Used to iterate over previously added memory dependencies in
adjustChainDeps() and iterateChainSucc().
SDep::isCtrl() was previously used in these places, that also gave
anti and output edges. The code may be worse if these are followed,
because MisNeedChainEdge() will conservatively return true since a
non-memory instruction has no memory operands, and a false chain dep
will be added. It is also unnecessary since all memory accesses of
interest will be reached by memory dependencies, and there is a budget
limit for the number of edges traversed.
This problem was found on an out-of-tree target with enabled alias
analysis. No test case for an in-tree target has been found.
Reviewed by Hal Finkel.
llvm-svn: 225351
The change in r225266 was reviewed under D6722. But the commit r225266 has a
typo, causing some MCHammer failures. This patch fixes it.
Change-Id: I573efcff25003af7478ac02548ebbe929fc7f5fd
llvm-svn: 225347
Even thouh gcc produces simialr instructions as Owen pointed out the two patterns aren’t equivalent in the case
where the original subtraction could have caused an overflow.
Reverting the same.
llvm-svn: 225341
passes too many time.
I think this is actually the issue that someone raised with me at the
developer's meeting and in an email, but that we never really got to the
bottom of. Having all the testing utilities made it much easier to dig
down and uncover the core issue.
When a pass manager is running many passes over a single function, we
need it to invalidate the analyses between each run so that they can be
re-computed as needed. We also need to track the intersection of
preserved higher-level analyses across all the passes that we run (for
example, if there is one module analysis which all the function analyses
preserve, we want to track that and propagate it). Unfortunately, this
interacted poorly with any enclosing pass adaptor between two IR units.
It would see the intersection of preserved analyses, and need to
invalidate any other analyses, but some of the un-preserved analyses
might have already been invalidated *and recomputed*! We would fail to
propagate the fact that the analysis had already been invalidated.
The solution to this struck me as really strange at first, but the more
I thought about it, the more natural it seemed. After a nice discussion
with Duncan about it on IRC, it seemed even nicer. The idea is that
invalidating an analysis *causes* it to be preserved! Preserving the
lack of result is trivial. If it is recomputed, great. Until something
*else* invalidates it again, we're good.
The consequence of this is that the invalidate methods on the analysis
manager which operate over many passes now consume their
PreservedAnalyses object, update it to "preserve" every analysis pass to
which it delivers an invalidation (regardless of whether the pass
chooses to be removed, or handles the invalidation itself by updating
itself). Then we return this augmented set from the invalidate routine,
letting the pass manager take the result and use the intersection of
*that* across each pass run to compute the final preserved set. This
accounts for all the places where the early invalidation of an analysis
has already "preserved" it for a future run.
I've beefed up the testing and adjusted the assertions to show that we
no longer repeatedly invalidate or compute the analyses across nested
pass managers.
llvm-svn: 225333
WillNotOverflowUnsignedAdd's smarts will live in ValueTracking as
computeOverflowForUnsignedAdd. It now returns a tri-state result:
never overflows, always overflows and sometimes overflows.
llvm-svn: 225329
Remove the README.txt entry regarding register allocation of CR logical ops,
and replace it with a FIXME in PPCInstrInfo.td. The text in the README.txt was
not really accurate, and thanks goes to Pat Haugen (and Bill Schmidt) from IBM
for clarifying what was intended and highlighting the relevant text in the ISA
specification.
llvm-svn: 225325
This is affecting the behavior of some ObjC++ / AArch64 test cases on Darwin.
Reverting to get the bots green while I track down the source of the changed
behavior.
llvm-svn: 225311
In order to make comdats always explicit in the IR, we decided to make
the syntax a bit more compact for the case of a GlobalObject in a
comdat with the same name.
Just dropping the $name causes problems for
@foo = globabl i32 0, comdat
$bar = comdat ...
and
declare void @foo() comdat
$bar = comdat ...
So the syntax is changed to
@g1 = globabl i32 0, comdat($c1)
@g2 = globabl i32 0, comdat
and
declare void @foo() comdat($c1)
declare void @foo() comdat
llvm-svn: 225302
int->fp conversions on PPC must be done through memory loads and stores. On a
modern core, this process begins by storing the int value to memory, then
loading it using a (sometimes special) FP load instruction. Unfortunately, we
would do this even when the value to be converted was itself a load, and we can
just use that same memory location instead of copying it to another first.
There is a slight complication when handling int_to_fp(fp_to_int(x)) pairs,
because the fp_to_int operand has not been lowered when the int_to_fp is being
lowered. We handle this specially by invoking fp_to_int's lowering logic
(partially) and getting the necessary memory location (some trivial refactoring
was done to make this possible).
