oota-llvm.git
11 years ago[stackprotector] Added significantly longer comment to FindPotentialTailCall to make...
Michael Gottesman [Tue, 20 Aug 2013 08:56:23 +0000 (08:56 +0000)]
[stackprotector] Added significantly longer comment to FindPotentialTailCall to make clear its relationship to llvm::isInTailCallPosition.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188770 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemoved trailing whitespace.
Michael Gottesman [Tue, 20 Aug 2013 08:46:16 +0000 (08:46 +0000)]
Removed trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188769 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[stackprotector] Removed stale TODO.
Michael Gottesman [Tue, 20 Aug 2013 08:46:13 +0000 (08:46 +0000)]
[stackprotector] Removed stale TODO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188768 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips][msa] Added and.v, bmnz.v, bmz.v, bsel.v, nor.v, or.v, xor.v
Daniel Sanders [Tue, 20 Aug 2013 08:38:21 +0000 (08:38 +0000)]
[mips][msa] Added and.v, bmnz.v, bmz.v, bsel.v, nor.v, or.v, xor.v

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188767 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[stackprotector] Added support for emitting the llvm intrinsic stack protector check.
Michael Gottesman [Tue, 20 Aug 2013 08:36:53 +0000 (08:36 +0000)]
[stackprotector] Added support for emitting the llvm intrinsic stack protector check.

rdar://13935163

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188766 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[stackprotector] Refactor out the end of isInTailCallPosition into the function retur...
Michael Gottesman [Tue, 20 Aug 2013 08:36:50 +0000 (08:36 +0000)]
[stackprotector] Refactor out the end of isInTailCallPosition into the function returnTypeIsEligibleForTailCall.

This allows me to use returnTypeIsEligibleForTailCall in the stack protector pass.

rdar://13935163

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188765 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove unused variables that crept in.
Michael Gottesman [Tue, 20 Aug 2013 07:17:27 +0000 (07:17 +0000)]
Remove unused variables that crept in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188761 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTeach selectiondag how to handle the stackprotectorcheck intrinsic.
Michael Gottesman [Tue, 20 Aug 2013 07:00:16 +0000 (07:00 +0000)]
Teach selectiondag how to handle the stackprotectorcheck intrinsic.

Previously, generation of stack protectors was done exclusively in the
pre-SelectionDAG Codegen LLVM IR Pass "Stack Protector". This necessitated
splitting basic blocks at the IR level to create the success/failure basic
blocks in the tail of the basic block in question. As a result of this,
calls that would have qualified for the sibling call optimization were no
longer eligible for optimization since said calls were no longer right in
the "tail position" (i.e. the immediate predecessor of a ReturnInst
instruction).

Then it was noticed that since the sibling call optimization causes the
callee to reuse the caller's stack, if we could delay the generation of
the stack protector check until later in CodeGen after the sibling call
decision was made, we get both the tail call optimization and the stack
protector check!

A few goals in solving this problem were:

  1. Preserve the architecture independence of stack protector generation.

  2. Preserve the normal IR level stack protector check for platforms like
     OpenBSD for which we support platform specific stack protector
     generation.

The main problem that guided the present solution is that one can not
solve this problem in an architecture independent manner at the IR level
only. This is because:

  1. The decision on whether or not to perform a sibling call on certain
     platforms (for instance i386) requires lower level information
     related to available registers that can not be known at the IR level.

  2. Even if the previous point were not true, the decision on whether to
     perform a tail call is done in LowerCallTo in SelectionDAG which
     occurs after the Stack Protector Pass. As a result, one would need to
     put the relevant callinst into the stack protector check success
     basic block (where the return inst is placed) and then move it back
     later at SelectionDAG/MI time before the stack protector check if the
     tail call optimization failed. The MI level option was nixed
     immediately since it would require platform specific pattern
     matching. The SelectionDAG level option was nixed because
     SelectionDAG only processes one IR level basic block at a time
     implying one could not create a DAG Combine to move the callinst.

To get around this problem a few things were realized:

  1. While one can not handle multiple IR level basic blocks at the
     SelectionDAG Level, one can generate multiple machine basic blocks
     for one IR level basic block. This is how we handle bit tests and
     switches.

  2. At the MI level, tail calls are represented via a special return
     MIInst called "tcreturn". Thus if we know the basic block in which we
     wish to insert the stack protector check, we get the correct behavior
     by always inserting the stack protector check right before the return
     statement. This is a "magical transformation" since no matter where
     the stack protector check intrinsic is, we always insert the stack
     protector check code at the end of the BB.

Given the aforementioned constraints, the following solution was devised:

  1. On platforms that do not support SelectionDAG stack protector check
     generation, allow for the normal IR level stack protector check
     generation to continue.

