Imagine we see:
EFLAGS = inst1
EFLAGS = inst2 FLAGS
gpr = inst3 EFLAGS
Previously, we would refuse to schedule inst2 because it clobbers
the EFLAGS of the predecessor. However, it also uses the EFLAGS
of the predecessor, so it is safe to emit. SDep edges ensure that
the right order happens already anyway.
This fixes 2 testsuite crashes with the X86 patch I'm going to
commit next.
llvm-svn: 122211
1. Fix pre-ra scheduler so it doesn't try to push instructions above calls to
"optimize for latency". Call instructions don't have the right latency and
this is more likely to use introduce spills.
2. Fix if-converter cost function. For ARM, it should use instruction latencies,
not # of micro-ops since multi-latency instructions is completely executed
even when the predicate is false. Also, some instruction will be "slower"
when they are predicated due to the register def becoming implicit input.
rdar://8598427
llvm-svn: 118135
operand and one of them has a single use that is a live out copy, favor the
one that is live out. Otherwise it will be difficult to eliminate the copy
if the instruction is a loop induction variable update. e.g.
BB:
sub r1, r3, #1
str r0, [r2, r3]
mov r3, r1
cmp
bne BB
=>
BB:
str r0, [r2, r3]
sub r3, r3, #1
cmp
bne BB
This fixed the recent 256.bzip2 regression.
llvm-svn: 117675
appropriate for targets without detailed instruction iterineries.
The scheduler schedules for increased instruction level parallelism in
low register pressure situation; it schedules to reduce register pressure
when the register pressure becomes high.
On x86_64, this is a win for all tests in CFP2000. It also sped up 256.bzip2
by 16%.
llvm-svn: 109300
it's too late to start backing off aggressive latency scheduling when most
of the registers are in use so the threshold should be a bit tighter.
- Correctly handle live out's and extract_subreg etc.
- Enable register pressure aware scheduling by default for hybrid scheduler.
For ARM, this is almost always a win on # of instructions. It's runtime
neutral for most of the tests. But for some kernels with high register
pressure it can be a huge win. e.g. 464.h264ref reduced number of spills by
54 and sped up by 20%.
llvm-svn: 109279
of getPhysicalRegisterRegClass with it.
If we want to make a copy (or estimate its cost), it is better to use the
smallest class as more efficient operations might be possible.
llvm-svn: 107140
implementing pop with a linear search for a "best" element. The priority
queue was a neat idea, but in practice the comparison functions depend
on dynamic information.
llvm-svn: 104718
pipeline stall. It's useful for targets like ARM cortex-a8. NEON has a lot
of long latency instructions so a strict register pressure reduction
scheduler does not work well.
Early experiments show this speeds up some NEON loops by over 30%.
llvm-svn: 104216
1. Introduce some enums and accessors in the InlineAsm class
that eliminate a ton of magic numbers when handling inline
asm SDNode.
2. Add a new MDNodeSDNode selection dag node type that holds
a MDNode (shocking!)
3. Add a new argument to ISD::INLINEASM nodes that hold !srcloc
metadata, propagating it to the instruction emitter, which
drops it.
No functionality change.
llvm-svn: 100605
into TargetOpcodes.h. #include the new TargetOpcodes.h
into MachineInstr. Add new inline accessors (like isPHI())
to MachineInstr, and start using them throughout the
codebase.
llvm-svn: 95687
the '-pre-RA-sched' flag. It actually makes more sense to do it this way. Also,
keep track of the SDNode ordering by default. Eventually, we would like to make
this ordering a way to break a "tie" in the scheduler. However, doing that now
breaks the "CodeGen/X86/abi-isel.ll" test for 32-bit Linux.
llvm-svn: 94308
basic blocks that are so long that their size overflows a short.
Also assert that overflow does not happen in the future, as requested by Evan.
This fixes PR4401.
llvm-svn: 83159
This adds location info for all llvm_unreachable calls (which is a macro now) in
!NDEBUG builds.