This is all somewhat ugly, and it would be nice if some later CodeGen stage
could just clean this stuff up, but because doing so would involve modifying
target-specific nodes (or instructions), it is not immediately clear how that
would work.
Also, remove a related entry from the README.txt for which we now generate
reasonable code.
llvm-svn: 225301
In DS write instructions, the address operand comes before the value
operand(s) which is reversed from every other instruction type.
The SIInsertWait assumed that the first use for each instruction
was the value, so for DS write it was protecting the address
operand with s_waitcnt instructions when it should have been
protecting the value operand.
llvm-svn: 225289
This is equivalent to the AMDGPUTargetMachine now, but it is the
starting point for separating R600 and GCN functionality into separate
targets.
It is recommened that users start using the gcn triple for GCN-based
GPUs, because using the r600 triple for these GPUs will be deprecated in
the future.
llvm-svn: 225277
This patch improves the logic added at revision 224899 (see review D6728) that
teaches the backend when it is profitable to speculate calls to cttz/ctlz.
The original algorithm conservatively avoided speculating more than one
instruction from a basic block in a control flow grap modelling an if-statement.
In particular, the only allowed instruction (excluding the terminator) was a
call to cttz/ctlz. However, there are cases where we could be less conservative
and still be able to speculate a call to cttz/ctlz.
With this patch, CodeGenPrepare now tries to speculate a cttz/ctlz if the
result is zero extended/truncated in the same basic block, and the zext/trunc
instruction is "free" for the target.
Added new test cases to CodeGen/X86/cttz-ctlz.ll
Differential Revision: http://reviews.llvm.org/D6853
llvm-svn: 225274
This also rolls in the changes discussed in http://reviews.llvm.org/D6766.
Defers migrating the debug info for new allocas until after all partitions
are created.
Thanks to Chandler for reviewing!
llvm-svn: 225272
The color scheme is the same as the one used by the colorize dwarfdump
script on Darwin.
A new --color option can be used to forcibly turn color on or off.
http://reviews.llvm.org/D6852
llvm-svn: 225269
No functional changes. Support for ARM's modified immediate syntax was added
in r223113 and r223115 (review: D6408). That patch introduced the mod_imm*
tblegen definitions which renders the existing so_imm* definitions redundant.
This patch gets rid of them completely.
Reviewed as: D6722
llvm-svn: 225266
This is already handled in general when it is known the
conversion can't lose bits with smaller integer types
casted into wider floating point types.
This pattern happens somewhat often in GPU programs that cast
workitem intrinsics to float, which are often compared with 0.
Specifically handle the special case of compares with zero which
should also be known to not lose information. I had a more general
version of this which allows equality compares if the casted float is
exactly representable in the integer, but I'm not 100% confident that
is always correct.
Also fold cases that aren't integers to true / false.
llvm-svn: 225265
Requires new AsmParserOperand types that detect 16-bit and 32/64-bit mode so that we choose the right instruction based on default sizing without predicates. This is necessary since predicates mess up the disassembler table building.
llvm-svn: 225256
Try harder to get rid of bitcast'd calls by ptrtoint/inttoptr'ing
arguments and return values when DataLayout says it is safe to do so.
llvm-svn: 225254
Overall this seems simpler. It reduces duplication of patterns between both modes and it simplifies the memory folding/unfolding tables as they don't need to create fake instructions just to keep track of 64-bitness.
llvm-svn: 225252
Because of how Clang represents structs as arrays (at least on non-Darwin
platforms), and what SROA does, etc. this is no longer a problem.
llvm-svn: 225251
"ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
relocation target a movq or addq instruction.
Prohibit the truncation of such loads to movl or addl.
This fixes PR22083.
Differential Revision: http://reviews.llvm.org/D6839
llvm-svn: 225250
The old target DAG combine that allowed for performing int_to_fp(fp_to_int(x))
without a load/store pair is updated here with support for unsigned integers,
and to support single-precision values without a third rounding step, on newer
cores with the appropriate instructions.
llvm-svn: 225248
a specific analysis result.
This is quite handy to test things, and will also likely be very useful
for debugging issues. You could narrow down pass validation failures by
walking these invalidate pass runs up and down the pass pipeline, etc.
I've added support to the pass pipeline parsing to be able to create one
of these for any analysis pass desired.
Just adding this class uncovered one latent bug where the
AnalysisManager CRTP base class had a hard-coded Module type rather than
using IRUnitT.
I've also added tests for invalidation and caching of analyses in
a basic way across all the pass managers. These in turn uncovered two
more bugs where we failed to correctly invalidate an analysis -- its
results were invalidated but the key for re-running the pass was never
cleared and so it was never re-run. Quite nasty. I'm very glad to debug
this here rather than with a full system.
