uops.info says these should be 15 cycle instructions. Uops.info also shows the 512-bit form uses port 0 and 5 for both register and memory. We had memory using 0 and 1.
Differential Revision: https://reviews.llvm.org/D75549
Based on uops.info these should have 5 cycle latency as they did on Haswell/Broadwell. I have no additional internal information from Intel.
This was also shown as a discrepancy in the spreadsheet that was sent with an early llvm-dev post about llvm-exegesis.
It also matches Agner Fog.
Differential Revision: https://reviews.llvm.org/D74357
The load ports need a cycle for each potentially loaded element just like Haswell and Skylake. Unlike Haswell and Broadwell, the number of uops does not scale with the number of elements. Instead the load uops run for multiple cycles.
I've taken the latency number from the uops.info. The port binding for the non-load uops is taken from the original IACA data I have.
Differential Revision: https://reviews.llvm.org/D74000
Broadwell was missing half the gather instructions. Both models
had some mixups in the resource costs and number of uops.
I've updated here based on what I think the original IACA source
says with some cross checking against the microcode.
I'm not sure about latency as the IACA source I have doesn't have
that information. So I'm using the latency from uops.info.
I plan to update Skylake models as well, but I'll do that in a
separate patch.
Differential Revision: https://reviews.llvm.org/D73844
Summary:
As determined with `llvm-exegesis`.
Some of these look like typos/misunderstandings of the sched model td
spec:
- latency defaults to `1` when not set => Maybe we can avoid
having a default ?
- problems with regexps not being anchored by default (XCHG matching
CMPXHG)
Note that this is not complete, it fixes only the most obvious mistakes,
and only for latency (not uops).
Reviewers: RKSimon, GGanesh
Subscribers: hiraditya, jfb, mstojanovic, hfinkel, craig.topper, andreadb, lebedev.ri, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73172
Based on exhaustive llvm-exegesis measurements.
There may still be some imperfections for LEA16r/LEA32r.
Much like was observed in D68646, i'm also measuring some outliers
with some specific registers.
Summary:
This patch fixes pr23772 [ARM] r226200 can emit illegal thumb2 instruction: "sub sp, r12, #80".
The violation was that SUB and ADD (reg, immediate) instructions can only write to SP if the source register is also SP. So the above instructions was unpredictable.
To enforce that the instruction t2(ADD|SUB)ri does not write to SP we now enforce the destination register to be rGPR (That exclude PC and SP).
Different than the ARM specification, that defines one instruction that can read from SP, and one that can't, here we inserted one that can't write to SP, and other that can only write to SP as to reuse most of the hard-coded size optimizations.
When performing this change, it uncovered that emitting Thumb2 Reg plus Immediate could not emit all variants of ADD SP, SP #imm instructions before so it was refactored to be able to. (see test/CodeGen/Thumb2/mve-stacksplot.mir where we use a subw sp, sp, Imm12 variant )
It also uncovered a disassembly issue of adr.w instructions, that were only written as SUBW instructions (see llvm/test/MC/Disassembler/ARM/thumb2.txt).
Reviewers: eli.friedman, dmgreen, carwil, olista01, efriedma, andreadb
Reviewed By: efriedma
Subscribers: gbedwell, john.brawn, efriedma, ostannard, kristof.beyls, hiraditya, dmgreen, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70680
The patch gives out the details of the znver2 scheduler model.
There are few improvements with respect to execution units, latencies and
throughput when compared with znver1.
The tests that were present for znver1 for llvm-mca tool were replicated.
The latencies, execution units, timeline and throughput information are updated for znver2.
Reviewers: craig.topper, Simon Pilgrim
Differential Revision: https://reviews.llvm.org/D66088
Noticed while fixing the reduction costs for D59710 - the SLM model doesn't account for the poor throughput of v2i64 ops.
Numbers taken from Intel AOM (+ checked against Agner)
This patch introduces the following changes to the btver2 scheduling model:
- The number of micro opcodes for YMM loads and stores is now 2 (it was
incorrectly set to 1 for both aligned and misaligned loads/stores).
- Increased the number of AGU resource cycles for YMM loads and stores
to 2cy (instead of 1cy).
- Removed JFPU01 and JFPX from the list of resources consumed by pure
float/vector loads (no MMX).
I verified with llvm-exegesis that pure XMM/YMM loads are no-pipe. Those
are dispatched to the FPU but not really issues on JFPU01.
