fix off by 1 error in displacement computation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35602 91177308-0d34-0410-b5e6-96231b3b80d8

fix the CodeGen/ARM/2007-03-13-InstrSched.ll regression: allow IV's with scales
to be folded into non-store instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35601 91177308-0d34-0410-b5e6-96231b3b80d8

Ugh. Copy coalescer does not update register numbers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35600 91177308-0d34-0410-b5e6-96231b3b80d8

New test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35599 91177308-0d34-0410-b5e6-96231b3b80d8

add support for the 'w' inline asm register class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35598 91177308-0d34-0410-b5e6-96231b3b80d8

Update example notes to clarify system linker's role.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35597 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35596 91177308-0d34-0410-b5e6-96231b3b80d8

Check for .svn directories too to determine if a debug build is appropriate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35595 91177308-0d34-0410-b5e6-96231b3b80d8

1. Make use of APInt operation instead of using ConstantExpr::getXXX.
2. Use cheaper APInt methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35594 91177308-0d34-0410-b5e6-96231b3b80d8

Use uint32_t for bitwidth instead of unsigned.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35593 91177308-0d34-0410-b5e6-96231b3b80d8

Pass the type of the store access, not the type of the store, into the
target hook.  This allows us to codegen a loop as:

LBB1_1: @cond_next
        mov r2, #0
        str r2, [r0, +r3, lsl #2]
        add r3, r3, #1
        cmn r3, #1
        bne LBB1_1      @cond_next

instead of:

LBB1_1: @cond_next
        mov r2, #0
        str r2, [r0], #+4
        add r3, r3, #1
        cmn r3, #1
        bne LBB1_1      @cond_next

This looks the same, but has one fewer induction variable (and therefore,
one fewer register) live in the loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35592 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35591 91177308-0d34-0410-b5e6-96231b3b80d8

Fix a bug which caused us to never be able to use signed comparisons for
equality comparisons of a constant.  This allows us to codegen the 'sintzero'
loop in PR1288 as:

LBB1_1: ;cond_next
        li r4, 0
        addi r2, r2, 1
        stw r4, 0(r3)
        addi r3, r3, 4
        cmpwi cr0, r2, -1
        bne cr0, LBB1_1 ;cond_next

instead of:

LBB1_1: ;cond_next
        addi r2, r2, 1
        li r4, 0
        xoris r5, r2, 65535
        stw r4, 0(r3)
        addi r3, r3, 4
        cmplwi cr0, r5, 65535
        bne cr0, LBB1_1 ;cond_next

This implements CodeGen/PowerPC/compare-simm.ll, and also cuts 74
instructions out of kc++.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35590 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35589 91177308-0d34-0410-b5e6-96231b3b80d8

Wrap long line

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35588 91177308-0d34-0410-b5e6-96231b3b80d8

use more obvious function name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35587 91177308-0d34-0410-b5e6-96231b3b80d8

Treat xor of signbit like an add.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35586 91177308-0d34-0410-b5e6-96231b3b80d8

add a helper function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35585 91177308-0d34-0410-b5e6-96231b3b80d8

simplify (x+c)^signbit as (x+c+signbit), pointed out by PR1288.  This implements
test/Transforms/InstCombine/xor.ll:test28

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35584 91177308-0d34-0410-b5e6-96231b3b80d8

creative way to add one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35583 91177308-0d34-0410-b5e6-96231b3b80d8

my patch fixed the fixme.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35582 91177308-0d34-0410-b5e6-96231b3b80d8

Fix illegal assembly syntax.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35581 91177308-0d34-0410-b5e6-96231b3b80d8

bwsap -> bswap

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35580 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35579 91177308-0d34-0410-b5e6-96231b3b80d8

Upgrade the bit count intrinsics to have an i32 result.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35578 91177308-0d34-0410-b5e6-96231b3b80d8

Add a test case to make sure that constant folding of the bit counting
intrinsics works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35577 91177308-0d34-0410-b5e6-96231b3b80d8

reduce use of std::set

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35576 91177308-0d34-0410-b5e6-96231b3b80d8

Various passes before isel split edges and do other CFG-restructuring changes.

isel has its own particular features that it wants in the CFG, in order to
reduce the number of times a constant is computed, etc.  Make sure that we
clean up the CFG before doing any other things for isel.  Doing so can
dramatically reduce the number of split edges and reduce the number of
places that constants get computed.  For example, this shrinks
CodeGen/Generic/phi-immediate-factoring.ll from 44 to 37 instructions on X86,
and from 21 to 17 MBB's in the output.  This is primarily a code size win,
not a performance win.

