stored here is of a certain kind. This is in preparation for when an Attribute
object represents a single attribute, instead of a bitmask of attributes.
llvm-svn: 171247
propagating one of the values it simplified to a constant across
a myriad of instructions. Notably, ptrtoint instructions when we had
a constant pointer (say, 0) didn't propagate that, blocking a massive
number of down-stream optimizations.
This was uncovered when investigating why we fail to inline and delete
the boilerplate in:
void f() {
std::vector<int> v;
v.push_back(1);
}
It turns out most of the efforts I've made thus far to improve the
analysis weren't making it far purely because of this. After this is
fixed, the store-to-load forwarding patch enables LLVM to optimize the
above to an empty function. We still can't nuke a second push_back, but
for different reasons.
There is a very real chance this will cause somewhat noticable changes
in inlining behavior, so please let me know if you see regressions (or
improvements!) because of this patch.
llvm-svn: 171196
how to propagate constants through insert and extract value
instructions.
With the recent improvements to instsimplify, this allows inline cost
analysis to constant fold through intrinsic functions, including notably
the with.overflow intrinsic math routines which often show up inside of
STL abstractions. This is yet another piece in the puzzle of breaking
down the code for:
void f() {
std::vector<int> v;
v.push_back(1);
}
But it still isn't enough. There are a pile of bugs in inline cost still
blocking this.
llvm-svn: 171195
constant folding calls. Add the initial tests for this which show that
now instsimplify can simplify blindingly obvious code patterns expressed
with both intrinsics and library calls.
llvm-svn: 171194
are nice and decomposed so that we can simplify synthesized calls as
easily as actually call instructions. The internal utility still has the
same behavior, it just now operates on a more generic interface so that
I can extend the set of call simplifications that instsimplify knows
about.
llvm-svn: 171189
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
The vector truncs were scalarized during LegalizeVectorOps, later vectorized again by some DAGCombine optimization
and finally, lowered by a dagcombing optimization. Now, they are properly lowered during LegalizeVectorOps.
No new testcase because the original testcases still work.
llvm-svn: 171146
information doesn't return an addend for Rel relocations. Go ahead
and use this information to fix relocation handling inside dwarfdump
for 32-bit ELF REL.
llvm-svn: 171126
such as by a compiler warning, a check in clang -fsanitizer=undefined, being
optimized to unreachable, or a combination of the above. PR14722.
llvm-svn: 171119
For the time being this includes only some dummy test cases. Once the
generic implementation of the intrinsics cost function does something other
than assuming scalarization in all cases, or some target specializes the
interface, some real test cases can be added.
Also, for consistency, I changed the type of IID from unsigned to Intrinsic::ID
in a few other places.
llvm-svn: 171079
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
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
These are now generally used for all diagnostics from the backend, not just
for inline assembly, so this drops the "InlineAsm" from the names. No
functional change. (I've left aliases for the old names but only for long
enough to let me switch over clang to use the new ones.)
llvm-svn: 171047
"check-all" can be executed with 0 status, "check-all does nothing, no tools built."
LLVM_EXTERNAL_CLANG_BUILD=OFF LLVM_BUILD_TOOLS=OFF can reproduce this.
Oscar Fuentes reported this. Thank you.
llvm-svn: 171046
This affords us to use std::string's allocation routines and use the destructor
for the memory management. Switching to that also means that we can use
operator==(const std::string&, const char *) to perform the string comparison
rather than resorting to libc functionality (i.e. strcmp).
Patch by Saleem Abdulrasool!
Differential Revision: http://llvm-reviews.chandlerc.com/D230
llvm-svn: 171042
When the backend is used from clang, it should produce proper diagnostics
instead of just printing messages to errs(). Other clients may also want to
register their own error handlers with the LLVMContext, and the same handler
should work for warnings in the same way as the existing emitError methods.
llvm-svn: 171041
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
the cost of arithmetic functions. We now assume that the cost of arithmetic
operations that are marked as Legal or Promote is low, but ops that are
marked as custom are higher.
llvm-svn: 171002
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
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
On MachO, sections also have segment names. When a tool looking at a .o file
prints a segment name, this is what they mean. In reality, a .o has only one
anonymous, segment.
