oota-llvm.git
10 years agoRemove projects/sample.
Rafael Espindola [Wed, 12 Mar 2014 22:40:22 +0000 (22:40 +0000)]
Remove projects/sample.

As an example that was not actually being used, it suffered from a slow bitrot.

The two main issues with it were that it had no cmake support and
included a copy of the autoconf directory. The reality is that
autoconf is not easily composable. The lack of composabilty is why we
have clang options in llvm's configure. Suggesting that users include
a copy of autoconf/ in their projects seems a bad idea.

We are also in the process of switching to cmake, so pushing autoconf
to new project is probably not what we want.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203728 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoMCDwarf: Remove unused parameter
David Blaikie [Wed, 12 Mar 2014 22:35:23 +0000 (22:35 +0000)]
MCDwarf: Remove unused parameter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203727 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoMCDwarf: Invert the Section+CU->LineEntries mapping so the CU is the primary dimension
David Blaikie [Wed, 12 Mar 2014 22:28:56 +0000 (22:28 +0000)]
MCDwarf: Invert the Section+CU->LineEntries mapping so the CU is the primary dimension

This makes the mapping consistent with other CU->X mappings in the
MCContext, helping pave the way to refactor all these values into a
single data structure per CU and thus a single map.

I haven't renamed the data structure as that would make the patch churn
even higher (the MCLineSection name no longer makes sense, as this
structure now contains lines for multiple sections covered by a single
CU, rather than lines for a single section in multiple CUs) and further
refactorings will follow that may remove this type entirely.

For convenience, I also gave the MCLineSection value semantics so we
didn't have to do the lazy construction, manual delete, etc.

(& for those playing at home, refactoring the line printing into a
single data structure will eventually alow that data structure to be
reused to own the debug_line.dwo line table used for type unit file name
resolution)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203726 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoThis test need the X86 backend, move it to the X86 sub directory.
Rafael Espindola [Wed, 12 Mar 2014 22:03:43 +0000 (22:03 +0000)]
This test need the X86 backend, move it to the X86 sub directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203725 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoBack out Profile library and dependent commits
Justin Bogner [Wed, 12 Mar 2014 22:00:57 +0000 (22:00 +0000)]
Back out Profile library and dependent commits

Chandler voiced some concern with checking this in without some
discussion first. Reverting for now.

This reverts r203703, r203704, r203708, and 203709.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203723 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoPR17473:
Michael Zolotukhin [Wed, 12 Mar 2014 21:31:05 +0000 (21:31 +0000)]
PR17473:
Don't normalize an expression during postinc transformation unless it's
invertible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203719 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[X86] Add peephole for masked rotate amount
Adam Nemet [Wed, 12 Mar 2014 21:20:55 +0000 (21:20 +0000)]
[X86] Add peephole for masked rotate amount

Extend what's currently done for shift because the HW performs this masking
implicitly:

   (rotl:i32 x, (and y, 31)) -> (rotl:i32 x, y)

I use the newly factored out multiclass that was only supporting shifts so
far.

For testing I extended my testcase for the new rotation idiom.

<rdar://problem/15295856>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203718 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFix the ocaml test to not create a alias to a declaration.
Rafael Espindola [Wed, 12 Mar 2014 21:20:42 +0000 (21:20 +0000)]
Fix the ocaml test to not create a alias to a declaration.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203717 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoTest commit
Michael Zolotukhin [Wed, 12 Mar 2014 21:15:56 +0000 (21:15 +0000)]
Test commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203716 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoProfile: Avoid an unnecessary __attribute__((packed))
Justin Bogner [Wed, 12 Mar 2014 20:40:14 +0000 (20:40 +0000)]
Profile: Avoid an unnecessary __attribute__((packed))

MSVC doesn't understand it, and it wasn't really necessary anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203709 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoProfile: Remove an inefficient and unnecessary API function
Justin Bogner [Wed, 12 Mar 2014 20:26:37 +0000 (20:26 +0000)]
Profile: Remove an inefficient and unnecessary API function

This was leftover from an approach I abandoned, but I forgot to update
it before committing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203708 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoResubmit "[SLPV] Recognize vectorizable intrinsics during SLP vectorization ..."
Raul E. Silvera [Wed, 12 Mar 2014 20:21:50 +0000 (20:21 +0000)]
Resubmit "[SLPV] Recognize vectorizable intrinsics during SLP vectorization ..."

This reverts commit 86cb795388643710dab34941ddcb5a9470ac39d8.
The problems previously found have been resolved through other CLs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203707 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoAdd a triple to fix the test on OS X.
Rafael Espindola [Wed, 12 Mar 2014 20:21:35 +0000 (20:21 +0000)]
Add a triple to fix the test on OS X.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203706 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoReject alias to undefined symbols in the verifier.
Rafael Espindola [Wed, 12 Mar 2014 20:15:49 +0000 (20:15 +0000)]
Reject alias to undefined symbols in the verifier.

