through the use of 'fpmath' metadata. Currently this only provides a 'fpaccuracy'
value, which may be a number in ULPs or the keyword 'fast', however the intent is
that this will be extended with additional information about NaN's, infinities
etc later. No optimizations have been hooked up to this so far.
llvm-svn: 154822
This is mostly to test the waters. I'd like to get results from FNT
build bots and other bots running on non-x86 platforms.
This feature has been pretty heavily tested over the last few months by
me, and it fixes several of the execution time regressions caused by the
inlining work by preventing inlining decisions from radically impacting
block layout.
I've seen very large improvements in yacr2 and ackermann benchmarks,
along with the expected noise across all of the benchmark suite whenever
code layout changes. I've analyzed all of the regressions and fixed
them, or found them to be impossible to fix. See my email to llvmdev for
more details.
I'd like for this to be in 3.1 as it complements the inliner changes,
but if any failures are showing up or anyone has concerns, it is just
a flag flip and so can be easily turned off.
I'm switching it on tonight to try and get at least one run through
various folks' performance suites in case SPEC or something else has
serious issues with it. I'll watch bots and revert if anything shows up.
llvm-svn: 154816
rotation. When there is a loop backedge which is an unconditional
branch, we will end up with a branch somewhere no matter what. Try
placing this backedge in a fallthrough position above the loop header as
that will definitely remove at least one branch from the loop iteration,
where whole loop rotation may not.
I haven't seen any benchmarks where this is important but loop-blocks.ll
tests for it, and so this will be covered when I flip the default.
llvm-svn: 154812
laid out in a form with a fallthrough into the header and a fallthrough
out of the bottom. In that case, leave the loop alone because any
rotation will introduce unnecessary branches. If either side looks like
it will require an explicit branch, then the rotation won't add any, do
it to ensure the branch occurs outside of the loop (if possible) and
maximize the benefit of the fallthrough in the bottom.
llvm-svn: 154806
This is a complex change that resulted from a great deal of
experimentation with several different benchmarks. The one which proved
the most useful is included as a test case, but I don't know that it
captures all of the relevant changes, as I didn't have specific
regression tests for each, they were more the result of reasoning about
what the old algorithm would possibly do wrong. I'm also failing at the
moment to craft more targeted regression tests for these changes, if
anyone has ideas, it would be welcome.
The first big thing broken with the old algorithm is the idea that we
can take a basic block which has a loop-exiting successor and a looping
successor and use the looping successor as the layout top in order to
get that particular block to be the bottom of the loop after layout.
This happens to work in many cases, but not in all.
The second big thing broken was that we didn't try to select the exit
which fell into the nearest enclosing loop (to which we exit at all). As
a consequence, even if the rotation worked perfectly, it would result in
one of two bad layouts. Either the bottom of the loop would get
fallthrough, skipping across a nearer enclosing loop and thereby making
it discontiguous, or it would be forced to take an explicit jump over
the nearest enclosing loop to earch its successor. The point of the
rotation is to get fallthrough, so we need it to fallthrough to the
nearest loop it can.
The fix to the first issue is to actually layout the loop from the loop
header, and then rotate the loop such that the correct exiting edge can
be a fallthrough edge. This is actually much easier than I anticipated
because we can handle all the hard parts of finding a viable rotation
before we do the layout. We just store that, and then rotate after
layout is finished. No inner loops get split across the post-rotation
backedge because we check for them when selecting the rotation.
That fix exposed a latent problem with our exitting block selection --
we should allow the backedge to point into the middle of some inner-loop
chain as there is no real penalty to it, the whole point is that it
*won't* be a fallthrough edge. This may have blocked the rotation at all
in some cases, I have no idea and no test case as I've never seen it in
practice, it was just noticed by inspection.
Finally, all of these fixes, and studying the loops they produce,
highlighted another problem: in rotating loops like this, we sometimes
fail to align the destination of these backwards jumping edges. Fix this
by actually walking the backwards edges rather than relying on loopinfo.
