Oops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34698 91177308-0d34-0410-b5e6-96231b3b80d8

reapply

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34697 91177308-0d34-0410-b5e6-96231b3b80d8

*** empty log message ***

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34696 91177308-0d34-0410-b5e6-96231b3b80d8

Implement countLeadingOnes() and getMinSignedBits(). This helps to minimize
the bit width of negative numbers by computing the minimum bit width for a
negative value. E.g. 0x1800000000000000 could be just 0x8000000000000000

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34695 91177308-0d34-0410-b5e6-96231b3b80d8

Back out previous commit temporarily.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34694 91177308-0d34-0410-b5e6-96231b3b80d8

Backing out
CodeGenTarget.cpp updated: 1.82 -> 1.83
Record.cpp updated: 1.55 -> 1.56
Record.h updated: 1.59 -> 1.60
TableGen.cpp updated: 1.47 -> 1.48
It's missing CallingConvEmitter.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34693 91177308-0d34-0410-b5e6-96231b3b80d8

Let MRegisterInfo now owns RegScavenger; eliminateFrameIndex must preserve register kill info.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34692 91177308-0d34-0410-b5e6-96231b3b80d8

Let MRegisterInfo now owns RegScavenger.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34691 91177308-0d34-0410-b5e6-96231b3b80d8

RegScavenger interface change to make it more flexible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34690 91177308-0d34-0410-b5e6-96231b3b80d8

Let MRegisterInfo owns RegScavenger.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34689 91177308-0d34-0410-b5e6-96231b3b80d8

Pass IsImp, IsKill, and IsDead to ChangeToRegister.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34688 91177308-0d34-0410-b5e6-96231b3b80d8

This test passes on x86/Linux now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34686 91177308-0d34-0410-b5e6-96231b3b80d8

implement CCMatchIfCC in terms of CCMatchIf

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34685 91177308-0d34-0410-b5e6-96231b3b80d8

build cc info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34684 91177308-0d34-0410-b5e6-96231b3b80d8

target for generating CC info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34683 91177308-0d34-0410-b5e6-96231b3b80d8

initial support for calling convention generation, still unfinished.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34682 91177308-0d34-0410-b5e6-96231b3b80d8

Adjust to changes in the APInt interface.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34681 91177308-0d34-0410-b5e6-96231b3b80d8

Improve APInt interface:
1. Add unsigned and signed versions of methods so a "bool" argument doesn't
   need to be passed in.
2. Make the various getMin/getMax functions all be inline since they are
   so simple.
3. Simplify sdiv and srem code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34680 91177308-0d34-0410-b5e6-96231b3b80d8

Conform to single person attribution in file comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34678 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1205:
Implement review feedback:
1. Use new APInt::RoundDoubleToAPInt interface to specify the bit width so
   that we don't have to truncate or extend in constant folding.
2. Fix a pasteo in SDiv that prevented a check for overflow.
3. Fix the shift operators: undef happens when the shift amount is equal
   to the bitwidth.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34677 91177308-0d34-0410-b5e6-96231b3b80d8

Update for constant folding now generating undef and overflow correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34676 91177308-0d34-0410-b5e6-96231b3b80d8

Shifting by the bit width now produces undef, not 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34675 91177308-0d34-0410-b5e6-96231b3b80d8

1. Fix three serious bugs in the comparison code.
2. Change RoundDoubleToAPInt to take a bit width parameter. Use that
   parameter to limit the bit width of the result.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34673 91177308-0d34-0410-b5e6-96231b3b80d8

Allow the RoundDoubleToAPInt function to specify a width to use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34672 91177308-0d34-0410-b5e6-96231b3b80d8

a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34670 91177308-0d34-0410-b5e6-96231b3b80d8

Make getPassManagerType() const.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34669 91177308-0d34-0410-b5e6-96231b3b80d8

Fix parenthesis for BCTRL_{ELF|Macho} test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34668 91177308-0d34-0410-b5e6-96231b3b80d8

Differentiate between the MachO and the ELF ABI the CALL instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34667 91177308-0d34-0410-b5e6-96231b3b80d8

