Commit Graph

6773 Commits

Author SHA1 Message Date
Nadav Rotem 977e0be4a0 Efficient lowering of vector sdiv when the divisor is a splatted power of two constant.
PR 14848. The lowered sequence is based on the existing sequence the target-independent
DAG Combiner creates for the scalar case.

Patch by Zvi Rackover.

llvm-svn: 171953
2013-01-09 05:14:33 +00:00
Andrew Trick 9f0b95f260 MIsched: add an ILP window property to machine model.
This was an experimental option, but needs to be defined
per-target. e.g. PPC A2 needs to aggressively hide latency.

I converted some in-order scheduling tests to A2. Hal is working on
more test cases.

llvm-svn: 171946
2013-01-09 03:36:49 +00:00
Tim Northover 90fb75d859 Specify complete triple for fp128 tests.
This avoids FileCheck failing over different comment characters in
assembly (notably powerpc64 on Linux vs Darwin) and should fix David's
build-bot.

llvm-svn: 171886
2013-01-08 19:36:33 +00:00
Preston Gurd a01daace88 Pad Short Functions for Intel Atom
The current Intel Atom microarchitecture has a feature whereby
when a function returns early then it is slightly faster to execute
a sequence of NOP instructions to wait until the return address is ready,
as opposed to simply stalling on the ret instruction until
the return address is ready.

When compiling for X86 Atom only, this patch will run a pass,
called "X86PadShortFunction" which will add NOP instructions where less
than four cycles elapse between function entry and return.

It includes tests.

This patch has been updated to address Nadav's review comments
- Optimize only at >= O1 and don't do optimization if -Os is set
- Stores MachineBasicBlock* instead of BBNum
- Uses DenseMap instead of std::map
- Fixes placement of braces

Patch by Andy Zhang.

llvm-svn: 171879
2013-01-08 18:27:24 +00:00
Tim Northover 7bb9992cce Allow the asm printer to print fp128 values properly.
llvm-svn: 171866
2013-01-08 16:56:23 +00:00
Bill Schmidt 9b1e3e25dc This patch addresses bug 14678 by fixing two problems in medium code model
code generation.  Variables addressed through a GlobalAlias were not being
handled, and variables with available_externally linkage were treated
incorrectly.  The patch contains two new tests to verify the correct code
generation for these cases.

llvm-svn: 171778
2013-01-07 19:29:18 +00:00
Silviu Baranga a055aab506 Make the MergeGlobals pass correctly handle the address space qualifiers of the global variables. We partition the set of globals by their address space, and apply the same the trasnformation as before to merge them.
llvm-svn: 171730
2013-01-07 12:31:25 +00:00
Craig Topper 4f1c7256f9 Fix suffix handling for parsing and printing of cvtsi2ss, cvtsi2sd, cvtss2si, cvttss2si, cvtsd2si, and cvttsd2si to match gas behavior.
cvtsi2* should parse with an 'l' or 'q' suffix or no suffix at all. No suffix should be treated the same as 'l' suffix. Printing should always print a suffix. Previously we didn't parse or print an 'l' suffix.
cvtt*2si/cvt*2si should parse with an 'l' or 'q' suffix or not suffix at all. No suffix should use the destination register size to choose encoding. Printing should not print a suffix.

Original 'l' suffix issue with cvtsi2* pointed out by Michael Kuperstein.

llvm-svn: 171668
2013-01-06 20:39:29 +00:00
Evan Cheng 3fb03e23a4 Fix for PR14739. It's not safe to fold a load into a call across a store. Thanks to Nick Lewycky for the initial patch.
llvm-svn: 171665
2013-01-06 19:00:15 +00:00
Craig Topper 92a70b1e65 Recommit r171461 which was incorrectly reverted. Mark DIV/IDIV instructions hasSideEffects=1 because they can trap when dividing by 0. This is needed to keep early if conversion from moving them across basic blocks.
llvm-svn: 171608
2013-01-05 07:39:25 +00:00
Nadav Rotem 478b6a47ec Revert revision 171524. Original message:
URL: http://llvm.org/viewvc/llvm-project?rev=171524&view=rev
Log:
The current Intel Atom microarchitecture has a feature whereby when a function
returns early then it is slightly faster to execute a sequence of NOP
instructions to wait until the return address is ready,
as opposed to simply stalling on the ret instruction
until the return address is ready.

When compiling for X86 Atom only, this patch will run a pass, called
"X86PadShortFunction" which will add NOP instructions where less than four
cycles elapse between function entry and return.

It includes tests.

Patch by Andy Zhang.

llvm-svn: 171603
2013-01-05 05:42:48 +00:00
Preston Gurd e36b685a94 The current Intel Atom microarchitecture has a feature whereby when a function
returns early then it is slightly faster to execute a sequence of NOP
instructions to wait until the return address is ready,
as opposed to simply stalling on the ret instruction
until the return address is ready.