  2. On platforms that do support SelectionDAG stack protector check
     generation:

    a. Use the IR level stack protector pass to decide if a stack
       protector is required/which BB we insert the stack protector check
       in by reusing the logic already therein. If we wish to generate a
       stack protector check in a basic block, we place a special IR
       intrinsic called llvm.stackprotectorcheck right before the BB's
       returninst or if there is a callinst that could potentially be
       sibling call optimized, before the call inst.

    b. Then when a BB with said intrinsic is processed, we codegen the BB
       normally via SelectBasicBlock. In said process, when we visit the
       stack protector check, we do not actually emit anything into the
       BB. Instead, we just initialize the stack protector descriptor
       class (which involves stashing information/creating the success
       mbbb and the failure mbb if we have not created one for this
       function yet) and export the guard variable that we are going to
       compare.

    c. After we finish selecting the basic block, in FinishBasicBlock if
       the StackProtectorDescriptor attached to the SelectionDAGBuilder is
       initialized, we first find a splice point in the parent basic block
       before the terminator and then splice the terminator of said basic
       block into the success basic block. Then we code-gen a new tail for
       the parent basic block consisting of the two loads, the comparison,
       and finally two branches to the success/failure basic blocks. We
       conclude by code-gening the failure basic block if we have not
       code-gened it already (all stack protector checks we generate in
       the same function, use the same failure basic block).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188755 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix formatting. No functional change.
Craig Topper [Tue, 20 Aug 2013 05:23:59 +0000 (05:23 +0000)]
Fix formatting. No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188746 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd AVX-512 and related features to the CPUID detection code.
Craig Topper [Tue, 20 Aug 2013 05:22:42 +0000 (05:22 +0000)]
Add AVX-512 and related features to the CPUID detection code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188745 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMove AVX and non-AVX replication inside a couple multiclasses to avoid repeating...
Craig Topper [Tue, 20 Aug 2013 04:24:14 +0000 (04:24 +0000)]
Move AVX and non-AVX replication inside a couple multiclasses to avoid repeating each instruction for both individually.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188743 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd an error check for a typo I accidentally made in a td file that caused an assert...
Craig Topper [Tue, 20 Aug 2013 04:22:09 +0000 (04:22 +0000)]
Add an error check for a typo I accidentally made in a td file that caused an assert to fire.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188742 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[PowerPC] More refactoring prior to real PPC emitPrologue/Epilogue changes.
Bill Schmidt [Tue, 20 Aug 2013 03:12:23 +0000 (03:12 +0000)]
[PowerPC] More refactoring prior to real PPC emitPrologue/Epilogue changes.

(Patch committed on behalf of Mark Minich, whose log entry follows.)

This is a continuation of the refactorings performed in svn rev 188573
(see that rev's comments for more detail).

This is my stage 2 refactoring: I combined the emitPrologue() &
emitEpilogue() PPC32 & PPC64 code into a single flow, simplifying a
lot of the code since in essence the PPC32 & PPC64 code generation
logic is the same, only the instruction forms are different (in most
cases). This simplification is necessary because my functional changes
(yet to come) add significant complexity, and without the
simplification of my stage 2 refactoring, the overall complexity of
both emitPrologue() & emitEpilogue() would have become almost
intractable for most mortal programmers (like me).

This submission was intended to be a pure refactoring (no functional
changes whatsoever). However, in the process of combining the PPC32 &
PPC64 flows, I spotted a difference that I believe is a bug (see svn
rev 186478 line 863, or svn rev 188573 line 888): This line appears to
be restoring the BP with the original FP content, not the original BP
content. When I merged the 32-bit and 64-bit code, I used the
corresponding code from the 64-bit flow, which I believe uses the
correct offset (BPOffset) for this operation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188741 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMarking MCJIT PIC tests as XFAIL on AArch64
Andrew Kaylor [Tue, 20 Aug 2013 01:50:50 +0000 (01:50 +0000)]
Marking MCJIT PIC tests as XFAIL on AArch64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188740 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[Sparc] Use HWEncoding instead of unused Num field in Sparc register definitions...
Venkatraman Govindaraju [Tue, 20 Aug 2013 01:26:14 +0000 (01:26 +0000)]
[Sparc] Use HWEncoding instead of unused Num field in Sparc register definitions. Also, correct the definitions of RETL and RET instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188738 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFixing XPASSes among MCJIT PIC test on i686
Andrew Kaylor [Tue, 20 Aug 2013 00:37:33 +0000 (00:37 +0000)]
Fixing XPASSes among MCJIT PIC test on i686

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188736 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSecond attempt to mark Large/PIC MCJIT test as XFAIL for PowerPC64
Andrew Kaylor [Tue, 20 Aug 2013 00:22:03 +0000 (00:22 +0000)]
Second attempt to mark Large/PIC MCJIT test as XFAIL for PowerPC64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188735 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMarking two MCJIT PIC tests as XFAIL on Darwin
Andrew Kaylor [Tue, 20 Aug 2013 00:14:50 +0000 (00:14 +0000)]
Marking two MCJIT PIC tests as XFAIL on Darwin

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188734 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTrying again with PIC tests for MCJIT
Andrew Kaylor [Mon, 19 Aug 2013 23:52:53 +0000 (23:52 +0000)]
Trying again with PIC tests for MCJIT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188730 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd a llvm.copysign intrinsic
Hal Finkel [Mon, 19 Aug 2013 23:35:46 +0000 (23:35 +0000)]
Add a llvm.copysign intrinsic

This adds a llvm.copysign intrinsic; We already have Libfunc recognition for
copysign (which is turned into the FCOPYSIGN SDAG node). In order to
autovectorize calls to copysign in the loop vectorizer, we need a corresponding
intrinsic as well.