This implements CodeGen/Generic/phi-immediate-factoring.ll and PR1296.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35575 91177308-0d34-0410-b5e6-96231b3b80d8

new testcase for PR1296

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35574 91177308-0d34-0410-b5e6-96231b3b80d8

- Divides the comparisons in two types: comparisons that only use N and Z
flags (ARMISD::CMPNZ) and comparisons that use all flags (ARMISD::CMP).
- Defines the instructions: TST, TEQ (ARM) and TST (Thumb).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35573 91177308-0d34-0410-b5e6-96231b3b80d8

add range version of insert()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35572 91177308-0d34-0410-b5e6-96231b3b80d8

Regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35571 91177308-0d34-0410-b5e6-96231b3b80d8

Don't upgrade functions that look like a bswap intrinsic but aren't.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35570 91177308-0d34-0410-b5e6-96231b3b80d8

Let llvm-upgrade upgrade the bswap intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35569 91177308-0d34-0410-b5e6-96231b3b80d8

Let llvm-upgrade upgrade bswap intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35568 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Make sure that the CTPOP result is casted to i32 as the bit counting
intrinsics all return i32 now (this affects CTLZ and CTTZ as well).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35567 91177308-0d34-0410-b5e6-96231b3b80d8

Revert the name changes for llvm.bswap to allow (and test) llvm-upgrade of
this intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35566 91177308-0d34-0410-b5e6-96231b3b80d8

Handle upgrade of llvm.bswap.iXX to llvm.bswap.iXX.iXX per new naming
rules for overloaded intrinsic functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35565 91177308-0d34-0410-b5e6-96231b3b80d8

Fix some bswap typos that Chris noticed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35564 91177308-0d34-0410-b5e6-96231b3b80d8

print the type of an inserted IV in -debug mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35563 91177308-0d34-0410-b5e6-96231b3b80d8

Add testcase for correct sret handling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35562 91177308-0d34-0410-b5e6-96231b3b80d8

simplify this code, make it work for ap ints

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35561 91177308-0d34-0410-b5e6-96231b3b80d8

fix breakage from last night, simplify code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35560 91177308-0d34-0410-b5e6-96231b3b80d8

gcc 3.4.4 is known-bad on x86-64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35559 91177308-0d34-0410-b5e6-96231b3b80d8

Preview documentation for additional intrinsic functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35558 91177308-0d34-0410-b5e6-96231b3b80d8

Unbreak VC++ build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35557 91177308-0d34-0410-b5e6-96231b3b80d8

The bit counting intrinsics return i32 not the operand type. This fixes
last night's regression in SingleSource/UnitTests/2005-05-11-Popcount-ffs-fls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35556 91177308-0d34-0410-b5e6-96231b3b80d8

Avoid unnecessary APInt construction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35555 91177308-0d34-0410-b5e6-96231b3b80d8

Testcase for the problem worked around in
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070319/046204.html

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35554 91177308-0d34-0410-b5e6-96231b3b80d8

Remove unused methods.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35553 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Preview documentation for the llvm.bit.concat intrinsic (yet to be
implemented).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35552 91177308-0d34-0410-b5e6-96231b3b80d8

Add i16 address mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35551 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Document changes to bswap and bit counting intrinsics. bswap's name now
requires two types in the suffix per overloaded intrinsic naming rules.
The ctpop, cttz, and ctlz intrinsics were changed to always return i32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35550 91177308-0d34-0410-b5e6-96231b3b80d8

Mention how to create overloaded intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35549 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Update these test cases to use proper signatures for bswap which is now
and overloaded intrinsic. Its name must be of the form llvm.bswap.i32.i32
since both the parameter and the result or of type "iAny". Also, the
bit counting intrinsics changed to always return i32.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35548 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Support overloaded intrinsics bswap, ctpop, cttz, ctlz.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35547 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Adjust for changes in the bit counting intrinsics. They all return i32
now so we have to trunc/zext the DAG node accordingly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35546 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Change getOperationName to return std::string instead of const char*

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35545 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Implement bswap, ctpop, ctlz, and cttz utilizing the new overloaded
intrinsic feature of tblgen. By defining llvm_int_ty to have an ValueType
of iAny, any intrinsic using that type becomes an overloaded intrinsic.
This signals that a) any integer type could be used for that argument and
b) the intrinsic needs a suffix to match each such argument for the
actual intrinsic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35544 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Intrinsic functions might need to be generated (if they are overloaded) so
returning a const char* is no longer possible. Change getOperationName to
return a std::string instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35543 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Change the interface to the getName, getType, getDeclaration methods to
include the "actual" types of the result/arguments. This is necessary to
support overloaded intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35542 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Implement "actual" argument types for the Intrinsic member functions. This
involves changing the getName, getType, and getDeclaration methods to have
optional parameters for the actual types. These are necessary in order for
the type/name to be constructed properly for overloaded intrinsics. Only
the caller knows the actual argument types desired.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35541 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
1. Clear up confusion between "GotBits" and "ExpectBits". GotBits is the
   type actually provided. ExpectedBits is the type expected for the
   intrinsics. Before this patch, it was reversed!
2. Implement checks for overloaded intrinsics. This involves computing the
   suffix expected and making sure the suffix matches the function name. It
   also includes some intrinsic-specific checks such as ensuring that the
   bswap parameter and result are the same width and an even number of bytes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35540 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Implement code generation for overloaded intrinsic functions. The basic
difference is that "actual" argument types must be provided when
constructing intrinsic names and types. Also, for recognition, only the
prefix is examined. If it matches, the suffix is assumed to match. The
suffix is checked by the Verifier, however.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35539 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1297:
Add an "iAny" ValueType as a place holder for an integer type of any
bit width. This is used for intrinsics overloading on integer type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35538 91177308-0d34-0410-b5e6-96231b3b80d8

not not!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35537 91177308-0d34-0410-b5e6-96231b3b80d8