Differential Revision: https://reviews.llvm.org/D68871
llvm-svn: 374765
Summary:
As disscused in https://bugs.llvm.org/show_bug.cgi?id=43219,
i believe it may be somewhat useful to show //some// aggregates
over all the sea of statistics provided.
Example:
```
Average Wait times (based on the timeline view):
[0]: Executions
[1]: Average time spent waiting in a scheduler's queue
[2]: Average time spent waiting in a scheduler's queue while ready
[3]: Average time elapsed from WB until retire stage
[0] [1] [2] [3]
0. 3 1.0 1.0 4.7 vmulps %xmm0, %xmm1, %xmm2
1. 3 2.7 0.0 2.3 vhaddps %xmm2, %xmm2, %xmm3
2. 3 6.0 0.0 0.0 vhaddps %xmm3, %xmm3, %xmm4
3 3.2 0.3 2.3 <total>
```
I.e. we average the averages.
Reviewers: andreadb, mattd, RKSimon
Reviewed By: andreadb
Subscribers: gbedwell, arphaman, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68714
llvm-svn: 374361
Before this patch, loads and stores were only tracked by their corresponding
queues in the LSUnit from dispatch until execute stage. In practice we should be
more conservative and assume that memory opcodes leave their queues at
retirement stage.
Basically, loads should leave the load queue only when they have completed and
delivered their data. We conservatively assume that a load is completed when it
is retired. Stores should be tracked by the store queue from dispatch until
retirement. In practice, stores can only leave the store queue if their data can
be written to the data cache.
This is mostly a mechanical change. With this patch, the retire stage notifies
the LSUnit when a memory instruction is retired. That would triggers the release
of LDQ/STQ entries. The only visible change is in memory tests for the bdver2
model. That is because bdver2 is the only model that defines the load/store
queue size.
This patch partially addresses PR39830.
Differential Revision: https://reviews.llvm.org/D68266
llvm-svn: 374034
This adds a -mattr flag to llvm-mca, for cases where the -mcpu option does not
contain all optional features.
Differential Revision: https://reviews.llvm.org/D68190
llvm-svn: 373358
This patch introduces a cut-off threshold for dependency edge frequences with
the goal of simplifying the critical sequence computation. This patch also
removes the cost normalization for loop carried dependencies. We didn't really
need to artificially amplify the cost of loop-carried dependencies since it is
already computed as the integral over time of the delay (in cycle).
In the absence of backend stalls there is no need for computing a critical
sequence. With this patch we early exit from the critical sequence computation
if no bottleneck was reported during the simulation.
llvm-svn: 372337
On BtVer2 conditional SIMD stores are heavily microcoded.
The latency is directly proportional to the number of packed elements extracted
from the input vector. Also, according to micro-benchmarks, most of the
computation seems to be done in the integer unit.
Only a minority of the uOPs is executed by the FPU. The observed behaviour on
the FPU looks similar to this:
- The input MASK value is moved to the Integer Unit
-- [ a VMOVMSK-like uOP-executed on JFPU0].
- In parallel, each element of the input XMM/YMM is extracted and then sent to
the IntegerUnit through JFPU1.
As expected, a (conditional) store is executed for every extracted element.
Interestingly, a (speculative) load is executed for every extracted element too.
It is as-if a "LOAD - BIT_EXTRACT- CMOV" sequence of uOPs is repeated by the
integer unit for every contionally stored element.
VMASKMOVDQU is a special case: the number of speculative loads is always 2
(presumably, one load per quadword). That means, extra shifts and masking is
performed on (one of) the loaded quadwords before each conditional store (that
also explains the big number of non-FP uOPs retired).
This patch replaces the existing writes for conditional SIMD stores (i.e.
WriteFMaskedStore, and WriteFMaskedStoreY) with the following new writes:
WriteFMaskedStore32 [ XMM Packed Single ]
WriteFMaskedStore32Y [ YMM Packed Single ]
WriteFMaskedStore64 [ XMM Packed Double ]
WriteFMaskedStore64Y [ YMM Packed Double ]
Added a wrapper class named X86SchedWriteMaskMove in X86Schedule.td to describe
both RM and MR variants for conditional SIMD moves in a single tablegen
definition.
Instances of that class are then passed in input to multiclass avx_movmask_rm
when constructing MASKMOVPS/PD definitions.
Since this patch introduces new writes, I had to update all the X86 scheduling
models.
