Split the fp and integer vector logical instruction scheduler classes - older CPUs especially often handled these on different pipes.
This unearthed a couple of things that are also handled in this patch:
(1) We were tagging avx512 fp logic ops as WriteFAdd, probably because of the lack of WriteFLogic
(2) SandyBridge had integer logic ops only using Port5, when afaict they can use Ports015.
(3) Cleaned up x86 FCHS/FABS scheduling as they are typically treated as fp logic ops.
Differential Revision: https://reviews.llvm.org/D45629
llvm-svn: 330480
Three new instructions:
umonitor - Sets up a linear address range to be
monitored by hardware and activates the monitor.
The address range should be a writeback memory
caching type.
umwait - A hint that allows the processor to
stop instruction execution and enter an
implementation-dependent optimized state
until occurrence of a class of events.
tpause - Directs the processor to enter an
implementation-dependent optimized state
until the TSC reaches the value in EDX:EAX.
Also modifying the description of the mfence
instruction, as the rep prefix (0xF3) was allowed
before, which would conflict with umonitor during
disassembly.
Before:
$ echo 0xf3,0x0f,0xae,0xf0 | llvm-mc -disassemble
.text
mfence
After:
$ echo 0xf3,0x0f,0xae,0xf0 | llvm-mc -disassemble
.text
umonitor %rax
Reviewers: craig.topper, zvi
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D45253
llvm-svn: 330462
Silvermont and Goldmont have the same issue on popcnt as Sandy Bridge, Haswell, Broadwell, and Skylake. Believe it is fixed in Goldmont Plus.
llvm-svn: 330358
The XCHG16rr/XCHG32rr/XCHG64rr instructions should be 3 uops just like XCHG8rr. I believe they're just implemented as 3 move uops with a temporary register.
XADD is probably 2 moves and an add also using a temporary register.
Change the latency for both from 2 cycles to 3 cycles. Only 2 of the uops are serialized in their execution, the move into the temporary and the move out of the temporary. The move from one GPR to the other should be able to go in parallel with this if there are ALU resources available.
llvm-svn: 330349
This is the patch that lowers x86 intrinsics to native IR
in order to enable optimizations. The patch also includes folding
of previously missing saturation patterns so that IR emits the same
machine instructions as the intrinsics.
Patch by tkrupa
Differential Revision: https://reviews.llvm.org/D44785
llvm-svn: 330322
This removes a bunch of unnecessary InstRW overrides. It also cleans up the missing information from the Sandy Bridge model. Other fixes to other models.
llvm-svn: 330308
Summary:
ASSERT_SORTED checks if a table is sorted, and uses a boolean to
prevent the check from being run again if it was earlier determined
that the table is in fact sorted. Unsynchronized reads and writes of
that boolean triggered ThreadSanitizer's data race detection. This
change rewrites the code to use std::atomic<bool> instead.
Fixes PR36922.
Reviewers: rnk
Reviewed By: rnk
Subscribers: llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D45742
llvm-svn: 330301
The compiler only emits the locked version of these which use different instruction definitions. The versions fixed here are only used by the assembler/disassembler.
llvm-svn: 330287
a zero register.
Previously I tried this and saw LLVM unable to transform this to fold
with memory operands such as spill slot rematerialization. However, it
clearly works as shown in this patch. We turn these into `cmpb $0,
<mem>` when useful for folding a memory operand without issue. This form
has no disadvantage compared to `testb $-1, <mem>`. So overall, this is
likely no worse and may be slightly smaller in some cases due to the
`testb %reg, %reg` form.
Differential Revision: https://reviews.llvm.org/D45475
llvm-svn: 330269
across basic blocks in the limited cases where it is very straight
forward to do so.