When compiling for X86 Atom only, this patch will run a pass, called
"X86PadShortFunction" which will add NOP instructions where less than four
cycles elapse between function entry and return.

It includes tests.

Patch by Andy Zhang.

llvm-svn: 171524
2013-01-04 20:54:54 +00:00
Akira Hatanaka b13b33359b [mips] MipsTargetLowering::getSetCCResultType should return a vector type if
vectors are being compared.

llvm-svn: 171517
2013-01-04 20:06:01 +00:00
Nadav Rotem c616a5408a Revert revision: 171467. This transformation is incorrect and makes some tests fail. Original message:
Simplified TRUNCATE operation that comes after SETCC. It is possible since SETCC result is 0 or -1.
Added a test.

llvm-svn: 171468
2013-01-04 17:35:21 +00:00
Elena Demikhovsky 5f2f06d2d9 Simplified TRUNCATE operation that comes after SETCC. It is possible since SETCC result is 0 or -1.
Added a test.

llvm-svn: 171467
2013-01-03 08:48:33 +00:00
Michael Gottesman 820aac1c78 Revert "Mark DIV/IDIV instructions hasSideEffects=1 because they can trap when dividing by 0. This is needed to keep early if conversion from moving them across basic blocks."
This reverts commit r171461 since it breaks the following tests:

Clang :: Analysis/outofbound-notwork.c
Clang :: Analysis/string-fail.c
Clang :: CXX/basic/basic.lookup/basic.lookup.qual/p6-0x.cpp
Clang :: CXX/basic/basic.lookup/basic.lookup.unqual/p15.cpp
Clang :: CXX/dcl.dcl/dcl.spec/dcl.fct.spec/p4.cpp
Clang :: CXX/dcl.dcl/dcl.spec/dcl.stc/p10.cpp
Clang :: CXX/temp/temp.param/p14.cpp
Clang :: CXX/temp/temp.res/temp.dep.res/temp.point/p1.cpp
Clang :: CodeGen/2009-02-13-zerosize-union-field-ppc.c
Clang :: CodeGen/blocks-2.c
Clang :: CodeGen/libcalls-d.c
Clang :: CodeGen/libcalls-ld.c
Clang :: CodeGenCXX/conversion-function.cpp
Clang :: CodeGenCXX/debug-info-limit-type.cpp
Clang :: CodeGenCXX/inheriting-constructor.cpp
Clang :: FixIt/fixit-errors.c
Clang :: FixIt/fixit-pmem.cpp
Clang :: Modules/namespaces.cpp
Clang :: PCH/changed-files.c
Clang :: PCH/pr4489.c
Clang :: PCH/source-manager-stack.c
Clang :: Parser/cxx-ambig-decl-expr-xfail.cpp
Clang :: SemaCXX/switch-implicit-fallthrough-cxx98.cpp
Clang :: SemaTemplate/instantiate-function-1.mm

llvm-svn: 171466
2013-01-03 08:18:30 +00:00
Craig Topper 7c27cc9fd0 Mark DIV/IDIV instructions hasSideEffects=1 because they can trap when dividing by 0. This is needed to keep early if conversion from moving them across basic blocks.
llvm-svn: 171461
2013-01-03 06:40:20 +00:00
Jakob Stoklund Olesen 725d57682b Fix PR14732 by handling all kinds of IMPLICIT_DEF live ranges.
Most IMPLICIT_DEF instructions are removed by the ProcessImplicitDefs
pass, and a few are reinserted by PHIElimination when a PHI argument is
<undef>.

RegisterCoalescer was assuming that all IMPLICIT_DEF live ranges look
like those created by PHIElimination, and that their live range never
leaves the basic block.

The PR14732 test case does tricks with PHI nodes that causes a longer
IMPLICIT_DEF live range to appear. This happens very rarely, but
RegisterCoalescer should be able to handle it.

llvm-svn: 171435
2013-01-03 00:47:51 +00:00
Tom Stellard 567f886eb0 DAGCombiner: Avoid generating illegal vector INT_TO_FP nodes
DAGCombiner::reduceBuildVecConvertToConvertBuildVec() was making two
mistakes:

1. It was checking the legality of scalar INT_TO_FP nodes and then generating
vector nodes.

2. It was passing the result value type to
TargetLoweringInfo::getOperationAction() when it should have been
passing the value type of the first operand.

llvm-svn: 171420
2013-01-02 22:13:01 +00:00
Nadav Rotem c8d7047fa9 AVX: Fix a bug in WidenMaskArithmetic.
llvm-svn: 171397
2013-01-02 17:40:39 +00:00
Hal Finkel 6dbdd4307b Support ppcf128 in SelectionDAG::getConstantFP
Fixes pr14751.

