properly account for the *global* probability of the edge being taken.
This manifested as a very large number of unconditional branches to
blocks being merged against the CFG even though they weren't
particularly hot within the CFG.
The fix is to check whether the edge being merged is both locally hot
relative to other successors for the source block, and globally hot
compared to other (unmerged) predecessors of the destination block.
This introduces a new crasher on GCC single-source, but it's currently
behind a flag, and Ben has offered to work on the reduction. =]
llvm-svn: 145010
is actually being tested. Also add some FileCheck goodness to much more
carefully ensure that the result is the desired result. Before this test
would only have failed through an assert failure if the underlying fix
were reverted.
Also, add some weight metadata and a comment explaining exactly what is
going on to a trick section of the test case. Originally, we were
getting very unlucky and trying to form a block chain that isn't
actually profitable. I'm working on a fix to avoid forming these
unprofitable chains, and that would also have masked any failure from
this test case. The easy solution is to add some metadata that makes it
*really* profitable to form the bad chain here.
llvm-svn: 145006
formation phase and into the initial walk of the basic blocks. We
essentially pre-merge all blocks where unanalyzable fallthrough exists,
as we won't be able to update the terminators effectively after any
reorderings. This is quite a bit more principled as there may be CFGs
where the second half of the unanalyzable pair has some analyzable
predecessor that gets placed first. Then it may get placed next,
implicitly breaking the unanalyzable branch even though we never even
looked at the part that isn't analyzable. I've included a test case that
triggers this (thanks Benjamin yet again!), and I'm hoping to synthesize
some more general ones as I dig into related issues.
Also, to make this new scheme work we have to be able to handle branches
into the middle of a chain, so add this check. We always fallback on the
incoming ordering.
Finally, this starts to really underscore a known limitation of the
current implementation -- we don't consider broken predecessors when
merging successors. This can caused major missed opportunities, and is
something I'm planning on looking at next (modulo more bug reports).
llvm-svn: 144994
has a reference to it. Unfortunately, that doesn't work for codegen passes
since we don't get notified of MBB's being deleted (the original BB stays).
Use that fact to our advantage and after printing a function, check if
any of the IL BBs corresponds to a symbol that was not printed. This fixes
pr11202.
llvm-svn: 144674
block sequence when recovering from unanalyzable control flow
constructs, *always* use the function sequence. I'm not sure why I ever
went down the path of trying to use the loop sequence, it is
fundamentally not the correct sequence to use. We're trying to preserve
the incoming layout in the cases of unreasonable control flow, and that
is only encoded at the function level. We already have a filter to
select *exactly* the sub-set of blocks within the function that we're
trying to form into a chain.
The resulting code layout is also significantly better because of this.
In several places we were ending up with completely unreasonable control
flow constructs due to the ordering chosen by the loop structure for its
internal storage. This change removes a completely wasteful vector of
basic blocks, saving memory allocation in the common case even though it
costs us CPU in the fairly rare case of unnatural loops. Finally, it
fixes the latest crasher reduced out of GCC's single source. Thanks
again to Benjamin Kramer for the reduction, my bugpoint skills failed at
it.
llvm-svn: 144627
Two new TargetInstrInfo hooks lets the target tell ExecutionDepsFix
about instructions with partial register updates causing false unwanted
dependencies.
The ExecutionDepsFix pass will break the false dependencies if the
updated register was written in the previoius N instructions.
The small loop added to sse-domains.ll runs twice as fast with
dependency-breaking instructions inserted.
llvm-svn: 144602
instructions of the two-address operands) in order to avoid inserting copies.
This fixes the few regressions introduced when the two-address hack was
disabled (without regressing the improvements).
rdar://10422688
llvm-svn: 144559
the sum of the edge weights not overflowing uint32, and crashed when
they did. This is generally safe as BranchProbabilityInfo tries to
provide this guarantee. However, the CFG can get modified during codegen
in a way that grows the *sum* of the edge weights. This doesn't seem
unreasonable (imagine just adding more blocks all with the default
weight of 16), but it is hard to come up with a case that actually
triggers 32-bit overflow. Fortuately, the single-source GCC build is
good at this. The solution isn't very pretty, but its no worse than the
previous code. We're already summing all of the edge weights on each
query, we can sum them, check for an overflow, compute a scale, and sum
them again.
I've included a *greatly* reduced test case out of the GCC source that
triggers it. It's a pretty lame test, as it clearly is just barely
triggering the overflow. I'd like to have something that is much more
definitive, but I don't understand the fundamental pattern that triggers
an explosion in the edge weight sums.