This fixes regressions on heapsort if block placement is enabled as well
as lots of other cases where the previous logic would introduce an
abundance of unnecessary branches into the execution.
llvm-svn: 154783
As an example, attach range info to the "invalid instruction" message:
$ clang -arch arm -c asm.c
asm.c:2:11: error: invalid instruction
__asm__("foo r0");
^
<inline asm>:1:2: note: instantiated into assembly here
foo r0
^~~
llvm-svn: 154765
thinking of generalizing it to be able to specify other freedoms beyond accuracy
(such as that NaN's don't have to be respected). I'd like the 3.1 release (the
first one with this metadata) to have the more generic name already rather than
having to auto-upgrade it in 3.2.
llvm-svn: 154744
When vectorizing pointer types it is important to realize that potential
pairs cannot be connected via the address pointer argument of a load or store.
This is because even after vectorization, the address is still a scalar because
the address of the higher half of the pair is implicit from the address of the
lower half (it need not be, and should not be, explicitly computed).
llvm-svn: 154735
This is a special flag for targets that really want their block
terminators in the DAG. The default scheduler cannot handle this
correctly, so it becomes the specialized scheduler's responsibility to
schedule terminators.
llvm-svn: 154712
- Don't copy offsets into HashData, the underlying vector won't change once the table is finalized.
- Allocate HashData and HashDataContents in a BumpPtrAllocator.
- Allocate string map entries in the same allocator.
- Random cleanups.
llvm-svn: 154694
As has been suggested by Duncan and others, Early-CSE and GVN should
do similar redundancy elimination, but Early-CSE is much less expensive.
Most of my autovectorization benchmarks show a performance regresion, but
all of these are < 0.1%, and so I think that it is still worth using
the less expensive pass.
llvm-svn: 154673
obviously cannot know that this code is present, let alone used. So prevent the
internalize pass from internalizing those global values which code-gen may
insert.
llvm-svn: 154645
directly instead of a user Instruction. This allows them to test
whether a def dominates a particular operand if the user instruction
is a PHI.
llvm-svn: 154631
of zero-initialized sections, virtual sections and common symbols
and preventing the loading of sections which are not required for
execution such as debug information.
Patch by Andy Kaylor!
llvm-svn: 154610
- FCOPYSIGN nodes that have operands of different types were not handled.
- Different code was generated depending on the endianness of the target.
Additionally, code is added that emits INS and EXT instructions, if they are
supported by target (they are R2 instructions).
llvm-svn: 154540
While there is an encoding for it in VUZP, the result of that is undefined,
so we should avoid it. Define the instruction as a pseudo for VTRN.32
instead, as the ARM ARM indicates.
rdar://11222366
llvm-svn: 154511
While there is an encoding for it in VZIP, the result of that is undefined,
so we should avoid it. Define the instruction as a pseudo for VTRN.32
instead, as the ARM ARM indicates.
rdar://11221911
llvm-svn: 154505
FoldingSet is implemented as a chained hash table. When there is a hash
collision during insertion, which is common as we fill the table until a
load factor of 2.0 is hit, we walk the chained elements, comparing every
operand with the new element's operands. This can be very expensive if the
MDNode has many operands.
We sacrifice a word of space in MDNode to cache the full hash value, reducing
compares on collision to a minimum. MDNode grows from 28 to 32 bytes + operands
on x86. On x86_64 the new bits fit nicely into existing padding, not growing
the struct at all.
The actual speedup depends a lot on the test case and is typically between
1% and 2% for C++ code with clang -c -O0 -g.
llvm-svn: 154497
Fix a dagcombine optimization which assumes that the vsetcc result type is always
of the same size as the compared values. This is ture for SSE/AVX/NEON but not
for all targets.
llvm-svn: 154490
Original message:
Modify the code that lowers shuffles to blends from using blendvXX to vblendXX.
blendV uses a register for the selection while Vblend uses an immediate.
On sandybridge they still have the same latency and execute on the same execution ports.
llvm-svn: 154483
predicates.
