oota-llvm.git
17 years agoDan points out that mem2reg also promotes vectors: be more precise.
Chris Lattner [Mon, 5 Nov 2007 17:45:54 +0000 (17:45 +0000)]
Dan points out that mem2reg also promotes vectors: be more precise.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43715 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoMove SimpleRegisterCoalescing.h to lib/CodeGen since there is now a common
Evan Cheng [Mon, 5 Nov 2007 17:41:38 +0000 (17:41 +0000)]
Move SimpleRegisterCoalescing.h to lib/CodeGen since there is now a common
register coalescer interface: RegisterCoalescing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43714 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoclarify
Chris Lattner [Mon, 5 Nov 2007 17:39:26 +0000 (17:39 +0000)]
clarify

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43712 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agofix typo
Chris Lattner [Mon, 5 Nov 2007 17:38:34 +0000 (17:38 +0000)]
fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43711 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix some typos.
Duncan Sands [Mon, 5 Nov 2007 16:04:58 +0000 (16:04 +0000)]
Fix some typos.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43709 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoTrivial fixes.
Duncan Sands [Mon, 5 Nov 2007 15:15:50 +0000 (15:15 +0000)]
Trivial fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43707 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoUse movups to spill / restore SSE registers on targets where stacks alignment is
Evan Cheng [Mon, 5 Nov 2007 07:30:01 +0000 (07:30 +0000)]
Use movups to spill / restore SSE registers on targets where stacks alignment is
less than 16. This is a temporary solution until dynamic stack alignment is
implemented.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43703 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoadd a link.
Chris Lattner [Mon, 5 Nov 2007 07:02:49 +0000 (07:02 +0000)]
add a link.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43702 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agofinish the tutorial, yaay.
Chris Lattner [Mon, 5 Nov 2007 07:00:54 +0000 (07:00 +0000)]
finish the tutorial, yaay.

comments and feedback welcome.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43701 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoSkip over deleted val#'s.
Evan Cheng [Mon, 5 Nov 2007 06:46:45 +0000 (06:46 +0000)]
Skip over deleted val#'s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43700 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoHandle cases where a register and one of its super-register are both marked as
Evan Cheng [Mon, 5 Nov 2007 03:11:55 +0000 (03:11 +0000)]
Handle cases where a register and one of its super-register are both marked as
defined on the same instruction. This fixes PR1767.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43699 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix test case. Chris didn't do make check. :-)
Evan Cheng [Mon, 5 Nov 2007 03:04:26 +0000 (03:04 +0000)]
Fix test case. Chris didn't do make check. :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43698 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdded support for PIC code with "explicit relocations" *only*.
Bruno Cardoso Lopes [Mon, 5 Nov 2007 03:02:32 +0000 (03:02 +0000)]
Added support for PIC code with "explicit relocations" *only*.
Removed all macro code for PIC (goodbye "la").
Support tested with shootout bench.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43697 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoBringing Passes.html up-to-date with the deletion of the
Gordon Henriksen [Mon, 5 Nov 2007 02:05:35 +0000 (02:05 +0000)]
Bringing Passes.html up-to-date with the deletion of the
-emitbitcode option and the addition of -preverify.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43696 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd the first section of chapter 8.
Chris Lattner [Mon, 5 Nov 2007 01:58:13 +0000 (01:58 +0000)]
Add the first section of chapter 8.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43695 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoDeleting redundant copy of block extractor pass. See also PR1775.
Gordon Henriksen [Mon, 5 Nov 2007 01:54:05 +0000 (01:54 +0000)]
Deleting redundant copy of block extractor pass. See also PR1775.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43694 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoDoh. PR1187 -> PR1766.
Evan Cheng [Mon, 5 Nov 2007 01:00:44 +0000 (01:00 +0000)]
Doh. PR1187 -> PR1766.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43693 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix PR1187.
Evan Cheng [Mon, 5 Nov 2007 00:59:10 +0000 (00:59 +0000)]
Fix PR1187.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43692 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoIf a long double is in a packed struct, it may be
Duncan Sands [Mon, 5 Nov 2007 00:35:07 +0000 (00:35 +0000)]
If a long double is in a packed struct, it may be
that there is no padding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43691 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agofinish the chapter.
Chris Lattner [Mon, 5 Nov 2007 00:23:57 +0000 (00:23 +0000)]
finish the chapter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43689 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoEliminate the remaining uses of getTypeSize. This
Duncan Sands [Mon, 5 Nov 2007 00:04:43 +0000 (00:04 +0000)]
Eliminate the remaining uses of getTypeSize.  This
should only effect x86 when using long double.  Now
12/16 bytes are output for long double globals (the
exact amount depends on the alignment).  This brings
globals in line with the rest of LLVM: the space
reserved for an object is now always the ABI size.
One tricky point is that only 10 bytes should be
output for long double if it is a field in a packed
struct, which is the reason for the additional
argument to EmitGlobalConstant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43688 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAnother step of stronger PHI elimination down.
Owen Anderson [Sun, 4 Nov 2007 22:33:26 +0000 (22:33 +0000)]
Another step of stronger PHI elimination down.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43684 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoDeleting -emitbitcode option which did nothing.
Gordon Henriksen [Sun, 4 Nov 2007 20:28:31 +0000 (20:28 +0000)]
Deleting -emitbitcode option which did nothing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43683 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agofix typos
Chris Lattner [Sun, 4 Nov 2007 19:42:13 +0000 (19:42 +0000)]
fix typos

