These can be lowered to code sequences using CMPFP and CMPFPE which then get
selected to VCMP and VCMPE. The implementation isn't fully correct, as the chain
operand isn't handled correctly, but resolving that looks like it would involve
changes around FPSCR-handling instructions and how the FPSCR is modelled.
The fp-intrinsics test was already testing some of this but as the entire test
was being XFAILed it wasn't noticed. Un-XFAIL the test and instead leave the
cases where we aren't generating the right instruction sequences as FIXME.
Differential Revision: https://reviews.llvm.org/D73194
Summary:
In big-endian MVE, the simple vector load/store instructions (i.e.
both contiguous and non-widening) don't all store the bytes of a
register to memory in the same order: it matters whether you did a
VSTRB.8, VSTRH.16 or VSTRW.32. Put another way, the in-register
formats of different vector types relate to each other in a different
way from the in-memory formats.
So, if you want to 'bitcast' or 'reinterpret' one vector type as
another, you have to carefully specify which you mean: did you want to
reinterpret the //register// format of one type as that of the other,
or the //memory// format?
The ACLE `vreinterpretq` intrinsics are specified to reinterpret the
register format. But I had implemented them as LLVM IR bitcast, which
is specified for all types as a reinterpretation of the memory format.
So a `vreinterpretq` intrinsic, applied to values already in registers,
would code-generate incorrectly if compiled big-endian: instead of
emitting no code, it would emit a `vrev`.
To fix this, I've introduced a new IR intrinsic to perform a
register-format reinterpretation: `@llvm.arm.mve.vreinterpretq`. It's
implemented by a trivial isel pattern that expects the input in an
MQPR register, and just returns it unchanged.
In the clang codegen, I only emit this new intrinsic where it's
actually needed: I prefer a bitcast wherever it will have the right
effect, because LLVM understands bitcasts better. So we still generate
bitcasts in little-endian mode, and even in big-endian when you're
casting between two vector types with the same lane size.
For testing, I've moved all the codegen tests of vreinterpretq out
into their own file, so that they can have a different set of RUN
lines to check both big- and little-endian.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73786
Summary:
These instructions generate a vector of consecutive elements starting
from a given base value and incrementing by 1, 2, 4 or 8. The `wdup`
versions also wrap the values back to zero when they reach a given
limit value. The instruction updates the scalar base register so that
another use of the same instruction will continue the sequence from
where the previous one left off.
At the IR level, I've represented these instructions as a family of
target-specific intrinsics with two return values (the constructed
vector and the updated base). The user-facing ACLE API provides a set
of intrinsics that throw away the written-back base and another set
that receive it as a pointer so they can update it, plus the usual
predicated versions.
Because the intrinsics return two values (as do the underlying
instructions), the isel has to be done in C++.
This is the first family of MVE intrinsics that use the `imm_1248`
immediate type in the clang Tablegen framework, so naturally, I found
I'd given it the wrong C integer type. Also added some tests of the
check that the immediate has a legal value, because this is the first
time those particular checks have been exercised.
Finally, I also had to fix a bug in MveEmitter which failed an
assertion when I nested two `seq` nodes (the inner one used to extract
the two values from the pair returned by the IR intrinsic, and the
outer one put on by the predication multiclass).
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73357
Summary:
The unpredicated case of this is trivial: the clang codegen just makes
a vector splat of the input, and LLVM isel is already prepared to
handle that. For the predicated version, I've generated a `select`
between the same vector splat and the `inactive` input parameter, and
added new Tablegen isel rules to match that pattern into a predicated
`MVE_VDUP` instruction.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73356
Summary: There are no counters for individual ports, but this is already
enough to find a lot of issues in the current model (upcoming patch).
Reviewers: dblaikie, gchatelet
Subscribers: hiraditya, tschuett, RKSimon, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72032
Summary:
D68092 introduced a new SIRemoveShortExecBranches optimization pass and
broke some graphics shaders. The problem is that it was removing
branches over KILL pseudo instructions, and the fix is to explicitly
check for that in mustRetainExeczBranch.
Reviewers: critson, arsenm, nhaehnle, cdevadas, hakzsam
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73771
We only need to call this on floating point comparisons. In this
case these are known to be integer compares. One of them even
has a SUB opcode instead of CMP.
This reverts commit e34801c8e6 and the followup due to multiple
problems.
I've tried to keep the tests and RDA parts where possible, as those
still seem useful.