Differential Revision: https://reviews.llvm.org/D66801
llvm-svn: 370649
This is a follow up of r369642.
This patch assigns a ReadAfterLd to every implicit register use of instruction
CMPXCHG8B and instruction CMPXCHG16B. Perf micro-benchmarks show that implicit
registers are read after 3cy from the start of execution.
llvm-svn: 369750
Excluding ADC/SBB and the bit-test instructions (BTR/BTS/BTC), the observed
latency of all other RMW integer arithmetic/logic instructions is 6cy and not
5cy.
Example (ADD):
```
addb $0, (%rsp) # Latency: 6cy
addb $7, (%rsp) # Latency: 6cy
addb %sil, (%rsp) # Latency: 6cy
addw $0, (%rsp) # Latency: 6cy
addw $511, (%rsp) # Latency: 6cy
addw %si, (%rsp) # Latency: 6cy
addl $0, (%rsp) # Latency: 6cy
addl $511, (%rsp) # Latency: 6cy
addl %esi, (%rsp) # Latency: 6cy
addq $0, (%rsp) # Latency: 6cy
addq $511, (%rsp) # Latency: 6cy
addq %rsi, (%rsp) # Latency: 6cy
```
The same latency profile applies to SUB/AND/OR/XOR/INC/DEC.
The observed latency of ADC/SBB is 7-8cy. So we need a different write to model
those. Latency of BTS/BTR/BTC is not fixed by this patch (they are much slower
than what the model for btver2 currently reports).
Differential Revision: https://reviews.llvm.org/D66636
llvm-svn: 369748
On Jaguar, XCHG has a latency of 1cy and decodes to 2 macro-opcodes. Maximum
throughput for XCHG is 1 IPC. The byte exchange has worse latency and decodes to
1 extra uOP; maximum observed throughput is 0.5 IPC.
```
xchgb %cl, %dl # Latency: 2cy - uOPs: 3 - 2 ALU
xchgw %cx, %dx # Latency: 1cy - uOPs: 2 - 2 ALU
xchgl %ecx, %edx # Latency: 1cy - uOPs: 2 - 2 ALU
xchgq %rcx, %rdx # Latency: 1cy - uOPs: 2 - 2 ALU
```
The reg-mem forms of XCHG are atomic operations with an observed latency of
16cy. The resource usage is similar to the XCHGrr variants. The biggest
difference is obviously the bus-locking, which prevents the LS to issue other
memory uOPs in parallel until the unlocking store uOP is executed.
```
xchgb %cl, (%rsp) # Latency: 16cy - uOPs: 3 - ECX latency: 11cy
xchgw %cx, (%rsp) # Latency: 16cy - uOPs: 3 - ECX latency: 11cy
xchgl %ecx, (%rsp) # Latency: 16cy - uOPs: 3 - ECX latency: 11cy
xchgq %rcx, (%rsp) # Latency: 16cy - uOPs: 3 - ECX latency: 11cy
```
The exchanged in/out register operand becomes available after 11cy from the
start of execution. Added test xchg.s to verify that we correctly see that
register write committed in 11cy (and not 16cy).
Reg-reg XADD instructions have the same latency/throughput than the byte
exchange (register-register variant).
```
xaddb %cl, %dl # latency: 2cy - uOPs: 3 - 3 ALU
xaddw %cx, %dx # latency: 2cy - uOPs: 3 - 3 ALU
xaddl %ecx, %edx # latency: 2cy - uOPs: 3 - 3 ALU
xaddq %rcx, %rdx # latency: 2cy - uOPs: 3 - 3 ALU
```
The non-atomic RM variants have a latency of 11cy, and decode to 4
macro-opcodes. They still consume 2 ALU pipes, and the exchange in/out register
operand becomes available in 3cy (it matches the 'load-to-use latency').
```
xaddb %cl, (%rsp) # latency: 11cy - uOPs: 4 - 3 ALU
xaddw %cx, (%rsp) # latency: 11cy - uOPs: 4 - 3 ALU
xaddl %ecx, (%rsp) # latency: 11cy - uOPs: 4 - 3 ALU
xaddq %rcx, (%rsp) # latency: 11cy - uOPs: 4 - 3 ALU
```
The atomic XADD variants execute in 16cy. The in/out register operand is
available after 11cy from the start of execution.