Also, yes, the naming here is horrid. I'm going to update some of the
names to be slightly less awful shortly. But really, I've no "good"
ideas for naming. I'll be satisfied if I can get it to "not bad".
llvm-svn: 225246
The assembler backend will relax to the long form if necessary. This removes a swap from long form to short form in the MCInstLowering code. Selecting the long form used to be required by the old JIT.
llvm-svn: 225242
manager tests to use them and be significantly more comprehensive.
This, naturally, uncovered a bug where the CGSCC pass manager wasn't
printing analyses when they were run.
The only remaining core manipulator is I think an invalidate pass
similar to the require pass. That'll be next. =]
llvm-svn: 225240
Now that `LLVMContextImpl` can call `MDNode::dropAllReferences()` to
prevent teardown madness, stop dropping uniquing just because an operand
drops to null.
Part of PR21532.
llvm-svn: 225223
We no longer generate horrible code for the stated function:
void f(signed char *a, _Bool b, _Bool c) {
signed char t = 0;
if (b) t = *a;
if (c) *a = t;
}
for which we now generate:
.L.f:
andi. 5, 5, 1
cmpldi 1, 4, 0
li 5, 0
beq 1, .LBB0_2
lbz 5, 0(3)
.LBB0_2: # %if.end
bclr 4, 1, 0
stb 5, 0(3)
blr
so we don't need the README.txt entry.
llvm-svn: 225217
We now produce the desired code as noted in the README.txt file (no spurious
or). Remove the README entry and improve the regression test.
llvm-svn: 225214
dsymutil would like to use all the AsmPrinter/MCStreamer infrastructure
to stream out the DWARF. In order to do so, it will reuse the DIE object
and so this header needs to be public.
The interface exposed here has some corners that cannot be used without a
DwarfDebug object, but clients that want to stream Dwarf can just avoid
these.
Differential Revision: http://reviews.llvm.org/D6695
llvm-svn: 225208
Consider this function from our README.txt file:
int foo(int a, int b) { return (a < b) << 4; }
We now explicitly track CR bits by default, so the comment in the README.txt
about not really having a SETCC is no longer accurate, but we did generate this
somewhat silly code:
cmpw 0, 3, 4
li 3, 0
li 12, 1
isel 3, 12, 3, 0
sldi 3, 3, 4
blr
which generates the zext as a select between 0 and 1, and then shifts the
result by a constant amount. Here we preprocess the DAG in order to fold the
results of operations on an extension of an i1 value into the SELECT_I[48]
pseudo instruction when the resulting constant can be materialized using one
instruction (just like the 0 and 1). This was not implemented as a DAGCombine
because the resulting code would have been anti-canonical and depends on
replacing chained user nodes, which does not fit well into the lowering
paradigm. Now we generate:
cmpw 0, 3, 4
li 3, 0
li 12, 16
isel 3, 12, 3, 0
blr
which is less silly.
llvm-svn: 225203
The 64-bit semantics of cntlzw are not special, the 32-bit population count is
stored as a 64-bit value in the range [0,32]. As a result, it is always zero
extended, and it can be added to the PPCISelDAGToDAG peephole optimization as a
frontier instruction for the removal of unnecessary zero extensions.
llvm-svn: 225192
lhbrx and lwbrx not only load their data with byte swapping, but also clear the
upper 32 bits (at least). As a result, they can be added to the PPCISelDAGToDAG
peephole optimization as frontier instructions for the removal of unnecessary
zero extensions.
llvm-svn: 225189
The swap implementation for iplist is currently unsupported. Simply splice the
old list into place, which achieves the same purpose. This is needed in order
to thread the -frewrite-map-file frontend option correctly. NFC.
llvm-svn: 225186
We used to generate code similar to:
umov.b w8, v0[2]
strb w8, [x0, x1]
because the STR*ro* patterns were preferred to ST1*.
Instead, we can avoid going through GPRs, and generate:
add x8, x0, x1
st1.b { v0 }[2], [x8]
This patch increases the ST1* AddedComplexity to achieve that.
rdar://16372710
Differential Revision: http://reviews.llvm.org/D6202
llvm-svn: 225183
For 0-lane stores, we used to generate code similar to:
fmov w8, s0
str w8, [x0, x1, lsl #2]
instead of:
str s0, [x0, x1, lsl #2]
To correct that: for store lane 0 patterns, directly match to STR <subreg>0.
Byte-sized instructions don't have the special case for a 0 index,
because FPR8s are defined to have untyped content.
rdar://16372710
Differential Revision: http://reviews.llvm.org/D6772
llvm-svn: 225181
Tag_compatibility takes two arguments, but before this patch it would
erroneously accept just one, it now produces an error in that case.
Change-Id: I530f918587620d0d5dfebf639944d6083871ef7d
llvm-svn: 225167
Claim conformance to version 2.09 of the ARM ABI.
