This reverts commit r197254.
This was an accidental merge of Juergen's patch. It will be checked in
shortly, but wasn't meant to go in quite yet.
Conflicts:
include/llvm/CodeGen/StackMaps.h
lib/CodeGen/StackMaps.cpp
test/CodeGen/X86/stackmap-liveness.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@197260 91177308-0d34-0410-b5e6-96231b3b80d8
This hook reverses the order of assignment for local live ranges. This
will generally allocate shorter local live ranges first. For targets with
many registers, this could reduce regalloc compile time by a large
factor. It should still achieve optimal coloring; however, it can change
register eviction decisions. It is disabled by default for two reasons:
(1) Top-down allocation is simpler and easier to debug for targets that
don't benefit from reversing the order.
(2) Bottom-up allocation could result in poor evicition decisions on some
targets affecting the performance of compiled code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@197001 91177308-0d34-0410-b5e6-96231b3b80d8
MCJIT needs to be able to run in hostile environments, even when PWD
is invalid. There's no need to crash MCJIT in this case.
The obvious fix is to simply leave MCContext's CompilationDir empty
when PWD can't be determined. This way, MCJIT clients,
and other clients that link with LLVM don’t need a valid working directory.
If we do want to guarantee valid CompilationDir, that should be done
only for clients of getCompilationDir(). This is as simple as checking
for an empty string.
The only current use of getCompilationDir is EmitGenDwarfInfo, which
won’t conceivably run with an invalid working dir. However, in the
purely hypothetically and untestable case that this happens, the
AT_comp_dir will be omitted from the compilation_unit DIE.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@196874 91177308-0d34-0410-b5e6-96231b3b80d8
These helper classes take care of the book-keeping the drives the
GenericScheduler heuristics. It is likely that developers writing
target-specific schedulers that work similarly to GenericScheduler
will want to use these helpers too. The immediate goal is to develop a
GenericPostScheduler that can run in place of the old PostRAScheduler,
but will use the new machine model.
No functionality change intended.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@196643 91177308-0d34-0410-b5e6-96231b3b80d8
This allows a target to use MI-Sched as an in-order scheduler that
will model strict resource conflicts without defining a processor
itinerary. Instead, the target can now use the new per-operand machine
model and define in-order resources with BufferSize=0. For example,
this would allow restricting the type of operations that can be formed
into a dispatch group. (Normally NumMicroOps is sufficient to enforce
dispatch groups).
If the intent is to model latency in in-order pipeline, as opposed to
resource conflicts, then a resource with BufferSize=1 should be
defined instead.
This feature is only casually tested as there are no in-tree targets
using it yet. However, Hal will be experimenting with POWER7.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@196517 91177308-0d34-0410-b5e6-96231b3b80d8
A Direct stack map location records the address of frame index. This
address is itself the value that the runtime requested. This differs
from IndirectMemRefOp locations, which refer to a stack locations from
which the requested values must be loaded. Direct locations can
directly communicate the address if an alloca, while IndirectMemRefOp
handle register spills.
For example:
entry:
%a = alloca i64...
llvm.experimental.stackmap(i32 <ID>, i32 <shadowBytes>, i64* %a)
Since both the alloca and stackmap intrinsic are in the entry block,
and the intrinsic takes the address of the alloca, the runtime can
assume that LLVM will not substitute alloca with any intervening
value. This must be verified by the runtime by checking that the stack
map's location is a Direct location type. The runtime can then
determine the alloca's relative location on the stack immediately after
compilation, or at any time thereafter. This differs from Register and
Indirect locations, because the runtime can only read the values in
those locations when execution reaches the instruction address of the
stack map.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@195712 91177308-0d34-0410-b5e6-96231b3b80d8
Hard-coded operand indices were scattered throughout lowering stages
and layers. It was super bug prone.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@195093 91177308-0d34-0410-b5e6-96231b3b80d8
Implementing this on bigendian platforms could get strange. I added a
target hook, getStackSlotRange, per Jakob's recommendation to make
this as explicit as possible.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194942 91177308-0d34-0410-b5e6-96231b3b80d8
This comes into play with patchpoint, which can fold multiple
operands. Since the patchpoint is already treated as a call, the
machine mem operands won't affect anything, and there's nothing to
test. But we still want to do the right thing here to be sure that our
MIs obey the rules.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194750 91177308-0d34-0410-b5e6-96231b3b80d8
If a null call target is provided, don't emit a dummy call. This
allows the runtime to reserve as little nop space as it needs without
the requirement of emitting a call.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194676 91177308-0d34-0410-b5e6-96231b3b80d8
We already know how to fold a reload from a frameindex without
analyzing the load instruction. Generalize this to handle any
frameindex load. This streamlines the logic for rematerializing loads
from stack arguments. As a side effect, it allows stackmaps to record
a stack argument location without spilling it.
Verified no effect on codegen for llvm test-suite.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194497 91177308-0d34-0410-b5e6-96231b3b80d8
Fixes <rdar://15432754> [JS] Assertion: "Folded a def to a non-store!"