This patch adds a MachO only function to fetch that segment name. I named it
getSectionFinalSegmentName since the main use for the name seems to be inform
the linker with segment this section should go to.
The patch also changes MachOObjectFile::getSectionName to return just the
section name instead of computing SegmentName,SectionName.
The main difference from the previous patch is that it doesn't use
InMemoryStruct. It is extremely dangerous: if the endians match it returns
a pointer to the file buffer, if not, it returns a pointer to an internal buffer
that is overwritten in the next API call.
We should change all of this code to use
support::detail::packed_endian_specific_integral like ELF, but since these
functions only handle strings, they work with big and little endian machines
as is.
I have tested this by installing ubuntu 12.10 ppc on qemu, that is why it took
so long :-)
llvm-svn: 170838
memory bound checks. Before the fix we were able to vectorize this loop from
the Livermore Loops benchmark:
for ( k=1 ; k<n ; k++ )
x[k] = x[k-1] + y[k];
llvm-svn: 170811
Instructions that are inserted in a basic block can still be decorated
with addOperand(MO).
Make the two-argument addOperand() function contain the actual
implementation. This function will now always have a valid MF reference
that it can use for memory allocation.
llvm-svn: 170798
This function is often used to decorate dangling instructions, so a
context reference is required to allocate memory for the operands.
Also add a corresponding MachineInstrBuilder method.
llvm-svn: 170797
Rename the AttributeImpl* from Attrs to pImpl to be consistent with other code.
Add comments where none were before. Or doxygen-ify other comments.
llvm-svn: 170767
Before if-conversion we could check if a value is loop invariant
if it was declared inside the basic block. Now that loops have
multiple blocks this check is incorrect.
This fixes External/SPEC/CINT95/099_go/099_go
llvm-svn: 170756
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
the script generating it. The test should never be modified manually. If anyone
needs to change it, please change the script and re-run it.
The script is placed into utils/testgen - I couldn't think of a better place,
and after some discussion on IRC this looked like a logical location.
llvm-svn: 170720
This is supposed to be a mechanical change with no functional effects.
InstrEmitter can generate all types of MachineOperands which revealed
that MachineInstrBuilder was missing a few methods, added by this patch.
Besides providing a context pointer to MI::addOperand(),
MachineInstrBuilder seems like a better fit for this code.
llvm-svn: 170712
Similarly inlining of the function is inhibited, if that would duplicate the call (in particular inlining is still allowed when there is only one callsite and the function has internal linkage).
llvm-svn: 170704
FYI, llvm and clang can be built deterministically between stage 2 and stage3, among iterative clean rebuilds, with GNU ar;
configure --disable-timestamps
make AR.Flags=crsD RANLIB=echo
llvm-svn: 170682
next few days but it's already tested a lot from test-suite and works fine.
This patch completes almost 100% pass of test-suite for mips 16.
llvm-svn: 170674
This has undefined behavior, because the classof implementation attempts to
access parts of the not-yet-constructed derived class. Found by clang
-fsanitize=vptr.
llvm-svn: 170658
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
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
This is a pretty lengthy document, so put the table of contents in your
face so that it's easier to scope out the content.
This document is a mess currently and needs to be
refactored/revised/split-up.
llvm-svn: 170646
For whatever reason the usage of '^^^' and '---' adornments were
reversed compared to the "canonical" style of the LLVM docs (which is
currently "the style used in SphinxQuickstartTemplate.rst"). This change
doesn't affect the document structure at all, I'm just doing it for
trivial stylistic consistency (the document content is *much* more
important---thanks Nadav for writing this up!).
Also, trim the adornments to be the same length as the section names.
llvm-svn: 170638
MC disassembler clients (LLDB) are interested in querying if an
instruction may affect control flow other than by virtue of being
an explicit branch instruction. For example, instructions which
write directly to the PC on some architectures.
llvm-svn: 170610
These were defined on TargetRegisterInfo, but they don't use any information
that's not available in MCRegisterInfo, so sink them down to be available
at the MC layer.
llvm-svn: 170608
Unlike SGPRs VGPRs doesn't need to be aligned.