On ELF and COFF an alias is just another name for a position in the file.
There is no way to refer to a position in another file, so an alias to
undefined is meaningless.

MachO currently doesn't support aliases. The spec has a N_INDR, which when
implemented will have a different set of restrictions. Adding support for
it shouldn't be harder than any other IR extension.

For now, having the IR represent what is actually possible with current
tools makes it easier to fix the design of GlobalAlias.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203705 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agollvm-profdata: Use the Profile library, implement show and generate
Justin Bogner [Wed, 12 Mar 2014 20:14:17 +0000 (20:14 +0000)]
llvm-profdata: Use the Profile library, implement show and generate

This replaces the llvm-profdata tool with a version that uses the
recently introduced Profile library. The new tool has the ability to
generate and summarize profdata files as well as merging them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203704 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoProfile: Add a library for the instrumentation based profiling format
Justin Bogner [Wed, 12 Mar 2014 20:14:05 +0000 (20:14 +0000)]
Profile: Add a library for the instrumentation based profiling format

This provides a library to work with the instrumentation based
profiling format that is used by clang's -fprofile-instr-* options and
by the llvm-profdata tool. This is a binary format, rather than the
textual one that's currently in use.

The tests are in the subsequent commits that use this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203703 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFix two thinkos in testcase and remove XFAIL.
Eric Christopher [Wed, 12 Mar 2014 20:12:02 +0000 (20:12 +0000)]
Fix two thinkos in testcase and remove XFAIL.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203702 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoUse -std=gnu++11 on cygwin and mingw.
Rafael Espindola [Wed, 12 Mar 2014 20:01:15 +0000 (20:01 +0000)]
Use -std=gnu++11 on cygwin and mingw.

Without this common features like off_t and strdup are missing.
This should bring back those bots.

Configure bits by Meador Inge.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203701 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoAllow exclamation and tilde to be parsed as a part of the ppc asm operand.
Roman Divacky [Wed, 12 Mar 2014 19:25:57 +0000 (19:25 +0000)]
Allow exclamation and tilde to be parsed as a part of the ppc asm operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203699 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoXFAIL this temporarily.
Eric Christopher [Wed, 12 Mar 2014 19:06:03 +0000 (19:06 +0000)]
XFAIL this temporarily.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203698 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoMove test to X86 only for now.
Eric Christopher [Wed, 12 Mar 2014 19:02:44 +0000 (19:02 +0000)]
Move test to X86 only for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203697 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoR600: Fix trunc store from i64 to i1
Matt Arsenault [Wed, 12 Mar 2014 18:45:52 +0000 (18:45 +0000)]
R600: Fix trunc store from i64 to i1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203695 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoAllow switch-to-lookup table for tables with holes by adding bitmask check
Hans Wennborg [Wed, 12 Mar 2014 18:35:40 +0000 (18:35 +0000)]
Allow switch-to-lookup table for tables with holes by adding bitmask check

This allows us to generate table lookups for code such as:

  unsigned test(unsigned x) {
    switch (x) {
      case 100: return 0;
      case 101: return 1;
      case 103: return 2;
      case 105: return 3;
      case 107: return 4;
      case 109: return 5;
      case 110: return 6;
      default: return f(x);
    }
  }

Since cases 102, 104, etc. are not constants, the lookup table has holes
in those positions. We therefore guard the table lookup with a bitmask check.

Patch by Jasper Neumann!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203694 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoWhen computing the size of a base type be conservative if the type
Eric Christopher [Wed, 12 Mar 2014 18:18:05 +0000 (18:18 +0000)]
When computing the size of a base type be conservative if the type
is a declaration and return the size of the type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203690 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoRevert r203488 and r203520.
Evan Cheng [Wed, 12 Mar 2014 18:09:37 +0000 (18:09 +0000)]
Revert r203488 and r203520.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203687 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoAvoid repeated calls to CE->getOperand(0). No functionality change.
Rafael Espindola [Wed, 12 Mar 2014 18:08:14 +0000 (18:08 +0000)]
Avoid repeated calls to CE->getOperand(0). No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203686 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[X86] Refactor peepholes for masked shift amount into a multiclass
Adam Nemet [Wed, 12 Mar 2014 18:02:33 +0000 (18:02 +0000)]
[X86] Refactor peepholes for masked shift amount into a multiclass

The peephole (shift x, (and y, 31)) -> (shift x, y) is repeated for each
integer type and each shift variant.

To improve this a new multiclass is added that covers all integer types.  The
shift patterns are now instantiated from this.  I am planning to add new
instances for rotates as well.