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43682 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix PR1761 by not printing (rip) suffix when in -static mode.
Chris Lattner [Sun, 4 Nov 2007 19:23:28 +0000 (19:23 +0000)]
Fix PR1761 by not printing (rip) suffix when in -static mode.
Evan, please review this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43680 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoValidation fix.
Gordon Henriksen [Sun, 4 Nov 2007 18:17:58 +0000 (18:17 +0000)]
Validation fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43679 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoTypo fix.
Gordon Henriksen [Sun, 4 Nov 2007 18:14:08 +0000 (18:14 +0000)]
Typo fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43678 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoCompleting Passes.html with the exception of -emitbitcode, which should be
Gordon Henriksen [Sun, 4 Nov 2007 18:10:18 +0000 (18:10 +0000)]
Completing Passes.html with the exception of -emitbitcode, which should be
removed. This document could still stand for significant improvement:

  * Editing the pass descriptions; most were lifted with minimal editing from
    comments. Although implementation details were elided, many of the were not
    written for the audience that would be interested in this document.
  * More "before and after" examples.
  * More implicit dependency details. (Perhaps listing transforms in
    -std-compile-opts order would help alleviate this.)
  * Adding documentation for how to invoke passes programmatically.
  * Rearranging the document into a more logical taxonomy. For instance, putting
    profiling passes together.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43677 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix crash before main on ppc/linux with static constructors. PR1771
Nick Lewycky [Sun, 4 Nov 2007 17:32:10 +0000 (17:32 +0000)]
Fix crash before main on ppc/linux with static constructors. PR1771

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43676 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix a validation error.
Gordon Henriksen [Sun, 4 Nov 2007 16:17:00 +0000 (16:17 +0000)]
Fix a validation error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43675 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFinishing initial docs for all transformations in Passes.html.
Gordon Henriksen [Sun, 4 Nov 2007 16:15:04 +0000 (16:15 +0000)]
Finishing initial docs for all transformations in Passes.html.