Duplicate use of LR, take 2.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34666 91177308-0d34-0410-b5e6-96231b3b80d8

Added myself

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34665 91177308-0d34-0410-b5e6-96231b3b80d8

Allow the AsmWriter to print out arbitrary precision integers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34664 91177308-0d34-0410-b5e6-96231b3b80d8

Correct the implelmentation of ConstantInt::getAllOnesValue() for bit
widths > 64 bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34663 91177308-0d34-0410-b5e6-96231b3b80d8

Get rid of extraneous // in file comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34662 91177308-0d34-0410-b5e6-96231b3b80d8

Add calling convention info

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34661 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1205:
Implement constant folding via APInt instead of uint64_t.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34660 91177308-0d34-0410-b5e6-96231b3b80d8

move target independent calling convention stuff to TargetCallingConv.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34659 91177308-0d34-0410-b5e6-96231b3b80d8

fill in some holes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34658 91177308-0d34-0410-b5e6-96231b3b80d8

switch x86-64 return value lowering over to using same mechanism as argument
lowering uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34657 91177308-0d34-0410-b5e6-96231b3b80d8

Minor refactoring of CC Lowering interfaces

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34656 91177308-0d34-0410-b5e6-96231b3b80d8

move CC Lowering stuff to its own public interface

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34655 91177308-0d34-0410-b5e6-96231b3b80d8

refactor x86-64 argument lowering yet again, this time eliminating templates,
'clients', etc, and adding CCValAssign instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34654 91177308-0d34-0410-b5e6-96231b3b80d8

ConstantFolding.cpp/h -> ConstantFold.cpp/h to avoid confusion with
lib/Analysis/ConstantFolding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34653 91177308-0d34-0410-b5e6-96231b3b80d8

Backing out Jim's LR spill changes. This was causing llvm-gcc bootstrapping
to infinite loop:
PPCMachineFunctionInfo.h updated: 1.2 -> 1.3
PPCRegisterInfo.cpp updated: 1.110 -> 1.111
PPCRegisterInfo.h updated: 1.28 -> 1.29

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34652 91177308-0d34-0410-b5e6-96231b3b80d8

XFAIL this on x86-Linux until its fixed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34651 91177308-0d34-0410-b5e6-96231b3b80d8

Remove test cases that produce undefined results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34650 91177308-0d34-0410-b5e6-96231b3b80d8

forward() should not increment internal iterator. Its client may insert instruction between now and next forward() call.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34649 91177308-0d34-0410-b5e6-96231b3b80d8

Simplify and document RoundDoubleToAPInt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34648 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1205:
Implement the first step towards arbitrary precision integer support in
LLVM. The APInt class provides arbitrary precision arithmetic and value
representation. This patch changes ConstantInt to use APInt as its value
representation without supporting bit widths > 64 yet. That change will
come after ConstantFolding handles bit widths > 64 bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34647 91177308-0d34-0410-b5e6-96231b3b80d8

1. Make sure all delete operators of arrays use the array form of delete.
2. Rewrite operator=(const APInt& RHS) to allow the RHS to be a different
   bit width than the LHS. This makes it possible to use APInt as the key
   of a DenseMap, as needed for the IntConstants map in Constants.cpp
3. Fix operator=(uint64_t) to clear unused bits in case the client assigns
   a value that has more bits than the APInt allows.
4. Assert that bit widths are equal in operator==
5. Revise getHashValue() to put the bit width in the low order six bits.
   This should help to make i1 0, i2 0, ... i64 0 all distinct in the
   IntConstants DenseMap.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34646 91177308-0d34-0410-b5e6-96231b3b80d8

Joining an interval with a dead copy instruction. Shorten the live range to the last use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34645 91177308-0d34-0410-b5e6-96231b3b80d8

Fix indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34643 91177308-0d34-0410-b5e6-96231b3b80d8

Implement the getHashValue method.
Fix toString use of getValue to use getZExtValue()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34642 91177308-0d34-0410-b5e6-96231b3b80d8