No functional change intended:

  * test/CodeGen/X86/shift-and.ll provides coverage

  * Compared the expanded tablegen output and matched up the defs for these
    Pat<>s before and after

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203685 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[X86] Set the scheduling resources of some of the FPStack instructions.
Quentin Colombet [Wed, 12 Mar 2014 17:33:42 +0000 (17:33 +0000)]
[X86] Set the scheduling resources of some of the FPStack instructions.
This is related to <rdar://problem/15607571>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203682 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoUse values we've already computed, update comment.
Eric Christopher [Wed, 12 Mar 2014 17:14:46 +0000 (17:14 +0000)]
Use values we've already computed, update comment.
No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203681 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoTurn on hashing by default for split dwarf compile units.
Eric Christopher [Wed, 12 Mar 2014 17:14:43 +0000 (17:14 +0000)]
Turn on hashing by default for split dwarf compile units.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203680 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agounittests: Fix -Werror build
Justin Bogner [Wed, 12 Mar 2014 17:00:52 +0000 (17:00 +0000)]
unittests: Fix -Werror build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203679 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoCorrect typo ("a entry" -> "an entry")
David Blaikie [Wed, 12 Mar 2014 16:56:05 +0000 (16:56 +0000)]
Correct typo ("a entry" -> "an entry")

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203678 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoTry harder to evaluate expressions when printing assembly.
Rafael Espindola [Wed, 12 Mar 2014 16:55:59 +0000 (16:55 +0000)]
Try harder to evaluate expressions when printing assembly.

When printing assembly we don't have a Layout object, but we can still
try to fold some constants.

Testcase by Ulrich Weigand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203677 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoDebugInfo: Use common line/file attribute construction code
David Blaikie [Wed, 12 Mar 2014 16:51:06 +0000 (16:51 +0000)]
DebugInfo: Use common line/file attribute construction code

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203676 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoRevive SizeOptLevel-explaining comments that were dropped in r203669
Eli Bendersky [Wed, 12 Mar 2014 16:44:17 +0000 (16:44 +0000)]
Revive SizeOptLevel-explaining comments that were dropped in r203669

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203675 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoAdd parens around && clauses in a || to appease the compiler.
Eli Bendersky [Wed, 12 Mar 2014 16:14:53 +0000 (16:14 +0000)]
Add parens around && clauses in a || to appease the compiler.

Otherwise gcc 4.8.2 generates a warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203671 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoAdd comment pointing to the binutils bugzilla entry
Hans Wennborg [Wed, 12 Mar 2014 16:14:23 +0000 (16:14 +0000)]
Add comment pointing to the binutils bugzilla entry

This is a follow-up to r203635 as suggested by Rafael.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203670 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoMove duplicated code into a helper function (exposed through overload).
Eli Bendersky [Wed, 12 Mar 2014 16:12:36 +0000 (16:12 +0000)]
Move duplicated code into a helper function (exposed through overload).

There's a bit of duplicated "magic" code in opt.cpp and Clang's CodeGen that
computes the inliner threshold from opt level and size opt level.

This patch moves the code to a function that lives alongside the inliner itself,
providing a convenient overload to the inliner creation.

A separate patch can be committed to Clang to use this once it's committed to
LLVM. Standalone tools that use the inlining pass can also avoid duplicating
this code and fearing it will go out of sync.

Note: this patch also restructures the conditinal logic of the computation to
be cleaner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203669 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoUpdate the datalayout string for ppc64LE.
Will Schmidt [Wed, 12 Mar 2014 14:59:17 +0000 (14:59 +0000)]
Update the datalayout string for ppc64LE.

Update the datalayout string for ppc64LE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203664 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoCloning a function now also clones its debug metadata if 'ModuleLevelChanges' is...
Alon Mishne [Wed, 12 Mar 2014 14:42:51 +0000 (14:42 +0000)]
Cloning a function now also clones its debug metadata if 'ModuleLevelChanges' is true.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203662 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[mips][fp64] Add an implicit def to MTHC1 claiming that it reads the lower 32-bits...
Daniel Sanders [Wed, 12 Mar 2014 13:35:43 +0000 (13:35 +0000)]
[mips][fp64] Add an implicit def to MTHC1 claiming that it reads the lower 32-bits of 64-bit FPR

Summary:
This is a white lie to workaround a widespread bug in the -mfp64
implementation.

The problem is that none of the 32-bit fpu ops mention the fact that they
clobber the upper 32-bits of the 64-bit FPR. This allows MTHC1 to be
scheduled on the wrong side of most 32-bit FPU ops, particularly MTC1.
Fixing that requires a major overhaul of the FPU implementation which can't
be done right now due to time constraints.

The testcase is SingleSource/Benchmarks/Misc/oourafft.c when given
TARGET_CFLAGS='-mips32r2 mfp64 -mmsa'.

Also correct the comment added in r203464 to indicate that two
instructions were affected.

Reviewers: matheusalmeida, jacksprat

Reviewed By: matheusalmeida

Differential Revision: http://llvm-reviews.chandlerc.com/D3029

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203659 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[mips] BSEL's and BINS[RL] operands are reversed compared to the vselect node used...
Daniel Sanders [Wed, 12 Mar 2014 11:54:00 +0000 (11:54 +0000)]
[mips] BSEL's and BINS[RL] operands are reversed compared to the vselect node used in the pattern.