In NDEBUG builds location info and the message is off (it only prints
"UREACHABLE executed").
llvm-svn: 75640
Make llvm_unreachable take an optional string, thus moving the cerr<< out of
line.
LLVM_UNREACHABLE is now a simple wrapper that makes the message go away for
NDEBUG builds.
llvm-svn: 75379
Massive check in. This changes the "-fast" flag to "-O#" in llc. If you want to
use the old behavior, the flag is -O0. This change allows for finer-grained
control over which optimizations are run at different -O levels.
Most of this work was pretty mechanical. The majority of the fixes came from
verifying that a "fast" variable wasn't used anymore. The JIT still uses a
"Fast" flag. I'll change the JIT with a follow-up patch.
llvm-svn: 70343
use the old behavior, the flag is -O0. This change allows for finer-grained
control over which optimizations are run at different -O levels.
Most of this work was pretty mechanical. The majority of the fixes came from
verifying that a "fast" variable wasn't used anymore. The JIT still uses a
"Fast" flag. I'm not 100% sure if it's necessary to change it there...
llvm-svn: 70270
add dependencies on nodes with exactly one successor which is a
COPY_TO_REGCLASS node. In the case that the copy is coalesced
away, the dependence should be on the user of the copy, rather
than the copy itself.
llvm-svn: 69309
help out the register pressure reduction heuristics in the case of
nodes with multiple uses. Currently this uses very conservative
heuristics, so it doesn't have a broad impact, but in cases where it
does help it can make a big difference.
llvm-svn: 67586
a data dependency on the load node, so it really needs a
data-dependence edge to the load node, even if the load previously
existed.
And add a few comments.
llvm-svn: 67554
with multiple chain operands. This can occur when the scheduler
has added chain operands to a node that already has a chain
operand, in order to handle physical register dependencies.
This fixes an llvm-gcc bootstrap failure on x86-64 introduced
in r66058.
llvm-svn: 66240
is determined by whether the node has a Flag operand. However, if the
node does have a Flag operand, it will be glued to its register's
def, so the heuristic would end up spuriously applying to whatever
node is the def.
llvm-svn: 64319
instruction index across each part. Instruction indices are used
to make live range queries, and live ranges can extend beyond
scheduling region boundaries.
Refactor the ScheduleDAGSDNodes class some more so that it
doesn't have to worry about this additional information.
llvm-svn: 64288
scheduling, and generalize is so that preserves state across
scheduling regions. This fixes incorrect live-range information around
terminators and labels, which are effective region boundaries.
In place of looking for terminators to anchor inter-block dependencies,
introduce special entry and exit scheduling units for this purpose.
llvm-svn: 64254
and into the ScheduleDAGInstrs class, so that they don't get
destructed and re-constructed for each block. This fixes a
compile-time hot spot in the post-pass scheduler.
To help facilitate this, tidy and do some minor reorganization
in the scheduler constructor functions.
llvm-svn: 62275
scheduling dependencies. Add assertion checks to help catch
this.
It appears the Mips target defaults to list-td, and it has a
regression test that uses a physreg dependence. Such code was
liable to be miscompiled, and now evokes an assertion failure.
llvm-svn: 62177
AddPseudoTwoAddrDeps. This lets the scheduling infrastructure
avoid recalculating node heights. In very large testcases this
was a major bottleneck. Thanks to Roman Levenstein for finding
this!
As a side effect, fold-pcmpeqd-0.ll is now scheduled better
and it no longer requires spilling on x86-32.
llvm-svn: 61778
instructions to avoid copies, because TwoAddressInstructionPass
also does this optimization. The scheduler's version didn't
account for live-out values, which resulted in spurious commutes
and missed opportunities.
Now, TwoAddressInstructionPass handles all the opportunities,
instead of just those that the scheduler missed. The result is
usually the same, though there are occasional trivial differences
resulting from the avoidance of spurious commutes.
llvm-svn: 61611
computation code. Also, avoid adding output-depenency edges when both
defs are dead, which frequently happens with EFLAGS defs.