Patch by: Christian König
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Christian König <deathsimple@vodafone.de>
llvm-svn: 170593
Branch if we have enough instructions so that it makes sense.
Also remove branches if they don't make sense.
Patch by: Christian König
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Christian König <deathsimple@vodafone.de>
llvm-svn: 170592
This patch replaces the control flow handling with a new
pass which structurize the graph before transforming it to
machine instruction. This has a couple of different advantages
and currently fixes 20 piglit tests without a single regression.
It is now a general purpose transformation that could be not
only be used for SI/R6xx, but also for other hardware
implementations that use a form of structurized control flow.
v2: further cleanup, fixes and documentation
Patch by: Christian König
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
llvm-svn: 170591
Use the version that also takes an MF reference instead.
It would technically be possible to extract an MF reference from the MI
as MI->getParent()->getParent(), but that would not work for MIs that
are not inserted into any basic block.
Given the reasonably small number of places this constructor was used at
all, I preferred the compile time check to a run time assertion.
llvm-svn: 170588
I introduced it in r166785. PR14291.
If TD is unavailable use getScalarSizeInBits, but don't optimize
pointers or vectors of pointers.
llvm-svn: 170586
Building Vectorizers.rst produces a few warnings of the form:
WARNING: Title underline too short.
Fixed by adding the extra needed dashes under the title.
llvm-svn: 170582
((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
When the least bit of C is greater than V, (x&C) must be greater than V
if it is not zero, so the comparison can be simplified.
Although this was suggested in Target/X86/README.txt, it benefits any
architecture with a directly testable form of AND.
Patch by Kevin Schoedel
llvm-svn: 170576
Just like for addMemOperand(), the function pointer provides a context
for allocating memory. This will make it possible to use a better memory
allocation strategy for the MI operand list, which is currently a slow
std::vector.
Most calls to addOperand() come from MachineInstrBuilder, so give that
class an MF reference as well. Code using BuildMI() won't need changing
at all since the MF reference is already required to allocate a
MachineInstr.
Future patches will fix code that calls MI::addOperand(Op) directly, as
well as code that uses the now deprecated MachineInstrBuilder(MI)
constructor.
llvm-svn: 170574
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
- An MVT can become an EVT when being split (e.g. v2i8 -> v1i8, the latter doesn't exist)
- Return the scalar value when an MVT is scalarized (v1i64 -> i64)
Fixes PR14639ff.
llvm-svn: 170546
I cannot reproduce it the failures locally, so I will keep an eye at the ppc
bots. This patch does add the change to the "Disassembly of section" message,
but that is not what was failing on the bots.
Original message:
Add a funciton to get the segment name of a section.
On MachO, sections also have segment names. When a tool looking at a .o file
prints a segment name, this is what they mean. In reality, a .o has only one
anonymous, segment.
This patch adds a MachO only function to fetch that segment name. I named it
getSectionFinalSegmentName since the main use for the name seems to be infor
the linker with segment this section should go to.
The patch also changes MachOObjectFile::getSectionName to return just the
section name instead of computing SegmentName,SectionName.
llvm-svn: 170545
This changes adds shadow and origin propagation for unknown intrinsics
by examining the arguments and ModRef behaviour. For now, only 3 classes
of intrinsics are handled:
- those that look like simple SIMD store
- those that look like simple SIMD load
- those that don't have memory effects and look like arithmetic/logic/whatever
operation on simple types.
llvm-svn: 170530
MapVector is a bit heavyweight, but I don't see a simpler way. Also the
InductionList is unlikely to be large. This should help 3-stage selfhost
compares (PR14647).
llvm-svn: 170528
bitwidth op back to the original size. If we reduce ANDs then this can cause
an endless loop. This patch changes the ZEXT to ANY_EXTEND if the demanded bits
are equal or smaller than the size of the reduced operation.
llvm-svn: 170505
instructions in the assembly code variant if one exists.
The intended use for this is so tools like lldb and darwin's otool(1)
can be switched to print Intel-flavored disassembly.
I discussed extensively this API with Jim Grosbach and we feel
while it may not be fully general, in reality there is only one syntax
for each assembly with the exception of X86 which has exactly
two for historical reasons.
rdar://10989182
llvm-svn: 170477
The bundle flags are now maintained by the slightly higher-level
functions bundleWithPred() / bundleWithSucc() which enforce consistent
bundle flags between neighboring instructions.