This will also be useful for other places where we do some limited
EFLAGS propagation across CFG edges and need to handle copy rewrites
afterward. I think this is rapidly approaching the maximum we can and
should be doing here. Everything else begins to require either heroic
analysis to prove how to do PHI insertion manually, or somehow managing
arbitrary PHI-ing of EFLAGS with general PHI insertion. Neither of these
seem at all promising so if those cases come up, we'll almost certainly
need to rewrite the parts of LLVM that produce those patterns.
We do now require dominator trees in order to reliably diagnose patterns
that would require PHI nodes. This is a bit unfortunate but it seems
better than the completely mysterious crash we would get otherwise.
Differential Revision: https://reviews.llvm.org/D45673
llvm-svn: 330264
Summary: Previously if a modifer was placed on a non-GPR register class we would hit an assert or crash.
Reviewers: echristo
Reviewed By: echristo
Subscribers: eraman, llvm-commits
Differential Revision: https://reviews.llvm.org/D45751
llvm-svn: 330238
Summary:
Add an LLVM intrinsic for type discriminated event logging with XRay.
Similar to the existing intrinsic for custom events, but also accepts
a type tag argument to allow plugins to be aware of different types
and semantically interpret logged events they know about without
choking on those they don't.
Relies on a symbol defined in compiler-rt patch D43668. I may wait
to submit before I can see demo everything working together including
a still to come clang patch.
Reviewers: dberris, pelikan, eizan, rSerge, timshen
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45633
llvm-svn: 330219
Split VCMP/VMAX/VMIN instructions off to WriteFCmp and VCOMIS instructions off to WriteFCom instead of assuming they match WriteFAdd
Differential Revision: https://reviews.llvm.org/D45656
llvm-svn: 330179
Using Goldmont's cost tables for these two upcoming
atom archs.
Reviewers: craig.topper
Reviewed By: craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45612
llvm-svn: 330109
The destination size of the movzx/movsx instruction is controlled by the normal operand size mechanisms. Only the input type is fixed.
This means that a 0x66 prefix on the encoding for zext/sext 16->32 should really produce a 16->16 instruction. Functionally this is equivalent to a GR16->GR16 move since bits 16 and above will be preserved. So nothing is actually extended.
llvm-svn: 330078
Similar to rL329834, don't rely on itinerary scheduler model to determine latencies for LEA thresholds, use the generic TargetSchedModel::computeInstrLatency call.
llvm-svn: 330030
Hint to hardware to move the cache line containing the
address to a more distant level of the cache without
writing back to memory.
Reviewers: craig.topper, zvi
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D45256
llvm-svn: 329992
This completes the work started in r329604 and r329605 when we changed clang to no longer use the intrinsics.
We lost some InstCombine SimplifyDemandedBit optimizations through this change as we aren't able to fold 'and', bitcast, shuffle very well.
llvm-svn: 329990
This removes the last of the x86 schedule itineraries, I'm intending to cleanup the remaining uses of NoItinerary/OpndItins/etc. before resolving PR37093.
llvm-svn: 329967
A previously missing intrinsic for an old instruction.
Reviewers: craig.topper, echristo
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D45312
llvm-svn: 329936
Similar to the wbinvd instruction, except this
one does not invalidate caches. Ring 0 only.
The encoding matches a wbinvd instruction with
an F3 prefix.
Reviewers: craig.topper, zvi, ashlykov
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D43816
llvm-svn: 329847
Atom is the only x86 target that still uses schedule itineraries, if we can remove this then we can begin the work on removing x86 itineraries. I've also found that it will help with PR36550.
I've focussed on matching the existing model as closely as possible (relying on the schedule tests), PR36895 indicated a lot of these were incorrect but we can just as easily fix these after this patch as before. Hopefully we can get llvm-exegesis to help here,
There are a few instructions that rely on itinerary scheduling (mainly push/pop/return) of multiple resource stages, but I don't think any of these are show stoppers.
There are also a few codegen changes that seem related to the post-ra scheduler acting a little differently, I haven't tracked these down but they don't seem critical.
NOTE: I don't have access to any Atom hardware, so this hasn't been tested in the wild.