Also remove NEON2 since it's not really useful and it is confusing. If
NEON + VFP4 implies NEON2 but NEON2 doesn't imply NEON + VFP4, what does it
really mean?
rdar://10139676
llvm-svn: 154480
1. The new instruction itinerary entries are not properly described.
2. The asm parser can't handle vfms and vfnms.
3. There were no assembler, disassembler test cases.
4. HasNEON2 has the wrong assembler predicate.
rdar://10139676
llvm-svn: 154456
Allow cheap instructions to be hoisted if they are register pressure
neutral or better. This happens if the instruction is the last loop use
of another virtual register.
Only expensive instructions are allowed to increase loop register
pressure.
llvm-svn: 154455
Hoisting a value that is used by a PHI in the loop will introduce a
copy because the live range is extended to cross the PHI.
The same applies to PHIs in exit blocks.
Also use this opportunity to make HasLoopPHIUse() non-recursive.
llvm-svn: 154454
- don't isntrument reads from constant globals.
Saves ~1.5% of instrumented instructions on CPU2006
(counting static instructions, not their execution).
- don't insrument reads from vtable (which is a global constant too).
Saves ~5%.
I did not measure the run-time impact of this,
but it is certainly non-negative.
llvm-svn: 154444
StringMap. This was redundant and unnecessarily bloated the MDString class.
Because the MDString class is a "Value" and will never have a "name", and
because the Name field in the Value class is a pointer to a StringMap entry, we
repurpose the Name field for an MDString. It stores the StringMap entry in the
Name field, and uses the normal methods to get the string (name) back.
PR12474
llvm-svn: 154429
a write to the same temp follows in the same BB.
Also add stats printing.
On Spec CPU2006 this optimization saves roughly 4% of instrumented reads
(which is 3% of all instrumented accesses):
Writes : 161216
Reads : 446458
Reads-before-write: 18295
llvm-svn: 154418
We were incorrectly conflating some add variants which don't have a
cc_out operand with the mirroring sub encodings, which do. Part of the
awesome non-orthogonality legacy of thumb1. Similarly, handling of
add/sub of an immediate was sometimes incorrectly removing the cc_out
operand for add/sub register variants.
rdar://11216577
llvm-svn: 154411
blendv uses a register for the selection while vblend uses an immediate.
On sandybridge they still have the same latency and execute on the same execution ports.
llvm-svn: 154396
the loop header has a non-loop predecessor which has been pre-fused into
its chain due to unanalyzable branches. In this case, rotating the
header into the body of the loop in order to place a loop exit at the
bottom of the loop is a Very Bad Idea as it makes the loop
non-contiguous.
I'm working on a good test case for this, but it's a bit annoynig to
craft. I should get one shortly, but I'm submitting this now so I can
begin the (lengthy) performance analysis process. An initial run of LNT
looks really, really good, but there is too much noise there for me to
trust it much.
llvm-svn: 154395
Take this opportunity to generalize the indirectbr bailout logic for
loop transformations. CFG transformations will never get indirectbr
right, and there's no point trying.
llvm-svn: 154386
legalizer always use the DAG entry node. This is wrong when the libcall is
emitted as a tail call since it effectively folds the return node. If
the return node's input chain is not the entry (i.e. call, load, or store)
use that as the tail call input chain.
PR12419
rdar://9770785
rdar://11195178
llvm-svn: 154370
in-register, such that we can use a single vector store rather then a
series of scalar stores.
For func_4_8 the generated code
vldr d16, LCPI0_0
vmov d17, r0, r1
vadd.i16 d16, d17, d16
vmov.u16 r0, d16[3]
strb r0, [r2, #3]
vmov.u16 r0, d16[2]
strb r0, [r2, #2]
vmov.u16 r0, d16[1]
strb r0, [r2, #1]
vmov.u16 r0, d16[0]
strb r0, [r2]
bx lr
becomes
vldr d16, LCPI0_0
vmov d17, r0, r1
vadd.i16 d16, d17, d16
vuzp.8 d16, d17
vst1.32 {d16[0]}, [r2, :32]
bx lr
I'm not fond of how this combine pessimizes 2012-03-13-DAGCombineBug.ll,
but I couldn't think of a way to judiciously apply this combine.