See also MIBundleBuilder for an even higher-level approach to building
bundles.
llvm-svn: 170475
The bundle_iterator::operator++ function now doesn't need to dig out the
basic block and check against end(). It can use the isBundledWithSucc()
flag to find the last bundled instruction safely.
Similarly, MachineInstr::isBundled() no longer needs to look at
iterators etc. It only has to look at flags.
llvm-svn: 170473
To not over constrain the scheduler for ARM in thumb mode, some optimizations for code size reduction, specific to ARM thumb, are blocked when they add a dependency (like write after read dependency).
Disables this check when code size is the priority, i.e., code is compiled with -Oz.
llvm-svn: 170462
The bundle-related MI flags need to be kept in sync with the neighboring
instructions. Don't allow the bulk flag-setting setFlags() function to
change them.
Also don't copy MI flags when cloning an instruction. The clone's bundle
flags will be set when it is explicitly inserted into a bundle.
llvm-svn: 170459
Remove the instr_iterator versions of the splice() functions. It doesn't
seem useful to be able to splice sequences of instructions that don't
consist of full bundles.
The normal splice functions that take MBB::iterator arguments are not
changed, and they can move whole bundles around without any problems.
llvm-svn: 170456
The single-element ilist::splice() function supports a noop move:
List.splice(I, List, I);
The corresponding std::list function doesn't allow that, so add a unit
test to document that behavior.
This also means that
List.splice(I, List, F);
is somewhat surprisingly not equivalent to
List.splice(I, List, F, next(F));
This patch adds an assertion to catch the illegal case I == F above.
Alternatively, we could make I == F a legal noop, but that would make
ilist differ even more from std::list.
llvm-svn: 170443
The normal insert() function takes an MBB::iterator position, and
inserts a stand-alone MachineInstr as before.
The insert() function that takes an MBB::instr_iterator position can
insert instructions inside a bundle, and will now update the bundle
flags correctly when that happens.
When the insert position is between two bundles, it is unclear whether
the instruction should be appended to the previous bundle, prepended to
the next bundle, or stand on its own. The MBB::insert() function doesn't
bundle the instruction in that case, use the MIBundleBuilder class for
that.
llvm-svn: 170437
A register can be associated with several distinct register classes.
For example, on PPC, the floating point registers are each associated with
both F4RC (which holds f32) and F8RC (which holds f64). As a result, this code
would fail when provided with a floating point register and an f64 operand
because it would happen to find the register in the F4RC class first and
return that. From the F4RC class, SDAG would extract f32 as the register
type and then assert because of the invalid implied conversion between
the f64 value and the f32 register.
Instead, search all register classes. If a register class containing the
the requested register has the requested type, then return that register
class. Otherwise, as before, return the first register class found that
contains the requested register.
llvm-svn: 170436
instruction.
This isn't strictly necessary at the moment because Thumb2SizeReduction
also copies all MI flags from the old instruction to the new. However, a
future patch will make that kind of direct flag tampering illegal.
llvm-svn: 170395
Most code is oblivious to bundles and uses the MBB::iterator which only
visits whole bundles. MBB::erase() operates on whole bundles at a time
as before.
MBB::remove() now refuses to remove bundled instructions. It is not safe
to remove all instructions in a bundle without deleting them since there
is no way of returning pointers to all the removed instructions.
MBB::remove_instr() and MBB::erase_instr() will now update bundle flags
correctly, lifting individual instructions out of bundles while leaving
the remaining bundle intact.
The MachineInstr convenience functions are updated so
eraseFromParent() erases a whole bundle as before
eraseFromBundle() erases a single instruction, leaving the rest of its bundle.
removeFromParent() refuses to operate on bundled instructions, and
removeFromBundle() lifts a single instruction out of its bundle.
These functions will no longer accidentally split or coalesce bundles -
bundle flags are updated to preserve the existing bundling, and explicit
bundleWith* / unbundleFrom* functions should be used to change the
instruction bundling.
