This patch is a follow-up to r338702.
We don't need to use a map to model the wait/ready/issued sets. It is much more
efficient to use a vector instead.
This patch gives us an average 7.5% speedup (on top of the ~12% speedup obtained
after r338702).
llvm-svn: 338883
We don't need to use a map to store ResourceState objects. The number of
processor resources is known statically from the scheduling model. We can
therefore use a vector, and reserve a slot for each processor resource that we
want to simulate.
Every time the ResourceManager queries the ResourceState vector, the index to
the vector of ResourceState objects can be easily computed from the processor
resource mask.
This drastically reduces the time complexity of method ResourceManager::use() and
method ResourceManager::release(). This patch gives an average speedup of 12%.
llvm-svn: 338702
A detailed description of the tool has been recently added by Matt to
CommandGuide/llvm-mca.rst. File README.txt is now redundant and can be removed;
all the relevant user-guide information has been improved and then moved to
llvm-mca.rst.
In future, we should add another .rst for the "llvm-mca developer manual" to
provide infromation about:
- llvm-mca internals.
- How to add custom stages to the simulated pipeline.
- How to provide extra processor info in the scheduling model to improve the
analysis performed by llvm-mca.
llvm-svn: 338386
This patch teaches llvm-mca how to identify dependency breaking instructions on
btver2.
An example of dependency breaking instructions is the zero-idiom XOR (example:
`XOR %eax, %eax`), which always generates zero regardless of the actual value of
the input register operands.
Dependency breaking instructions don't have to wait on their input register
operands before executing. This is because the computation is not dependent on
the inputs.
Not all dependency breaking idioms are also zero-latency instructions. For
example, `CMPEQ %xmm1, %xmm1` is independent on
the value of XMM1, and it generates a vector of all-ones.
That instruction is not eliminated at register renaming stage, and its opcode is
issued to a pipeline for execution. So, the latency is not zero.
This patch adds a new method named isDependencyBreaking() to the MCInstrAnalysis
interface. That method takes as input an instruction (i.e. MCInst) and a
MCSubtargetInfo.
The default implementation of isDependencyBreaking() conservatively returns
false for all instructions. Targets may override the default behavior for
specific CPUs, and return a value which better matches the subtarget behavior.
In future, we should teach to Tablegen how to automatically generate the body of
isDependencyBreaking from scheduling predicate definitions. This would allow us
to expose the knowledge about dependency breaking instructions to the machine
schedulers (and, potentially, other codegen passes).
Differential Revision: https://reviews.llvm.org/D49310
llvm-svn: 338372
registers.
The goal of this patch is to improve the throughput analysis in llvm-mca for the
case where instructions perform partial register writes.
On x86, partial register writes are quite difficult to model, mainly because
different processors tend to implement different register merging schemes in
hardware.
When the code contains partial register writes, the IPC (instructions per
cycles) estimated by llvm-mca tends to diverge quite significantly from the
observed IPC (using perf).
Modern AMD processors (at least, from Bulldozer onwards) don't rename partial
registers. Quoting Agner Fog's microarchitecture.pdf:
" The processor always keeps the different parts of an integer register together.
For example, AL and AH are not treated as independent by the out-of-order
execution mechanism. An instruction that writes to part of a register will
therefore have a false dependence on any previous write to the same register or
any part of it."
This patch is a first important step towards improving the analysis of partial
register updates. It changes the semantic of RegisterFile descriptors in
tablegen, and teaches llvm-mca how to identify false dependences in the presence
of partial register writes (for more details: see the new code comments in
include/Target/TargetSchedule.h - class RegisterFile).
This patch doesn't address the case where a write to a part of a register is
followed by a read from the whole register. On Intel chips, high8 registers
(AH/BH/CH/DH)) can be stored in separate physical registers. However, a later
(dirty) read of the full register (example: AX/EAX) triggers a merge uOp, which
adds extra latency (and potentially affects the pipe usage).