1. Split getValue() into getSExtValue() and getZExtValue() to match
   ConstantInt better.
2. Add a getHashValue() method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34641 91177308-0d34-0410-b5e6-96231b3b80d8

Use efficient container SmallPtrSet

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34640 91177308-0d34-0410-b5e6-96231b3b80d8

always lower to RETFLAG, never leave it as just ret.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34639 91177308-0d34-0410-b5e6-96231b3b80d8

Do not unswitch loop on same value again and again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34638 91177308-0d34-0410-b5e6-96231b3b80d8

fix attribution

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34637 91177308-0d34-0410-b5e6-96231b3b80d8

Add a description of the X86-64 calling convention and the return
conventions.  This doesn't do anything yet, but may in the future.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34636 91177308-0d34-0410-b5e6-96231b3b80d8

Implement inline methods that make transition of ConstantInt to use APInt
easier to comprehend and might be useful elsewhere.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34635 91177308-0d34-0410-b5e6-96231b3b80d8

Re-enable this. The header was committed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34634 91177308-0d34-0410-b5e6-96231b3b80d8

switch to smallvector

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34633 91177308-0d34-0410-b5e6-96231b3b80d8

another missing header :( :( :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34632 91177308-0d34-0410-b5e6-96231b3b80d8

initial hack at splitting the x86-64 calling convention info out from the
mechanics that process it.  I'm still not happy with this, but it's a step
in the right direction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34631 91177308-0d34-0410-b5e6-96231b3b80d8

Make isNegative() a const function since it doesn't modify the APInt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34630 91177308-0d34-0410-b5e6-96231b3b80d8

1. Remove redundant calls to clearUsedBits().
2. Fix countTrailingZeros to use a faster algorithm.
3. Simplify sext() slightly by using isNegative().
4. Implement ashr using word-at-a-time logic instead of bit-at-a-time
5. Rename locals named isNegative so they don't clash with method name.
6. Fix fromString to compute negated value correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34629 91177308-0d34-0410-b5e6-96231b3b80d8

the truncate must always be done, it's only the assert that is conditional.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34628 91177308-0d34-0410-b5e6-96231b3b80d8

reapply my previous patch with a bugfix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34627 91177308-0d34-0410-b5e6-96231b3b80d8

revert my previous change, something strange is happening.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34626 91177308-0d34-0410-b5e6-96231b3b80d8

add an accessor.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34625 91177308-0d34-0410-b5e6-96231b3b80d8

emit an enum value for the # of target registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34624 91177308-0d34-0410-b5e6-96231b3b80d8

in X86-64 CCC, i8/i16 arguments are already properly zext/sext'd on input.
Capture this so that downstream zext/sext's are optimized out.  This
compiles:
  int test(short X) { return (int)X; }

to:

_test:
        movl %edi, %eax
        ret

instead of:

_test:
        movswl %di, %eax
        ret

GCC produces this bizarre code:

_test:
        movw    %di, -12(%rsp)
        movswl  -12(%rsp),%eax
        ret

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34623 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34622 91177308-0d34-0410-b5e6-96231b3b80d8

Fold (sext (truncate x)) more aggressively, by avoiding creation of a
sextinreg if not needed.   This is useful in two cases: before legalize,
it avoids creating a sextinreg that will be trivially removed.  After legalize
if the target doesn't support sextinreg, the trunc/sext would not have been
removed before.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34621 91177308-0d34-0410-b5e6-96231b3b80d8

track signedness of formal argument, though we have a fixme here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34620 91177308-0d34-0410-b5e6-96231b3b80d8

Add an isNegative method to determine if the APInt's value is negative.
This is much less expensive than a test against zero.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34619 91177308-0d34-0410-b5e6-96231b3b80d8

Rewrite lshr to not do bit by bit copy but to copy and shift whole words.
This makes it much more efficient.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34618 91177308-0d34-0410-b5e6-96231b3b80d8

Fix sext operation. Shifting by zero would leave an incorrect mask.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34617 91177308-0d34-0410-b5e6-96231b3b80d8