Differential Revision: https://reviews.llvm.org/D45486
llvm-svn: 329837
Pre-commit for D45486, don't rely on itinerary scheduler model to determine latencies for padding, use the generic TargetSchedModel::computeInstrLatency call.
Also, replace hard coded (atom specific) 2*uop creation per padding cycle with a version based on the scheduler model's issue width.
Differential Revision: https://reviews.llvm.org/D45486
llvm-svn: 329834
The 128/256-bit versions were no longer used by clang. It uses the legacy SSE/AVX2 version and a select. The 512-bit was changed to the same for consistency.
llvm-svn: 329774
The BroadwellModelProcResources had an entry for HWPort5, which is a Haswell
resource, and not a Broadwell processor resource. That entry was added to the
Broadwell model because variable blends were consuming it.
This was clearly a typo (the resource name should have been BWPort5), which
unfortunately was never caught before. It was not reported as an error because
HWPort5 is a resource defined by the Haswell model. It has been found when
testing some code with llvm-mca: the list of resources in the resource pressure
view was odd.
This patch fixes the issue; now variable blend instructions consume 2 cycles on
BWPort5 instead of HWPort5. This is enough to get rid of the extra (spurious)
entry in the BroadWellModelProcResources table.
llvm-svn: 329686
Summary:
Subtargets can define the libpfm counter names that can be used to
measure cycles and uops issued on ProcResUnits.
This allows making llvm-exegesis available on more targets.
Fixes PR36984.
Reviewers: gchatelet, RKSimon, andreadb, craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45360
llvm-svn: 329675
This cleans up a number of operations that only claimed te use EFLAGS
due to using DF. But no instructions which we think of us setting EFLAGS
actually modify DF (other than things like popf) and so this needlessly
creates uses of EFLAGS that aren't really there.
In fact, DF is so restrictive it is pretty easy to model. Only STD, CLD,
and the whole-flags writes (WRFLAGS and POPF) need to model this.
I've also somewhat cleaned up some of the flag management instruction
definitions to be in the correct .td file.
Adding this extra register also uncovered a failure to use the correct
datatype to hold X86 registers, and I've corrected that as necessary
here.
Differential Revision: https://reviews.llvm.org/D45154
llvm-svn: 329673
Prefer to use the 32-bit AND with immediate instead.
Primarily I'm doing this to ensure that immediates created by shrinkAndImmediate will always get absorbed into the AND. But I do believe this would be a reduction in the number of uops that need to execute. Ideally we should shrink the 'and' and the 'load' during DAG combine to re-enable the fold.
Fixes PR37063.
llvm-svn: 329667
The key idea is to lower COPY nodes populating EFLAGS by scanning the
uses of EFLAGS and introducing dedicated code to preserve the necessary
state in a GPR. In the vast majority of cases, these uses are cmovCC and
jCC instructions. For such cases, we can very easily save and restore
the necessary information by simply inserting a setCC into a GPR where
the original flags are live, and then testing that GPR directly to feed
the cmov or conditional branch.
However, things are a bit more tricky if arithmetic is using the flags.
This patch handles the vast majority of cases that seem to come up in
practice: adc, adcx, adox, rcl, and rcr; all without taking advantage of
partially preserved EFLAGS as LLVM doesn't currently model that at all.
There are a large number of operations that techinaclly observe EFLAGS
currently but shouldn't in this case -- they typically are using DF.
Currently, they will not be handled by this approach. However, I have
never seen this issue come up in practice. It is already pretty rare to
have these patterns come up in practical code with LLVM. I had to resort
to writing MIR tests to cover most of the logic in this pass already.
I suspect even with its current amount of coverage of arithmetic users
of EFLAGS it will be a significant improvement over the current use of
pushf/popf. It will also produce substantially faster code in most of
the common patterns.
This patch also removes all of the old lowering for EFLAGS copies, and
the hack that forced us to use a frame pointer when EFLAGS copies were
found anywhere in a function so that the dynamic stack adjustment wasn't
a problem. None of this is needed as we now lower all of these copies
directly in MI and without require stack adjustments.