The primary purpose of anyregcc is to prevent a patchpoint's call
arguments and return value from being spilled. They must be available
in a register, although the calling convention does not pin the
register. It's up to the front end to avoid using this convention for
calls with more arguments than allocatable registers.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194428 91177308-0d34-0410-b5e6-96231b3b80d8
Patch by Michele Scandale!
Rewrite of the functions used to compute the backedge taken count of a
loop on LT and GT comparisons.
I decided to split the handling of LT and GT cases becasue the trick
"a > b == -a < -b" in some cases prevents the trip count computation
due to the multiplication by -1 on the two operands of the
comparison. This issue comes from the conservative computation of
value range of SCEVs: taking the negative SCEV of an expression that
have a small positive range (e.g. [0,31]), we would have a SCEV with a
fullset as value range.
Indeed, in the new rewritten function I tried to better handle the
maximum backedge taken count computation when MAX/MIN expression are
used to handle the cases where no entry guard is found.
Some test have been modified in order to check the new value correctly
(I manually check them and reasoning on possible overflow the new
values seem correct).
I finally added a new test case related to the multiplication by -1
issue on GT comparisons.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194116 91177308-0d34-0410-b5e6-96231b3b80d8
MorphNodeTo is not safe to call during DAG building. It eagerly
deletes dependent DAG nodes which invalidates the NodeMap. We could
expose a safe interface for morphing nodes, but I don't think it's
worth it. Just create a new MachineNode and replaceAllUsesWith.
My understaning of the SD design has been that we want to support
early target opcode selection. That isn't very well supported, but
generally works. It seems reasonable to rely on this feature even if
it isn't widely used.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194102 91177308-0d34-0410-b5e6-96231b3b80d8
Partial fix for PR17459: wrong code at -O3 on x86_64-linux-gnu
(affecting trunk and 3.3)
When SCEV expands a recurrence outside of a loop it attempts to scale
by the stride of the recurrence. Chained recurrences don't work that
way. We could compute binomial coefficients, but would hve to
guarantee that the chained AddRec's are in a perfectly reduced form.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193438 91177308-0d34-0410-b5e6-96231b3b80d8
Partial fix for PR17459: wrong code at -O3 on x86_64-linux-gnu
(affecting trunk and 3.3)
ScalarEvolutionNormalization was attempting to normalize by adding and
subtracting strides. Chained recurrences don't work that way.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193437 91177308-0d34-0410-b5e6-96231b3b80d8
SCEV currently fails to compute loop counts for nonunit stride
loops. This comes up frequently. It prevents loop optimization and
forces vectorization to insert extra loop checks.
For example:
void foo(int n, int *x) {
for (int i = 0; i < n; i += 3) {
x[i] = i;
x[i+1] = i+1;
x[i+2] = i+2;
}
}
We need to properly handle the case in which limit > INT_MAX-stride. In
the above case: n > INT_MAX-3. In this case the loop counter will step
beyond the limit and overflow at the same time. However, knowing that
signed integer overlow in undefined, we can assume the loop test
behavior is arbitrary after overflow. This obeys both C undefined
behavior rules, and the more strict LLVM poison value rules.
I'm finally fixing this in response to Hal Finkel's persistence.
The most probable reason that we never optimized this before is that
we were being careful to handle case where the developer expected a
side-effect free infinite loop relying on overflow:
for (int i = 0; i < n; i += s) {
++j;
}
return j;
If INT_MAX+1 is a multiple of s and n > INT_MAX-s, then we might
expect an infinite loop. However there are plenty of ways to achieve
this effect without relying on undefined behavior of signed overflow.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193015 91177308-0d34-0410-b5e6-96231b3b80d8
This changes the SelectionDAG scheduling preference to source
order. Soon, the SelectionDAG scheduler can be bypassed saving
a nice chunk of compile time.
Performance differences that result from this change are often a
consequence of register coalescing. The register coalescer is far from
perfect. Bugs can be filed for deficiencies.
On x86 SandyBridge/Haswell, the source order schedule is often
preserved, particularly for small blocks.
Register pressure is generally improved over the SD scheduler's ILP
mode. However, we are still able to handle large blocks that require
latency hiding, unlike the SD scheduler's BURR mode. MI scheduler also
attempts to discover the critical path in single-block loops and
adjust heuristics accordingly.
The MI scheduler relies on the new machine model. This is currently
unimplemented for AVX, so we may not be generating the best code yet.
Unit tests are updated so they don't depend on SD scheduling heuristics.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@192750 91177308-0d34-0410-b5e6-96231b3b80d8
rdar:15221834 False AVX register dependencies cause 5x slowdown on
flops-5/6 and significant slowdown on several others.
This was blocking the switch to MI-Sched.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@192669 91177308-0d34-0410-b5e6-96231b3b80d8
This pass is needed to break false dependencies. Without it, unlucky
register assignment can result in wild (5x) swings in
performance. This pass was trying to handle AVX but not getting it
right. AVX doesn't have partial register defs, it has unused register
reads in which the high bits of a source operand are copied into the
unused bits of the dest.