```
lock xaddb %cl, (%rsp) # latency: 16cy - uOPs: 4 - 3 ALU -- ECX latency: 11cy
lock xaddw %cx, (%rsp) # latency: 16cy - uOPs: 4 - 3 ALU -- ECX latency: 11cy
lock xaddl %ecx, (%rsp) # latency: 16cy - uOPs: 4 - 3 ALU -- ECX latency: 11cy
lock xaddq %rcx, (%rsp) # latency: 16cy - uOPs: 4 - 3 ALU -- ECX latency: 11cy
```
Added test xadd.s to verify those latencies as well as read-advance values.
Differential Revision: https://reviews.llvm.org/D66535
llvm-svn: 369642
Latency and throughput of LOCK INC/DEC/NEG/NOT is always 19cy.
Number of uOPs is still 1.
Differential Revision: https://reviews.llvm.org/D66469
llvm-svn: 369388
On Jaguar, CMPXCHG has a latency of 11cy, and a maximum throughput of 0.33 IPC.
Throughput is superiorly limited to 0.33 because of the implicit in/out
dependency on register EAX. In the case of repeated non-atomic CMPXCHG with the
same memory location, store-to-load forwarding occurs and values for sequent
loads are quickly forwarded from the store buffer.
Interestingly, the functionality in LLVM that computes the reciprocal throughput
doesn't seem to know about RMW instructions. That functionality only looks at
the "consumed resource cycles" for the throughput computation. It should be
fixed/improved by a future patch. In particular, for RMW instructions, that
logic should also take into account for the write latency of in/out register
operands.
An atomic CMPXCHG has a latency of ~17cy. Throughput is also limited to
~17cy/inst due to cache locking, which prevents other memory uOPs to start
executing before the "lock releasing" store uOP.
CMPXCHG8rr and CMPXCHG8rm are treated specially because they decode to one less
macro opcode. Their latency tend to be the same as the other RR/RM variants. RR
variants are relatively fast 3cy (but still microcoded - 5 macro opcodes).
CMPXCHG8B is 11cy and unfortunately doesn't seem to benefit from store-to-load
forwarding. That means, throughput is clearly limited by the in/out dependency
on GPR registers. The uOP composition is sadly unknown (due to the lack of PMCs
for the Integer pipes). I have reused the same mix of consumed resource from the
other CMPXCHG instructions for CMPXCHG8B too.
LOCK CMPXCHG8B is instead 18cycles.
CMPXCHG16B is 32cycles. Up to 38cycles when the LOCK prefix is specified. Due to
the in/out dependencies, throughput is limited to 1 instruction every 32 (or 38)
cycles dependeing on whether the LOCK prefix is specified or not.
I wouldn't be surprised if the microcode for CMPXCHG16B is similar to 2x
microcode from CMPXCHG8B. So, I have speculatively set the JALU01 consumption to
2x the resource cycles used for CMPXCHG8B.
The two new hasLockPrefix() functions are used by the btver2 scheduling model
check if a MCInst/MachineInst has a LOCK prefix. Calls to hasLockPrefix() have
been encoded in predicates of variant scheduling classes that describe lat/thr
of CMPXCHG.
Differential Revision: https://reviews.llvm.org/D66424
llvm-svn: 369365
In D66424 it has been requested to move all the new tests added by r369278 into
resources-x86_64.s. That is because only the 8b/16 ops should be tested by
resources-cmpxchg.s. This partially reverts r369278.
llvm-svn: 369288
Flag -show-encoding enables the printing of instruction encodings as part of the
the instruction info view.
Example (with flags -mtriple=x86_64-- -mcpu=btver2):
Instruction Info:
[1]: #uOps
[2]: Latency
[3]: RThroughput
[4]: MayLoad
[5]: MayStore
[6]: HasSideEffects (U)
[7]: Encoding Size
[1] [2] [3] [4] [5] [6] [7] Encodings: Instructions:
1 2 1.00 4 c5 f0 59 d0 vmulps %xmm0, %xmm1, %xmm2
1 4 1.00 4 c5 eb 7c da vhaddps %xmm2, %xmm2, %xmm3
1 4 1.00 4 c5 e3 7c e3 vhaddps %xmm3, %xmm3, %xmm4
In this example, column Encoding Size is the size in bytes of the instruction
encoding. Column Encodings reports the actual instruction encodings as byte
sequences in hex (objdump style).
The computation of encodings is done by a utility class named mca::CodeEmitter.
In future, I plan to expose the CodeEmitter to the instruction builder, so that
information about instruction encoding sizes can be used by the simulator. That
would be a first step towards simulating the throughput from the decoders in the
hardware frontend.