In addition to the expected changes to the language reference, the loop
vectorizer, BasicTTI, and the SDAG builder (the intrinsic is transformed into
an FCOPYSIGN node, just like the function call), this also adds FCOPYSIGN to a
few lists in LegalizeVector{Ops,Types} so that vector copysigns can be
expanded.

In TargetLoweringBase::initActions, I've made the default action for FCOPYSIGN
be Expand for vector types. This seems correct for all in-tree targets, and I
think is the right thing to do because, previously, there was no way to generate
vector-values FCOPYSIGN nodes (and most targets don't specify an action for
vector-typed FCOPYSIGN).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188728 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDon't form PPC CTR-based loops around a copysignl call
Hal Finkel [Mon, 19 Aug 2013 23:35:24 +0000 (23:35 +0000)]
Don't form PPC CTR-based loops around a copysignl call

copysign/copysignf never become function calls (because the SDAG expansion code
does not lower to the corresponding function call, but rather directly
implements the associated logic), but copysignl almost always is lowered into a
call to the requested libm functon (and, thus, might clobber CTR).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188727 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdding PIC support for ELF on x86_64 platforms
Andrew Kaylor [Mon, 19 Aug 2013 23:27:43 +0000 (23:27 +0000)]
Adding PIC support for ELF on x86_64 platforms

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188726 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoIntroduce non-const overloads for GlobalAlias::{get,resolve}AliasedGlobal.
Peter Collingbourne [Mon, 19 Aug 2013 23:13:33 +0000 (23:13 +0000)]
Introduce non-const overloads for GlobalAlias::{get,resolve}AliasedGlobal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188725 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse pop_back_val() instead of both back() and pop_back().
Jakub Staszak [Mon, 19 Aug 2013 22:47:55 +0000 (22:47 +0000)]
Use pop_back_val() instead of both back() and pop_back().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188723 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTeach InstCombine visitGetElementPtr about address spaces
Matt Arsenault [Mon, 19 Aug 2013 22:17:40 +0000 (22:17 +0000)]
Teach InstCombine visitGetElementPtr about address spaces

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188721 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoCleanup visitGetElementPtr to make address space change easier
Matt Arsenault [Mon, 19 Aug 2013 22:17:34 +0000 (22:17 +0000)]
Cleanup visitGetElementPtr to make address space change easier

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188720 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agocommonPointerCast cleanups to make address space change easier
Matt Arsenault [Mon, 19 Aug 2013 22:17:18 +0000 (22:17 +0000)]
commonPointerCast cleanups to make address space change easier

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188719 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMake sure that pop_back_val() result is used.
Jakub Staszak [Mon, 19 Aug 2013 22:12:00 +0000 (22:12 +0000)]
Make sure that pop_back_val() result is used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188717 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoReverting r188709 until I can figure out the proper way to XFAIL it.
Andrew Kaylor [Mon, 19 Aug 2013 22:05:07 +0000 (22:05 +0000)]
Reverting r188709 until I can figure out the proper way to XFAIL it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188715 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix assert with GEP ptr vector indexing structs
Matt Arsenault [Mon, 19 Aug 2013 21:43:16 +0000 (21:43 +0000)]
Fix assert with GEP ptr vector indexing structs

Also fix it calculating the wrong value. The struct index
is not a ConstantInt, so it was being interpreted as an array
index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188713 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse less verbose code and update comments.
Eric Christopher [Mon, 19 Aug 2013 21:41:38 +0000 (21:41 +0000)]
Use less verbose code and update comments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188711 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert non-test parts of r188507
Matt Arsenault [Mon, 19 Aug 2013 21:40:31 +0000 (21:40 +0000)]
Revert non-test parts of r188507

Re-add the inboundsless tests I didn't add originally

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188710 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdding tests for PIC with MCJIT
Andrew Kaylor [Mon, 19 Aug 2013 21:08:35 +0000 (21:08 +0000)]
Adding tests for PIC with MCJIT

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188709 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTurn on pubnames by default on linux.
Eric Christopher [Mon, 19 Aug 2013 21:07:38 +0000 (21:07 +0000)]
Turn on pubnames by default on linux.