The buggy code is duplicated within this file. I'll colapse them into
a single implementation in a subsequent commit.
llvm-svn: 144526
get loop info structures associated with them, and so we need some way
to make forward progress selecting and placing basic blocks. The
technique used here is pretty brutal -- it just scans the list of blocks
looking for the first unplaced candidate. It keeps placing blocks like
this until the CFG becomes tractable.
The cost is somewhat unfortunate, it requires allocating a vector of all
basic block pointers eagerly. I have some ideas about how to simplify
and optimize this, but I'm trying to get the logic correct first.
Thanks to Benjamin Kramer for the reduced test case out of GCC. Sadly
there are other bugs that GCC is tickling that I'm reducing and working
on now.
llvm-svn: 144516
second algorithm, but only loosely. It is more heavily based on the last
discussion I had with Andy. It continues to walk from the inner-most
loop outward, but there is a key difference. With this algorithm we
ensure that as we visit each loop, the entire loop is merged into
a single chain. At the end, the entire function is treated as a "loop",
and merged into a single chain. This chain forms the desired sequence of
blocks within the function. Switching to a single algorithm removes my
biggest problem with the previous approaches -- they had different
behavior depending on which system triggered the layout. Now there is
exactly one algorithm and one basis for the decision making.
The other key difference is how the chain is formed. This is based
heavily on the idea Andy mentioned of keeping a worklist of blocks that
are viable layout successors based on the CFG. Having this set allows us
to consistently select the best layout successor for each block. It is
expensive though.
The code here remains very rough. There is a lot that needs to be done
to clean up the code, and to make the runtime cost of this pass much
lower. Very much WIP, but this was a giant chunk of code and I'd rather
folks see it sooner than later. Everything remains behind a flag of
course.
I've added a couple of tests to exercise the issues that this iteration
was motivated by: loop structure preservation. I've also fixed one test
that was exhibiting the broken behavior of the previous version.
llvm-svn: 144495
It was off by default.
The new register allocators don't have the problems that made it
necessary to reallocate registers during stack slot coloring.
llvm-svn: 144481
instruction lower optimization" in the pre-RA scheduler.
The optimization, rather the hack, was done before MI use-list was available.
Now we should be able to implement it in a better way, perhaps in the
two-address pass until a MI scheduler is available.
Now that the scheduler has to backtrack to handle call sequences. Adding
artificial scheduling constraints is just not safe. Furthermore, the hack
is not taking all the other scheduling decisions into consideration so it's just
as likely to pessimize code. So I view disabling this optimization goodness
regardless of PR11314.
llvm-svn: 144267
The TII.foldMemoryOperand hook preserves implicit operands from the
original instruction. This is not what we want when those implicit
operands refer to the register being spilled.
Implicit operands referring to other registers are preserved.
This fixes PR11347.
llvm-svn: 144247
dragonegg self-host buildbot will recover (it is complaining about object
files differing between different build stages). Original commit message:
Add a hack to the scheduler to disable pseudo-two-address dependencies in
basic blocks containing calls. This works around a problem in which
these artificial dependencies can get tied up in calling seqeunce
scheduling in a way that makes the graph unschedulable with the current
approach of using artificial physical register dependencies for calling
sequences. This fixes PR11314.
llvm-svn: 144188
During the initial RPO traversal of the basic blocks, remember the ones
that are incomplete because of back-edges from predecessors that haven't
been visited yet.
After the initial RPO, revisit all those loop headers so the incoming
DomainValues on the back-edges can be properly collapsed.
This will properly fix execution domains on software pipelined code,
like the included test case.
llvm-svn: 144151
basic blocks containing calls. This works around a problem in which
these artificial dependencies can get tied up in calling seqeunce
scheduling in a way that makes the graph unschedulable with the current
approach of using artificial physical register dependencies for calling
sequences. This fixes PR11314.
llvm-svn: 144124
DomainValues that are only used by "don't care" instructions are now
collapsed to the first possible execution domain after all basic blocks
have been processed. This typically means the PS domain on x86.
For example, the vsel_i64 and vsel_double functions in sse2-blend.ll are
completely collapsed to the PS domain instead of containing a mix of
execution domains created by isel.
llvm-svn: 144037
The xorps instruction is smaller than pxor, so prefer that encoding.