This build attribute must be emitted first amongst the build attributes when
written to an object file. This is to simplify conformance detection by
consumers.
Change-Id: If9eddcfc416bc9ad6e5cc8cdcb05d0031af7657e
llvm-svn: 225166
This patch lowers patterns such as-
sub v0.4s, v0.4s, v1.4s
abs v0.4s, v0.4s
to
sabd v0.4s, v0.4s, v1.4s
on AArch64.
Review: http://reviews.llvm.org/D6781
llvm-svn: 225165
when all are being preserved.
We want to short-circuit this for a couple of reasons. One, I don't
really want passes to grow a dependency on actually receiving their
invalidate call when they've been preserved. I'm thinking about removing
this entirely. But more importantly, preserving everything is likely to
be the common case in a lot of scenarios, and it would be really good to
bypass all of the invalidation and preservation machinery there.
Avoiding calling N opaque functions to try to invalidate things that are
by definition still valid seems important. =]
This wasn't really inpsired by much other than seeing the spam in the
logging for analyses, but it seems better ot get it checked in rather
than forgetting about it.
llvm-svn: 225163
manager.
This starts to allow us to test analyses more easily, but it's really
only the beginning. Some of the code here is still untestable without
manual changes to create analysis passes, but I wanted to factor it into
a small of chunks as possible.
Next up in order to be able to test things are, in no particular order:
- No-op analyses passes so we don't have to use real ones to exercise
the pass maneger itself.
- Automatic way of generating dummy passes that require an analysis be
run, including a variant that calls a 'print' method on a pass to make
it even easier to print out the results of an analysis.
- Dummy passes that invalidate all analyses for their IR unit so we can
test invalidation and re-runs.
- Automatic way to print each analysis pass as it is re-run.
- Automatic but optional verification of analysis passes everywhere
possible.
I'm not claiming I'll get to all of these immediately, but that's what
is in the pipeline at some stage. I'm fleshing out exactly what I need
and what to prioritize by working on converting analyses and then trying
to test the conversion. =]
llvm-svn: 225162
{code}
// loop body
... = a[i] (1)
... = a[i+1] (2)
.......
a[i+1] = .... (3)
a[i] = ... (4)
{code}
The algorithm tries to collect memory access candidates from AliasSetTracker, and then check memory dependences one another. The memory accesses are unique in AliasSetTracker, and a single memory access in AliasSetTracker may map to multiple entries in AccessAnalysis, which could cover both 'read' and 'write'. Originally the algorithm only checked 'write' entry in Accesses if only 'write' exists. This is incorrect and the consequence is it ignored all read access, and finally some RAW and WAR dependence are missed.
For the case given above, if we ignore two reads, the dependence between (1) and (3) would not be able to be captured, and finally this loop will be incorrectly vectorized.
The fix simply inserts a new loop to find all entries in Accesses. Since it will skip most of all other memory accesses by checking the Value pointer at the very beginning of the loop, it should not increase compile-time visibly.
llvm-svn: 225159
PPC has an instruction for ctlz with defined zero behavior, and our lowering of
cttz (provided by DAGCombine) is also efficient and branchless, so speculating
these makes sense.
llvm-svn: 225150
assert out of the new pre-splitting in SROA.
This fix makes the code do what was originally intended -- when we have
a store of a load both dealing in the same alloca, we force them to both
be pre-split with identical offsets. This is really quite hard to do
because we can keep discovering problems as we go along. We have to
track every load over the current alloca which for any resaon becomes
invalid for pre-splitting, and go back to remove all stores of those
loads. I've included a couple of test cases derived from PR22093 that
cover the different ways this can happen. While that PR only really
triggered the first of these two, its the same fundamental issue.
The other challenge here is documented in a FIXME now. We end up being
quite a bit more aggressive for pre-splitting when loads and stores
don't refer to the same alloca. This aggressiveness comes at the cost of
introducing potentially redundant loads. It isn't clear that this is the
right balance. It might be considerably better to require that we only
do pre-splitting when we can presplit every load and store involved in
the entire operation. That would give more consistent if conservative
results. Unfortunately, it requires a non-trivial change to the actual
pre-splitting operation in order to correctly handle cases where we end
up pre-splitting stores out-of-order. And it isn't 100% clear that this
is the right direction, although I'm starting to suspect that it is.
llvm-svn: 225149
r225135 added the ability to materialize i64 constants using rotations in order
to reduce the instruction count. Sometimes we can use a rotation only with some
extra masking, so that we take advantage of the fact that generating a bunch of
extra higher-order 1 bits is easy using li/lis.
llvm-svn: 225147
units.
This was debated back and forth a bunch, but using references is now
clearly cleaner. Of all the code written using pointers thus far, in
only one place did it really make more sense to have a pointer. In most
cases, this just removes immediate dereferencing from the code. I think
it is much better to get errors on null IR units earlier, potentially
at compile time, than to delay it.