Until gdb supports the new accelerator tables we should add the
pubnames section so that gdb_index can be generated from gold
at link time. On darwin we already emit the accelerator tables
and so don't need to worry about pubnames.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188708 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSuppress an annoying CMake warning in ChooseMSVCCRT.cmake
Reid Kleckner [Mon, 19 Aug 2013 20:25:26 +0000 (20:25 +0000)]
Suppress an annoying CMake warning in ChooseMSVCCRT.cmake

Warning was:
  Argument not separated from preceding token by whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188701 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoImprove the widening of integral binary vector operations
Paul Redmond [Mon, 19 Aug 2013 20:01:35 +0000 (20:01 +0000)]
Improve the widening of integral binary vector operations

- split WidenVecRes_Binary into WidenVecRes_Binary and WidenVecRes_BinaryCanTrap
  - WidenVecRes_BinaryCanTrap preserves the original behaviour for operations
    that can trap
  - WidenVecRes_Binary simply widens the operation and improves codegen for
    3-element vectors by allowing widening and promotion on x86 (matches the
    behaviour of unary and ternary operation widening)
- use WidenVecRes_Binary for operations on integers.

Reviewed by: nrotem

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188699 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdding comments to document RuntimeDyld relocation handling
Andrew Kaylor [Mon, 19 Aug 2013 19:38:06 +0000 (19:38 +0000)]
Adding comments to document RuntimeDyld relocation handling

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188697 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Fix instruction definitions that were incorrectly marked as code-gen-only.
Akira Hatanaka [Mon, 19 Aug 2013 19:08:03 +0000 (19:08 +0000)]
[mips] Fix instruction definitions that were incorrectly marked as code-gen-only.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188690 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd definition of __warn_unused_result__ attribute. It will be used in the
Jakub Staszak [Mon, 19 Aug 2013 19:02:33 +0000 (19:02 +0000)]
Add definition of __warn_unused_result__ attribute. It will be used in the
futher commits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188689 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoIntroduce SpecialCaseList::isIn overload for GlobalAliases.
Peter Collingbourne [Mon, 19 Aug 2013 19:00:35 +0000 (19:00 +0000)]
Introduce SpecialCaseList::isIn overload for GlobalAliases.

Differential Revision: http://llvm-reviews.chandlerc.com/D1437

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188688 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoThumb2 add immediate alias for SP
Mihai Popa [Mon, 19 Aug 2013 15:02:25 +0000 (15:02 +0000)]
Thumb2 add immediate alias for SP

The Thumb2 add immediate is in fact defined for SP. The manual is misleading as it points to a different section for add immediate with SP, however the encoding is the same as for add immediate with register only with the SP operand hard coded. As such add immediate with SP and add immediate with register can safely be treated as the same instruction.

All the patch does is adjust a register constraint on an instruction alias.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188676 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAVX-512: added arithmetic and logical operations.
Elena Demikhovsky [Mon, 19 Aug 2013 13:26:14 +0000 (13:26 +0000)]
AVX-512: added arithmetic and logical operations.
ADD, SUB, MUL integer and FP types. OR, AND, XOR.
Added embeded broadcast form for these instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188673 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Add negative integer absolute (load negative)
Richard Sandiford [Mon, 19 Aug 2013 12:56:58 +0000 (12:56 +0000)]
[SystemZ] Add negative integer absolute (load negative)

For now this matches the equivalent of (neg (abs ...)), which did hit a few
times in projects/test-suite.  We should probably also match cases where
absolute-like selects are used with reversed arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188671 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Add integer absolute (load positive)
Richard Sandiford [Mon, 19 Aug 2013 12:48:54 +0000 (12:48 +0000)]
[SystemZ] Add integer absolute (load positive)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188670 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Add support for sibling calls
Richard Sandiford [Mon, 19 Aug 2013 12:42:31 +0000 (12:42 +0000)]
[SystemZ] Add support for sibling calls

This first cut is pretty conservative.  The final argument register (R6)
is call-saved, so we would need to make sure that the R6 argument to a
sibling call is the same as the R6 argument to the calling function,
which seems worth keeping as a separate patch.

Saying that integer truncations are free means that we no longer
use the extending instructions LGF and LLGF for spills in int-conv-09.ll
and int-conv-10.ll.  Instead we treat the registers as 64 bits wide and
truncate them to 32-bits where necessary.  I think it's unlikely we'd
use LGF and LLGF for spills in other situations for the same reason,
so I'm removing the tests rather than replacing them.  The associated
code is generic and applies to many more instructions than just
LGF and LLGF, so there is no corresponding code removal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188669 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdds missing TLI check for library simplification of
Michael Kuperstein [Mon, 19 Aug 2013 06:55:47 +0000 (06:55 +0000)]
Adds missing TLI check for library simplification of
* pow(x, 0.5) -> fabs(sqrt(x))
* pow(2.0, x) -> exp2(x)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188656 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd ExpandFloatOp_FCOPYSIGN to handle ppcf128-related expansions
Hal Finkel [Mon, 19 Aug 2013 06:55:37 +0000 (06:55 +0000)]
Add ExpandFloatOp_FCOPYSIGN to handle ppcf128-related expansions