Fixing this requires conservative liveness analysis. This is awkard
because the pass already has its own pseudo-liveness. However, proper
liveness is expensive, and we would like to use a generic utility to
compute it. The fix only invokes liveness on-demand. It is rare to
detect a case that needs undef-read dependence breaking, but when it
happens, it can be needed many times within a very large block.
I think the existing heuristic which uses a register window of 16 is
too conservative for loop-carried false dependencies. If the loop is a
reduction. The out-of-order engine may be able to execute several loop
iterations in parallel. However, I'll leave this tuning exercise for
next time.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@192635 91177308-0d34-0410-b5e6-96231b3b80d8
Clobbering is exclusive not inclusive on register units.
For liveness, we need to consider all the preserved registers.
e.g. A regmask that clobbers YMM0 may preserve XMM0.
Units are only clobbered when all super-registers are clobbered.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@192623 91177308-0d34-0410-b5e6-96231b3b80d8
Some clients may add block live ins and may track liveness over a
large scope. This guarantees an efficient implementation in all cases
with no memory allocation/deallocation, independent of the number of
target registers. It could be slightly less convenient but is fine in
the expected case.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@192622 91177308-0d34-0410-b5e6-96231b3b80d8
This was only working because AVX had cheaper rules in all cases.
I'm sure there are other places in this file where predicates are missing.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@192276 91177308-0d34-0410-b5e6-96231b3b80d8
Ideally, the machinel model is added at the time the instructions are
defined. But many instructions in X86InstrSSE.td still need a model.
Without this workaround the scheduler asserts because x86 already has
itinerary classes for these instructions, indicating they should be
modeled by the scheduler. Since we use the new machine model for other
instructions, it expects a new machine model for these too.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191391 91177308-0d34-0410-b5e6-96231b3b80d8
This is being disabled because it is no longer needed for
performance. It is only used by postRAscheduler which is also planned
for removal, and it is implemented with an out-dated view of register
liveness. It consideres aliases instead of register units, assumes
valid kill flags, and assumes implicit uses on partial register
defs. Kill flags and implicit operands are error prone and impossible
to verify. We should gradually eliminate dependence on them in the
postRA phases.
Targets that still benefit from this should move to the MI
scheduler. If that doesn't solve the problem, then we should add a
hook to regalloc to optimize reload placement.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191348 91177308-0d34-0410-b5e6-96231b3b80d8
The global registry is used to allow command line override of the
scheduler selection, but does not work well as the normal selection
API. For example, the same LLVM process should be able to target
multiple targets or subtargets.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191071 91177308-0d34-0410-b5e6-96231b3b80d8
This was an experimental scheduler a year ago. It's now used by
several subtargets, both in-order and out-of-order, and it
is about to be enabled by default for x86 and armv7. It will be the
new GenericScheduler for subtargets that don't provide their own
SchedulingStrategy.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191051 91177308-0d34-0410-b5e6-96231b3b80d8
This puts all the global PassManager debugging flags, like
-print-after-all and -time-passes, behind a managed static. This
eliminates their static initializers and, more importantly, exit-time
destructors.
The only behavioral change I anticipate is that tools need to
initialize the PassManager before parsing the command line in order to
export these options, which makes sense. Tools that already initialize
the standard passes (opt/llc) don't need to do anything new.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@190974 91177308-0d34-0410-b5e6-96231b3b80d8
Arnold's idea.
I generally try to avoid stateful heuristics because it can make
debugging harder. However, we need a way to prevent the latency
priority from dominating, and it somewhat makes sense to schedule
aggressively for latency only within an issue group.
Swift in particular likes this, and it doesn't hurt anyone else:
| Benchmarks/MiBench/consumer-lame | 10.39% |
| Benchmarks/Misc/himenobmtxpa | 9.63% |
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@190360 91177308-0d34-0410-b5e6-96231b3b80d8
Allow subtargets to customize the generic scheduling strategy.
This is convenient for targets that don't need to add new heuristics
by specializing the strategy.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@190176 91177308-0d34-0410-b5e6-96231b3b80d8
Fast register pressure tracking currently only takes effect during
bottom up scheduling. Forcing this is a bit faster and simpler for
targets that don't have many scheduling constraints and don't need
top-down scheduling.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@190014 91177308-0d34-0410-b5e6-96231b3b80d8
If the instruction window is < NumRegs/2, pressure tracking is not
likely to be effective. The scheduler has to process a very large
number of tiny blocks. We want this to be fast.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189991 91177308-0d34-0410-b5e6-96231b3b80d8
Register pressure tracking is half the complexity of the
scheduler. It's useful to be able to turn it off for compile time and
performance comparisons.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189987 91177308-0d34-0410-b5e6-96231b3b80d8
There was one case that we could hit a DebugValue where I didn't think
to check. DebugValues are evil. No checkinable test case, sorry. It's
an obvious fix.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189717 91177308-0d34-0410-b5e6-96231b3b80d8
This removes all expensive pressure tracking logic from the scheduling
critical path of node comparison.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189643 91177308-0d34-0410-b5e6-96231b3b80d8
Andrew Trick