Summary:
Correct the match patterns and the lowerings that made the CodeGen tests pass despite the mistakes.

The original testcase that discovered the problem was SingleSource/UnitTests/SignlessType/factor.c in test-suite.
During review, we also found that some of the existing CodeGen tests were incorrect and fixed them:
* bitwise.ll: In bsel_v16i8 the IfSet/IfClear were reversed because bsel and bmnz have different operand orders and the test didn't correctly account for this. bmnz goes 'IfClear, IfSet, CondMask', while bsel goes 'CondMask, IfClear, IfSet'.
* vec.ll: In the cases where a bsel is emitted as a bmnz (they are the same operation with a different input tied to the result) the operands were in the wrong order.
* compare.ll and compare_float.ll: The bsel operand order was correct for a greater-than comparison, but a greater-than comparison instruction doesn't exist. Lowering this operation inverts the condition so the IfSet/IfClear need to be swapped to match.

The differences between BSEL, BMNZ, and BMZ and how they map to/from vselect are rather confusing. I've therefore added a note to MSA.txt to explain this in a single place in addition to the comments that explain each case.

Reviewers: matheusalmeida, jacksprat

Reviewed By: matheusalmeida

Differential Revision: http://llvm-reviews.chandlerc.com/D3028

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203657 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoARM: correct Dwarf output for non-contiguous VFP saves.
Tim Northover [Wed, 12 Mar 2014 11:29:23 +0000 (11:29 +0000)]
ARM: correct Dwarf output for non-contiguous VFP saves.

When the list of VFP registers to be saved was non-contiguous (so multiple
vpush/vpop instructions were needed) these were being ordered oddly, as in:
    vpush {d8, d9}
    vpush {d11}

This led to the layout in memory being [d11, d8, d9] which is ugly and doesn't
match the CFI_INSTRUCTIONs we're generating either (so Dwarf info would be
broken).

This switches the order of vpush/vpop (in both prologue and epilogue,
obviously) so that the Dwarf locations are correct again.

rdar://problem/16264856

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203655 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoReplace '#include ValueTypes.h' with forward declarations.
Patrik Hagglund [Wed, 12 Mar 2014 08:00:24 +0000 (08:00 +0000)]
Replace '#include ValueTypes.h' with forward declarations.

In some cases the include is pushed "downstream" (or removed if
unused).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203644 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[ARM] Use DWARF register numbers for CFI directives in ELF assembly
Hans Wennborg [Wed, 12 Mar 2014 03:52:34 +0000 (03:52 +0000)]
[ARM] Use DWARF register numbers for CFI directives in ELF assembly

It seems gas can't handle CFI directives with VFP register names ("d12", etc.).
This broke us trying to build Chromium for Android after 201423.

A gas bug has been filed: https://sourceware.org/bugzilla/show_bug.cgi?id=16694

compnerd suggested making this conditional on whether we're using the integrated
assembler or not. I'll look into that in a follow-up patch.

Differential Revision: http://llvm-reviews.chandlerc.com/D3049

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203635 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoDebugInfo: Omit pubnames/pubtypes when compiling with -gmlt
David Blaikie [Wed, 12 Mar 2014 03:34:38 +0000 (03:34 +0000)]
DebugInfo: Omit pubnames/pubtypes when compiling with -gmlt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203634 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoDebugInfo: Do not emit pubnames/pubtypes sections if they are empty
David Blaikie [Tue, 11 Mar 2014 23:35:06 +0000 (23:35 +0000)]
DebugInfo: Do not emit pubnames/pubtypes sections if they are empty

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203622 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoTest for empty pubnames/pubtypes
David Blaikie [Tue, 11 Mar 2014 23:35:03 +0000 (23:35 +0000)]
Test for empty pubnames/pubtypes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203621 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoDebugInfo: Avoid re-looking up the DwarfUnit when emitting pubnames/pubtypes
David Blaikie [Tue, 11 Mar 2014 23:23:39 +0000 (23:23 +0000)]
DebugInfo: Avoid re-looking up the DwarfUnit when emitting pubnames/pubtypes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203620 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoDebugInfo: Refactor emitDebugPubNames/Types into a common implementation
David Blaikie [Tue, 11 Mar 2014 23:18:15 +0000 (23:18 +0000)]
DebugInfo: Refactor emitDebugPubNames/Types into a common implementation

I could fold the callers into their one call site, but the indirection
(given how verbose choosing the section is) seemed helpful.