Summary:
Virtual registers that are undef have an empty LiveInterval at this
point, which means beginIndex() and endIndex() cannot be used. We
only need those indices to determine the range in which to scan for
affected other NSA instructions, and undef operands cannot contribute
to that range.
Reviewers: arsenm, rampitec, mareko
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73831
We were checking that the original Value * for the compare operands
were null. But that can never happen.
I believe we intended to check for 0 registers here instead.
Fixes PR44749.
This is an alternate fix for the issue D73606 was trying to
solve.
The main issue here is that we bailed out of
foldOffsetIntoAddress if Offset is 0. But if we just found a
symbolic displacement and AM.Disp became non-zero
earlier, we still need to validate that AM.Disp with the symbolic
displacement.
This is my second attempt at committing this after failing
build bots previously. One thing I realized about the previous
attempt is that its possible that AM.Disp is already non-zero
and the new Offset changes it back to zero. In that case my
previous attempt failed to update AM.Disp to zero. So this patch
removes the early out for 0 and appropriately handle the 0 case
in each check so we still update AM.Disp at the end.
This is based on this llvm-dev thread http://lists.llvm.org/pipermail/llvm-dev/2019-December/137521.html
The current strategy for f16 is to promote type to float every except where the specific width is required like loads, stores, and bitcasts. This results in rounding occurring in odd places instead of immediately after arithmetic operations. This interacts in weird ways with the __fp16 type in clang which is a storage only type where arithmetic is always promoted to float. InstCombine can remove some fpext/fptruncs around such arithmetic and turn it into arithmetic on half. This wouldn't be so bad if SelectionDAG was able to put those fpext/fpround back in when it promotes.
It is also not obvious how to handle to make the existing strategy work with STRICT fp. We need to use STRICT versions of the conversions which require chain operands. But if the conversions are created for a bitcast, there is no place to get an appropriate chain from.
This patch implements a different strategy where conversions are emitted directly around arithmetic operations. And otherwise its passed around as an i16 including in arguments and return values. This can result in more conversions between arithmetic operations, but is closer to matching the IR the frontend generates for __fp16. And it will allow us to use the chain from constrained arithmetic nodes to link the STRICT_FP_TO_FP16/STRICT_FP16_TO_FP that will need to be added. I've set it up so that each target can opt into the new behavior. Converting all the targets myself was more than I was able to handle.
Differential Revision: https://reviews.llvm.org/D73749
This fixes legalizations of global stores > 128-bits. It seems work is
needed on how this split actually occurs. For example, we get the
right code for s160, with an s128 and s32 load, but get 5 s32 loads
for <5 x s32>.
Summary:
Implements the jump pseudo-instruction, which is used in e.g. the Linux kernel.
Reviewers: asb, lenary
Reviewed By: lenary
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73178
When you encounter a G_TRUNC, you are moving from a larger type to a smaller
type.
Asking for the i-th bit on a larger value is the same as asking for the i-th
bit on a smaller value.
So, we should always be able to walk through G_TRUNC when computing the bit
for a TB(N)Z.
Differential Revision: https://reviews.llvm.org/D73748
Summary: This is a first step before changing the types to llvm::Align and introduce functions to ease client code.
Reviewers: courbet
Subscribers: arsenm, sdardis, nemanjai, jvesely, nhaehnle, hiraditya, kbarton, jrtc27, atanasyan, jsji, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73785
Try out using combine definition rules.
This really should be a post-legalizer combine, but the combiner pass
is currently pre-legalize. Most of the target combines are really
post-legalize, so we should probably move the pass.
For a MC_GlobalAddress reference to a dso_local external GlobalValue with a definition, emit .Lfoo$local to avoid a relocation.
-fno-pic and -fpie can infer dso_local but -fpic cannot. In the future,
we can explore the possibility of inferring dso_local with -fpic. As the
description of D73228 says, LLVM's existing IPO optimization behaviors
(like -fno-semantic-interposition) and a previous assembly behavior give
us enough license to be aggressive here.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D73230
I believe this also fixes bugs with CI 32-bit handling, which was
incorrectly skipping offsets that look like signed 32-bit values. Also
validate the offsets are dword aligned before folding.
This is similar to the code in getTestBitOperand in AArch64ISelLowering. Instead
of implementing all of the TB(N)Z optimizations at once, this patch implements
the simplest case first. The way that this is set up should make it fairly easy
to add the rest as we go along.
The idea here is that after determining that we can use a TB(N)Z, we can
continue looking through instructions and perform further folding.