This is a very interesting article on the subject with a very informative answer
from Peter Cordes:
https://stackoverflow.com/questions/45660139/how-exactly-do-partial-registers-on-haswell-skylake-perform-writing-al-seems-to
In future, the definition of RegisterFile can be extended with extra information
that may be used to identify delays caused by merge opcodes triggered by a dirty
read of a partial write.
Differential Revision: https://reviews.llvm.org/D49196
llvm-svn: 337123
Summary:
This patch converts the InstructionTables class into a subclass of mca::Stage. This change allows us to use the Stage's inherited Listeners for event notifications. This also allows us to create a simple pipeline for viewing the InstructionTables report.
I have been working on a follow on patch that should cleanup addView in InstructionTables. Right now, addView adds the view to both the Listener list and Views list. The follow-on patch addresses the fact that we don't really need two lists in this case. That change is not specific to just InstructionTables, so it will be a separate patch.
Reviewers: andreadb, courbet, RKSimon
Reviewed By: andreadb
Subscribers: tschuett, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D49329
llvm-svn: 337113
Summary:
This patch clears up some of the semantics within the Stage class. Now, preExecute
can be called multiple times per simulated cycle. Previously preExecute was
only called once per cycle, and postExecute could have been called multiple
times.
Now, cycleStart/cycleEnd are called only once per simulated cycle.
preExecute/postExecute can be called multiple times per cycle. This
occurs because multiple execution events can occur during a single cycle.
When stages are executed (Pipeline::runCycle), the postExecute hook will
be called only if all Stages return a success from their 'execute' callback.
Reviewers: andreadb, courbet, RKSimon
Reviewed By: andreadb
Subscribers: tschuett, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D49250
llvm-svn: 336959
Summary:
This patch eliminates some redundancy in iterating across Listeners for the
Instruction and Stall HWEvents, by introducing a template onEvent routine.
This change was suggested by @courbet in https://reviews.llvm.org/D48576. I
hope that this patch addresses that suggestion appropriately. I do like this
change better than what we had previously.
Reviewers: andreadb, courbet, RKSimon
Reviewed By: andreadb, courbet
Subscribers: javed.absar, tschuett, gbedwell, llvm-commits, courbet
Differential Revision: https://reviews.llvm.org/D48672
llvm-svn: 336916
This makes easier to identify changes in the instruction info flags. It also
helps spotting potential regressions similar to the one recently introduced at
r336728.
Using the same character to mark MayLoad/MayStore/HasSideEffects is problematic
for llvm-lit. When pattern matching substrings, llvm-lit consumes tabs and
spaces. A change in position of the flag marker may not trigger a test failure.
This patch only changes the character used for flag `hasSideEffects`. The reason
why I didn't touch other flags is because I want to avoid spamming the mailing
because of the massive diff due to the numerous tests affected by this change.
In future, each instruction flag should be associated with a different character
in the Instruction Info View.
llvm-svn: 336797
This is a short-term fix for PR38093.
For now, we llvm::report_fatal_error if the instruction builder finds an
unsupported instruction in the instruction stream.
We need to revisit this fix once we start addressing PR38101.
Essentially, we need a better framework for error handling.
llvm-svn: 336543
This patch moves the construction of the default backend from llvm-mca.cpp and
into mca::Context. The Context class is responsible for holding ownership of
the simulated hardware components. These components are subclasses of
HardwareUnit. Right now the HardwareUnit is pretty bare-bones, but eventually
we might want to add some common functionality across all hardware components,
such as isReady() or something similar.
I have a feeling this patch will probably need some updates, but it's a start.
One thing I am not particularly fond of is the rather large interface for
createDefaultPipeline. That convenience routine takes a rather large set of
inputs from the llvm-mca driver, where many of those inputs are generated via
command line options.
One item I think we might want to change is the separating of ownership of
hardware components (owned by the context) and the pipeline (which owns
Stages). In short, a Pipeline owns Stages, a Context (currently) owns hardware.