This
ldrh r0, [r0, #4]
strh r0, [r1]
becomes
vldr d16, [r0]
vmov.u16 r0, d16[2]
vmov.32 d16[0], r0
vuzp.16 d16, d17
vst1.32 {d16[0]}, [r1, :32]
PR11158
rdar://10703339
llvm-svn: 154340
This patch restores TwoAddressInstructionPass's pre-r153892 behaviour when
rescheduling instructions in TryInstructionTransform. Hopefully this will fix
PR12493. To refix PR11861, lowering of INSERT_SUBREGS is deferred until after
the copy that unties the operands is emitted (this seems to be a more
appropriate fix for that issue anyway).
llvm-svn: 154338
A couple of cases where we were accidentally creating constant conditions by
something like "x == a || b" instead of "x == a || x == b". In one case a
conditional & then unreachable was used - I transformed this into a direct
assert instead.
llvm-svn: 154324
x86 addressing modes. This allows PIE-based TLS offsets to fit directly
into an addressing mode immediate offset, which is the last remaining
code quality issue from PR12380. With this patch, that PR is completely
fixed.
To understand why this patch is correct to match these offsets into
addressing mode immediates, break it down by cases:
1) 32-bit is trivially correct, and unmodified here.
2) 64-bit non-small mode is unchanged and never matches.
3) 64-bit small PIC code which is RIP-relative is handled specially in
the match to try to fit RIP into the base register. If it fails, it
now early exits. This behavior is unchanged by the patch.
4) 64-bit small non-PIC code which is not RIP-relative continues to work
as it did before. The reason these immediates are safe is because the
ABI ensures they fit in small mode. This behavior is unchanged.
5) 64-bit small PIC code which is *not* using RIP-relative addressing.
This is the only case changed by the patch, and the primary place you
see it is in TLS, either the win64 section offset TLS or Linux
local-exec TLS model in a PIC compilation. Here the ABI again ensures
that the immediates fit because we are in small mode, and any other
operations required due to the PIC relocation model have been handled
externally to the Wrapper node (extra loads etc are made around the
wrapper node in ISelLowering).
I've tested this as much as I can comparing it with GCC's output, and
everything appears safe. I discussed this with Anton and it made sense
to him at least at face value. That said, if there are issues with PIC
code after this patch, yell and we can revert it.
llvm-svn: 154304
when -ffast-math, i.e. don't just always do it if the reciprocal can
be formed exactly. There is already an IR level transform that does
that, and it does it more carefully.
llvm-svn: 154296
optimizations which are valid for position independent code being linked
into a single executable, but not for such code being linked into
a shared library.
I discussed the design of this with Eric Christopher, and the decision
was to support an optional bit rather than a completely separate
relocation model. Fundamentally, this is still PIC relocation, its just
that certain optimizations are only valid under a PIC relocation model
when the resulting code won't be in a shared library. The simplest path
to here is to expose a single bit option in the TargetOptions. If folks
have different/better designs, I'm all ears. =]
I've included the first optimization based upon this: changing TLS
models to the *Exec models when PIE is enabled. This is the LLVM
component of PR12380 and is all of the hard work.
llvm-svn: 154294
in TargetLowering. There was already a FIXME about this location being
odd. The interface is simplified as a consequence. This will also make
it easier to change TLS models when compiling with PIE.
llvm-svn: 154292
where a chain outside of the loop block-set ended up in the worklist for
scheduling as part of the contiguous loop. However, asserting the first
block in the chain is in the loop-set isn't a valid check -- we may be
forced to drag a chain into the worklist due to one block in the chain
being part of the loop even though the first block is *not* in the loop.
This occurs when we have been forced to form a chain early due to
un-analyzable branches.