Also cleaned up some comments in source files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43674 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoRemoving a dead reference from LLVM.xcodeproj.
Gordon Henriksen [Sun, 4 Nov 2007 16:12:17 +0000 (16:12 +0000)]
Removing a dead reference from LLVM.xcodeproj.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43673 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoChange uses of getTypeSize to getABITypeSize, getTypeStoreSize
Duncan Sands [Sun, 4 Nov 2007 14:43:57 +0000 (14:43 +0000)]
Change uses of getTypeSize to getABITypeSize, getTypeStoreSize
or getTypeSizeInBits as appropriate in ScalarReplAggregates.
The right change to make was not always obvious, so it would
be good to have an sroa guru review this.  While there I noticed
some bugs, and fixed them: (1) arrays of x86 long double have
holes due to alignment padding, but this wasn't being spotted
by HasStructPadding (renamed to HasPadding).  The same goes
for arrays of oddly sized ints.  Vectors also suffer from this,
in fact the problem for vectors is much worse because basic
vector assumptions seem to be broken by vectors of type with
alignment padding.   I didn't try to fix any of these vector
problems.  (2) The code for extracting smaller integers from
larger ones (in the "int union" case) was wrong on big-endian
machines for integers with size not a multiple of 8, like i1.
Probably this is impossible to hit via llvm-gcc, but I fixed
it anyway while there and added a testcase.  I also got rid of
some trailing whitespace and changed a function name which
had an obvious typo in it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43672 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoIf an interval is being undone clear its preference as well since the source interval...
Evan Cheng [Sun, 4 Nov 2007 08:32:21 +0000 (08:32 +0000)]
If an interval is being undone clear its preference as well since the source interval may have been undone as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43670 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix PR1763 by allowing the 'q' constraint to work with 64-bit
Chris Lattner [Sun, 4 Nov 2007 06:51:12 +0000 (06:51 +0000)]
Fix PR1763 by allowing the 'q' constraint to work with 64-bit
regs on x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43669 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoDisable tail duplication of call instructions. The cost
Chris Lattner [Sun, 4 Nov 2007 06:37:55 +0000 (06:37 +0000)]
Disable tail duplication of call instructions.  The cost
metric is way off for these in general, and this works around
buggy code like that in PR1764.  we'll see if there is a big
performance impact of this.  If so, I'll revert it tomorrow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43668 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agofinish the 'Memory in LLVM' section
Chris Lattner [Sat, 3 Nov 2007 22:22:30 +0000 (22:22 +0000)]
finish the 'Memory in LLVM' section

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43667 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agohack and slash the first 20% of chapter seven.
Chris Lattner [Sat, 3 Nov 2007 08:55:29 +0000 (08:55 +0000)]
hack and slash the first 20% of chapter seven.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43663 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoThere are times when the coalescer would not coalesce away a copy but the copy
Evan Cheng [Sat, 3 Nov 2007 07:20:12 +0000 (07:20 +0000)]
There are times when the coalescer would not coalesce away a copy but the copy
can be eliminated by the allocator is the destination and source targets the
same register. The most common case is when the source and destination registers
are in different class. For example, on x86 mov32to32_ targets GR32_ which
contains a subset of the registers in GR32.

The allocator can do 2 things:
1. Set the preferred allocation for the destination of a copy to that of its source.
2. After allocation is done, change the allocation of a copy destination (if
   legal) so the copy can be eliminated.

This eliminates 443 extra moves from 403.gcc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43662 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd std:: to sort calls.
Dan Gohman [Fri, 2 Nov 2007 22:24:01 +0000 (22:24 +0000)]
Add std:: to sort calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43652 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoChange illegal uses of ++ to uses of STLExtra.h's next function.
Dan Gohman [Fri, 2 Nov 2007 22:22:02 +0000 (22:22 +0000)]
Change illegal uses of ++ to uses of STLExtra.h's next function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43651 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdded overloaded version of Deserializer::ReadOwnedPtr which allows
Ted Kremenek [Fri, 2 Nov 2007 18:04:20 +0000 (18:04 +0000)]
Added overloaded version of Deserializer::ReadOwnedPtr which allows
the target pointer to be passed by reference.  This can result in less
typing, as the object to be deserialized can be inferred from the
argument.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43647 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoUnbreak tailcall opt.
Evan Cheng [Fri, 2 Nov 2007 17:45:40 +0000 (17:45 +0000)]
Unbreak tailcall opt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43646 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd run line.
Evan Cheng [Fri, 2 Nov 2007 17:36:58 +0000 (17:36 +0000)]
Add run line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43645 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoOne more extract_subreg coalescing bug.
Evan Cheng [Fri, 2 Nov 2007 17:35:08 +0000 (17:35 +0000)]
One more extract_subreg coalescing bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43644 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoadd a note
Chris Lattner [Fri, 2 Nov 2007 17:04:20 +0000 (17:04 +0000)]
add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43642 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix a thinko.
Duncan Sands [Fri, 2 Nov 2007 15:18:06 +0000 (15:18 +0000)]
Fix a thinko.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43639 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoRemove some unnecessary C-style statics.
Neil Booth [Fri, 2 Nov 2007 15:10:05 +0000 (15:10 +0000)]
Remove some unnecessary C-style statics.
Restore an assertion that arithmetic can be performed on this format.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43638 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agofix typos
Chris Lattner [Fri, 2 Nov 2007 05:54:25 +0000 (05:54 +0000)]
fix typos