We had previously been asserting when faced with a FCOPYSIGN f64, ppcf128 node
because there was no way to expand the FCOPYSIGN node. Because ppcf128 is the
sum of two doubles, and the first double must have the larger magnitude, we
can take the sign from the first double. As a result, in addition to fixing the
crash, this is also an optimization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188655 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAVX-512: compiler intrinsics
Elena Demikhovsky [Mon, 19 Aug 2013 06:55:01 +0000 (06:55 +0000)]
AVX-512: compiler intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188654 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd the PPC fcpsgn instruction
Hal Finkel [Mon, 19 Aug 2013 05:01:02 +0000 (05:01 +0000)]
Add the PPC fcpsgn instruction

Modern PPC cores support a floating-point copysign instruction, and we can use
this to lower the FCOPYSIGN node (which is created from calls to the libm
copysign function). A couple of extra patterns are necessary because the
operand types of FCOPYSIGN need not agree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188653 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agollvm-dwarfdump: Do not include address offsets for attributes, only for tags
David Blaikie [Mon, 19 Aug 2013 03:36:23 +0000 (03:36 +0000)]
llvm-dwarfdump: Do not include address offsets for attributes, only for tags

This reduces the noise in diffs making it more likely that, at least for
LLVM revision-over-revision, diffs will actually yield usable results.

This is consistent with objdump's DWARF dumping behavior.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188650 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDebugInfo: don't emit zero-length names for parameters
David Blaikie [Mon, 19 Aug 2013 03:34:03 +0000 (03:34 +0000)]
DebugInfo: don't emit zero-length names for parameters

We check this in many/all other cases, just missed this one it seems.
Perhaps it'd be worth unifying this so we never emit zero-length
DW_AT_names.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188649 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove SpecialCaseList::findCategory.
Peter Collingbourne [Mon, 19 Aug 2013 00:24:20 +0000 (00:24 +0000)]
Remove SpecialCaseList::findCategory.

It turned out that I didn't need this for DFSan.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188646 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoARM: make sure we keep inline asm operands tied.
Tim Northover [Sun, 18 Aug 2013 18:06:03 +0000 (18:06 +0000)]
ARM: make sure we keep inline asm operands tied.

When patching inlineasm nodes to use GPRPair for 64-bit values, we
were dropping the information that two operands were tied, which
effectively broke the live-interval of vregs affected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188643 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAVX-512: Added VMOVD, VMOVQ, VMOVSS, VMOVSD instructions.
Elena Demikhovsky [Sun, 18 Aug 2013 13:08:57 +0000 (13:08 +0000)]
AVX-512: Added VMOVD, VMOVQ, VMOVSS, VMOVSD instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188637 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMake more of the lowering helpers static. Also use MVT instead of EVT in a couple...
Craig Topper [Sun, 18 Aug 2013 08:53:01 +0000 (08:53 +0000)]
Make more of the lowering helpers static. Also use MVT instead of EVT in a couple places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188629 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agodocs: command guide: cleanups, no text changes
Dmitri Gribenko [Sun, 18 Aug 2013 08:32:32 +0000 (08:32 +0000)]
docs: command guide: cleanups, no text changes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188627 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove unused stdio.h includes
Dmitri Gribenko [Sun, 18 Aug 2013 08:29:51 +0000 (08:29 +0000)]
Remove unused stdio.h includes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188626 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMakefile.rules: Avoid -fomit-frame-pointer also on cygwin due to PR14646.
NAKAMURA Takumi [Sun, 18 Aug 2013 03:38:40 +0000 (03:38 +0000)]
Makefile.rules: Avoid -fomit-frame-pointer also on cygwin due to PR14646.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188620 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMakefile.rules: Simplify nested if(s) on OmitFramePointer.
NAKAMURA Takumi [Sun, 18 Aug 2013 02:46:21 +0000 (02:46 +0000)]
Makefile.rules: Simplify nested if(s) on OmitFramePointer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188619 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoPort the detection of zlib from the main autoconf system to the sample
Chandler Carruth [Sun, 18 Aug 2013 01:55:15 +0000 (01:55 +0000)]
Port the detection of zlib from the main autoconf system to the sample
project's autoconf. This is the last of the missing optional checks used
by libSupport that seemed to be missing from the sample project, but
I could easily have missed some as this was done by inspection when
Craig asked me to add the terminfo support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188618 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support for linking librt and using clock_gettime to the sample
Chandler Carruth [Sun, 18 Aug 2013 01:46:34 +0000 (01:46 +0000)]
Add support for linking librt and using clock_gettime to the sample
autoconf setup.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188617 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUpdate the sample project autoconf setup to include support for
Chandler Carruth [Sun, 18 Aug 2013 01:43:57 +0000 (01:43 +0000)]
Update the sample project autoconf setup to include support for
detecting terminfo. Requested by Craig Topper, and probably should be
done much more systematically.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188616 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoGo through the really awkward dance required to delete the memory
Chandler Carruth [Sun, 18 Aug 2013 01:20:32 +0000 (01:20 +0000)]
Go through the really awkward dance required to delete the memory
allocated by setupterm. Without this, some folks are seeing leaked
memory whenever this routine is called more than once. Thanks to Craig
Topper for the report.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188615 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix SCEVExpander creating distinct duplicate PHI entries
Hal Finkel [Sun, 18 Aug 2013 00:16:23 +0000 (00:16 +0000)]
Fix SCEVExpander creating distinct duplicate PHI entries

This fixes SCEVExpander so that it does not create multiple distinct induction
variables for duplicate PHI entries. Specifically, given some code like this:

do.body6:                                         ; preds = %do.body6, %do.body6, %if.then5
  %end.0 = phi i8* [ undef, %if.then5 ], [ %incdec.ptr, %do.body6 ], [ %incdec.ptr, %do.body6 ]
...