No test case here as I have no idea how to even begin reducing one, and
it will be hopelessly fragile. We have to somehow end up with a loop
header of an inner loop which is a successor of a basic block with an
unanalyzable pair of branch instructions. Ow. Self-host triggers it so
it is unlikely it will regress.
This at least gets block placement back to passing selfhost and the test
suite. There are still a lot of slowdown that I don't like coming out of
block placement, although there are now also a lot of speedups. =[ I'm
seeing swings in both directions up to 10%. I'm going to try to find
time to dig into this and see if we can turn this on for 3.1 as it does
a really good job of cleaning up after some loops that degraded with the
inliner changes.
llvm-svn: 154287
GEPs, bit casts, and stores reaching it but no other instructions. These
often show up during the iterative processing of the inliner, SROA, and
DCE. Once we hit this point, we can completely remove the alloca. These
were actually showing up in the final, fully optimized code in a bunch
of inliner tests I've been working on, and notably they show up after
LLVM finishes optimizing away all function calls involved in
hash_combine(a, b).
llvm-svn: 154285
Previously we used three instructions to broadcast an immediate value into a
vector register.
On Sandybridge we continue to load the broadcasted value from the constant pool.
llvm-svn: 154284
An MDNode has a list of MDNodeOperands allocated directly after it as part of
its allocation. Therefore, the Parent of the MDNodeOperands can be found by
walking back through the operands to the beginning of that list. Mark the first
operand's value pointer as being the 'first' operand so that we know where the
beginning of said list is.
This saves a *lot* of space during LTO with -O0 -g flags.
llvm-svn: 154280
shuffle node because it could introduce new shuffle nodes that were not
supported efficiently by the target.
2. Add a more restrictive shuffle-of-shuffle optimization for cases where the
second shuffle reverses the transformation of the first shuffle.
llvm-svn: 154266
reciprocal if converting to the reciprocal is exact. Do it even if inexact
if -ffast-math. This substantially speeds up ac.f90 from the polyhedron
benchmarks.
llvm-svn: 154265
speculate. Without this, loop rotate (among many other places) would
suddenly stop working in the presence of debug info. I found this
looking at loop rotate, and have augmented its tests with a reduction
out of a very hot loop in yacr2 where failing to do this rotation costs
sometimes more than 10% in runtime performance, perturbing numerous
downstream optimizations.
This should have no impact on performance without debug info, but the
change in performance when debug info is enabled can be extreme. As
a consequence (and this how I got to this yak) any profiling of
performance problems should be treated with deep suspicion -- they may
have been wildly innacurate of debug info was enabled for profiling. =/
Just a heads up.
llvm-svn: 154263
The tLDRr instruction with the last register operand set to the zero register
prints in assembly as if no register was specified, and the assembler encodes
it as a tLDRi instruction with a zero immediate. With the integrated assembler,
that zero register gets emitted as "r0", so we get "ldr rx, [ry, r0]" which
is broken. Emit the instruction as tLDRi with a zero immediate. I don't
know if there's a good way to write a testcase for this. Suggestions welcome.
Opportunities for follow-up work:
1) The asm printer should complain if a non-optional register operand is set
to the zero register, instead of silently dropping it.
2) The integrated assembler should complain in the same situation, instead of
silently emitting the operand as "r0".
llvm-svn: 154261
Cygwin-1.7 supports dw2. Some recent mingw distros support one, too.
I have confirmed test-suite/SingleSource/Benchmarks/Shootout-C++/except.cpp can pass on Cygwin.
llvm-svn: 154247
by default.
This is a behaviour configurable in the MCAsmInfo. I've decided to turn
it on by default in (possibly optimistic) hopes that most assemblers are
reasonably sane. If this proves a problem, switching to default seems
reasonable.
I'm not sure if this is the opportune place to test, but it seemed good
to make sure it was tested somewhere.
llvm-svn: 154235
After register masks were introdruced to represent the call clobbers, it
is no longer necessary to have duplicate instruction for iOS.
llvm-svn: 154209
disassembler requires a MCSubtargetInfo and a
MCInstrInfo to exist in order to initialize the
instruction printer and disassembler; however,
although the printer and disassembler keep
references to these objects they do not own them.