The ExecutionDepsFix pass will switch the encoding to pxor and xorpd
when appropriate.
llvm-svn: 143996
On x86: (shl V, 1) -> add V,V
Hardware support for vector-shift is sparse and in many cases we scalarize the
result. Additionally, on sandybridge padd is faster than shl.
llvm-svn: 143311
fixes: Use a separate register, instead of SP, as the
calling-convention resource, to avoid spurious conflicts with
actual uses of SP. Also, fix unscheduling of calling sequences,
which can be triggered by pseudo-two-address dependencies.
llvm-svn: 143206
Don't assume APInt::getRawData() would hold target-aware endianness nor host-compliant endianness. rawdata[0] holds most lower i64, even on big endian host.
FIXME: Add a testcase for big endian target.
FIXME: Ditto on CompileUnit::addConstantFPValue() ?
llvm-svn: 143194
it fixes the dragonegg self-host (it looks like gcc is miscompiled).
Original commit messages:
Eliminate LegalizeOps' LegalizedNodes map and have it just call RAUW
on every node as it legalizes them. This makes it easier to use
hasOneUse() heuristics, since unneeded nodes can be removed from the
DAG earlier.
Make LegalizeOps visit the DAG in an operands-last order. It previously
used operands-first, because LegalizeTypes has to go operands-first, and
LegalizeTypes used to be part of LegalizeOps, but they're now split.
The operands-last order is more natural for several legalization tasks.
For example, it allows lowering code for nodes with floating-point or
vector constants to see those constants directly instead of seeing the
lowered form (often constant-pool loads). This makes some things
somewhat more complicated today, though it ought to allow things to be
simpler in the future. It also fixes some bugs exposed by Legalizing
using RAUW aggressively.
Remove the part of LegalizeOps that attempted to patch up invalid chain
operands on libcalls generated by LegalizeTypes, since it doesn't work
with the new LegalizeOps traversal order. Instead, define what
LegalizeTypes is doing to be correct, and transfer the responsibility
of keeping calls from having overlapping calling sequences into the
scheduler.
Teach the scheduler to model callseq_begin/end pairs as having a
physical register definition/use to prevent calls from having
overlapping calling sequences. This is also somewhat complicated, though
there are ways it might be simplified in the future.
This addresses rdar://9816668, rdar://10043614, rdar://8434668, and others.
Please direct high-level questions about this patch to management.
Delete #if 0 code accidentally left in.
llvm-svn: 143188
on every node as it legalizes them. This makes it easier to use
hasOneUse() heuristics, since unneeded nodes can be removed from the
DAG earlier.
Make LegalizeOps visit the DAG in an operands-last order. It previously
used operands-first, because LegalizeTypes has to go operands-first, and
LegalizeTypes used to be part of LegalizeOps, but they're now split.
The operands-last order is more natural for several legalization tasks.
For example, it allows lowering code for nodes with floating-point or
vector constants to see those constants directly instead of seeing the
lowered form (often constant-pool loads). This makes some things
somewhat more complicated today, though it ought to allow things to be
simpler in the future. It also fixes some bugs exposed by Legalizing
using RAUW aggressively.
Remove the part of LegalizeOps that attempted to patch up invalid chain
operands on libcalls generated by LegalizeTypes, since it doesn't work
with the new LegalizeOps traversal order. Instead, define what
LegalizeTypes is doing to be correct, and transfer the responsibility
of keeping calls from having overlapping calling sequences into the
scheduler.
Teach the scheduler to model callseq_begin/end pairs as having a
physical register definition/use to prevent calls from having
overlapping calling sequences. This is also somewhat complicated, though
there are ways it might be simplified in the future.
This addresses rdar://9816668, rdar://10043614, rdar://8434668, and others.
Please direct high-level questions about this patch to management.
llvm-svn: 143177
MORESTACK_RET_RESTORE_R10; which are lowered to a RET and a RET
followed by a MOV respectively. Having a fake instruction prevents
the verifier from seeing a MachineBasicBlock end with a
non-terminator (MOV). It also prevents the rather eccentric case of a
MachineBasicBlock ending with RET but having successors nevertheless.
Patch by Sanjoy Das.
llvm-svn: 143062
discussions with Andy. Fundamentally, the previous algorithm is both
counter productive on several fronts and prioritizing things which
aren't necessarily the most important: static branch prediction.
The new algorithm uses the existing loop CFG structure information to
walk through the CFG itself to layout blocks. It coalesces adjacent
blocks within the loop where the CFG allows based on the most likely
path taken. Finally, it topologically orders the block chains that have
been formed. This allows it to choose a (mostly) topologically valid
ordering which still priorizes fallthrough within the structural
constraints.
