Add 2-addr to 3-addr promotion code that allows 32-bit LEA to be used via subregisters when 16-bit LEA is disabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41007 91177308-0d34-0410-b5e6-96231b3b80d8

Move isSubRegOf into MRegisterInfo. Fix a missed move elimination in LowerSubregs and add more debugging output there.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41005 91177308-0d34-0410-b5e6-96231b3b80d8

Do not overuse std::string.  Pass around char * directly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41001 91177308-0d34-0410-b5e6-96231b3b80d8

Clone loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40998 91177308-0d34-0410-b5e6-96231b3b80d8

Add utility to clone loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40997 91177308-0d34-0410-b5e6-96231b3b80d8

Add #ifndef guard.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40991 91177308-0d34-0410-b5e6-96231b3b80d8

add #ifndef guards

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40990 91177308-0d34-0410-b5e6-96231b3b80d8

add Value::getNameStart/getNameLen() accessors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40989 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test script for counting lines. This can be used instead of the
"| grep foo | wc -l | grep 2" idiom used by many tests, so that, for
example, tests don't mistakenly accept a count of 12 when 2 is
expected. Also, the new form is more consice: "| grep foo | count 2".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40988 91177308-0d34-0410-b5e6-96231b3b80d8

Fix EXTRACT_ELEMENT, EXTRACT_SUBVECTOR, and EXTRACT_VECTOR_ELT to
use an intptr ValueType instead of i32 for the index operand in
getCopyToParts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40987 91177308-0d34-0410-b5e6-96231b3b80d8

propagate struct size and alignment of byval arguments to the DAG

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40986 91177308-0d34-0410-b5e6-96231b3b80d8

For kicks, I though it would be fun to use the correct opcode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40985 91177308-0d34-0410-b5e6-96231b3b80d8

memcpy with zero length is hugely expensive, so avoid it.  This speeds up coallescing from 1.17s to 0.88s on siod.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40984 91177308-0d34-0410-b5e6-96231b3b80d8

small speedup in the case where a smallvector is default ctor'd from
an empty vector.  This speeds up llc slightly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40983 91177308-0d34-0410-b5e6-96231b3b80d8

Adding SSSE3 intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40982 91177308-0d34-0410-b5e6-96231b3b80d8

minor simplifications.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40981 91177308-0d34-0410-b5e6-96231b3b80d8

avoid copying strings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40980 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unncessary duplication.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40979 91177308-0d34-0410-b5e6-96231b3b80d8

Calculate exit and start value of true loop and false loop respectively.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40978 91177308-0d34-0410-b5e6-96231b3b80d8

ExitCondition and Induction variable are loop constraints
not split condition constraints.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40977 91177308-0d34-0410-b5e6-96231b3b80d8

unbreak the build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40976 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fix. ~1U marks the val# dead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40975 91177308-0d34-0410-b5e6-96231b3b80d8

Patch 10 for long double.  Doing constants right needs expanding ConstantFP
to handle values bigger than double.  If we assume host==target and host
long double works correctly, this is not too bad, but we don't want to
have that limitation longterm.  I could implement accepting double
constants as long double or something like that, which would lead to
incorrect codegen with no errors; the more I think about that the worse
it seems.  Rather than do such a hack that would be backed out later,
I'm settling for giving reasonable error messages, for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40974 91177308-0d34-0410-b5e6-96231b3b80d8

Temporarily backing out this change until we know why some dejagnu tests are failing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40973 91177308-0d34-0410-b5e6-96231b3b80d8

divb / mulb outputs to ah. Under x86-64 it's not legal to read ah if the instruction requires a rex prefix (i.e. outputs to r8b, etc.). So issue shift right by 8 on AX and then truncate it to 8 bits instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40972 91177308-0d34-0410-b5e6-96231b3b80d8

when we see a unaligned load from an insufficiently aligned global or
alloca, increase the alignment of the load, turning it into an aligned load.

This allows us to compile:

#include <xmmintrin.h>
__m128i foo(__m128i x){
 static const unsigned int c_0[4] = { 0, 0, 0, 0 };
  __m128i v_Zero = _mm_loadu_si128((__m128i*)c_0);
  x  = _mm_unpacklo_epi8(x,  v_Zero);
  return x;
}

into:

_foo:
punpcklbw _c_0.5944, %xmm0
ret
.data
.lcomm _c_0.5944,16,4 # c_0.5944

instead of:

_foo:
movdqu _c_0.5944, %xmm1
punpcklbw %xmm1, %xmm0
ret
.data
.lcomm _c_0.5944,16,2 # c_0.5944

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40971 91177308-0d34-0410-b5e6-96231b3b80d8

GR16_ sub-register class should be GR8_, not GR8. That is, it should only be 8-bit registers in 32-bit mode. Ditto for GR32_.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40970 91177308-0d34-0410-b5e6-96231b3b80d8

Update per review comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40965 91177308-0d34-0410-b5e6-96231b3b80d8