Compute Depth and Height lazily, and always in terms of edge latency
values. For the schedulers that don't care about latency, edge latencies
are set to 1.
Eliminate Cycle and CycleBound, and LatencyPriorityQueue's Latencies array.
These are all subsumed by the Depth and Height fields.
llvm-svn: 61073
The Cost field is removed. It was only being used in a very limited way,
to indicate when the scheduler should attempt to protect a live register,
and it isn't really needed to do that. If we ever want the scheduler to
start inserting copies in non-prohibitive situations, we'll have to
rethink some things anyway.
A Latency field is added. Instead of giving each node a single
fixed latency, each edge can have its own latency. This will eventually
be used to model various micro-architecture properties more accurately.
The PointerIntPair class and an internal union are now used, which
reduce the overall size.
llvm-svn: 60806
introduce any new spilling; it just uses unused registers.
Refactor the SUnit topological sort code out of the RRList scheduler and
make use of it to help with the post-pass scheduler.
llvm-svn: 59999
schedulers. This doesn't have much immediate impact because
targets that use these schedulers by default don't yet provide
pipeline information.
This code also didn't have the benefit of register pressure
information. Also, removing it will avoid problems with list-burr
suddenly starting to do latency-oriented scheduling on x86 when we
start providing pipeline data, which would increase spilling.
llvm-svn: 59775
the RR scheduler actually does look at latency values, but it
doesn't use a hazard recognizer so it has no way to know when
a no-op is needed, as opposed to just stalling and incrementing
the cycle count.
llvm-svn: 59759
dedicated "fast" scheduler in -fast mode instead, which is
faster. This speeds up llc -fast by a few percent on some
testcases -- the speedup only happens for code not handled by
fast-isel.
llvm-svn: 59700
is currently off by default, and can be enabled with
-disable-post-RA-scheduler=false.
This doesn't have a significant impact on most code yet because it doesn't
yet do anything to address anti-dependencies and it doesn't attempt to
disambiguate memory references. Also, several popular targets
don't have pipeline descriptions yet.
The majority of the changes here are splitting the SelectionDAG-specific
code out of ScheduleDAG, so that ScheduleDAG can be moved to
libLLVMCodeGen.a. The interface between ScheduleDAG-using code and
the rest of the scheduling code is somewhat rough and will evolve.
llvm-svn: 59676
new CycleBound value. Instead, just update CycleBound on each call.
Also, make ReleasePred and ReleaseSucc methods more consistent accross
the various schedulers.
This also happens to make ScheduleDAGRRList's CycleBound computation
somewhat more interesting, though it still doesn't have any noticeable
effect, because no current targets that use the register-pressure
reduction scheduler provide pipeline models.
llvm-svn: 59475
to carry a SmallVector of flagged nodes, just calculate the flagged nodes
dynamically when they are needed.
The local-liveness change is due to a trivial scheduling change where
the scheduler arbitrary decision differently.
llvm-svn: 59273
before creating the SUnit for the operation that it was unfolded from. This
allows each SUnit to have all of its predecessor SUnits available at the time
it is created. I don't know yet if this will be absolutely required, but it
is a little tidier to do it this way.
llvm-svn: 59083
instead of requiring all "short description" strings to begin with
two spaces. This makes these strings less mysterious, and it fixes
some cases where short description strings mistakenly did not
begin with two spaces.
llvm-svn: 57521
of two, and to not need a scratch std::vector. Also, compute the ordering
immediately in the result array, instead of in another scratch std::vector
that is copied to the result array.
llvm-svn: 55421
replacement of multiple values. This is slightly more efficient
than doing multiple ReplaceAllUsesOfValueWith calls, and theoretically
could be optimized even further. However, an important property of this
new function is that it handles the case where the source value set and
destination value set overlap. This makes it feasible for isel to use
SelectNodeTo in many very common cases, which is advantageous because
SelectNodeTo avoids a temporary node and it doesn't require CSEMap
updates for users of values that don't change position.