As a final twist in the algorithm, it does violate the CFG when it
discovers a "hot" edge, that is an edge that is more than 4x hotter than
the competing edges in the CFG. These are forcibly merged into
a fallthrough chain.
Future transformations that need te be added are rotation of loop exit
conditions to be fallthrough, and better isolation of cold block chains.
I'm also planning on adding statistics to model how well the algorithm
does at laying out blocks based on the probabilities it receives.
The old tests mostly still pass, and I have some new tests to add, but
the nested loops are still behaving very strangely. This almost seems
like working-as-intended as it rotated the exit branch to be
fallthrough, but I'm not convinced this is actually the best layout. It
is well supported by the probabilities for loops we currently get, but
those are pretty broken for nested loops, so this may change later.
llvm-svn: 142743
ZExtPromotedInteger and SExtPromotedInteger based on the operation we legalize.
SetCC return type needs to be legalized via PromoteTargetBoolean.
llvm-svn: 142660
it's a bit more plausible to use this instead of CodePlacementOpt. The
code for this was shamelessly stolen from CodePlacementOpt, and then
trimmed down a bit. There doesn't seem to be much utility in returning
true/false from this pass as we may or may not have rewritten all of the
blocks. Also, the statistic of counting how many loops were aligned
doesn't seem terribly important so I removed it. If folks would like it
to be included, I'm happy to add it back.
This was probably the most egregious of the missing features, and now
I'm going to start gathering some performance numbers and looking at
specific loop structures that have different layout between the two.
Test is updated to include both basic loop alignment and nested loop
alignment.
llvm-svn: 142645
canonical example I used when developing it, and is one of the primary
motivating real-world use cases for __builtin_expect (when burried under
a macro).
I'm working on more test cases here, but I'm trying to make sure both
that the pass is doing the right thing with the test cases and that they
aren't too brittle to changes elsewhere in the code generation pipeline.
Feedback and/or suggestions on how to test this are very welcome.
Especially feedback on whether testing the block comments is a good
strategy; I couldn't find any good examples to steal from but all the
other ideas I had were a lot uglier or more fragile.
llvm-svn: 142644
When checking the availability of instructions using the TLI, a 'promoted'
instruction IS available. It means that the value is bitcasted to another type
for which there is an operation. The correct check for the availablity of an
instruction is to check if it should be expanded.
llvm-svn: 142542
.file filenumber "directory" "filename"
This removes one join+split of the directory+filename in MC internals. Because
bitcode files have independent fields for directory and filenames in debug info,
this patch may change the .o files written by existing .bc files.
llvm-svn: 142300
When spilling around an instruction with a dead def, remember to add a
value number for the def.
The missing value number wouldn't normally create problems since there
would be an incoming live range as well. However, due to another bug
we could spill a dead V_SET0 instruction which doesn't read any values.
The missing value number caused an empty live range to be created which
is dangerous since it doesn't interfere with anything.
This fixes part of PR11125.
llvm-svn: 141923
http://lab.llvm.org:8011/builders/llvm-x86_64-linux/builds/101
--- Reverse-merging r141854 into '.':
U test/MC/Disassembler/X86/x86-32.txt
U test/MC/Disassembler/X86/simple-tests.txt
D test/CodeGen/X86/bmi.ll
U lib/Target/X86/X86InstrInfo.td
U lib/Target/X86/X86ISelLowering.cpp
U lib/Target/X86/X86.td
U lib/Target/X86/X86Subtarget.h
llvm-svn: 141857
Now that MI->getRegClassConstraint() can also handle inline assembly,
don't bail when recomputing the register class of a virtual register
used by inline asm.
This fixes PR11078.
llvm-svn: 141836
our current machine instruction defines a register with the same register class
as what's being replaced. This showed up in the SPEC 403.gcc benchmark, where it
would ICE because a tail call was expecting one register class but was given
another. (The machine instruction verifier catches this situation.)
<rdar://problem/10270968>
llvm-svn: 141830
file. Since it should only be used when necessary propagate it through
the backend code generation and tweak testcases accordingly.
This helps with code like in clang's test/CodeGen/debug-info-line.c where
we have multiple #line directives within a single lexical block and want
to generate only a single block that contains each file change.
Part of rdar://10246360
llvm-svn: 141729
In 64-bit mode, sub_8bit_hi sub-registers can only be used by NOREX
instructions. The COPY created from the EXTRACT_SUBREG DAG node cannot
target all GR8 registers, only those in GR8_NOREX.