The Pipeline's Stages make use of the components, and thus there is a lifetime
dependency generated. The components must outlive the pipeline. We could solve
this by having the Context also own the Pipeline, and not return a
unique_ptr<Pipeline>. Now that I think about it, I like that idea more.
Differential Revision: https://reviews.llvm.org/D48691
llvm-svn: 336456
This patch modifies the Scheduler heuristic used to select the next instruction
to issue to the pipelines.
The motivating example is test X86/BtVer2/add-sequence.s, for which llvm-mca
wrongly reported an estimated IPC of 1.50. According to perf, the actual IPC for
that test should have been ~2.00.
It turns out that an IPC of 2.00 for test add-sequence.s cannot possibly be
predicted by a Scheduler that only prioritizes instructions based on their
"age". A similar issue also affected test X86/BtVer2/dependent-pmuld-paddd.s,
for which llvm-mca wrongly estimated an IPC of 0.84 instead of an IPC of 1.00.
Instructions in the ReadyQueue are now ranked based on two factors:
- The "age" of an instruction.
- The number of unique users of writes associated with an instruction.
The new logic still prioritizes older instructions over younger instructions to
minimize the pressure on the reorder buffer. However, the number of users of an
instruction now also affects the overall rank. This potentially increases the
ability of the Scheduler to extract instruction level parallelism. This patch
fixes the problem with the wrong IPC reported for test add-sequence.s and test
dependent-pmuld-paddd.s.
llvm-svn: 336420
Different CodeBlocks don't overlap. The same MCInst cannot appear in more than
one code block because all blocks are instantiated before the simulation is run.
We should always clear the content of map VariantDescriptors before every
simulation, since VariantDescriptors cannot possibly store useful information
for the next blocks. It is also "safer" to clear its content because `MCInst*`
is used as the key type for map VariantDescriptors.
llvm-svn: 336142
On darwin, all virtual sections have zerofill type, and having a
.zerofill directive in a non-virtual section is not allowed. Instead of
asserting, show a nicer error.
In order to use the equivalent of .zerofill in a non-virtual section,
the usage of .zero of .space is required.
This patch replaces the assert with an error.
Differential Revision: https://reviews.llvm.org/D48517
llvm-svn: 336127
This simplifies the logic that updates RAW dependencies in the DispatchStage.
There is no advantage in storing that flag in the ReadDescriptor; we should
simply rely on the call to `STI.getReadAdvanceCycles()` to obtain the
ReadAdvance cycles. If there are no read-advance entries, then method
`getReadAdvanceCycles()` quickly returns 0.
No functional change intended.
llvm-svn: 335977
This patch introduces a new class named WriteRef. A WriteRef is used by the
RegisterFile to keep track of register definitions. Internally it wraps a
WriteState, as well as the source index of the defining instruction.
This patch allows the tool to propagate additional information to support future
analysis on data dependencies.
llvm-svn: 335867
Rather than calling std::find in a loop, just sort the vector and remove
duplicate entries at the end of the function.
Also, move the debug print at the end of the function, and query the
MCRegisterInfo to print register names rather than physreg IDs.
No functional change intended.
llvm-svn: 335837
Summary:
This patch removes a few callbacks from Pipeline. It comes at the cost of
registering Listeners with all Stages. Not all stages need listeners or issue
callbacks, this registration is a bit redundant. However, as we build-out the
API, this redundancy can disappear.
The main purpose here is to move callback code from the Pipeline and into the
stages that actually issue those callbacks. This removes the back-pointer to
the Pipeline that was put into a few Stage subclasses.
Reviewers: andreadb, courbet, RKSimon
Reviewed By: andreadb, courbet
Subscribers: tschuett, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D48576
llvm-svn: 335748
When promoting instructions from the wait queue to the ready queue, we should
check if an instruction has already reached the IS_READY state before
calling method update().
llvm-svn: 335722
Summary:
This change renames the Backend and BackendPrinter to Pipeline and PipelinePrinter respectively.
Variables and comments have also been updated to reflect this change.
The reason for this rename, is to be slightly more correct about what MCA is modeling. MCA models a Pipeline, which implies some logical sequence of stages.