In this case, when we have a G_ZEXT or G_ANYEXT where the extended bits are not
used, we can fold it into the TB(N)Z.
Differential Revision: https://reviews.llvm.org/D73673
Found by inspection, but there's no test for this yet because G_PTR_ADD is
currently illegal for vectors. I'll add the test at a later time when the
legalizer support has landed.
There's not much value to this separate node from the intrinsic. Make
the operand structure the same as the intrinsic, so we can reuse the
same pattern for GlobalISel.
Summary:
Added file headers for files which implement iterative lightweight scheduling
strategies. Which is basically an exercise which I undertook in order to get
used to LLVM development process.
Reviewers: arsenm, vpykhtin, cdevadas
Reviewed By: vpykhtin
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, javed.absar, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73417
Summary:
For -fpatchable-function-entry=N,0 -mbranch-protection=bti, after
9a24488cb6, we place the NOP sled after
the initial BTI.
```
.Lfunc_begin0:
bti c
nop
nop
.section __patchable_function_entries,"awo",@progbits,f,unique,0
.p2align 3
.xword .Lfunc_begin0
```
This patch adds a label after the initial BTI and changes the __patchable_function_entries entry to reference the label:
```
.Lfunc_begin0:
bti c
.Lpatch0:
nop
nop
.section __patchable_function_entries,"awo",@progbits,f,unique,0
.p2align 3
.xword .Lpatch0
```
This placement is compatible with the resolution in
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92424 .
A local linkage function whose address is not taken does not need a BTI.
Placing the patch label after BTI has the advantage that code does not
need to differentiate whether the function has an initial BTI.
Reviewers: mrutland, nickdesaulniers, nsz, ostannard
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73680
fadd/fmul reductions without reassoc are lowered to
VECREDUCE_STRICT_FADD/FMUL nodes, which don't have legalization
support. Until that is in place, expand these intrinsics on
ARM and AArch64. Other targets always expand the vector reduction
intrinsics.
Additionally expand fmax/fmin reductions without nonan flag on
AArch64, as the backend asserts that the flag is present when
lowering VECREDUCE_FMIN/FMAX.
This fixes https://bugs.llvm.org/show_bug.cgi?id=44600.
Differential Revision: https://reviews.llvm.org/D73135
The recommended optimization level for BPF programs
is O2 since (1). BPF is running inside the kernel and
linux kernel won't work at -O0 level, and (2). Verifier
is not able to handle O0 code properly, e.g., potential
large stack size and a lot of spills.
But we should keep -O0 at least compiling.
This patch fixed a bug in BPFMISimplifyPatchable phase
where with -O0, a segmentation fault will happen for a
simple program like:
int test(int a, int b) { return a + b; }
A test case is added to capture such a case.
Differential Revision: https://reviews.llvm.org/D73681
Summary:
This patch intends to support three most common relocation type
on AIX: R_POS, R_TOC, R_RBR.
These three relocation type will be needed for object file generation
on AIX for small code model.
We will have follow up patches to bring relocation support for
large code model on AIX.
Reviewers: hubert.reinterpretcast, daltenty, DiggerLin
Differential Revision: https://reviews.llvm.org/D72027
By adding the prefixed instructions the branch distances are no longer
computed correctly. Since prefixed instructions cannot cross a 64 byte
boundary we have to assume that a prefixed instruction may have a nop
prepended to it. This patch tries to take that nop into consideration
when computing the size of basic blocks.
Differential Revision: https://reviews.llvm.org/D72572
The commit https://reviews.llvm.org/rG14fc20ca6 added some options to the X86
back end that cause the help text for opt/llc to become much harder to read.
The issue is that the cl::value_desc is part of the option name and is used to
compute the indentation of the description text (i.e. the maximum length option
name is what everything aligns to). Since the commit puts a large number of
characters into that text, everything is aligned to that width.
This patch just reformats the option so that the description is contained in the
description and the list of possible values is within the angle brackets.
Note: the readability issue of the helptext was fixed in commit
70cbf8c71c, but the re-formatting wasn't
added on that commit so I am still committing this.
Differential revision: https://reviews.llvm.org/D73267
Strict fp-to-int and int-to-fp conversions can be handled in the same way that
the non-strict versions are (by using the appropriate instruction or converting
to a function call when we have no instruction).
Differential Revision: https://reviews.llvm.org/D73625
On targets that don't have the normal packed f16 layout, handle these
during legalization. Directly modify the register types. We can infer
this was a d16 load based on the mem operand size during selection.