Patch by Kai; Thanks!

llvm-svn: 171261
2012-12-30 19:03:32 +00:00
Dmitri Gribenko 56bf2e1830 Tests: rewrite 'opt ... %s' to 'opt ... < %s' so that opt does not emit a ModuleID
This is done to avoid odd test failures, like the one fixed in r171243.

llvm-svn: 171250
2012-12-30 02:33:22 +00:00
Nadav Rotem 3da9ac72fa AVX: Move the ZEXT/ANYEXT DAGCo optimizations to the lowering of these optimizations. The old test cases still cover all of these lowering/optimizations. The single change that we have is that now anyext does not need to zero a register, because it does not use the exact code path as the zero_extend.
llvm-svn: 171178
2012-12-28 05:45:24 +00:00
Nadav Rotem 2a054b4475 On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
register. In most cases we actually compare or select YMM-sized registers
and mixing the two types creates horrible code. This commit optimizes
some of the transition sequences.

PR14657.

llvm-svn: 171148
2012-12-27 08:15:45 +00:00
NAKAMURA Takumi 40aa3285f4 llvm/test/CodeGen/X86: FileCheck-ize two tests in r171083.
llvm-svn: 171084
2012-12-26 03:19:30 +00:00
NAKAMURA Takumi 334f685328 llvm/test/CodeGen/X86: Disable avx in two tests corresponding to r171082.
llvm-svn: 171083
2012-12-26 03:08:55 +00:00
Hal Finkel 2ebe6d08cd Loosen scheduling restrictions on the PPC dcbt intrinsic
As with the prefetch intrinsic to which it maps, simply have dcbt
marked as reading from and writing to its arguments instead of having
unmodeled side effects. While this might cause unwanted code motion
(because aliasing checks don't really capture cache-line sharing),
it is more important that prefetches in unrolled loops don't block
the scheduler from rearranging the unrolled loop body.

llvm-svn: 171073
2012-12-25 18:51:18 +00:00
Hal Finkel 1b5ff08d43 Expand PPC64 atomic load and store
Use of store or load with the atomic specifier on 64-bit types would
cause instruction-selection failures. As with the 32-bit case, these
can use the default expansion in terms of cmp-and-swap.

llvm-svn: 171072
2012-12-25 17:22:53 +00:00
Benjamin Kramer a9f265ee98 Harden test so it's not affected by changes to compare lowering.
This only failed on hosts that don't have SSE41.

llvm-svn: 171066
2012-12-25 13:23:23 +00:00
Benjamin Kramer 81b5a8fd2e X86: Shave off one shuffle from the pcmpeqq sequence for SSE2 by making use of and commutativity.
llvm-svn: 171064
2012-12-25 13:09:08 +00:00
Benjamin Kramer df4af41b9b X86: Custom lower <2 x i64> eq and ne when SSE41 is not available.
pcmpeqd, pshufd, pshufd, pand is cheaper than unpack + cmpq, sbbq, cmpq, sbbq + pack.
Small speedup on loop-vectorized viterbi (-march=core2).

llvm-svn: 171063
2012-12-25 12:54:19 +00:00
NAKAMURA Takumi 1b18db7ea3 llvm/test/CodeGen/X86/fold-vex.ll: Add explicit triple.
llvm-svn: 171029
2012-12-24 11:14:06 +00:00
Nadav Rotem dc0ad92b64 Some x86 instructions can load/store one of the operands to memory. On SSE, this memory needs to be aligned.
When these instructions are encoded in VEX (on AVX) there is no such requirement. This changes the folding
tables and removes the alignment restrictions from VEX-encoded instructions.

llvm-svn: 171024
2012-12-24 09:40:33 +00:00
Benjamin Kramer 76268ac682 X86: Turn mul of <4 x i32> into pmuludq when no SSE4.1 is available.
pmuludq is slow, but it turns out that all the unpacking and packing of the
scalarized mul is even slower. 10% speedup on loop-vectorized paq8p.

llvm-svn: 170985
2012-12-22 16:07:56 +00:00
Benjamin Kramer b2f0a2bd4b X86: Emit vector sext as shuffle + sra if vpmovsx is not available.
Also loosen the SSSE3 dependency a bit, expanded pshufb + psra is still better
than scalarized loads. Fixes PR14590.

llvm-svn: 170984
2012-12-22 11:34:28 +00:00
Nadav Rotem d5aae980cb In some cases, due to scheduling constraints we copy the EFLAGS.
The only way to read the eflags is using push and pop. If we don't
adjust the stack then we run over the first frame index. This is
not something that we want to do, so we have to make sure that
our machine function does not copy the flags. If it does then
we have to emit the prolog that adjusts the stack.