Lots of thanks to Reid who came up with several aspects of this
approach, and Craig who helped me work out a couple of things tripping
me up while working on this.
Differential Revision: https://reviews.llvm.org/D45146
llvm-svn: 329657
LowerIntUnary as its name says has an assert for integer types. But for the bitcast case one side might be an FP type.
Rather than making sure the function really works for fp types and renaming it. Just do really basic splitting directly. The LowerIntUnary has the advantage that it can peek through BUILD_VECTOR because every other call is during Lowering. But these calls are during legalization and will be followed by a DAG combine round.
Revert some change to LowerVectorIntUnary that were originally made just to make these two calls work even in pure integer cases.
This was found purely by compiling the avx512f-builtins.c test from clang so I've copied over the offending function from that.
llvm-svn: 329616
While it appears to be correct information based on Intel's optimization manual and Agner's data, it causes perf regressions on a couple of the benchmarks in our internal list.
llvm-svn: 329593
The TargetSchedModel is always initialized using the TargetSubtargetInfo's
MCSchedModel and TargetInstrInfo, so we don't need to extract those and
pass 3 parameters to init().
Differential Revision: https://reviews.llvm.org/D44789
llvm-svn: 329540
Summary:
Cmov and setcc previously used WriteALU, but on Intel processors at least they are more restricted than basic ALU ops.
This patch adds new SchedWrites for them and removes the InstRWs. I had to leave some InstRWs for CMOVA/CMOVBE and SETA/SETBE because those have an extra uop relative to the other condition codes on Intel CPUs.
The test changes are due to fixing a missing ZnAGU dependency on the memory form of setcc.
Reviewers: RKSimon, andreadb, GGanesh
Reviewed By: RKSimon
Subscribers: GGanesh, llvm-commits
Differential Revision: https://reviews.llvm.org/D45380
llvm-svn: 329539
Summary:
This removes the InstRWs for BLENDVPS/PD in favor of WriteFVarBlend. The latency listed was 3 cycles but WriteFVarBlend is defined as 1 cycle latency. The 1 cycle latency matches Agner Fog's data.
The patterns were missing the VEX forms which is why there are no test changes. We don't test "-mcpu=znver1 -mattr=-avx"
Reviewers: RKSimon, GGanesh
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D44841
llvm-svn: 329538
Summary:
r327219 added wrappers to std::sort which randomly shuffle the container before sorting.
This will help in uncovering non-determinism caused due to undefined sorting
order of objects having the same key.
To make use of that infrastructure we need to invoke llvm::sort instead of std::sort.
Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort.
Refer the comments section in D44363 for a list of all the required patches.
Reviewers: chandlerc, craig.topper, RKSimon
Reviewed By: chandlerc, craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D44874
llvm-svn: 329534
Previously we used a custom lowering for this because of the AVX1 splitting requirement. But we can do the split during DAG combine if we check the types and subtarget
llvm-svn: 329510
Summary:
This patch removes InstRW overrides for basic arithmetic/logic instructions. To do this I've added the store address port to RMW. And used a WriteSequence to make the latency additive. It does not cover ADC/SBB because they have different latency.
Apparently we were inconsistent about whether the store has latency or not thus the test changes.
I've also left out Sandy Bridge because the load latency there is currently 4 cycles and should be 5.
Reviewers: RKSimon, andreadb
Reviewed By: andreadb
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45351
llvm-svn: 329416
As mentioned on D44647, this patch increases the default memory latency to +5cy , which more closely matches what most custom cases are doing for reg-mem instructions.
I've bumped LoadLatency, ReadAfterLd and WriteLoad values to 5cy to be consistent.
As Sandy Bridge is currently our default generic model, this affects a lot of scheduling tests...
Differential Revision: https://reviews.llvm.org/D44654
llvm-svn: 329388
Simon Pilgrim