A16 operands should possibly be handled here as well, but don't worry
about that yet.
This trivially avoids violating the constant bus restriction.
Previously this was allowing one SGPR in the first source
operand, which technically also avoided violating this for most
operations (but not for special cases reading vcc).
We do need to write some new, smarter operand folds to pick the
optimal SGPR to use in some kind of post-isel fold, but that's purely
an optimization.
I was originally thinking we would pick which operands should be SGPRs
in RegBankSelect, but I think this isn't really manageable. There
would be additional complexity to handle every G_* instruction, and
then any nontrivial instruction patterns would need to know when to
avoid violating it, which is likely to be very error prone.
I think having all inputs being canonically copies to VGPRs will
simplify the operand folding logic. The current folding we do is
backwards, and only considers one operand at a time, relative to
operands it already has. It therefore poorly handles the case where
there is already a constant bus operand user. If all operands are
copies, it's somewhat simpler to consider all input operands at once
to choose the optimal constant bus user.
Since the failure mode for constant bus violations is now a verifier
error and not an selection failure, this moves towards a place where
we can turn on the fallback mode. The SGPR copy folding optimizations
can be left for later.
A known limitation for Future CPU is that the new prefixed instructions may
not cross 64 Byte boundaries.
All instructions are already 4 byte aligned so the only situation where this
can occur is when the prefix is in one 64 byte block and the instruction that
is prefixed is at the top of the next 64 byte block. To fix this case
PPCELFStreamer was added to intercept EmitInstruction. When a prefixed
instruction is emitted we try to align it to 64 Bytes by adding a maximum of
4 bytes. If the prefixed instruction crosses the 64 Byte boundary then the
alignment would trigger and a 4 byte nop would be added to push the
instruction into the next 64 byte block.
Differential Revision: https://reviews.llvm.org/D72570
This gets selected to the appropriate fcvt instruction. Handling from there on
isn't fully correct yet, as we need to model fcvt reading and writing to fpsr
and fpcr.
Differential Revision: https://reviews.llvm.org/D73201
These become STRICT_FCMP and STRICT_FCMPE, which then get selected to the
corresponding FCMP and FCMPE instructions, though the handling from there on
isn't fully correct as we don't model reads and writes to FPCR and FPSR.
Differential Revision: https://reviews.llvm.org/D73368
Summary:
The code was assuming in a few places that if there was only one exit
from the function that it was a normal return, which is invalid. It
could be an infinite loop, in which case we still need to insert the
usual fake edge so that the null export happens. This fixes shaders that
end with an infinite loop that discards.
Reviewers: arsenm, nhaehnle, critson
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71192
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
This lowering tries to look for G_PTR_ADD instructions and then converts
them to a standard G_ADD with a COPY on the source, and G_INTTOPTR on the
result. This is ok for address space 0 on AArch64 as p0 can be treated as
s64.
The motivation behind this is to expose the add semantics to the imported
tablegen patterns. We shouldn't need to check for uses being loads/stores,
because the selector works bottom up, uses before defs. By the time we
end up trying to select a G_PTR_ADD, we should have already attempted to
fold this into addressing modes and were therefore unsuccessful.
This gives some performance and code size improvements across the board.
Differential Revision: https://reviews.llvm.org/D73673
Currently some prefixes are emitted as instructions, to distinguish them
from real instruction, fuction isPrefix() is added. The kinds of prefix
are consistent with X86GenInstrInfo.inc.
Differential Revision: https://reviews.llvm.org/D73013
This is an alternate fix for the issue D73606 was trying to
solve.
The main issue here is that we bailed out of
foldOffsetIntoAddress if Offset is 0. But if we just found a
symbolic displacement and AM.Disp became non-zero
earlier, we still need to validate that AM.Disp with the symbolic
displacement.
This passes fold-add-pcrel.ll.
Differential Revision: https://reviews.llvm.org/D73608
We seem to be inheriting the cost from sse4.1. But if we have 256-bit registers we should be able to do this with just one extract to split the 16i16 and two v8i16->v8i32 operations so our cost should be 3 not 4.
Differential Revision: https://reviews.llvm.org/D73646
This is passed to legalizeCustom, but not intrinsic. Also remove the
MRI argument, since you can get that from the MachineIRBuilder.
I'm not sure why MachineIRBuilder has a private observer member, and
this is passed separately.
Rename Destructive enumerator in preparation for a larger set of patches to
support prefixing destructive oeprations with MOVPRFX.