Note that it is legal to have multiple entries for a basic block so long as the
associated value is the same. So the above input is okay, but expanding an
AddRec in this loop could produce code like this:

do.body6:                                         ; preds = %do.body6, %do.body6, %if.then5
  %indvar = phi i64 [ %indvar.next, %do.body6 ], [ %indvar.next1, %do.body6 ], [ 0, %if.then5 ]
  %end.0 = phi i8* [ undef, %if.then5 ], [ %incdec.ptr, %do.body6 ], [ %incdec.ptr, %do.body6 ]
...
  %indvar.next = add i64 %indvar, 1
  %indvar.next1 = add i64 %indvar, 1

And this is not legal because there are two PHI entries for %do.body6 each with
a distinct value.

Unfortunately, I don't have an in-tree test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188614 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoThe vbroadcastsi256 intrinsic does not exactly resemble the GCC
Juergen Ributzka [Sat, 17 Aug 2013 16:38:37 +0000 (16:38 +0000)]
The vbroadcastsi256 intrinsic does not exactly resemble the GCC
builtin. The GCC builtin expects the arguments to be passed by val,
whereas the LLVM intrinsic expects a pointer instead.

This is related to PR 16581 and rdar:14747994.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188608 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRecognize NetBSD's terminfo implementation.
Joerg Sonnenberger [Sat, 17 Aug 2013 11:06:00 +0000 (11:06 +0000)]
Recognize NetBSD's terminfo implementation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188606 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoPR 16899: Do not modify the basic block using the iterator, but keep the
Joerg Sonnenberger [Sat, 17 Aug 2013 11:04:47 +0000 (11:04 +0000)]
PR 16899: Do not modify the basic block using the iterator, but keep the
next value. This avoids crashes due to invalidation.

Patch by Joey Gouly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188605 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Fix possible use of an uninitialized variable
Tom Stellard [Sat, 17 Aug 2013 00:06:51 +0000 (00:06 +0000)]
R600: Fix possible use of an uninitialized variable

Spotted by Nick Lewycky!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188599 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Expand vector FRINT ops
Tom Stellard [Fri, 16 Aug 2013 23:51:33 +0000 (23:51 +0000)]
R600: Expand vector FRINT ops

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188598 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Expand vector FFLOOR ops
Tom Stellard [Fri, 16 Aug 2013 23:51:29 +0000 (23:51 +0000)]
R600: Expand vector FFLOOR ops

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188597 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Expand vector float operations for both SI and R600
Tom Stellard [Fri, 16 Aug 2013 23:51:24 +0000 (23:51 +0000)]
R600: Expand vector float operations for both SI and R600

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188596 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoARM: Properly constrain comparison fastisel register classes.
Jim Grosbach [Fri, 16 Aug 2013 23:37:40 +0000 (23:37 +0000)]
ARM: Properly constrain comparison fastisel register classes.

Ongoing 'make the verifier happy' improvements to ARM fast-isel.

rdar://12594152

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188595 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoARM: Fast-isel register class constrain for extends.
Jim Grosbach [Fri, 16 Aug 2013 23:37:36 +0000 (23:37 +0000)]
ARM: Fast-isel register class constrain for extends.

Properly constrain the operand register class for instructions used
in [sz]ext expansion. Update more tests to use the verifier now that
we're getting the register classes correct.

rdar://12594152

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188594 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoARM: Fix more fast-isel verifier failures.
Jim Grosbach [Fri, 16 Aug 2013 23:37:31 +0000 (23:37 +0000)]
ARM: Fix more fast-isel verifier failures.

Teach the generic instruction selection helper functions to constrain
the register classes of their input operands. For non-physical register
references, the generic code needs to be careful not to mess that up
when replacing references to result registers. As the comment indicates
for MachineRegisterInfo::replaceRegWith(), it's important to call
constrainRegClass() first.

rdar://12594152

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188593 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoARM: Clean up fast-isel machine verifier errors.
Jim Grosbach [Fri, 16 Aug 2013 23:37:23 +0000 (23:37 +0000)]
ARM: Clean up fast-isel machine verifier errors.

Lots of machine verifier errors result from using a plain GPR regclass
for incoming argument copies. A more restrictive rGPR class is more
appropriate since it more accurately represents what's happening, plus
it lines up better with isel later on so the verifier is happier.
Reduces the number of ARM fast-isel tests not running with the verifier
enabled by over half.

rdar://12594152

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188592 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[lit] Rewrite TODO list, and elaborate on some things.
Daniel Dunbar [Fri, 16 Aug 2013 23:30:23 +0000 (23:30 +0000)]
[lit] Rewrite TODO list, and elaborate on some things.