Differential Revision: https://reviews.llvm.org/D65948
llvm-svn: 368432
The upper 4 bits of the immediate byte are used to encode a
register. We need to limit the explicit immediate to fit in the
remaining 4 bits.
Fixes PR42899.
llvm-svn: 368123
This patch adds a new llvm-mca flag named -print-imm-hex.
By default, the instruction printer prints immediate operands as decimals. Flag
-print-imm-hex enables the instruction printer to print those operands in hex.
This patch also adds support for MASM binary and hex literal numbers (example
0FFh, 101b).
Added tests to verify the behavior of the new flag. Tests also verify that masm
numeric literal operands are now recognized.
Differential Revision: https://reviews.llvm.org/D65588
llvm-svn: 367671
This patch teaches the bottleneck analysis how to identify and print the most
expensive sequence of instructions according to the simulation. Fixes PR37494.
The goal is to help users identify the sequence of instruction which is most
critical for performance.
A dependency graph is internally used by the bottleneck analysis to describe
data dependencies and processor resource interferences between instructions.
There is one node in the graph for every instruction in the input assembly
sequence. The number of nodes in the graph is independent from the number of
iterations simulated by the tool. It means that a single node of the graph
represents all the possible instances of a same instruction contributed by the
simulated iterations.
Edges are dynamically "discovered" by the bottleneck analysis by observing
instruction state transitions and "backend pressure increase" events generated
by the Execute stage. Information from the events is used to identify critical
dependencies, and materialize edges in the graph. A dependency edge is uniquely
identified by a pair of node identifiers plus an instance of struct
DependencyEdge::Dependency (which provides more details about the actual
dependency kind).
The bottleneck analysis internally ranks dependency edges based on their impact
on the runtime (see field DependencyEdge::Dependency::Cost). To this end, each
edge of the graph has an associated cost. By default, the cost of an edge is a
function of its latency (in cycles). In practice, the cost of an edge is also a
function of the number of cycles where the dependency has been seen as
'contributing to backend pressure increases'. The idea is that the higher the
cost of an edge, the higher is the impact of the dependency on performance. To
put it in another way, the cost of an edge is a measure of criticality for
performance.
Note how a same edge may be found in multiple iteration of the simulated loop.
The logic that adds new edges to the graph checks if an equivalent dependency
already exists (duplicate edges are not allowed). If an equivalent dependency
edge is found, field DependencyEdge::Frequency of that edge is incremented by
one, and the new cost is cumulatively added to the existing edge cost.
At the end of simulation, costs are propagated to nodes through the edges of the
graph. The goal is to identify a critical sequence from a node of the root-set
(composed by node of the graph with no predecessors) to a 'sink node' with no
successors. Note that the graph is intentionally kept acyclic to minimize the
complexity of the critical sequence computation algorithm (complexity is
currently linear in the number of nodes in the graph).
The critical path is finally computed as a sequence of dependency edges. For
edges describing processor resource interferences, the view also prints a
so-called "interference probability" value (by dividing field
DependencyEdge::Frequency by the total number of iterations).
Examples of critical sequence computations can be found in tests added/modified
by this patch.
On output streams that support colored output, instructions from the critical
sequence are rendered with a different color.
Strictly speaking the analysis conducted by the bottleneck analysis view is not
a critical path analysis. The cost of an edge doesn't only depend on the
dependency latency. More importantly, the cost of a same edge may be computed
differently by different iterations.
The number of dependencies is discovered dynamically based on the events
generated by the simulator. However, their number is not fixed. This is
especially true for edges that model processor resource interferences; an
interference may not occur in every iteration. For that reason, it makes sense
to also print out a "probability of interference".
By construction, the accuracy of this analysis (as always) is strongly dependent
on the simulation (and therefore the quality of the information available in the
scheduling model).
That being said, the critical sequence effectively identifies a performance
criticality. Instructions from that sequence are expected to have a very big
impact on performance. So, users can take advantage of this information to focus
their attention on specific interactions between instructions.
In my experience, it works quite well in practice, and produces useful
output (in a reasonable amount time).
Differential Revision: https://reviews.llvm.org/D63543
llvm-svn: 364045
Summary:
llvm.x86.sse.stmxcsr only writes to memory.
llvm.x86.sse.ldmxcsr only reads from memory, and might generate an FPE.
Reviewers: craig.topper, RKSimon
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62896
llvm-svn: 363773