The use of a member function pointer's a bit "tricky", but seems limited
enough, the call sites are simple/clean/clear, and there's only one use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203619 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoAccept Twine's to AsmPrinter::getTempSymbol (refactoring for an incoming change)
David Blaikie [Tue, 11 Mar 2014 23:12:08 +0000 (23:12 +0000)]
Accept Twine's to AsmPrinter::getTempSymbol (refactoring for an incoming change)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203617 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoClean up test/DebugInfo/empty.ll now that we have an alias for "llc with dwarf output"
David Blaikie [Tue, 11 Mar 2014 22:46:12 +0000 (22:46 +0000)]
Clean up test/DebugInfo/empty.ll now that we have an alias for "llc with dwarf output"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203616 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agomove WITH_POLLY option before add_subdirectory(tools)
Sebastian Pop [Tue, 11 Mar 2014 22:42:07 +0000 (22:42 +0000)]
move WITH_POLLY option before add_subdirectory(tools)

the first run of the polly buildbot failed, and then it started passing.
This is due to the fact that the buildbot re-builds in an existing directory,
and the first run does not have WITH_POLLY set when it enters tools/.
Thus, cmake ignores the tools/polly dir in the first run, and then because
it reuses the CMakeCache.txt of the previous run, it has the WITH_POLLY set
by the previous run, and so it passes the second time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203615 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoDebugInfo: Remove unused labels now that we just emit DW_AT_gnu_pubnames as a flag...
David Blaikie [Tue, 11 Mar 2014 22:24:33 +0000 (22:24 +0000)]
DebugInfo: Remove unused labels now that we just emit DW_AT_gnu_pubnames as a flag (as of r203082)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203612 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agosupport: add a utility function to normalise path separators
Saleem Abdulrasool [Tue, 11 Mar 2014 22:05:42 +0000 (22:05 +0000)]
support: add a utility function to normalise path separators

Add a utility function to convert the Windows path separator to Unix style path
separators.  This is used by a subsequent change in clang to enable the use of
Windows SDK headers on Linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203611 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[mips] Implement NaCl sandboxing of function calls:
Sasa Stankovic [Tue, 11 Mar 2014 21:23:40 +0000 (21:23 +0000)]
[mips] Implement NaCl sandboxing of function calls:

  * Add masking instructions before indirect calls (in MC layer).
  * Align call + branch delay to the bundle end (in MC layer).

Differential Revision: http://llvm-reviews.chandlerc.com/D3032

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203606 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoSimplify a really complicated check for Arch == X86_64.
Rafael Espindola [Tue, 11 Mar 2014 21:22:57 +0000 (21:22 +0000)]
Simplify a really complicated check for Arch == X86_64.

The function hasReliableSymbolDifference had exactly one use in the MachO
writer. It is also only true for X86_64. In fact, the comments refers to
"Darwin x86_64" and everything else, so this makes the code match the
comment.

If this is to be abstracted again, it should be a property of
TargetObjectWriter, like useAggressiveSymbolFolding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203605 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoRemove dead code.
Rafael Espindola [Tue, 11 Mar 2014 21:10:44 +0000 (21:10 +0000)]
Remove dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203604 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoCleanup the interface for creating soft or hard links.
Rafael Espindola [Tue, 11 Mar 2014 18:40:24 +0000 (18:40 +0000)]
Cleanup the interface for creating soft or hard links.

Before this patch the unix code for creating hardlinks was unused. The code
for creating symbolic links was implemented in lib/Support/LockFileManager.cpp
and the code for creating hard links in lib/Support/*/Path.inc.

The only use we have for these is in LockFileManager.cpp and it can use both
soft and hard links. Just have a create_link function that creates one or the
other depending on the platform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203596 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoDon't assume an empty stderr.
Rafael Espindola [Tue, 11 Mar 2014 18:25:33 +0000 (18:25 +0000)]
Don't assume an empty stderr.

GuardMalloc can print info to stderr, causing these tests to fail.
Since FileCheck errors on empty inputs, just add a bit of dummy
data to make it happy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203595 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoRange-ify a loop.
Owen Anderson [Tue, 11 Mar 2014 17:37:48 +0000 (17:37 +0000)]
Range-ify a loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203590 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoX86: Don't generate 64-bit movd after cmpneqsd in 32-bit mode (PR19059)
Hans Wennborg [Tue, 11 Mar 2014 15:49:24 +0000 (15:49 +0000)]
X86: Don't generate 64-bit movd after cmpneqsd in 32-bit mode (PR19059)

This fixes the bug where we would bitcast the 64-bit floating point result
of cmpneqsd to a 64-bit integer even on 32-bit targets.

Differential Revision: http://llvm-reviews.chandlerc.com/D3009

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203581 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[ppc64] Patch in TOC restore code after all external function calls
Ulrich Weigand [Tue, 11 Mar 2014 15:26:27 +0000 (15:26 +0000)]
[ppc64] Patch in TOC restore code after all external function calls

When resolving a function call to an external routine, the dynamic
loader must patch the "nop" after the branch instruction to a load
that restores the TOC register.

Current code does that, but only with the *first* instance of a call
to any particular external routine, i.e. at the point where it also
allocates the call stub.  With subsequent calls to the same routine,
current code neglects to patch in the TOC restore code.  This is a
bug, and leads to corrupt TOC pointers in those cases.