Most notably, the legacy pass manager uses references for its routines
and so as more and more code works with both, the use of pointers was
likely to become really annoying. I noticed this when I ported the
domtree analysis over and wrote the entire thing with references only to
have it fail to compile. =/ It seemed better to switch now than to
delay. We can, of course, revisit this is we learn that references are
really problematic in the API.
llvm-svn: 225145
from before I removed thet non-const use of the function.
The unused variable that held the const_cast was already kindly removed
by Michael.
llvm-svn: 225143
Materializing full 64-bit constants on PPC64 can be expensive, requiring up to
5 instructions depending on the locations of the non-zero bits. Sometimes
materializing a rotated constant, and then applying the inverse rotation, requires
fewer instructions than the direct method. If so, do that instead.
In r225132, I added support for forming constants using bit inversion. In
effect, this reverts that commit and replaces it with rotation support. The bit
inversion is useful for turning constants that are mostly ones into ones that
are mostly zeros (thus enabling a more-efficient shift-based materialization),
but the same effect can be obtained by using negative constants and a rotate,
and that is at least as efficient, if not more.
llvm-svn: 225135
Materializing full 64-bit constants on PPC64 can be expensive, requiring up to
5 instructions depending on the locations of the non-zero bits. Sometimes
materializing the bit-reversed constant, and then flipping the bits, requires
fewer instructions than the direct method. If so, do that instead.
llvm-svn: 225132
a cache of assumptions for a single function, and an immutable pass that
manages those caches.
The motivation for this change is two fold. Immutable analyses are
really hacks around the current pass manager design and don't exist in
the new design. This is usually OK, but it requires that the core logic
of an immutable pass be reasonably partitioned off from the pass logic.
This change does precisely that. As a consequence it also paves the way
for the *many* utility functions that deal in the assumptions to live in
both pass manager worlds by creating an separate non-pass object with
its own independent API that they all rely on. Now, the only bits of the
system that deal with the actual pass mechanics are those that actually
need to deal with the pass mechanics.
Once this separation is made, several simplifications become pretty
obvious in the assumption cache itself. Rather than using a set and
callback value handles, it can just be a vector of weak value handles.
The callers can easily skip the handles that are null, and eventually we
can wrap all of this up behind a filter iterator.
For now, this adds boiler plate to the various passes, but this kind of
boiler plate will end up making it possible to port these passes to the
new pass manager, and so it will end up factored away pretty reasonably.
llvm-svn: 225131
We assumed the output of a match was a Value, this would cause us to
assert because we would fail a cast<>. Instead, use a helper in the
Operator family to hide the distinction between Value and Constant.
This fixes PR22087.
llvm-svn: 225127
PHI nodes can have zero operands in the middle of a transform. It is
expected that utilities in Analysis don't freak out when this happens.
Note that it is considered invalid to allow these misshapen phi nodes to
make it to another pass.
This fixes PR22086.
llvm-svn: 225126
This patch adds a check for underflow when truncating results back to lower
precision at the end of an FMA. The additional sign handling logic in
APFloat::fusedMultiplyAdd should only be performed when the result of the
addition step of the FMA (in full precision) is exactly zero, not when the
result underflows to zero.
Unit tests for this case and related signed zero FMA results are included.
Fixes <rdar://problem/18925551>.
llvm-svn: 225123
Weak externals are resolved statically, so we can actually generate the tail
call on PE/COFF targets without breaking the requirements. It is questionable
whether we want to propagate the current behaviour for MachO as the requirements
are part of the ARM ELF specifications, and it seems that prior to the SVN
r215890, we would have tail'ed the call. For now, be conservative and only
permit it on PE/COFF where the call will always be fully resolved.
llvm-svn: 225119
The existing code provided for specifying a global loop alignment preference.
However, the preferred loop alignment might depend on the loop itself. For
recent POWER cores, loops between 5 and 8 instructions should have 32-byte
alignment (while the others are better with 16-byte alignment) so that the
entire loop will fit in one i-cache line.
To support this, getPrefLoopAlignment has been made virtual, and can be
provided with an optional MachineLoop* so the target can inspect the loop
before answering the query. The default behavior, as before, is to return the
value set with setPrefLoopAlignment. MachineBlockPlacement now queries the
target for each loop instead of only once per function. There should be no
functional change for other targets.
llvm-svn: 225117
Most modern PowerPC cores prefer that functions and loops start on
16-byte-aligned boundaries (*), so instruct block placement, etc. to make this
happen. The branch selector has also been adjusted so account for the extra
nops that might now be inserted before loop headers.