Previously, the MCSubtargetInfo and MCInstrInfo
objects were just leaked.
I have extended LLVMDisasmContext to own these
objects and delete them when it is destroyed.
llvm-svn: 154192
simplification has been performed. This is a bit less efficient
(requires another ilist walk of the basic blocks) but shouldn't matter
in practice. More importantly, it's just too much work to keep track of
all the various ways the return instructions can be mutated while
simplifying them. This fixes yet another crasher, reported by Daniel
Dunbar.
llvm-svn: 154179
dead code, including dead return instructions in some cases. Otherwise,
we end up having a bogus poniter to a return instruction that blows up
much further down the road.
It turns out that this pattern is both simpler to code, easier to update
in the face of enhancements to the inliner cleanup, and likely cheaper
given that it won't add dead instructions to the list.
Thanks to John Regehr's numerous test cases for teasing this out.
llvm-svn: 154157
We had special instructions for iOS because r9 is call-clobbered, but
that is represented dynamically by the register mask operands now, so
there is no need for the pseudo-instructions.
llvm-svn: 154144
The load/store optimizer splits LDRD/STRD into two instructions when the
register pairing doesn't work out. For negative offsets in Thumb2, it uses
t2STRi8 to do that. That's fine, except for the case when the offset is in
the range [-4,-1]. In that case, we'll also form a second t2STRi8 with
the original offset plus 4, resulting in a t2STRi8 with a non-negative
offset, which ends up as if it were an STRT, which is completely bogus.
Similarly for loads.
No testcase, unfortunately, as any I've been able to construct is both large
and extremely fragile.
rdar://11193937
llvm-svn: 154141
'add r2, #-1024' should just use 'sub r2, #1024' rather than erroring out.
Thumb1 aliases for adding a negative immediate to the stack pointer,
also.
rdar://11192734
llvm-svn: 154123
LSR always tries to make the ICmp in the loop latch use the incremented
induction variable. This allows the induction variable to be kept in a
single register.
When the induction variable limit is equal to the stride,
SimplifySetCC() would break LSR's hard work by transforming:
(icmp (add iv, stride), stride) --> (cmp iv, 0)
This forced us to use lea for the IC update, preventing the simpler
incl+cmp.
<rdar://problem/7643606>
<rdar://problem/11184260>
llvm-svn: 154119
of the BBVectorizePass without using command line option. As pointed out
by Hal, we can ask the TargetLoweringInfo for the architecture specific
VectorizeConfig to perform vectorizing with architecture specific
information.
llvm-svn: 154096
the caller requested a null-terminated one.
When mapping the file there could be a racing issue that resulted in the file being larger
than the FileSize passed by the caller. We already have an assertion
for this in MemoryBuffer::init() but have a runtime guarantee that
the buffer will be null-terminated, so do a copy that adds a null-terminator.
Protects against crash of rdar://11161822.
llvm-svn: 154082
LSR can fold three addressing modes into its ICmpZero node:
ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset
ICmpZero -1*ScaleReg + Offset => ICmp ScaleReg, Offset
ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
The first two cases are only used if TLI->isLegalICmpImmediate() likes
the offset.
Make sure the right Offset sign is passed to this method in the second
case. The ARM version is not symmetric.
<rdar://problem/11184260>
llvm-svn: 154079
A MOVCCr instruction can be commuted by inverting the condition. This
can help reduce register pressure and remove unnecessary copies in some
cases.
<rdar://problem/11182914>
llvm-svn: 154033
This allows us to keep passing reduced masks to SimplifyDemandedBits, but
know about all the bits if SimplifyDemandedBits fails. This allows instcombine
to simplify cases like the one in the included testcase.
llvm-svn: 154011
When folding X == X we need to check getBooleanContents() to determine if the
result is a vector of ones or a vector of negative ones.
I tried creating a test case, but the problem seems to only be exposed on a
much older version of clang (around r144500).
rdar://10923049
llvm-svn: 153966