 - If anyone is interested in lit's feature set, I'd appreciate any comments on
   the elaborated items.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188590 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[typo] An LLVM.
Daniel Dunbar [Fri, 16 Aug 2013 23:30:19 +0000 (23:30 +0000)]
[typo] An LLVM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188589 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix a subtle difference between running clang vs llc for mips16.
Reed Kotler [Fri, 16 Aug 2013 23:05:18 +0000 (23:05 +0000)]
Fix a subtle difference between running clang vs llc for mips16.
This regards how mips16 is viewed. It's not really a target type but
there has always been a target for it in the td files. It's more properly
-mcpu=mips32 -mattr=+mips16 . This is how clang treats it but we have
always had the -mcpu=mips16 which I probably should delete now but it will
require updating all the .ll test cases for mips16. In this case it changed
how we decide if we have a count bits instruction and whether instruction
lowering should then expand ctlz. Now that we have dual mode compilation,
-mattr=+mips16 really just indicates the inital processor mode that
we are compiling for. (It is also possible to have -mcpu=64 -mattr=+mips16
but as far as I know, nobody has even built such a processor, though there
is an architecture manual for this).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188586 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoActually, use GNU inline asm for cpuid with clang
Reid Kleckner [Fri, 16 Aug 2013 22:42:42 +0000 (22:42 +0000)]
Actually, use GNU inline asm for cpuid with clang

Clang doesn't support the MSVC __cpuid intrinsic yet, and fixing that is
blocked on some fairly complicated issues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188584 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoI'm told that != is not ==
Chris Lattner [Fri, 16 Aug 2013 22:29:44 +0000 (22:29 +0000)]
I'm told that != is not ==

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188583 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoallow != to compare PointerUnion, we already support ==.
Chris Lattner [Fri, 16 Aug 2013 22:09:02 +0000 (22:09 +0000)]
allow != to compare PointerUnion, we already support ==.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188582 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd difference_type to ImmutableMap/Set iterators so they have a complete set of...
Benjamin Kramer [Fri, 16 Aug 2013 21:55:56 +0000 (21:55 +0000)]
Add difference_type to ImmutableMap/Set iterators so they have a complete set of typedefs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188579 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDebugInfo: Allow the addition of other (such as static data) members to a record...
David Blaikie [Fri, 16 Aug 2013 20:42:14 +0000 (20:42 +0000)]
DebugInfo: Allow the addition of other (such as static data) members to a record type after construction

Plus a type cleanup & minor fix to enumerate members of declarations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188577 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[PowerPC] Preparatory refactoring for making prologue and epilogue
Bill Schmidt [Fri, 16 Aug 2013 20:05:04 +0000 (20:05 +0000)]
[PowerPC] Preparatory refactoring for making prologue and epilogue
safe on PPC32 SVR4 ABI

[Patch and following text by Mark Minich; committing on his behalf.]

There are FIXME's in PowerPC/PPCFrameLowering.cpp, method
PPCFrameLowering::emitPrologue() related to "negative offsets of R1"
on PPC32 SVR4. They're true, but the real issue is that on PPC32 SVR4
(and any ABI without a Red Zone), no spills may be made until after
the stackframe is claimed, which also includes the LR spill which is
at a positive offset. The same problem exists in emitEpilogue(),
though there's no FIXME for it. I intend to fix this issue, making
LLVM-compiled code finally safe for use on SVR4/EABI/e500 32-bit
platforms (including in particular, OS-free embedded systems & kernel
code, where interrupts may share the same stack as user code).

In preparation for making these changes, to make the diffs for the
functional changes less cluttered, I am providing the non-functional
refactorings in two stages:

Stage 1 does some minor fluffy refactorings to pull multiple method
calls up into a single bool, creating named bools for repeated uses of
obscure logic, moving some code up earlier because either stage 2 or
my final version will require it earlier, and rewording/adding some
comments. My stage 1 changes can be characterized as primarily fluffy
cleanup, the purpose of which may be unclear until the stage 2 or
final changes are made.

My stage 2 refactorings combine the separate PPC32 & PPC64 logic,
which is currently performed by largely duplicate code, into a single
flow, with the differences handled by a group of constants initialized
early in the methods.

This submission is for my stage 1 changes. There should be no
functional changes whatsoever; this is a pure refactoring.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188573 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFixed RuntimeDyldELF absolute relocations.
Richard Mitton [Fri, 16 Aug 2013 18:54:26 +0000 (18:54 +0000)]
Fixed RuntimeDyldELF absolute relocations.

If an ELF relocation is pointed at an absolute address, it will have a symbol ID of zero.
RuntimeDyldELF::processRelocationRef was not previously handling this case, and was instead trying to handle it as a section-relative fixup.