Fixed by patching in restore code every time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203580 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoARM: honour -f{no-,}optimize-sibling-calls
Saleem Abdulrasool [Tue, 11 Mar 2014 15:09:54 +0000 (15:09 +0000)]
ARM: honour -f{no-,}optimize-sibling-calls

Use the options in the ARMISelLowering to control whether tail calls are
optimised or not.  Previously, this option was entirely ignored on the ARM
target and only honoured on x86.

This option is mostly useful in profiling scenarios.  The default remains that
tail call optimisations will be applied.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203577 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoARM: remove ancient -arm-tail-calls option
Saleem Abdulrasool [Tue, 11 Mar 2014 15:09:49 +0000 (15:09 +0000)]
ARM: remove ancient -arm-tail-calls option

This option is from 2010, designed to work around a linker issue on Darwin for
ARM.  According to grosbach this is no longer an issue and this option can
safely be removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203576 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoARM: enable tail call optimisation on Thumb 2
Saleem Abdulrasool [Tue, 11 Mar 2014 15:09:44 +0000 (15:09 +0000)]
ARM: enable tail call optimisation on Thumb 2

Tail call optimisation was previously disabled on all targets other than
iOS5.0+.  This enables the tail call optimisation on all Thumb 2 capable
platforms.

The test adjustments are to remove the IR hint "tail" to function invocation.
The tests were designed assuming that tail call optimisations would not kick in
which no longer holds true.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203575 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFix crash in PRE.
Erik Verbruggen [Tue, 11 Mar 2014 15:07:32 +0000 (15:07 +0000)]
Fix crash in PRE.

After r203553 overflow intrinsics and their non-intrinsic (normal)
instruction get hashed to the same value. This patch prevents PRE from
moving an instruction into a predecessor block, and trying to add a phi
node that gets two different types (the intrinsic result and the
non-intrinsic result), resulting in a failing assert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203574 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[lit] Fix non-function style print statement.
Daniel Dunbar [Tue, 11 Mar 2014 14:05:49 +0000 (14:05 +0000)]
[lit] Fix non-function style print statement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203573 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoARM: simplify EmitAtomicBinary64
Tim Northover [Tue, 11 Mar 2014 13:19:55 +0000 (13:19 +0000)]
ARM: simplify EmitAtomicBinary64

ATOMIC_STORE operations always get here as a lowered ATOMIC_SWAP, so there's no
need for any code to handle them specially.

There should be no functionality change so no tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203567 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoRemove copy ctors that did the same thing as the default one.
Benjamin Kramer [Tue, 11 Mar 2014 11:32:49 +0000 (11:32 +0000)]
Remove copy ctors that did the same thing as the default one.

The code added nothing but potentially disabled move semantics and made
types non-trivially copyable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203563 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoIR: add a second ordering operand to cmpxhg for failure
Tim Northover [Tue, 11 Mar 2014 10:48:52 +0000 (10:48 +0000)]
IR: add a second ordering operand to cmpxhg for failure

The syntax for "cmpxchg" should now look something like:

cmpxchg i32* %addr, i32 42, i32 3 acquire monotonic

where the second ordering argument gives the required semantics in the case
that no exchange takes place. It should be no stronger than the first ordering
constraint and cannot be either "release" or "acq_rel" (since no store will
have taken place).

rdar://problem/15996804

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203559 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoGVN: fix hashing of extractvalue.
Erik Verbruggen [Tue, 11 Mar 2014 10:21:30 +0000 (10:21 +0000)]
GVN: fix hashing of extractvalue.

My last commit did not add the indexes to the hashed value for
extractvalue. Adding that back in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203558 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoGVN: merge overflow intrinsics with non-overflow instructions.
Erik Verbruggen [Tue, 11 Mar 2014 09:36:48 +0000 (09:36 +0000)]
GVN: merge overflow intrinsics with non-overflow instructions.

When an overflow intrinsic is followed by a non-overflow instruction,
replace the latter with an extract. For example:

  %sadd = tail call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
  %sadd3 = add i32 %a, %b

Here the add statement will be replaced by an extract.

When an overflow intrinsic follows a non-overflow instruction, a clone
of the intrinsic is inserted before the normal instruction, which makes
it the same as the previous case. Subsequent runs of GVN can then clean
up the duplicate instructions and insert the extract.

