tuning, so
    it is not enabled by default.
X. The instruction selector framework now supports DAG-to-DAG instruction
    selection, where the instruction selector does pattern matching, but
    no code emission (necessary for scheduling & .td file autogeneration).

Other Code Generator Changes:

X. Duraid contributed many improvements to the Itanium backend
    (details??).
X. Andrew Lenharth contributed a major change to the varargs support,
    allowing LLVM to work with targets whose va_list type is a struct.
X. The instruction framework that debuted in LLVM 1.5 is far more mature
    and robust, and is able to handle many more strange target features.
X. Andrew added initial JIT support to the Alpha backend, which can run
    some simple programs. It is not fully complete yet though.
X. Jim Laskey contributed patches to improve the instruction selection in
    the PowerPC backend, matching more RLWIMI cases for example.
X. Nate implemented most of the PowerPC DAG-to-DAG instruction selector.
X. The tblgen tool & code generator now have more assertions and checking,
    which catch errors early, making it easier to work on the backend.
X. The default register allocator is now far faster on some testcases,
    particularly on targets with a large number of registers (e.g. IA64
    and PPC).
X. Jim extended tblgen to allow description of subtarget features in the
    .td files for the target.
X. There have been several minor improvements to the register allocator to
    coallesce more aggressively and coallesce spill code with copies more
    effectively.

Miscellaneous Improvements:

X. Andrew added support to the JIT to built a Global Offset Table if
    needed by a target.
X. Alexander Friedman improved the .ll file parser to be able to read
    from a text buffer in addition to a file.
X. Bryan Turner tried LLVM with Quest, a random testcase generator, and
    exposed several bugs (which are now fixed).
X. The llvm-test suite can now use the NAG Fortran to C compiler to compile
    SPEC FP programs if available (allowing us to test all of SPEC'95 & 2K).
X. The JIT-debugger mode of bugpoint is now much faster than before.
X. When bugpoint is grinding away and the user hits ctrl-C, it now
    gracefully stops and gives what it has reduced so far, instead of
    giving up completely.
X. LLVM now includes Xcode project files.
X. Jim Laskey added bitvector support to the command line option parsing
    library. See CommandLine 2.0 Library Manual — LLVM 16.0.0git documentation
X. We no longer build two versions of most LLVM libraries, which reduces
    the time required to build LLVM.
X. Documentation for the code generator is improving, though it is
    still incomplete: The LLVM Target-Independent Code Generator — LLVM 16.0.0git documentation
X. The code produced when exception handling is enabled is far more
    efficient in some cases, particularly on Mac OS/X.

Portability Improvements:

X. Nate added support for Mac OS/X on Intel.
X. X86 tail calls now work with the JIT and Jeff Cohen added code
    to support them under Visual C++.
X. Nate contributed a patch to allow LLVM to build with GCC 4.x, and Reid
    contributed several cleanup patches to silence GCC 4 warnings.
X. The llvm-test suite is now fully compatible with Mac OS/X
    (non-portabilities in the programs have been fixed).
X. Jeff Cohen contributed portability fixes to build on AMD64.

In addition to the new features and infrastructure we have built, we have also fixed many minor bugs in the C/C++ front-end, optimizers, and code generator including 54 bugzilla bugs (search for target milestone = 1.6). LLVM 1.6 is by far the best release we've had yet!

As usual, if you have any questions or comments about LLVM or any of the features in this status update, please feel free to contact the LLVMdev mailing list (llvmdev at cs.uiuc.edu)!

Finally, here is the previous status report, the LLVM 1.5 announcement:
http://lists.cs.uiuc.edu/pipermail/llvm-announce/2005-May/000016.html

-Chris

X. Reid added guards to the JIT, enabling it to JIT multithreaded code
  (on systems with pthreads). Threads support can be disabled by
  configuring LLVM with the --disable-threads swith.

swith -> switch

Everything else looks good to me.

-Tanya

Chris Lattner wrote:

I can't think of anything you missed; most of my comments below are about commas!
:frowning:

FWIW,

-- John T.

Hi All,

I'm putting together the announcement for the LLVM 1.6 release. Here is what I have so far. Because so much has been done, it is very likely that I have forgotten something. If you have done something that is not on the list, please send me a private email so I can add it (also, please tell me if I've made a mistake or miscredited something)!

My next project is to start hacking on the release notes. I would appreciate it if people could look into the documentation and let us know about (or fix!) anything that is out of date, misleading, or could be improved.

I believe we are still on track to start rolling the release next monday (Oct 31), which makes the likely 'official' release to be about a week from then. Once the release is officially out, it would be great to get help making binary distros for additional platforms (e.g. Cygwin) and for packages to be built (e.g. for the various Linux distros).

Thanks all!