I think this is the right fix here, but my elf-fu is poor on some of the more exotic platforms, so I'd appreciate it if anyone with greater knowledge could verify this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188572 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTest commit.
Richard Mitton [Fri, 16 Aug 2013 18:09:06 +0000 (18:09 +0000)]
Test commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188568 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[tests] Another attempt to workaround broken misched-copy.s test on some buildbots.
Daniel Dunbar [Fri, 16 Aug 2013 18:01:18 +0000 (18:01 +0000)]
[tests] Another attempt to workaround broken misched-copy.s test on some buildbots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188567 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSwitching to using a helper function instead of manually converting the string to...
Aaron Ballman [Fri, 16 Aug 2013 17:53:28 +0000 (17:53 +0000)]
Switching to using a helper function instead of manually converting the string to UTF-8.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188566 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemoving unused functionality.
Aaron Ballman [Fri, 16 Aug 2013 17:33:57 +0000 (17:33 +0000)]
Removing unused functionality.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188565 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFileCheck: Fix stray quote in CHECK-LABEL error message.
Stephen Lin [Fri, 16 Aug 2013 17:29:01 +0000 (17:29 +0000)]
FileCheck: Fix stray quote in CHECK-LABEL error message.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188564 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoInstCombine: Use isAllOnesValue() instead of explicit -1.
Jim Grosbach [Fri, 16 Aug 2013 17:03:36 +0000 (17:03 +0000)]
InstCombine: Use isAllOnesValue() instead of explicit -1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188563 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600/SI: Add pattern for xor of i1
Michel Danzer [Fri, 16 Aug 2013 16:19:31 +0000 (16:19 +0000)]
R600/SI: Add pattern for xor of i1

Fixes two recent piglit regressions with radeonsi.

Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188559 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600/SI: Fix broken encoding of DS_WRITE_B32
Michel Danzer [Fri, 16 Aug 2013 16:19:24 +0000 (16:19 +0000)]
R600/SI: Fix broken encoding of DS_WRITE_B32

The logic in SIInsertWaits::getHwCounts() only really made sense for SMRD
instructions, and trying to shoehorn it into handling DS_WRITE_B32 caused
it to corrupt the encoding of that by clobbering the first operand with
the second one.

Undo that damage and only apply the SMRD logic to that.

Fixes some derivates related piglit regressions with radeonsi.

Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188558 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoReverted test commit (r188556)
Daniel Sanders [Fri, 16 Aug 2013 15:27:12 +0000 (15:27 +0000)]
Reverted test commit (r188556)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188557 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTest commit. Just a blank line
Daniel Sanders [Fri, 16 Aug 2013 15:26:36 +0000 (15:26 +0000)]
Test commit. Just a blank line

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188556 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Allocate memoperand in the MachienFunction so it doesn't leak.
Benjamin Kramer [Fri, 16 Aug 2013 14:48:09 +0000 (14:48 +0000)]
R600: Allocate memoperand in the MachienFunction so it doesn't leak.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188555 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUpdating function comments; no functional changes intended.
Aaron Ballman [Fri, 16 Aug 2013 14:33:07 +0000 (14:33 +0000)]
Updating function comments; no functional changes intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188554 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoWhen initializing the PIC global base register on ARM/ELF add pc to fix the address.
Benjamin Kramer [Fri, 16 Aug 2013 12:52:08 +0000 (12:52 +0000)]
When initializing the PIC global base register on ARM/ELF add pc to fix the address.

This unbreaks PIC with fast isel on ELF targets (PR16717). The output matches
what GCC and SDag do for PIC but may not cover all of the many flavors of PIC
that exist.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188551 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support for Thumb2 literal loads with negative zero offset
Mihai Popa [Fri, 16 Aug 2013 12:03:00 +0000 (12:03 +0000)]
Add support for Thumb2 literal loads with negative zero offset

Thumb2 literal loads use an offset encoding which allows for
negative zero. This fixes parsing and encoding so that #-0
is correctly processed. The parser represents #-0 as INT32_MIN.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188549 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix Thumb2 aliasing complementary instructions taking modified immediates
Mihai Popa [Fri, 16 Aug 2013 11:55:44 +0000 (11:55 +0000)]
Fix Thumb2 aliasing complementary instructions taking modified immediates

There are many Thumb instructions which take 12-bit immediates encoded in a special
8-byte value + 4-byte rotator form. Not all numbers are represented, and it's legal
to transform an assembly instruction to be able to encode the immediate.

For example: AND and BIC are complementary instructions; one can switch the AND
to a BIC as long as the immediate is complemented.

The intent is to switch one instruction into its complementary one when the immediate
cannot be encoded in the form requested in the original assembly and when the
complementary immediate is encodable.

The patch addresses two issues:
1. definition of t2SOImmNot immediate - it has to check that the orignal value is
not encoded naturally
2. t2AND and t2BIC instruction aliases which should use the Thumb2 SOImm operand
rather than the ARM one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@188548 91177308-0d34-0410-b5e6-96231b3b80d8