This fixes PR8817.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203553 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFix fixme: remove unused method.
Erik Verbruggen [Tue, 11 Mar 2014 09:32:50 +0000 (09:32 +0000)]
Fix fixme: remove unused method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203552 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoObject: rename ARMV7 to ARMNT
Saleem Abdulrasool [Tue, 11 Mar 2014 03:08:37 +0000 (03:08 +0000)]
Object: rename ARMV7 to ARMNT

The official specifications state the name to be ARMNT (as per the Microsoft
Portable Executable and Common Object Format Specification v8.3).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203530 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoCleanup whitespace
Duncan P. N. Exon Smith [Tue, 11 Mar 2014 02:44:45 +0000 (02:44 +0000)]
Cleanup whitespace

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203529 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoR600: Calculate store mask instead of using switch.
Matt Arsenault [Tue, 11 Mar 2014 01:38:53 +0000 (01:38 +0000)]
R600: Calculate store mask instead of using switch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203527 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoAdd helpers for getting scalar sizes of vector value types.
Matt Arsenault [Tue, 11 Mar 2014 01:38:48 +0000 (01:38 +0000)]
Add helpers for getting scalar sizes of vector value types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203526 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoX86: Enable ISel of 16-bit MOVBE instructions.
Jim Grosbach [Tue, 11 Mar 2014 00:44:14 +0000 (00:44 +0000)]
X86: Enable ISel of 16-bit MOVBE instructions.

When the MOVBE instructions are available, use them for 16-bit endian
swapping as well as for 32 and 64 bit.

The patterns were already present on the instructions, but weren't being
matched because the operation was unconditionally marked to 'Expand.'
Change that to be conditional on whether the MOVBE instructions are
available. Use 'rolw' to implement the in-register version (32 and 64
bit have the dedicated 'bswap' instruction for that).

Patch by Louis Gerbarg <lgg@apple.com>.

rdar://15479984

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203524 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoMultiJITTest.cpp: Tweak getPointerToNamedFunction() to be aware of also Windows x64.
NAKAMURA Takumi [Tue, 11 Mar 2014 00:34:38 +0000 (00:34 +0000)]
MultiJITTest.cpp: Tweak getPointerToNamedFunction() to be aware of also Windows x64.

In import thunk, jmp is:
  - On x86, 0xFF 0x25 [disp32].
  - On x64, 0xFF 0x25 [pcrel32].

See also my r144178.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203523 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFollow up to r203488. Code clean up to eliminate a lot of copy+paste.
Evan Cheng [Tue, 11 Mar 2014 00:24:20 +0000 (00:24 +0000)]
Follow up to r203488. Code clean up to eliminate a lot of copy+paste.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203520 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFix undefined behavior in vector shift tests.
Matt Arsenault [Tue, 11 Mar 2014 00:01:41 +0000 (00:01 +0000)]
Fix undefined behavior in vector shift tests.

These were all shifting the same amount as the bitwidth.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203519 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoRemove incomplete comment
Matt Arsenault [Tue, 11 Mar 2014 00:01:37 +0000 (00:01 +0000)]
Remove incomplete comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203518 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoMove trivial getter into header.
Matt Arsenault [Tue, 11 Mar 2014 00:01:34 +0000 (00:01 +0000)]
Move trivial getter into header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203517 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoUse .data() instead of &x[0]
Matt Arsenault [Tue, 11 Mar 2014 00:01:31 +0000 (00:01 +0000)]
Use .data() instead of &x[0]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203516 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFix indentation
Matt Arsenault [Tue, 11 Mar 2014 00:01:27 +0000 (00:01 +0000)]
Fix indentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203515 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFix non 2-space indentation.
Matt Arsenault [Tue, 11 Mar 2014 00:01:25 +0000 (00:01 +0000)]
Fix non 2-space indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203514 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoModule: Don't rename in getOrInsertFunction()
Duncan P. N. Exon Smith [Mon, 10 Mar 2014 23:42:28 +0000 (23:42 +0000)]
Module: Don't rename in getOrInsertFunction()

During LTO, user-supplied definitions of C library functions often
exist.  -instcombine uses Module::getOrInsertFunction() to get a handle
on library functions (e.g., @puts, when optimizing @printf).

Previously, Module::getOrInsertFunction() would rename any matching
functions with local linkage, and create a new declaration.  In LTO,
this is the opposite of desired behaviour, as it skips by the
user-supplied version of the library function and creates a new
undefined reference which the linker often cannot resolve.

After some discussing with Rafael on the list, it looks like it's
undesired behaviour.  If a consumer actually *needs* this behaviour, we
should add new API with a more explicit name.

I added two testcases: one specifically for the -instcombine behaviour
and one for the LTO flow.

<rdar://problem/16165191>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203513 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoWhen analyzing vectors of element type that require legalization,
Raul E. Silvera [Mon, 10 Mar 2014 22:59:13 +0000 (22:59 +0000)]
When analyzing vectors of element type that require legalization,
the legalization cost must be included to get an accurate
estimation of the total cost of the scalarized vector.
The inaccurate cost triggered unprofitable SLP vectorization on
32-bit X86.

Summary:
Include legalization overhead when computing scalarization cost

Reviewers: hfinkel, nadav

CC: chandlerc, rnk, llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2992

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203509 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoUse discriminator information in sample profiles.
Diego Novillo [Mon, 10 Mar 2014 22:41:28 +0000 (22:41 +0000)]
Use discriminator information in sample profiles.

Summary:
When the sample profiles include discriminator information,
use the discriminator values to distinguish instruction weights
in different basic blocks.