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43637 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFinish chapter 6, and add a spiffy demo that shows off the language.
Chris Lattner [Fri, 2 Nov 2007 05:42:52 +0000 (05:42 +0000)]
Finish chapter 6, and add a spiffy demo that shows off the language.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43636 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoVAArgInst does, in fact, read memory.
Owen Anderson [Fri, 2 Nov 2007 04:01:21 +0000 (04:01 +0000)]
VAArgInst does, in fact, read memory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43633 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoUpdated VC++ build system
Hartmut Kaiser [Fri, 2 Nov 2007 01:44:08 +0000 (01:44 +0000)]
Updated VC++ build system

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43631 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoMissing a getNumOperands check.
Evan Cheng [Fri, 2 Nov 2007 01:26:22 +0000 (01:26 +0000)]
Missing a getNumOperands check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43630 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd back line whose removal somehow crept into prior patch
Neil Booth [Thu, 1 Nov 2007 22:51:07 +0000 (22:51 +0000)]
Add back line whose removal somehow crept into prior patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43627 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoWhen converting to integer, do bit manipulations in the destination
Neil Booth [Thu, 1 Nov 2007 22:43:37 +0000 (22:43 +0000)]
When converting to integer, do bit manipulations in the destination
memory rather than in a copy of the APFloat.  This avoids problems
when the destination is wider than our significand and is cleaner.

Also provide deterministic values in all cases where conversion
fails, namely zero for NaNs and the minimal or maximal value
respectively for underflow or overflow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43626 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoRemoved ReadVal from SerializeTrait<T>, and also removed it from
Ted Kremenek [Thu, 1 Nov 2007 22:23:34 +0000 (22:23 +0000)]
Removed ReadVal from SerializeTrait<T>, and also removed it from
Deserializer.

There were issues with Visual C++ barfing when instantiating
SerializeTrait<T> when "T" was an abstract class AND
SerializeTrait<T>::ReadVal was *never* called:

template <typename T>
struct SerializeTrait {
 <SNIP>
  static inline T ReadVal(Deserializer& D) { T::ReadVal(D); }
 <SNIP>
};

Visual C++ would complain about "T" being an abstract class, even
though ReadVal was never instantiated (although one of the other
member functions were).

Removing this from the trait is not a big deal.  It was used hardly
ever, and users who want "read-by-value" deserialization can simply
call the appropriate methods directly instead of relying on
trait-based-dispatch.  The trait dispatch for
serialization/deserialization is simply sugar in many cases (like this
one).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43624 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoExecutive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
Duncan Sands [Thu, 1 Nov 2007 20:53:16 +0000 (20:53 +0000)]
Executive summary: getTypeSize -> getTypeStoreSize / getABITypeSize.
The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment.  This gives a primitive type for
which getTypeSize differed from getABITypeSize.  For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).

This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition).  Instead there is:

(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type.  For a primitive type, this is the minimum number
of bits.  For an i36 this is 36 bits.  For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.

(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it).  For an
i36 this is 40 bits, for an x86 long double it is 80 bits.  This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes).  There doesn't seem to be anything
corresponding to this in gcc.

(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment.  For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS.  This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes).  This is
TYPE_SIZE in gcc.

Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize.  This means that the size of an array
is the length times the getABITypeSize.  It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize.  Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case.  So alloca's and mallocs should use getABITypeSize.  Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.

Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.

In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases).  I will get around to auditing these too at some point,
but I could do with some help.

Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize.  I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers.  If someone wants to pack these types more
tightly they can always use a packed struct.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43620 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoDon't barf on empty basic blocks. Do not rely on assert
Duncan Sands [Thu, 1 Nov 2007 10:50:26 +0000 (10:50 +0000)]
Don't barf on empty basic blocks.  Do not rely on assert
doing something - this needs to work for release builds
too.  I chose to just abort rather than following the
fancy logic of abortIfBroken, because (1) it is a pain
to do otherwise, and (2) nothing is going to work if the
module is this broken.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43611 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoSilence a warning saying that the variables always resolve to "true" in an
Bill Wendling [Thu, 1 Nov 2007 09:38:19 +0000 (09:38 +0000)]
Silence a warning saying that the variables always resolve to "true" in an
expression.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43610 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoSilence, accersed warning
Bill Wendling [Thu, 1 Nov 2007 08:51:44 +0000 (08:51 +0000)]
Silence, accersed warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43609 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoGet rid of compilation warning during release builds
Bill Wendling [Thu, 1 Nov 2007 08:24:40 +0000 (08:24 +0000)]
Get rid of compilation warning during release builds

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43608 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd the start of chapter 6, still much to go.
Chris Lattner [Thu, 1 Nov 2007 06:49:54 +0000 (06:49 +0000)]
Add the start of chapter 6, still much to go.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43607 91177308-0d34-0410-b5e6-96231b3b80d8

17 years ago- Coalesce extract_subreg when both intervals are relatively small.
Evan Cheng [Thu, 1 Nov 2007 06:22:48 +0000 (06:22 +0000)]
- Coalesce extract_subreg when both intervals are relatively small.
- Some code clean up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43606 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix test/Transforms/DeadStoreElimination/PartialStore.ll, which had been
Owen Anderson [Thu, 1 Nov 2007 05:29:16 +0000 (05:29 +0000)]
Fix test/Transforms/DeadStoreElimination/PartialStore.ll, which had been
silently failing because of an incorrect run line for some time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43605 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoremove verifier command line option: this should be part of the API, not
Chris Lattner [Thu, 1 Nov 2007 04:43:13 +0000 (04:43 +0000)]
remove verifier command line option: this should be part of the API, not
a command line optn.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43603 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoadd a broken gcc version, thanks to P. Durante for pointing this out.
Chris Lattner [Thu, 1 Nov 2007 04:20:16 +0000 (04:20 +0000)]
add a broken gcc version, thanks to P. Durante for pointing this out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43602 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoNow with less tabs!
Owen Anderson [Thu, 1 Nov 2007 03:54:23 +0000 (03:54 +0000)]
Now with less tabs!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43601 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoTest that expand_vector_elt(v2i64) works in 32-bit mode.
Dale Johannesen [Thu, 1 Nov 2007 02:38:24 +0000 (02:38 +0000)]
Test that expand_vector_elt(v2i64) works in 32-bit mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43598 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix InstCombine/2007-10-31-StringCrash.ll by removing an obvious
Chris Lattner [Thu, 1 Nov 2007 02:30:35 +0000 (02:30 +0000)]
Fix InstCombine/2007-10-31-StringCrash.ll by removing an obvious
(in hindsight) infinite recursion.  Simplify the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43597 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix InstCombine/2007-10-31-RangeCrash.ll
Chris Lattner [Thu, 1 Nov 2007 02:18:41 +0000 (02:18 +0000)]
Fix InstCombine/2007-10-31-RangeCrash.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43596 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoRewrote backpatcher. Backpatcher now stores the "has final pointer"
Ted Kremenek [Thu, 1 Nov 2007 00:57:37 +0000 (00:57 +0000)]
Rewrote backpatcher.  Backpatcher now stores the "has final pointer"
flag in the **key** of the backpatch map, as opposed to the mapped
value which contains either the final pointer, or a pointer to a chain
of pointers that need to be backpatched.  The bit flag was moved to
the key because we were erroneously assuming that the backpatched
pointers would be at an alignment of >= 2 bytes, which obviously
doesn't work for character strings.  Now we just steal the bit from the key.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43595 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdded typedef "value_type" to DenseMap (similar typedef appears in std::map).
Ted Kremenek [Thu, 1 Nov 2007 00:54:57 +0000 (00:54 +0000)]
Added typedef "value_type" to DenseMap (similar typedef appears in std::map).