Revamp MorphNodeTo, which is what does all the work of SelectNodeTo, to
handle operand lists more efficiently, and to correctly handle a number
of corner cases to which its new wider use exposes it.
This commit also includes a change to the encoding of post-isel opcodes
in SDNodes; now instead of being sandwiched between the target-independent
pre-isel opcodes and the target-dependent pre-isel opcodes, post-isel
opcodes are now represented as negative values. This makes it possible
to test if an opcode is pre-isel or post-isel without having to know
the size of the current target's post-isel instruction set.
These changes speed up llc overall by 3% and reduce memory usage by 10%
on the InstructionCombining.cpp testcase with -fast and -regalloc=local.
llvm-svn: 53728
to DenseMap<SDNode*, SUnit*>, and adjust the way cloned SUnit nodes are
handled so that only the original node needs to be in the map.
This speeds up llc on 447.dealII.llvm.bc by about 2%.
llvm-svn: 52576
and better control the abstraction. Rename the type
to MVT. To update out-of-tree patches, the main
thing to do is to rename MVT::ValueType to MVT, and
rewrite expressions like MVT::getSizeInBits(VT) in
the form VT.getSizeInBits(). Use VT.getSimpleVT()
to extract a MVT::SimpleValueType for use in switch
statements (you will get an assert failure if VT is
an extended value type - these shouldn't exist after
type legalization).
This results in a small speedup of codegen and no
new testsuite failures (x86-64 linux).
llvm-svn: 52044
This removes the existing bottleneck related to the removal of elements from
the middle of the queue.
Also fixes a subtle bug in ScheduleDAGRRList::CapturePred:
It was updating the state of the SUnit before removing it. As a result, the
comparison operators were working incorrectly and this SUnit could not be removed
from the queue properly.
Reviewed by Evan and Dan. Approved by Dan.
llvm-svn: 50412
ScheduleDAG; they don't correspond to any actual instructions so they
don't need to be scheduled.
This fixes a bug where the EntryToken was being scheduled multiple
times in some cases, though it ended up not causing any trouble because
EntryToken doesn't expand into anything. With this fixed the schedulers
reliably schedule the expected number of units, so we can check this
with an assertion.
This requires a tweak to test/CodeGen/X86/loop-hoist.ll because it
ends up getting scheduled differently in a trivial way, though it was
enough to fool the prcontext+grep that the test does.
llvm-svn: 49701
called LimitedSumOfUnscheduledPredsOfSuccs. It terminates the computation
after a given treshold is reached. This new function is always faster, but
brings real wins only on bigger test-cases.
The old function SumOfUnscheduledPredsOfSuccs is left in-place for now and therefore a warning about an unused static function is produced.
llvm-svn: 48872
This fixes Bugzilla #1835 (http://llvm.org/bugs/show_bug.cgi?id=1835).
This patched is reviewed by Tanya and Dan. Dan tested and approved it.
The reason for the bad performance of the old algorithm is that it is very naive and scans every
time all nodes of the DAG in the worst case.
This patch introduces a new algorithm based on the paper "Online algorithms
for maintaining the topological order of a directed acyclic graph" by
David J.Pearce and Paul H.J.Kelly. This is the MNR algorithm. It has a
linear time worst-case and performs much better in most situations.
The paper can be found here:
http://fano.ics.uci.edu/cites/Document/Online-algorithms-for-maintaining-the-topological-order-of-a-directed-acyclic-graph.html
The main idea of the new algorithm is to compute the topological ordering of the SNodes in the
DAG and to maintain it even after DAG modifications. The topological ordering allows for very fast
node reachability checks.
Tests on very big input files with tens of thousands of instructions in a BB indicate huge
speed-ups (up to 10x compilation time improvement) compared to the old version.
llvm-svn: 48817
all clients over to using predicates instead of these flags directly.