This modifies the BodySamples mapping to map <line, discriminator> pairs
to weights. Instructions on the same line but different blocks, will
use different discriminator values. This, in turn, means that the blocks
may have different weights.

Other changes in this patch:

- Add tests for positive values of line offset, discriminator and samples.
- Change data types from uint32_t to unsigned and int and do additional
  validation.

Reviewers: chandlerc

CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2857

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203508 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoTest commit: Remove trailing whitespace.
Manuel Jacob [Mon, 10 Mar 2014 22:24:07 +0000 (22:24 +0000)]
Test commit: Remove trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203502 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFix a couple typos.
Mark Lacey [Mon, 10 Mar 2014 21:59:28 +0000 (21:59 +0000)]
Fix a couple typos.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203499 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[lit] Bump dev version number.
Daniel Dunbar [Mon, 10 Mar 2014 21:58:16 +0000 (21:58 +0000)]
[lit] Bump dev version number.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203498 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[lit] Add a README.txt.
Daniel Dunbar [Mon, 10 Mar 2014 21:58:12 +0000 (21:58 +0000)]
[lit] Add a README.txt.
 - Also, update MANIFEST.in and utils/check-sdist.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203497 91177308-0d34-0410-b5e6-96231b3b80d8

10 years ago[lit] Add --version option.
Daniel Dunbar [Mon, 10 Mar 2014 21:57:48 +0000 (21:57 +0000)]
[lit] Add --version option.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203496 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agofix polly buildbot
Sebastian Pop [Mon, 10 Mar 2014 21:27:04 +0000 (21:27 +0000)]
fix polly buildbot

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203492 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoIR: Slightly more verbose error in Verifier
Justin Bogner [Mon, 10 Mar 2014 21:22:44 +0000 (21:22 +0000)]
IR: Slightly more verbose error in Verifier

Extend the error message generated by the Verifier when an intrinsic
name does not match the expected mangling to include the expected
name.  Simplifies debugging.

Patch by Philip Reames!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203490 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoMemCpyOpt: When merging memsets also merge the trivial case of two memsets with the...
Benjamin Kramer [Mon, 10 Mar 2014 21:05:13 +0000 (21:05 +0000)]
MemCpyOpt: When merging memsets also merge the trivial case of two memsets with the same destination.

The testcase is from PR19092, but I think the bug described there is actually a clang issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203489 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agoFor functions with ARM target specific calling convention, when simplify-libcall
Evan Cheng [Mon, 10 Mar 2014 20:49:45 +0000 (20:49 +0000)]
For functions with ARM target specific calling convention, when simplify-libcall
optimize a call to a llvm intrinsic to something that invovles a call to a C
library call, make sure it sets the right calling convention on the call.

e.g.
extern double pow(double, double);
double t(double x) {
  return pow(10, x);
}

Compiles to something like this for AAPCS-VFP:
define arm_aapcs_vfpcc double @t(double %x) #0 {
entry:
  %0 = call double @llvm.pow.f64(double 1.000000e+01, double %x)
  ret double %0
}

declare double @llvm.pow.f64(double, double) #1

Simplify libcall (part of instcombine) will turn the above into:
define arm_aapcs_vfpcc double @t(double %x) #0 {
entry:
  %__exp10 = call double @__exp10(double %x) #1
  ret double %__exp10
}

declare double @__exp10(double)

The pre-instcombine code works because calls to LLVM builtins are special.
Instruction selection will chose the right calling convention for the call.
However, the code after instcombine is wrong. The call to __exp10 will use
the C calling convention.

I can think of 3 options to fix this.

1. Make "C" calling convention just work since the target should know what CC
   is being used.

   This doesn't work because each function can use different CC with the "pcs"
   attribute.

2. Have Clang add the right CC keyword on the calls to LLVM builtin.

   This will work but it doesn't match the LLVM IR specification which states
   these are "Standard C Library Intrinsics".

3. Fix simplify libcall so the resulting calls to the C routines will have the
   proper CC keyword. e.g.
   %__exp10 = call arm_aapcs_vfpcc double @__exp10(double %x) #1

   This works and is the solution I implemented here.

Both solutions #2 and #3 would work. After carefully considering the pros and
cons, I decided to implement #3 for the following reasons.

1. It doesn't change the "spec" of the intrinsics.
2. It's a self-contained fix.

There are a couple of potential downsides.
1. There could be other places in the optimizer that is broken in the same way
   that's not addressed by this.
2. There could be other calling conventions that need to be propagated by
   simplify-libcall that's not handled.

But for now, this is the fix that I'm most comfortable with.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203488 91177308-0d34-0410-b5e6-96231b3b80d8

10 years agofix PR13550: add a cmake WITH_POLLY option
Sebastian Pop [Mon, 10 Mar 2014 20:47:39 +0000 (20:47 +0000)]
fix PR13550: add a cmake WITH_POLLY option

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203486 91177308-0d34-0410-b5e6-96231b3b80d8