Reviewers: andreadb, courbet, RKSimon
Reviewed By: andreadb, courbet
Subscribers: mgorny, javed.absar, tschuett, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D48496
llvm-svn: 335496
The DispatchUnit is no longer a dependency of RCU, so this patch removes a
stale include and forward decl. This patch also cleans up some comments.
llvm-svn: 335392
Summary:
Remove explicit stages and introduce a list of stages.
A pipeline should be composed of an arbitrary list of stages, and not any
predefined list of stages in the Backend. The Backend should not know of any
particular stage, rather it should only be concerned that it has a list of
stages, and that those stages will fulfill the contract of what it means to be
a Stage (namely pre/post/execute a given instruction).
For now, we leave the original set of stages defined in the Backend ctor;
however, I imagine these will be moved out at a later time.
This patch makes an adjustment to the semantics of Stage::isReady.
Specifically, what the Backend really needs to know is if a Stage has
unfinished work. With that said, it is more appropriately renamed
Stage::hasWorkToComplete(). This change will clean up the check in
Backend::run(), allowing us to query each stage to see if there is unfinished
work, regardless of what subclass a stage might be. I feel that this change
simplifies the semantics too, but that's a subjective statement.
Given how RetireStage and ExecuteStage handle data in their preExecute(), I've
had to change the order of Retire and Execute in our stage list. Retire must
complete any of its preExecute actions before ExecuteStage's preExecute can
take control. This is mainly because both stages utilize the RCU. In the
meantime, I want to see if I can adjust that or remove that coupling.
Reviewers: andreadb, RKSimon, courbet
Reviewed By: andreadb
Subscribers: tschuett, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D46907
llvm-svn: 335361
This patch teaches llvm-mca how to identify register writes that implicitly zero
the upper portion of a super-register.
On X86-64, a general purpose register is implemented in hardware as a 64-bit
register. Quoting the Intel 64 Software Developer's Manual: "an update to the
lower 32 bits of a 64 bit integer register is architecturally defined to zero
extend the upper 32 bits". Also, a write to an XMM register performed by an AVX
instruction implicitly zeroes the upper 128 bits of the aliasing YMM register.
This patch adds a new method named clearsSuperRegisters to the MCInstrAnalysis
interface to help identify instructions that implicitly clear the upper portion
of a super-register. The rest of the patch teaches llvm-mca how to use that new
method to obtain the information, and update the register dependencies
accordingly.
I compared the kernels from tests clear-super-register-1.s and
clear-super-register-2.s against the output from perf on btver2. Previously
there was a large discrepancy between the estimated IPC and the measured IPC.
Now the differences are mostly in the noise.
Differential Revision: https://reviews.llvm.org/D48225
llvm-svn: 335113
Summary:
While that is indeed a quite interesting summary stat,
there are cases where it does not really add anything
other than consuming extra lines.
Declutters the output of D48190.
Reviewers: RKSimon, andreadb, courbet, craig.topper
Reviewed By: andreadb
Subscribers: javed.absar, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D48209
llvm-svn: 334833
Summary: This patch transforms the Scheduler class into the ExecuteStage. Most of the logic remains.
Reviewers: andreadb, RKSimon, courbet
Reviewed By: andreadb
Subscribers: mgorny, javed.absar, tschuett, gbedwell, llvm-commits
Differential Revision: https://reviews.llvm.org/D47246
llvm-svn: 334679
Fixes PR37790.
In some (very rare) cases, the LSUnit (Load/Store unit) was wrongly marking a
load (or store) as "ready to execute" effectively bypassing older memory barrier
instructions.
To reproduce this bug, the memory barrier must be the first instruction in the
input assembly sequence, and it doesn't have to perform any register writes.
llvm-svn: 334633
Not sure why, but it breaks buildbot clang-cmake-armv8-full.
It causes a failure in TEST 'Xray-armhf-linux :: TestCases/Posix/profiling-single-threaded.cc'.
llvm-svn: 334617
Andrea Di Biagio