These are now private values which are only to be used to statically
initialize the tables.
llvm-svn: 45692
that "machine" classes are used to represent the current state of
the code being compiled. Given this expanded name, we can start
moving other stuff into it. For now, move the UsedPhysRegs and
LiveIn/LoveOuts vectors from MachineFunction into it.
Update all the clients to match.
This also reduces some needless #includes, such as MachineModuleInfo
from MachineFunction.
llvm-svn: 45467
take a deleted nodes vector, instead of requiring it.
One more significant change: Implement the start of a legalizer that
just works on types. This legalizer is designed to run before the
operation legalizer and ensure just that the input dag is transformed
into an output dag whose operand and result types are all legal, even
if the operations on those types are not.
This design/impl has the following advantages:
1. When finished, this will *significantly* reduce the amount of code in
LegalizeDAG.cpp. It will remove all the code related to promotion and
expansion as well as splitting and scalarizing vectors.
2. The new code is very simple, idiomatic, and modular: unlike
LegalizeDAG.cpp, it has no 3000 line long functions. :)
3. The implementation is completely iterative instead of recursive, good
for hacking on large dags without blowing out your stack.
4. The implementation updates nodes in place when possible instead of
deallocating and reallocating the entire graph that points to some
mutated node.
5. The code nicely separates out handling of operations with invalid
results from operations with invalid operands, making some cases
simpler and easier to understand.
6. The new -debug-only=legalize-types option is very very handy :),
allowing you to easily understand what legalize types is doing.
This is not yet done. Until the ifdef added to SelectionDAGISel.cpp is
enabled, this does nothing. However, this code is sufficient to legalize
all of the code in 186.crafty, olden and freebench on an x86 machine. The
biggest issues are:
1. Vectors aren't implemented at all yet
2. SoftFP is a mess, I need to talk to Evan about it.
3. No lowering to libcalls is implemented yet.
4. Various operations are missing etc.
5. There are FIXME's for stuff I hax0r'd out, like softfp.
Hey, at least it is a step in the right direction :). If you'd like to help,
just enable the #ifdef in SelectionDAGISel.cpp and compile code with it. If
this explodes it will tell you what needs to be implemented. Help is
certainly appreciated.
Once this goes in, we can do three things:
1. Add a new pass of dag combine between the "type legalizer" and "operation
legalizer" passes. This will let us catch some long-standing isel issues
that we miss because operation legalization often obfuscates the dag with
target-specific nodes.
2. We can rip out all of the type legalization code from LegalizeDAG.cpp,
making it much smaller and simpler. When that happens we can then
reimplement the core functionality left in it in a much more efficient and
non-recursive way.
3. Once the whole legalizer is non-recursive, we can implement whole-function
selectiondags maybe...
llvm-svn: 42981
(almost) a register copy. However, it always coalesced to the register of the
RHS (the super-register). All uses of the result of a EXTRACT_SUBREG are sub-
register uses which adds subtle complications to load folding, spiller rewrite,
etc.
llvm-svn: 42899
scheduler will try a number of tricks in order to avoid generating the
copies. This may not be possible in case the node produces a chain value
that prevent movement. Try unfolding the load from the node before to allow
it to be moved / cloned.
llvm-svn: 42625
it as a late BURR scheduling tie-breaker.
Intuitively, it's good to push down instructions whose results are liveout so
their long live ranges won't conflict with other values which are needed inside
the BB. Further prioritize liveout instructions by the number of operands which
are calculated within the BB.
llvm-svn: 35109
e.g.
t1 = op t2, c1
t3 = op t4, c2
and the following instructions are both ready.
t2 = op c3
t4 = op c4
Then schedule t2 = op first.
i.e.
t4 = op c4
t2 = op c3
t1 = op t2, c1
t3 = op t4, c2
This creates more short live intervals which work better with the register
allocator.
llvm-svn: 35089
- Fixed bugs in sethi-ullman number computation and priority queue comparison
functions.
- Separate code that handles priority computation special cases from SU number computation.
llvm-svn: 33025
Andrew Trick