-Chris

----------------- 8< ----------------- 8< --------------------

Insert high-level description/overview blurb here :slight_smile:
Themes: Major code generator work. New excellent performance/stability on PPC/darwin. Going back to 3-month release cycle. LLVM being used by many different people projects, e.g. papers being published.

Major New Features:

X. Reid added guards to the JIT, enabling it to JIT multithreaded code
   (on systems with pthreads). Threads support can be disabled by
   configuring LLVM with the --disable-threads swith.
X. LLVM includes a new optimization to statically evaluate C++ static
   variable constructors when they are simple enough. For example, it can
   now statically initialize "struct X { int a; X() : a(4) {} } g;".
X. The Loop Strength Reduction pass is completely rewritten, far more
   aggressive and enabled by default on most RISC targets. On PPC, we
   find that it often speeds up programs from 10-40% depending on the
   program.
X. LLVM now includes support for auto-generating large portions of the
   instruction selectors from target descriptions. This allows us to
   write patterns in the target .td file, instead of writing lots of

There should be no comma in the line above.

   nasty C++ code. Most of the PowerPC instruction selector is now
   generated from the PowerPC target description files and other targets
   are adding support that will be live for LLVM 1.7.
X. Andrew has dramatically improved the LLVM Alpha backend, to the point

Again, no comma.

   where it is now considered fully functional and off the 'beta' list.

Major Code Generator Changes:

X. Nate wrote a new component for the backend, a DAG Combiner. This
   allows the backend to take advantage of identities and do low-level
   peephole-style optimizations on the DAG.
X. Nate added support for a new TargetSubtarget interface, which
   determines which parts of the target to enable based on the
   target-triple (e.g., whether to use GAS or Intel asm printers on X86).
X. Jim Laskey extended sub-target support to include -mcpu and -mattr,
   allowing the target to think about what to do when particular features
   are enabled, but allowing the end-user to think about what CPU they
   have.
X. Jim also contributed a new light-weight instruction scheduler,
   available to targets that use DAG-to-DAG instruction selectors. In
   this release, the scheduler is fully operational but needs tuning, so
   it is not enabled by default.
X. The instruction selector framework now supports DAG-to-DAG instruction
   selection, where the instruction selector does pattern matching, but

Remove the second comma above.

   no code emission (necessary for scheduling & .td file autogeneration).

Other Code Generator Changes:

X. Duraid contributed many improvements to the Itanium backend
   (details??).
X. Andrew Lenharth contributed a major change to the varargs support,
   allowing LLVM to work with targets whose va_list type is a struct.
X. The instruction framework that debuted in LLVM 1.5 is far more mature
   and robust, and is able to handle many more strange target features.

Zap the comma above.
Also, do you mean instruction selector framework in the above line?

X. Andrew added initial JIT support to the Alpha backend, which can run
   some simple programs. It is not fully complete yet though.

Comma before "though"

X. Jim Laskey contributed patches to improve the instruction selection in
   the PowerPC backend, matching more RLWIMI cases for example.

Comman before "for example."

X. Nate implemented most of the PowerPC DAG-to-DAG instruction selector.
X. The tblgen tool & code generator now have more assertions and checking,
   which catch errors early, making it easier to work on the backend.
X. The default register allocator is now far faster on some testcases,
   particularly on targets with a large number of registers (e.g. IA64
   and PPC).
X. Jim extended tblgen to allow description of subtarget features in the
   .td files for the target.

Was this mentioned earlier, or is it subtly different from the last comment on target triples support?

X. There have been several minor improvements to the register allocator to
   coallesce more aggressively and coallesce spill code with copies more
   effectively.

Miscellaneous Improvements:

X. Andrew added support to the JIT to built a Global Offset Table if
   needed by a target.
X. Alexander Friedman improved the .ll file parser to be able to read
   from a text buffer in addition to a file.
X. Bryan Turner tried LLVM with Quest, a random testcase generator, and
   exposed several bugs (which are now fixed).
X. The llvm-test suite can now use the NAG Fortran to C compiler to compile
   SPEC FP programs if available (allowing us to test all of SPEC'95 & 2K).

Comma before "if available"

X. The JIT-debugger mode of bugpoint is now much faster than before.
X. When bugpoint is grinding away and the user hits ctrl-C, it now
   gracefully stops and gives what it has reduced so far, instead of
   giving up completely.

Cool.

The vector LLVA extension will not be merged into the 1.6 release branch?

It will make me have to merge twice: one for 1.6 and one for vector LLVA.

When do you plan to merge the vector LLVA to the main trunk, please?

The vector LLVA extension will not be merged into the 1.6 release branch?
It will make me have to merge twice: one for 1.6 and one for vector LLVA.
When do you plan to merge the vector LLVA to the main trunk, please?

No, the Vector LLVA extensions are not ready yet. When they are ready, they can be merged into LLVM. Maybe this will happen for LLVM 1.7. In the meantime, to use them, I would suggest using Rob's branch.

-Chris