TO enforce this, we ensure that all instructions using the
EXTRACT_SUBREG are GR8_NOREX constrained.
This fixes PR11088.
llvm-svn: 141499
and the alignment is 0 (i.e., it's defined globally in one file and declared in
another file) it could get an alignment which is larger than the ABI allows for
that type, resulting in aligned moves being used for unaligned loads.
For instance, in file A.c:
struct S s;
In file B.c:
struct {
// something long
};
extern S s;
void foo() {
struct S p = s;
// ...
}
this copy is a 'memcpy' which is turned into a series of 'movaps' instructions
on X86. But this is wrong, because 'struct S' has alignment of 4, not 16.
llvm-svn: 140902
Rewriting the entire loop nest now requires -enable-lsr-nested.
See PR11035 for some performance data.
A few unit tests specifically test nested LSR, and are now under a flag.
llvm-svn: 140762
This also enables domain swizzling for AVX code which required a few
trivial test changes.
The pass will be moved to lib/CodeGen shortly.
llvm-svn: 140659
Few weeks ago, llvm completely inverted the debug info graph. Earlier each debug info node used to keep track of its compile unit, now compile unit keeps track of important nodes. One impact of this change is that the global variable's do not have any context, which should be checked before deciding to use AT_specification DIE.
llvm-svn: 140282
On Windows x64, 128-bit arguments are not passed by reg but by indirect. eg.
maxpd:
vmovapd (%rcx), %xmm0
vmaxpd (%rdx), %xmm0, %xmm0
FIXME: I don't care YMM on x64 for now.
llvm-svn: 140143
dag-combine optimization to implement the ext-load efficiently (using shuffles).
For example the type <4 x i8> is stored in memory as i32, but it needs to
find its way into a <4 x i32> register. Previously we scalarized the memory
access, now we use shuffles.
llvm-svn: 139995
maxps and maxpd). This broke the sse41-blend.ll testcase by causing
maxpd to be produced rather than a cmp+blend pair, which is the reason
I tweaked it. Gives a small speedup on doduc with dragonegg when the
GCC vectorizer is used.
llvm-svn: 139986
take into consideration the presence of AVX. This change, together with
the SSEDomainFix enabled for AVX, makes AVX codegen to always (hopefully)
emit the same code as SSE for 128-bit vector ops. I don't
have a testcase for this, but AVX now beats SSE in performance for
128-bit ops in the majority of programas in the llvm testsuite
llvm-svn: 139817
However with this fix it does now.
Basically the operand order for the x86 target specific node
is not the same as the instruction, but since the intrinsic need that
specific order at the instruction definition, just change the order
during legalization. Also, there were some wrong invertions of condition
codes, such as GE => LE, GT => LT, fix that too. Fix PR10907.
llvm-svn: 139528
assert("not implemented for target shuffle node");
to:
assert(0 && "not implemented for target shuffle node");
This causes a test failure in CodeGen/X86/palignr.ll which has
been marked as XFAIL for the time being.
Test failure filed at PR10901.
llvm-svn: 139454
in Nadav's r139285 and r139287 commits.
1) Rename vsel.ll to a more descriptive name
2) Change the order of BLEND operands to "Op1, Op2, Cond", this is
necessary because PBLENDVB is already used in different places with
this order, and it was being emitted in the wrong way for vselect
3) Add AVX patterns and tests for the same SSE41 instructions
llvm-svn: 139305
The explanation about a 0 argument being materialized as xor is no
longer valid. Rematerialization will check if EFLAGS is live before
clobbering it.
The code produced by X86TargetLowering::EmitLoweredSelect does not
clobber EFLAGS.
This causes one less testb instruction to be generated in the cmov.ll
test case.
llvm-svn: 139057
Added canClobberReachingPhysRegUse() to handle a particular pattern in
which a two-address instruction could be forced to interfere with
EFLAGS, causing a compare to be unnecessarilly cloned.
Fixes rdar://problem/5875261
llvm-svn: 138924
- Duplicate some store patterns to their AVX forms!
- Catched a bug while restricting the patterns subtarget, fix it
and update a testcase to check it properly
llvm-svn: 138851
SSE transition penalty. The pass is enabled through the "x86-use-vzeroupper"
llc command line option. This is only the first step (very naive and
conservative one) to sketch out the idea, but proper DFA is coming next
to allow smarter decisions. Comments and ideas now and in further commits
will be very appreciated.
llvm-svn: 138317