(*) Some cores actually prefer other alignments for small loops, but that will
be addressed in a follow-up commit.
llvm-svn: 225115
Make sure they all have llvm_unreachable on the default path out of the switch. Remove unnecessary "default: break". Remove a 'return' after unreachable. Fix some indentation.
llvm-svn: 225114
We would sometimes leave the out-param APInts untouched while going
through computeKnownBits. While I don't know of a way to trigger a bug
involving this in practice, it goes against the overall design of
computeKnownBits.
Found via code inspection.
llvm-svn: 225109
Newer POWER cores, and the A2, support the cmpb instruction. This instruction
compares its operands, treating each of the 8 bytes in the GPRs separately,
returning a 'mask' result of 0 (for false) or -1 (for true) in each byte.
Code generation support is added, in the form of a PPCISelDAGToDAG
DAG-preprocessing routine, that recognizes patterns close to what the
instruction computes (either exactly, or related by a constant masking
operation), and generates the cmpb instruction (along with any necessary
constant masking operation). This can be expanded if use cases arise.
llvm-svn: 225106
WillNotOverflowUnsignedMul's smarts will live in ValueTracking as
computeOverflowForUnsignedMul. It now returns a tri-state result:
never overflows, always overflows and sometimes overflows.
llvm-svn: 225076
This is necessary to allow the disassembler to be able to handle AdSize32 instructions in 64-bit mode when address size prefix is used.
Eventually we should probably also support 'addr32' and 'addr16' in the assembler to override the address size on some of these instructions. But for now we'll just use special operand types that will lookup the current mode size to select the right instruction.
llvm-svn: 225075
a pre-splitting pass over loads and stores.
Historically, splitting could cause enough problems that I hamstrung the
entire process with a requirement that splittable integer loads and
stores must cover the entire alloca. All smaller loads and stores were
unsplittable to prevent chaos from ensuing. With the new pre-splitting
logic that does load/store pair splitting I introduced in r225061, we
can now very nicely handle arbitrarily splittable loads and stores. In
order to fully benefit from these smarts, we need to mark all of the
integer loads and stores as splittable.
However, we don't actually want to rewrite partitions with all integer
loads and stores marked as splittable. This will fail to extract scalar
integers from aggregates, which is kind of the point of SROA. =] In
order to resolve this, what we really want to do is only do
pre-splitting on the alloca slices with integer loads and stores fully
splittable. This allows us to uncover all non-integer uses of the alloca
that would benefit from a split in an integer load or store (and where
introducing the split is safe because it is just memory transfer from
a load to a store). Once done, we make all the non-whole-alloca integer
loads and stores unsplittable just as they have historically been,
repartition and rewrite.
The result is that when there are integer loads and stores anywhere
within an alloca (such as from a memcpy of a sub-object of a larger
object), we can split them up if there are non-integer components to the
aggregate hiding beneath. I've added the challenging test cases to
demonstrate how this is able to promote to scalars even a case where we
have even *partially* overlapping loads and stores.
This restores the single-store behavior for small arrays of i8s which is
really nice. I've restored both the little endian testing and big endian
testing for these exactly as they were prior to r225061. It also forced
me to be more aggressive in an alignment test to actually defeat SROA.
=] Without the added volatiles there, we actually split up the weird i16
loads and produce nice double allocas with better alignment.
This also uncovered a number of bugs where we failed to handle
splittable load and store slices which didn't have a begininng offset of
zero. Those fixes are included, and without them the existing test cases
explode in glorious fireworks. =]
I've kept support for leaving whole-alloca integer loads and stores as
splittable even for the purpose of rewriting, but I think that's likely
no longer needed. With the new pre-splitting, we might be able to remove
all the splitting support for loads and stores from the rewriter. Not
doing that in this patch to try to isolate any performance regressions
that causes in an easy to find and revert chunk.
llvm-svn: 225074
instructions.
I noticed this when working on dialing up how aggressively we can
pre-split loads and stores. My test case wasn't passing because dead
GEPs into the allocas persisted when they were built by this routine.
This isn't terribly harmful, we still rewrote and promoted the alloca
and I can't conceive of how to cause this to happen in a case where we
will keep the exact same alloca but rewrite and promote the uses of it.
If that ever happened, we'd get an assert out of mem2reg.
So I don't have a direct test case yet, but the subsequent commit's test
case wouldn't pass without this. There are other problems fixed by this
patch that I spotted purely by inspection such as the fact that
getAdjustedPtr could have actually deleted dead base pointers. I don't
know how to get a base pointer to go into getAdjustedPtr today, so
I think this bug could never have manifested (and I certainly can't
write a test case for it) but, it wasn't the intent of the code. The
code really just wanted to GC the new instructions built. That can be
done more directly by comparing with the base pointer which is the only
non-new instruction that this code can return.
llvm-svn: 225073
array. This prevents it from walking out of bounds on the splits array.