Added method FindAndConstruct() to DenseMap, which does the same thing as
operator[], except that it refers value_type& (a reference to both the
key and mapped data pair).  This method is useful for clients that wish
to access the stored key value, as opposed to the key used to do the
actual lookup (these need not always be the same).

Redefined operator[] to use FindAndConstruct() (same logic).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43594 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoconstified several pointer arguments for methods in the Deserializer.
Ted Kremenek [Wed, 31 Oct 2007 22:42:03 +0000 (22:42 +0000)]
constified several pointer arguments for methods in the Deserializer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43583 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix a regression in test/CodeGen/X86/2007-04-24-VectorCrash.ll introduced
Dan Gohman [Wed, 31 Oct 2007 21:36:31 +0000 (21:36 +0000)]
Fix a regression in test/CodeGen/X86/2007-04-24-VectorCrash.ll introduced
by r43510. Gracefully handle constants with vector type that aren't
ConstantVector or ConstantAggregateZero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43579 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd a preverifier pass to check that every basic block ends in a terminator, so that...
Owen Anderson [Wed, 31 Oct 2007 21:04:18 +0000 (21:04 +0000)]
Add a preverifier pass to check that every basic block ends in a terminator, so that we don't segfault when verifying invalid code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43578 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoImplemented deserialization of references. References are handled
Ted Kremenek [Wed, 31 Oct 2007 19:58:32 +0000 (19:58 +0000)]
Implemented deserialization of references.  References are handled
just like pointers, except that they cannot be backpatched.  This
means that references are essentially non-owning pointers where the
referred object must be deserialized prior to the reference being
deserialized.  Because of the nature of references, this ordering of
objects is always possible.

Fixed a bug in backpatching code (returning the backpatched pointer
would accidentally include a bit flag).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43570 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdded Serializer::EmitRef to deal with emitting arbitrary references.
Ted Kremenek [Wed, 31 Oct 2007 18:23:21 +0000 (18:23 +0000)]
Added Serializer::EmitRef to deal with emitting arbitrary references.
Modified Serializer::EmitPtr to handle const pointers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43565 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoChanged access control within FoldingSet for some ivars from "private"
Ted Kremenek [Wed, 31 Oct 2007 17:12:47 +0000 (17:12 +0000)]
Changed access control within FoldingSet for some ivars from "private"
to "protected".  This allows iterators to work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43559 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoMake ARM and X86 LowerMEMCPY identical by moving the isThumb check into getMaxInlineS...
Rafael Espindola [Wed, 31 Oct 2007 14:39:58 +0000 (14:39 +0000)]
Make ARM and X86 LowerMEMCPY identical by moving the isThumb check into getMaxInlineSizeThreshold
and by restructuring the X86 version.

New I just have to move this to a common place :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43554 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix a typo in a comment.
Dan Gohman [Wed, 31 Oct 2007 14:35:39 +0000 (14:35 +0000)]
Fix a typo in a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43553 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoMake ARM an X86 memcpy expansion more similar to each other.
Rafael Espindola [Wed, 31 Oct 2007 11:52:06 +0000 (11:52 +0000)]
Make ARM an X86 memcpy expansion more similar to each other.
Now both subtarget define getMaxInlineSizeThreshold and the expansion uses it.

This should not change generated code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43552 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoPromotion of sdiv/srem/udiv/urem.
Duncan Sands [Wed, 31 Oct 2007 08:57:43 +0000 (08:57 +0000)]
Promotion of sdiv/srem/udiv/urem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43551 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd a newline at the end of the file.
Duncan Sands [Wed, 31 Oct 2007 08:49:24 +0000 (08:49 +0000)]
Add a newline at the end of the file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43550 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agofix validation
Chris Lattner [Wed, 31 Oct 2007 07:30:39 +0000 (07:30 +0000)]
fix validation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43549 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agookay, fine, make me finish this chapter. :)
Chris Lattner [Wed, 31 Oct 2007 07:29:43 +0000 (07:29 +0000)]
okay, fine, make me finish this chapter. :)