Differential Revision: https://reviews.llvm.org/D73212
When the bit is <= 32, we have to use the W register variant for TB(N)Z.
This is because of the way the instruction is encoded.
Differential Revision: https://reviews.llvm.org/D73660
Some housekeeping for the DestructiveInstType enum before a larger set of patches to support prefixing destructive oeprations with MOVPRFX.
Differential Revision: https://reviews.llvm.org/D73141
A previous patch should have added pld and pstd and any support code in
the backend that is required for prefixed load and store type operations.
This patch adds a number of additional prefixed load and store type
instructions for the future CPU.
Differential Revision: https://reviews.llvm.org/D72577
Summary:
The initialization of RegisterBank needs to be done only once. The
logic of AlreadyInit has data race, use llvm::call_once instead.
This is continuing work of D73587.
Reviewers: arsenm, rovka, dsanders, t.p.northover, efriedma, apazos
Reviewed By: arsenm
Subscribers: wdng, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73605
Summary:
The initialization of RegisterBank needs to be done only once. The
logic of AlreadyInit has data race, use llvm::call_once instead.
This is continuing work of D73587.
Reviewers: arsenm, tstellar, ronlieb, efriedma, apazos, nhaehnle
Reviewed By: nhaehnle
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73604
Summary:
The initialization of RegisterBank needs to be done only once. The
logic of AlreadyInit has a data race, use llvm::call_once instead.
This issue was identified through thread sanitizer.
Reviewers: efriedma, apazos, qcolombet, dsanders
Reviewed By: efriedma
Subscribers: arsenm, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73587
ISD::FROUND is defined to round to nearest with ties rounding
away from 0. This mode isn't supported in hardware on X86.
But as long as we aren't compiling with trapping math, we can
emulate this with floor(X + copysign(nextafter(0.5, 0.0), X)).
We have to use nextafter to avoid some corner cases that adding
0.5 would have. For example, if X is nextafter(0.5, 0.0) it should
round to 0.0, but adding 0.5 would need one extra bit of mantissa
than can be stored so it rounds to 1.0. Adding nextafter(0.5, 0.0)
instead will just increase the exponent by 1 and leave the mantissa
as all 1s. This would be nextafter(1.0, 0.0) which will floor to 0.0.
Techically this requires -fno-trapping-math which isn't our default.
But if we care about exceptions we should be using constrained
intrinsics. Constrained intrinsics would use STRICT_FROUND which
won't go through this code.
Fixes PR42195.
Differential Revision: https://reviews.llvm.org/D73607
This code needs to map from the FPCW 2-bit encoding for rounding mode to the 2-bit encoding defined for FLT_ROUNDS. The previous implementation did some clever swapping of bits and adding 1 modulo 4 to do the mapping.
This patch instead uses an 8-bit immediate as a lookup table of four 2-bit values. Then we use the 2-bit FPCW encoding to index the lookup table by using a right shift and an AND. This requires extracting the 2-bit value from FPCW and multipying it by 2 to make it usable as a shift amount. But still results in less code.
Differential Revision: https://reviews.llvm.org/D73599
Fixes selection for scalar G_SMULH/G_UMULH. Also switches to using
tablegen selected add/sub, which switch to the signed version of the
opcode. This matches the current DAG behavior. We can't drop the
manual selection for add/sub yet, because it's still both for VALU
add/sub and for G_PTR_ADD.
Manually select this is as a tablegen workraound. Both SelectionDAG
and GlobalISel end up misplacing the copy to m0 when both instructions
in the output need it. Neither considers that both output instructions
depend on m0. I don't know of any other pattern we need to handle this
case, so it's less effort to just workaround this for now.
Summary:
Due to the fact that kill is just a normal intrinsic, even though it's
supposed to terminate the thread, we can end up with provably infinite
loops that are actually supposed to end successfully. The
AMDGPUUnifyDivergentExitNodes pass breaks up these loops, but because
there's no obvious place to make the loop branch to, it just makes it
return immediately, which skips the exports that are supposed to happen
at the end and hangs the GPU if all the threads end up being killed.
While it would be nice if the fact that kill terminates the thread were
modeled in the IR, I think that the structurizer as-is would make a mess if we
did that when the kill is inside control flow. For now, we just add a null
export at the end to make sure that it always exports something, which fixes
the immediate problem without penalizing the more common case. This means that
we sometimes do two "done" exports when only some of the threads enter the
discard loop, but from tests the hardware seems ok with that.