1. Fix the flip() method to correctly flip all words of the APInt.
2. Implement the trunc, sext, and zext operations.
3. Improve fromString to accept negative values as input.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34616 91177308-0d34-0410-b5e6-96231b3b80d8

Fix an X86-64 abi bug.  We now compile:

void foo(short);
void bar(unsigned short A) {
  foo(A);
}

into:

_bar:
        subq $8, %rsp
        movswl %di, %edi
        call _foo
        addq $8, %rsp
        ret

instead of:

_bar:
        subq $8, %rsp
        call _foo
        addq $8, %rsp
        ret

Testcase here: test/CodeGen/X86/x86-64-shortint.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34615 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34614 91177308-0d34-0410-b5e6-96231b3b80d8

fix CodeGen/X86/2007-02-25-FastCCStack.ll, a regression from my patch last
night:  fastcc returns should only go in XMM0 if we have SSE2 or above.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34613 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34612 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34611 91177308-0d34-0410-b5e6-96231b3b80d8

Fix for PR1224.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34610 91177308-0d34-0410-b5e6-96231b3b80d8

Test for PR1224.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34609 91177308-0d34-0410-b5e6-96231b3b80d8

Rework GlobalValue::removeDeadConstantUsers to always remove dead constant
exprs hanging off a global, even if the global is not otherwise dead.  This
requires some tricky iterator gymnastics.

This implements Transforms/GlobalOpt/constantexpr-dangle.ll by deleting a
constantexpr that made it appear that the address of the function was taken.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34608 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase.  @foo should be marked fastcc by globalopt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34607 91177308-0d34-0410-b5e6-96231b3b80d8

disable some noisy debug output

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34606 91177308-0d34-0410-b5e6-96231b3b80d8

no really, this is the right patch

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34605 91177308-0d34-0410-b5e6-96231b3b80d8

always promote float varargs to double.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34604 91177308-0d34-0410-b5e6-96231b3b80d8

1. Provide more detail in file comment.
2. Move comments for methods to .h file, delete them in .cpp file.
3. All places that were doing manual clear of high order bits now call the
   clearUnusedBits() method in order to not depend on undefined behavior
   of the >> operator when the number of bits shifted equals the word size.
4. Reduced # of loc by using the new result of clearUnusedBits() method.
5. Simplified logic (decreased indentation) in a few places.
6. Added code comments to larger functions that needed them.
7. Added FIXME notes about weak implementations of things (e.g. bit-by-bit
   shift right is sub-optimal).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34603 91177308-0d34-0410-b5e6-96231b3b80d8

Fix clearUnusedBits to not depend on "undefined behavior" of >> operator
when the bit size is equal to the word size. This happens to work out okay
on x86, but might not on other platforms. The change just detects when
there are no bits to clear (because BitWidth is a multiple of the word size)
and returns early.

Also, move some comments from .cpp file into header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34602 91177308-0d34-0410-b5e6-96231b3b80d8

one important bugfix: PPC32 didn't have both elf and macho support for
external symbols and global addresses.  Add the missing ones.

one important workaround: PPCISD::CALL is matched by both PPCcall_ELF
and PPCcall_Macho, disable the _ELF patterns for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34601 91177308-0d34-0410-b5e6-96231b3b80d8

add -enable-eh

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34600 91177308-0d34-0410-b5e6-96231b3b80d8

optimize duplicate ValueMap lookups

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34599 91177308-0d34-0410-b5e6-96231b3b80d8

cosmetic changes from review of last patch.  obvious

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34598 91177308-0d34-0410-b5e6-96231b3b80d8

A couple of more places where a register liveness has been extended and its last kill should be updated accordingly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34597 91177308-0d34-0410-b5e6-96231b3b80d8

Add an assertion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34596 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a couple of bugs related IsDead back propagation during coalescing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34595 91177308-0d34-0410-b5e6-96231b3b80d8

If the liveinterval of the source instruction has been extended, remove the IsKill marker.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@34594 91177308-0d34-0410-b5e6-96231b3b80d8