Feedback appreciated!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43548 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoadd the code for expression code that we'll add, though most of the
Chris Lattner [Wed, 31 Oct 2007 06:47:39 +0000 (06:47 +0000)]
add the code for expression code that we'll add, though most of the
description is missing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43547 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd the first half of chapter 5: if/then/else.
Chris Lattner [Wed, 31 Oct 2007 06:30:21 +0000 (06:30 +0000)]
Add the first half of chapter 5: if/then/else.
To come: for statement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43546 91177308-0d34-0410-b5e6-96231b3b80d8

17 years ago__main is dead along with llvm-gcc3.
Chris Lattner [Wed, 31 Oct 2007 05:53:31 +0000 (05:53 +0000)]
__main is dead along with llvm-gcc3.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43545 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd the skeleton of a better PHI elimination pass.
Owen Anderson [Wed, 31 Oct 2007 03:37:57 +0000 (03:37 +0000)]
Add the skeleton of a better PHI elimination pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43542 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoSome fixes to get MachineDomTree working better.
Owen Anderson [Wed, 31 Oct 2007 03:30:14 +0000 (03:30 +0000)]
Some fixes to get MachineDomTree working better.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43541 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoMake i64=expand_vector_elt(v2i64) work in 32-bit mode.
Dale Johannesen [Wed, 31 Oct 2007 00:32:36 +0000 (00:32 +0000)]
Make i64=expand_vector_elt(v2i64) work in 32-bit mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43535 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAt end of LSR, replace uses of now constant (as result of SplitCriticalEdge) PHI...
Evan Cheng [Tue, 30 Oct 2007 23:45:15 +0000 (23:45 +0000)]
At end of LSR, replace uses of now constant (as result of SplitCriticalEdge) PHI node with the constant value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43533 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdded member function "size()" to FoldingSet to return the number of
Ted Kremenek [Tue, 30 Oct 2007 23:35:55 +0000 (23:35 +0000)]
Added member function "size()" to FoldingSet to return the number of
nodes in the FoldingSet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43531 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoNew test.
Devang Patel [Tue, 30 Oct 2007 23:07:47 +0000 (23:07 +0000)]
New test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43527 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoIt's not safe to tell SplitCriticalEdge to merge identical edges. It may delete the...
Evan Cheng [Tue, 30 Oct 2007 22:27:26 +0000 (22:27 +0000)]
It's not safe to tell SplitCriticalEdge to merge identical edges. It may delete the phi instruction that's being processed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43524 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd missing SSE builtins: CVTPD2PI, CVTPS2PI,
Dale Johannesen [Tue, 30 Oct 2007 22:15:38 +0000 (22:15 +0000)]
Add missing SSE builtins:  CVTPD2PI, CVTPS2PI,
CVTTPD2PI, CVTTPS2PI, CVTPI2PD, CVTPI2PS.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43523 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoTypo.
Evan Cheng [Tue, 30 Oct 2007 20:11:21 +0000 (20:11 +0000)]
Typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43511 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd support for folding binary operators with vector zero operands.
Dan Gohman [Tue, 30 Oct 2007 19:00:49 +0000 (19:00 +0000)]
Add support for folding binary operators with vector zero operands.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43510 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoFix for visibility warnings generated by gcc-4.2.
Duncan Sands [Tue, 30 Oct 2007 13:14:37 +0000 (13:14 +0000)]
Fix for visibility warnings generated by gcc-4.2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43500 91177308-0d34-0410-b5e6-96231b3b80d8

17 years agoAdd support for expanding trunc stores. Consider
Duncan Sands [Tue, 30 Oct 2007 12:50:39 +0000 (12:50 +0000)]
Add support for expanding trunc stores.  Consider
storing an i170 on a 32 bit machine.  This is first
promoted to a trunc-i170 store of an i256.  On a
little-endian machine this expands to a store of
an i128 and a trunc-i42 store of an i128.  The
trunc-i42 store is further expanded to a trunc-i42
store of an i64, then to a store of an i32 and a
trunc-i10 store of an i32.  At this point the operand
type is legal (i32) and expansion stops (legalization
of the trunc-i10 needs to be handled in LegalizeDAG.cpp).
On big-endian machines the high bits are stored first,
and some bit-fiddling is needed in order to generate
aligned stores.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43499 91177308-0d34-0410-b5e6-96231b3b80d8