rdar://12896831

llvm-svn: 170961
2012-12-21 23:48:49 +00:00
Benjamin Kramer b4688f84bd try to unbreak ppc buildbots.
llvm-svn: 170913
2012-12-21 18:11:45 +00:00
Benjamin Kramer 82d1c371e2 X86: Match pmin/pmax as a target specific dag combine. This occurs during vectorization.
Part of PR14667.

llvm-svn: 170908
2012-12-21 17:46:58 +00:00
Tom Stellard a8b0351720 R600: Expand vec4 INT <-> FP conversions
llvm-svn: 170901
2012-12-21 16:33:24 +00:00
Reed Kotler 93f778d2bd Add test case for r170674
llvm-svn: 170823
2012-12-21 00:55:10 +00:00
Eric Christopher 6e47b725ff Move these files over to the debug info directory.
llvm-svn: 170810
2012-12-21 00:03:42 +00:00
Bob Wilson 7bba4f8957 Revert "Adding support for llvm.arm.neon.vaddl[su].* and"
This reverts r170694.  The operations can be represented in IR without
adding any new intrinsics.

llvm-svn: 170765
2012-12-20 21:09:38 +00:00
Evan Cheng ddc0cb6dc5 On some ARM cpus, flags setting movs with shifter operand, i.e. lsl, lsr, asr,
are more expensive than the non-flag setting variant. Teach thumb2 size
reduction pass to avoid generating them unless we are optimizing for size.

rdar://12892707

llvm-svn: 170728
2012-12-20 19:59:30 +00:00
Rafael Espindola 642c7cd56e Simplify the testcase a bit.
I checked that it would still crash llc before the corresponding fix.

llvm-svn: 170709
2012-12-20 17:47:27 +00:00
Renato Golin 6b2ea4a48f Adding support for llvm.arm.neon.vaddl[su].* and
llvm.arm.neon.vsub[su].* intrinsics.

Patch by Pete Couperus <pjcoup@gmail.com>

llvm-svn: 170694
2012-12-20 13:52:11 +00:00
Reed Kotler d019dbf75e fix most of remaining issues with large frames.
these patches are tested a lot by test-suite but
make check tests are forthcoming once the next
few patches that complete this are committed.
with the next few patches the pass rate for mips16 is
near 100%

llvm-svn: 170656
2012-12-20 04:07:42 +00:00
Akira Hatanaka f423672117 [mips] Use "or $r0, $r1, $zero" instead of "addu $r0, $zero, $r1" to copy
physical register $r1 to $r0.

GNU disassembler recognizes an "or" instruction as a "move", and this change
makes the disassembled code easier to read.

Original patch by Reed Kotler.

llvm-svn: 170655
2012-12-20 04:06:06 +00:00
Bob Wilson 3365b80290 Do not introduce vector operations in functions marked with noimplicitfloat.
<rdar://problem/12879313>

llvm-svn: 170630
2012-12-20 01:36:20 +00:00
Evan Cheng eae6d2ccea LLVM sdisel normalize bit extraction of the form:
((x & 0xff00) >> 8) << 2
to
 (x >> 6) & 0x3fc

This is general goodness since it folds a left shift into the mask. However,
the trailing zeros in the mask prevents the ARM backend from using the bit
extraction instructions. And worse since the mask materialization may require
an addition instruction. This comes up fairly frequently when the result of 
the bit twiddling is used as memory address. e.g.

 = ptr[(x & 0xFF0000) >> 16]

We want to generate:
  ubfx   r3, r1, #16, #8
  ldr.w  r3, [r0, r3, lsl #2]

vs.
  mov.w  r9, #1020
  and.w  r2, r9, r1, lsr #14
  ldr    r2, [r0, r2]

Add a late ARM specific isel optimization to
ARMDAGToDAGISel::PreprocessISelDAG(). It folds the left shift to the
'base + offset' address computation; change the mask to one which doesn't have
trailing zeros and enable the use of ubfx.

Note the optimization has to be done late since it's target specific and we
don't want to change the DAG normalization. It's also fairly restrictive
as shifter operands are not always free. It's only done for lsh 1 / 2. It's
known to be free on some cpus and they are most common for address
computation.

This is a slight win for blowfish, rijndael, etc.

rdar://12870177

llvm-svn: 170581
2012-12-19 20:16:09 +00:00
Benjamin Kramer c5071466d4 PowerPC: Expand VSELECT nodes.
There's probably a better expansion for those nodes than the default for
altivec, but this is better than crashing. VSELECTs occur in loop vectorizer
output.

llvm-svn: 170551
2012-12-19 15:49:14 +00:00