Bug found with the existing tests by ASan and by the MSVC debug build.
llvm-svn: 225069
a +asserts bootstrap, but my bootstrap had asserts off. Oops.
Anyways, in some places it is reasonable to cast (as a sanity check) the
pointer operand to a load or store to an instruction within SROA --
namely when the pointer operand is expected to be derived from an
alloca, and thus always an instruction. However, the pre-splitting code
also deals with loads and stores to non-alloca pointers and there we
need to just use the Value*. Nothing about the code relied on the
instruction cast, it was only there essentially as an invariant
assertion. Remove the two that don't actually hold.
This should fix the proximate issue in PR22080, but I'm also doing an
asserts bootstrap myself to see if there are other issues lurking.
I'll craft a reduced test case in a moment, but I wanted to get the tree
healthy as quickly as possible.
llvm-svn: 225068
Attempting to fix PR22078 (building on 32-bit systems) by replacing my careless
use of 1ul to be a uint64_t constant with UINT64_C(1).
llvm-svn: 225066
of my new load and store splitting, and fix a bug where it logged
a totally irrelevant slice rather than the actual slice in question.
The logging here previously worked because we used to place new slices
onto the back of the core sequence, but that caused other problems.
I updated the actual code to store new slices in their own vector but
didn't update the logging. There isn't a good way to reuse the logging
any more, and frankly it wasn't needed. We can directly log this bit
more easily.
llvm-svn: 225063
stores.
When there are accesses to an entire alloca with an integer
load or store as well as accesses to small pieces of the alloca, SROA
splits up the large integer accesses. In order to do that, it uses bit
math to merge the small accesses into large integers. While this is
effective, it produces insane IR that can cause significant problems in
the rest of the optimizer:
- It can cause load and store mismatches with GVN on the non-alloca side
where we end up loading an i64 (or some such) rather than loading
specific elements that are stored.
- We can't always get rid of the integer bit math, which is why we can't
always fix the loads and stores to work well with GVN.
- This is especially bad when we have operations that mix poorly with
integer bit math such as floating point operations.
- It will block things like the vectorizer which might be able to handle
the scalar stores that underly the aggregate.
At the same time, we can't just directly split up these loads and stores
in all cases. If there is actual integer arithmetic involved on the
values, then using integer bit math is actually the perfect lowering
because we can often combine it heavily with the surrounding math.
The solution this patch provides is to find places where SROA is
partitioning aggregates into small elements, and look for splittable
loads and stores that it can split all the way to some other adjacent
load and store. These are uniformly the cases where failing to split the
loads and stores hurts the optimizer that I have seen, and I've looked
extensively at the code produced both from more and less aggressive
approaches to this problem.
However, it is quite tricky to actually do this in SROA. We may have
loads and stores to the same alloca, or other complex patterns that are
hard to handle. This complexity leads to the somewhat subtle algorithm
implemented here. We have to do this entire process as a separate pass
over the partitioning of the alloca, and split up all of the loads prior
to splitting the stores so that we can handle safely the cases of
overlapping, including partially overlapping, loads and stores to the
same alloca. We also have to reconstitute the post-split slice
configuration so we can avoid iterating again over all the alloca uses
(the slow part of SROA). But we also have to ensure that when we split
up loads and stores to *other* allocas, we *do* re-iterate over them in
SROA to adapt to the more refined partitioning now required.
With this, I actually think we can fix a long-standing TODO in SROA
where I avoided splitting as many loads and stores as probably should be
splittable. This limitation historically mitigated the fallout of all
the bad things mentioned above. Now that we have more intelligent
handling, I plan to remove the FIXME and more aggressively mark integer
loads and stores as splittable. I'll do that in a follow-up patch to
help with bisecting any fallout.
The net result of this change should be more fine-grained and accurate
scalars being formed out of aggregates. At the very least, Clang now
generates perfect code for this high-level test case using
std::complex<float>:
#include <complex>
void g1(std::complex<float> &x, float a, float b) {
x += std::complex<float>(a, b);
}
void g2(std::complex<float> &x, float a, float b) {
x -= std::complex<float>(a, b);
}
void foo(const std::complex<float> &x, float a, float b,
std::complex<float> &x1, std::complex<float> &x2) {
std::complex<float> l1 = x;
g1(l1, a, b);
std::complex<float> l2 = x;
g2(l2, a, b);
x1 = l1;
x2 = l2;
}
This code isn't just hypothetical either. It was reduced out of the hot
inner loops of essentially every part of the Eigen math library when
using std::complex<float>. Those loops would consistently and
pervasively hop between the floating point unit and the integer unit due
to bit math extraction and insertion of floating point values that were
"stored" in a 64-bit integer register around the loop backedge.