Make NonLocal and None const in the right way. :-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40961 91177308-0d34-0410-b5e6-96231b3b80d8

Traverse loop blocks' terminators to find split candidates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40960 91177308-0d34-0410-b5e6-96231b3b80d8

long double 9 of N.  This finishes up the X86-32 bits
(constants are still not handled).  Adds ConvertActions
to control fp-to-fp conversions (these are currently
defaulted for all other targets, so no changes there).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40958 91177308-0d34-0410-b5e6-96231b3b80d8

Fix arguments for some Altivec instructions.  From SWB.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40957 91177308-0d34-0410-b5e6-96231b3b80d8

If a target really needs to custom lower constants, it should be allowed
to do so.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40955 91177308-0d34-0410-b5e6-96231b3b80d8

Add more comments to memdep.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40953 91177308-0d34-0410-b5e6-96231b3b80d8

Add cost analysis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40952 91177308-0d34-0410-b5e6-96231b3b80d8

Make memdep fit in 80 cols.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40950 91177308-0d34-0410-b5e6-96231b3b80d8

Add one more comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40949 91177308-0d34-0410-b5e6-96231b3b80d8

Cleanup and comment-ize the memdep header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40948 91177308-0d34-0410-b5e6-96231b3b80d8

Preserve dom info while processing one iteration loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40947 91177308-0d34-0410-b5e6-96231b3b80d8

Change the None and NonLocal markers in memdep to be const.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40946 91177308-0d34-0410-b5e6-96231b3b80d8

Separate program name from error message with a :

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40945 91177308-0d34-0410-b5e6-96231b3b80d8

Clear split info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40944 91177308-0d34-0410-b5e6-96231b3b80d8

Terminate an error message with a newline.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40943 91177308-0d34-0410-b5e6-96231b3b80d8

Handle multiple split conditions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40941 91177308-0d34-0410-b5e6-96231b3b80d8

Make it clear that getSTDIN returns null when stdin is empty.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40940 91177308-0d34-0410-b5e6-96231b3b80d8

Make getSTDIN return null if the standard input is empty, as the header file
documentation implies and as its uses depend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40939 91177308-0d34-0410-b5e6-96231b3b80d8

Allow the filename "-" to be a place holder for stdin. This allows directing
stdin through llvm-ld and llvm-link.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40938 91177308-0d34-0410-b5e6-96231b3b80d8

Global values also don't undead-ify pointers in our dead alloca's set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40936 91177308-0d34-0410-b5e6-96231b3b80d8

Make handleEndBlock significantly faster with one trivial improvement,
and one hack to avoid hitting a bad case when the alias analysis is imprecise.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40935 91177308-0d34-0410-b5e6-96231b3b80d8

Small improvement: if a function doesn't access memory, we don't need to scan
it for potentially undeading pointers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40933 91177308-0d34-0410-b5e6-96231b3b80d8

Add some comments, remove a dead argument, and simplify some control flow.
No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40932 91177308-0d34-0410-b5e6-96231b3b80d8

eliminate redundant conditions from the signless types conversion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40927 91177308-0d34-0410-b5e6-96231b3b80d8

Handle functions with no name better.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40926 91177308-0d34-0410-b5e6-96231b3b80d8

Adding kill info to val#.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40925 91177308-0d34-0410-b5e6-96231b3b80d8

significantly speed up constant folding of calls (and thus all clients that use
ConstantFoldInstruction on calls) by avoiding Value::getName().  getName() constructs
and returns an std::string, which does heap allocation stuff.  This slightly speeds up
instcombine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40924 91177308-0d34-0410-b5e6-96231b3b80d8

Speed up updateDFSNumbers with two observations:
1. domtree is a tree, not a graph.  There is no need to avoid revisiting nodes with a set.
2. the worklist can contain the child iterator pointers so we don't get N^2 rescanning of children.

This speeds up updateDFSNumbers significantly, making it basically free.  On the testcase in PR1432,
this speeds up loopsimplify by another 3x, dropping it from the 12th most expensive pass to the to
the 30th. :)  It used to be #1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40923 91177308-0d34-0410-b5e6-96231b3b80d8

A few more small cleanups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40922 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up and bug fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40921 91177308-0d34-0410-b5e6-96231b3b80d8

reimplement dfs number computation to be significantly faster.  This speeds up
natural loop canonicalization (which does many cfg xforms) by 4.3x, for
example.  This also fixes a bug in postdom dfnumber computation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40920 91177308-0d34-0410-b5e6-96231b3b80d8

First round of cleanups from Chris'  feedback.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40919 91177308-0d34-0410-b5e6-96231b3b80d8

- Each val# can have multiple kills.
- Fix some minor bugs related to special markers on val# def. ~0U means
  undefined, ~1U means dead val#.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40916 91177308-0d34-0410-b5e6-96231b3b80d8

Embrace patch review feedback.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40915 91177308-0d34-0410-b5e6-96231b3b80d8

Remove a dead assertion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40914 91177308-0d34-0410-b5e6-96231b3b80d8