This API update is still a work in progress. I am going to update APIs
first so they maintain bundle flags automatically when possible. Then
I'll add stricter verification of the bundle flags.
llvm-svn: 170384
compilation directory.
This defaults to the current working directory, just as it always has,
but now an assembler can choose to override it with a custom directory.
I've taught llvm-mc about this option and added a test case.
llvm-svn: 170371
This was a silly oversight, we weren't pruning allocas which were used
by variable-length memory intrinsics from the set that could be widened
and promoted as integers. Fix that.
llvm-svn: 170353
They seem to work fine.
Patch by: Christian König
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Christian König <deathsimple@vodafone.de>
llvm-svn: 170343
Patch by: Christian König
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Christian König <deathsimple@vodafone.de>
llvm-svn: 170342
The Align parameter is a power of two, so 16 results in 64K
alignment. Additional to that even 16 byte alignment doesn't
make any sense, so just remove it.
Patch by: Christian König
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Christian König <deathsimple@vodafone.de>
llvm-svn: 170341
This also cleans up a bit of the memcpy call rewriting by sinking some
irrelevant code further down and making the call-emitting code a bit
more concrete.
Previously, memcpy of a subvector would actually miscompile (!!!) the
copy into a single vector element copy. I have no idea how this ever
worked. =/ This is the memcpy half of PR14478 which we probably weren't
noticing previously because it didn't actually assert.
The rewrite relies on the newly refactored insert- and extractVector
functions to do the heavy lifting, and those are the same as used for
loads and stores which makes the test coverage a bit more meaningful
here.
llvm-svn: 170338
TargetLowering::getRegClassFor).
Some isSimple() guards were missing, or getSimpleVT() were hoisted too
far, resulting in asserts on valid LLVM assembly input.
llvm-svn: 170336
Check whether a BB is known as reachable before adding it to the worklist.
This way BB's with multiple predecessors are added to the list no more than
once.
llvm-svn: 170335
The first half of fixing this bug was actually in r170328, but was
entirely coincidental. It did however get me to realize the nature of
the bug, and adapt the test case to test more interesting behavior. In
turn, that uncovered the rest of the bug which I've fixed here.
This should fix two new asserts that showed up in the vectorize nightly
tester.
llvm-svn: 170333
I noticed this while looking at r170328. We only ever do a vector
rewrite when the alloca *is* the vector type, so it's good to not paper
over bugs here by doing a convertValue that isn't needed.
llvm-svn: 170331
This will allow its use inside of memcpy rewriting as well. This routine
is more complex than extractVector, and some of its uses are not 100%
where I want them to be so there is still some work to do here.
While this can technically change the output in some cases, it shouldn't
be a change that matters -- IE, it can leave some dead code lying around
that prior versions did not, etc.
Yet another step in the refactorings leading up to the solution to the
last component of PR14478.
llvm-svn: 170328
The method helpers all implicitly act upon the alloca, and what we
really want is a fully generic helper. Doing memcpy rewrites is more
special than all other rewrites because we are at times rewriting
instructions which touch pointers *other* than the alloca. As
a consequence all of the helpers needed by memcpy rewriting of
sub-vector copies will need to be generalized fully.
Note that all of these helpers ({insert,extract}{Integer,Vector}) are
woefully uncommented. I'm going to go back through and document them
once I get the factoring correct.
No functionality changed.
llvm-svn: 170325
PR14478 highlights a serious problem in SROA that simply wasn't being
exercised due to a lack of vector input code mixed with C-library
function calls. Part of SROA was written carefully to handle subvector
accesses via memset and memcpy, but the rewriter never grew support for
this. Fixing it required refactoring the subvector access code in other
parts of SROA so it could be shared, and then fixing the splat formation
logic and using subvector insertion (this patch).
The PR isn't quite fixed yet, as memcpy is still broken in the same way.
I'm starting on that series of patches now.
Hopefully this will be enough to bring the bullet benchmark back to life
with the bb-vectorizer enabled, but that may require fixing memcpy as
well.
llvm-svn: 170301
No functionality changed. Refactoring leading up to the fix for PR14478
which requires some significant changes to the memset and memcpy
rewriting.
llvm-svn: 170299
Bill Wendling