So far, this change has passed a bootstrap and I have done some other
testing and so far, no issues. That doesn't mean there won't be though,
so I'll be prepared to help with any fallout. If you performance swings
in particular, please let me know. I'm very curious what all the impact
of this change will be. Stay tuned for the follow-up to also split more
integer loads and stores.
llvm-svn: 225061
This is the second installment of improvements to instruction selection for "bit
permutation" instruction sequences. r224318 added logic for instruction
selection for 32-bit bit permutation sequences, and this adds lowering for
64-bit sequences. The 64-bit sequences are more complicated than the 32-bit
ones because:
a) the 64-bit versions of the 32-bit rotate-and-mask instructions
work by replicating the lower 32-bits of the value-to-be-rotated into the
upper 32 bits -- and integrating this into the cost modeling for the various
bit group operations is non-trivial
b) unlike the 32-bit instructions in 32-bit mode, the rotate-and-mask instructions
cannot, in one instruction, specify the
mask starting index, the mask ending index, and the rotation factor. Also,
forming arbitrary 64-bit constants is more complicated than in 32-bit mode
because the number of instructions necessary is value dependent.
Plus, support for 'late masking' was added: it is sometimes more efficient to
treat the overall value as if it had no mandatory zero bits when planning the
bit-group insertions, and then mask them in at the very end. Unfortunately, as
the structure of the bit groups is different in the two cases, the more
feasible implementation technique was to generate both instruction sequences,
and then pick the shorter one.
And finally, we now generate reasonable code for i64 bswap:
rldicl 5, 3, 16, 0
rldicl 4, 3, 8, 0
rldicl 6, 3, 24, 0
rldimi 4, 5, 8, 48
rldicl 5, 3, 32, 0
rldimi 4, 6, 16, 40
rldicl 6, 3, 48, 0
rldimi 4, 5, 24, 32
rldicl 5, 3, 56, 0
rldimi 4, 6, 40, 16
rldimi 4, 5, 48, 8
rldimi 4, 3, 56, 0
vs. what we used to produce:
li 4, 255
rldicl 5, 3, 24, 40
rldicl 6, 3, 40, 24
rldicl 7, 3, 56, 8
sldi 8, 3, 8
sldi 10, 3, 24
sldi 12, 3, 40
rldicl 0, 3, 8, 56
sldi 9, 4, 32
sldi 11, 4, 40
sldi 4, 4, 48
andi. 5, 5, 65280
andis. 6, 6, 255
andis. 7, 7, 65280
sldi 3, 3, 56
and 8, 8, 9
and 4, 12, 4
and 9, 10, 11
or 6, 7, 6
or 5, 5, 0
or 3, 3, 4
or 7, 9, 8
or 4, 6, 5
or 3, 3, 7
or 3, 3, 4
which is 12 instructions, instead of 25, and seems optimal (at least in terms
of code size).
llvm-svn: 225056
The issues was that AArch64 has additional restrictions on when local
relocations can be used. We have to take those into consideration when
deciding to put a L symbol in the symbol table or not.
Original message:
Remove doesSectionRequireSymbols.
In an assembly expression like
bar:
.long L0 + 1
the intended semantics is that bar will contain a pointer one byte past L0.
In sections that are merged by content (strings, 4 byte constants, etc), a
single position in the section doesn't give the linker enough information.
For example, it would not be able to tell a relocation must point to the
end of a string, since that would look just like the start of the next.
The solution used in ELF to use relocation with symbols if there is a non-zero
addend.
In MachO before this patch we would just keep all symbols in some sections.
This would miss some cases (only cstrings on x86_64 were implemented) and was
inefficient since most relocations have an addend of 0 and can be represented
without the symbol.
This patch implements the non-zero addend logic for MachO too.
llvm-svn: 225048
We are allowed to move the 'B' to the right hand side if we an prove
there is no signed overflow and if the comparison itself is signed.
llvm-svn: 225034
GCC does this for non-zero discriminators and since GCC doesn't produce
column info, that was the only place it comes up there. For LLVM, since
we can emit discriminators and/or column info, it makes more sense to
invert the condition and just test for changes in line number.
This should resolve at least some of the GDB 7.5 test suite failures
created by recent Clang changes that increase the location fidelity
(which, since Clang defaults to including column info on Linux by
default created a bunch of cases that confused GDB).
In theory we could do this better/differently by grouping actual source
statements together in a similar manner to the way lexical scopes are
handled but given that GDB isn't really in a position to consume that (&
users are probably somewhat used to different lines being different
'statements') this seems the safest and cheapest change. (I'm concerned
that doing this 'right' would bloat the debugloc data even further -
something Duncan's working hard to address)
llvm-svn: 225011
Tom Stellard