- LiveInterval value#'s now have 3 components: def instruction #,
kill instruction #, and source register number (iff the value# is defined by a
copy).
- Now def instruction # is set for every value#, not just for copy defined ones.
- Update some outdated code related inactive live ranges.
- Kill info not yet set. That's next patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40913 91177308-0d34-0410-b5e6-96231b3b80d8

Fix new compare instruction's signness. Caught by Chris during review.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40912 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40911 91177308-0d34-0410-b5e6-96231b3b80d8

Don't insert nearly as many redundant phi nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40909 91177308-0d34-0410-b5e6-96231b3b80d8

Fix spelling of mtvscr and mfvscr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40908 91177308-0d34-0410-b5e6-96231b3b80d8

Long double patch 8 of N:  make it partially work in
SSE mode (all but conversions <-> other FP types, I think):
>>Do not mark all-80-bit operations as "Requires[FPStack]"
(which really means "not SSE").
>>Refactor load-and-extend to facilitate this.
>>Update comments.
>>Handle long double in SSE when computing FP_REG_KILL.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40906 91177308-0d34-0410-b5e6-96231b3b80d8

Who thought up this crazy formatting scheme?

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40905 91177308-0d34-0410-b5e6-96231b3b80d8

Okay, over/back tags don't next.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40904 91177308-0d34-0410-b5e6-96231b3b80d8

Use eraseFromParent().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40903 91177308-0d34-0410-b5e6-96231b3b80d8

Try an indent level for better formatting.
Add the -version option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40902 91177308-0d34-0410-b5e6-96231b3b80d8

Add the code generation options.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40900 91177308-0d34-0410-b5e6-96231b3b80d8

Make this code more efficient

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40899 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40898 91177308-0d34-0410-b5e6-96231b3b80d8

Fix GLIBCXX_DEBUG error triggered by incrementing erased iterator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40897 91177308-0d34-0410-b5e6-96231b3b80d8

Add a missing forward declaration.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40896 91177308-0d34-0410-b5e6-96231b3b80d8

Rearrange options into sections and add the last floating point related option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40895 91177308-0d34-0410-b5e6-96231b3b80d8

Add the -disable-excess-fp-precision option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40894 91177308-0d34-0410-b5e6-96231b3b80d8

Add another missing option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40893 91177308-0d34-0410-b5e6-96231b3b80d8

Add some more missing options.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40892 91177308-0d34-0410-b5e6-96231b3b80d8

Remove the -f option which is no longer supported, but add the -fake-argv0
option which is similar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40891 91177308-0d34-0410-b5e6-96231b3b80d8

Describe the global/local naming convention.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40890 91177308-0d34-0410-b5e6-96231b3b80d8

Add reference to problem report.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40889 91177308-0d34-0410-b5e6-96231b3b80d8

Initial JIT support for ARM by Raul Fernandes Herbster.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40887 91177308-0d34-0410-b5e6-96231b3b80d8

Long double patch 7 of N, unless I lost count:).
Last x87 bits for full functionality (not
thoroughly tested, and long doubles do not work
in SSE modes at all - use -mcpu=i486 for now)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40886 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of unnecessary #include.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40885 91177308-0d34-0410-b5e6-96231b3b80d8

Clean up a bunch of caching stuff in memdep.  This reduces the time to run GVN
on 403.gcc from ~15s to ~10s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40884 91177308-0d34-0410-b5e6-96231b3b80d8

Begin loop index split pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40883 91177308-0d34-0410-b5e6-96231b3b80d8

Improve the accuracy of memdep for determining the dependencies of loads.
This brings GVN to parity with GCSE+LoadVN.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40882 91177308-0d34-0410-b5e6-96231b3b80d8

Replace 4-line function with 10-line version per review comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40881 91177308-0d34-0410-b5e6-96231b3b80d8

remove #if 0 code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40880 91177308-0d34-0410-b5e6-96231b3b80d8

remove #if 0 code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40879 91177308-0d34-0410-b5e6-96231b3b80d8

Move lengthy conditional down 1 level per review comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40878 91177308-0d34-0410-b5e6-96231b3b80d8

Get X86 long double calling convention to work
(on Darwin, anyway).  Fix some table omissions for
LD arithmetic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40877 91177308-0d34-0410-b5e6-96231b3b80d8

regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40875 91177308-0d34-0410-b5e6-96231b3b80d8

Fix PR1577, a crash on invalid bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40874 91177308-0d34-0410-b5e6-96231b3b80d8

This resolves a regression of BasicAA which failed to find any memory information for overloaded intrinsics (PR1600). This resolves that issue, and improves the matching scheme to use a BitVector rather than a binary search.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40872 91177308-0d34-0410-b5e6-96231b3b80d8

This fixes resizing issues with BitVectors. It ensures that the BitWord type and type size is always used, and ensures completely correct clearing of unused high bits, and setting of bits when resizing. It should resolve PR1563.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40871 91177308-0d34-0410-b5e6-96231b3b80d8