Fix InstCombine/2007-03-31-InfiniteLoop.ll

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35536 91177308-0d34-0410-b5e6-96231b3b80d8

New testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35535 91177308-0d34-0410-b5e6-96231b3b80d8

Implement union of wrapped sets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35534 91177308-0d34-0410-b5e6-96231b3b80d8

Readme

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35533 91177308-0d34-0410-b5e6-96231b3b80d8

Consistency with native compilers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35532 91177308-0d34-0410-b5e6-96231b3b80d8

Fix comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35531 91177308-0d34-0410-b5e6-96231b3b80d8

add a note

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35530 91177308-0d34-0410-b5e6-96231b3b80d8

move a bunch of code out of the sdisel pass into its own opt pass "codegenprepare".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35529 91177308-0d34-0410-b5e6-96231b3b80d8

Split the sdisel code munging stuff out into its own opt-pass, CodeGenPrepare.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35528 91177308-0d34-0410-b5e6-96231b3b80d8

switch TL::getValueType to use MVT::getValueType.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35527 91177308-0d34-0410-b5e6-96231b3b80d8

add a method to turn a type into a VT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35526 91177308-0d34-0410-b5e6-96231b3b80d8

Delete dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35525 91177308-0d34-0410-b5e6-96231b3b80d8

Use APInt operators to calculate the carry bits, remove this loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35524 91177308-0d34-0410-b5e6-96231b3b80d8

Match GCC's MMX calling convention.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35523 91177308-0d34-0410-b5e6-96231b3b80d8

Add a -print-lsr-output option to LLC, to print the output of the LSR pass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35522 91177308-0d34-0410-b5e6-96231b3b80d8

implement the new addressing mode description hook.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35521 91177308-0d34-0410-b5e6-96231b3b80d8

add one addressing mode description hook to rule them all.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35520 91177308-0d34-0410-b5e6-96231b3b80d8

Missing '&&'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35519 91177308-0d34-0410-b5e6-96231b3b80d8

This is working again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35518 91177308-0d34-0410-b5e6-96231b3b80d8

Fix incorrect combination of different loads.  Reenable zext-over-truncate
combination.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35517 91177308-0d34-0410-b5e6-96231b3b80d8

add a testcase for x86

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35516 91177308-0d34-0410-b5e6-96231b3b80d8

New test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35515 91177308-0d34-0410-b5e6-96231b3b80d8

Don't add the same MI to register reuse "last def/use" twice if it reads the
register more than once.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35513 91177308-0d34-0410-b5e6-96231b3b80d8

Bug fix for PR1279. When isDead is propagate by copy coalescing, we keep length
of dead def live interval at 1 to avoid multiple def's targeting the same
register. The previous patch missed a case where the source operand is live-in.
In that case, remove the whole interval.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35512 91177308-0d34-0410-b5e6-96231b3b80d8

Test case for PR1279 (part 2).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35511 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure the use of ConstantInt::getZExtValue() for shift amount safe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35510 91177308-0d34-0410-b5e6-96231b3b80d8

Make sure this method just return value equal or less than Limit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35509 91177308-0d34-0410-b5e6-96231b3b80d8

For PR1293:
* Rename the FunctionType* parameter from M to FT on all the functions.
* Implement a fix for PR1293 by just asserting that library functions that
  must return pointers should have pointer typed results. This just makes
  sure that we don't attempt to use an uninitialized integer or something
  later on.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35508 91177308-0d34-0410-b5e6-96231b3b80d8

Remove reference to checkpoint.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35507 91177308-0d34-0410-b5e6-96231b3b80d8

Various hopefully correct easy fixes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35506 91177308-0d34-0410-b5e6-96231b3b80d8

1. Make sure the use of ConstantInt::getZExtValue() for getting shift
   amount is safe.
2. Use new method on ConstantInt instead of (? :) operator.
3. Use new method uge() on ConstantInt to simplify codes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35505 91177308-0d34-0410-b5e6-96231b3b80d8

Correct typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35504 91177308-0d34-0410-b5e6-96231b3b80d8

Use APInt operation instead of ConstantExpr::getXX.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35503 91177308-0d34-0410-b5e6-96231b3b80d8

Rename and correct the method "greaterOrEqual".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35502 91177308-0d34-0410-b5e6-96231b3b80d8