This fixes dEQP-VK.graphicsfuzz.while-inside-switch with radv.
Reviewers: arsenm, nhaehnle
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70781
SIMachineScheduler uses isHighLatencyInstruction with the same
sematincs, but TargetInstrInfo has virtual isHighLatencyDef
method, so override it instead.
Added FLAT to the list of high latency opcodes and a check for
mayLoad since stores are not technically high latency in terms
of data dependency.
This change did not produce any visible impact on our tests.
Differential Revision: https://reviews.llvm.org/D73582
Summary:
udiv/sdiv/urem/srem/mul integer isel patterns and tests.
Pretend for now that integer division were always cheap in HW.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D73623
This should be no problem to support with a pattern, but it turns out
there are just too many yaks to shave. The main problem is in the DAG
emitter, which I have no desire to sink effort into fixing.
If we had a bit to disable patterns in the DAG importer, fixing the
GlobalISelEmitter is more manageable.
Summary:
The code was assuming in a few places that if there was only one exit
from the function that it was a normal return, which is invalid. It
could be an infinite loop, in which case we still need to insert the
usual fake edge so that the null export happens. This fixes shaders that
end with an infinite loop that discards.
Reviewers: arsenm, nhaehnle, critson
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71192
Summary:
Due to the fact that kill is just a normal intrinsic, even though it's
supposed to terminate the thread, we can end up with provably infinite
loops that are actually supposed to end successfully. The
AMDGPUUnifyDivergentExitNodes pass breaks up these loops, but because
there's no obvious place to make the loop branch to, it just makes it
return immediately, which skips the exports that are supposed to happen
at the end and hangs the GPU if all the threads end up being killed.
While it would be nice if the fact that kill terminates the thread were
modeled in the IR, I think that the structurizer as-is would make a mess if we
did that when the kill is inside control flow. For now, we just add a null
export at the end to make sure that it always exports something, which fixes
the immediate problem without penalizing the more common case. This means that
we sometimes do two "done" exports when only some of the threads enter the
discard loop, but from tests the hardware seems ok with that.
This fixes dEQP-VK.graphicsfuzz.while-inside-switch with radv.
Reviewers: arsenm, nhaehnle
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70781
Summary:
Currently, sqdmulh_lane and friends from the ACLE (implemented in arm_neon.h),
are represented in LLVM IR as a (by vector) sqdmulh and a vector of (repeated)
indices, like so:
%shuffle = shufflevector <4 x i16> %v, <4 x i16> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
%vqdmulh2.i = tail call <4 x i16> @llvm.aarch64.neon.sqdmulh.v4i16(<4 x i16> %a, <4 x i16> %shuffle)
When %v's values are known, the shufflevector is optimized away and we are no
longer able to select the lane variant of sqdmulh in the backend.
This defeats a (hand-coded) optimization that packs several constants into a
single vector and uses the lane intrinsics to reduce register pressure and
trade-off materialising several constants for a single vector load from the
constant pool, like so:
int16x8_t v = {2,3,4,5,6,7,8,9};
a = vqdmulh_laneq_s16(a, v, 0);
b = vqdmulh_laneq_s16(b, v, 1);
c = vqdmulh_laneq_s16(c, v, 2);
d = vqdmulh_laneq_s16(d, v, 3);
[...]
In one microbenchmark from libjpeg-turbo this accounts for a 2.5% to 4%
performance difference.
We could teach the compiler to recover the lane variants, but this would likely
require its own pass. (Alternatively, "volatile" could be used on the constants
vector, but this is a bit ugly.)
This patch instead implements the following LLVM IR intrinsics for AArch64 to
maintain the original structure through IR optmization and into instruction
selection:
- sqdmulh_lane
- sqdmulh_laneq
- sqrdmulh_lane
- sqrdmulh_laneq.
These 'lane' variants need an additional register class. The second argument
must be in the lower half of the 64-bit NEON register file, but only when
operating on i16 elements.
Note that the existing patterns for shufflevector and sqdmulh into sqdmulh_lane
(etc.) remain, so code that does not rely on NEON intrinsics to generate these
instructions is not affected.
This patch also changes clang to emit these IR intrinsics for the corresponding
NEON intrinsics (AArch64 only).
Reviewers: SjoerdMeijer, dmgreen, t.p.northover, rovka, rengolin, efriedma
Reviewed By: efriedma
Subscribers: kristof.beyls, hiraditya, jdoerfert, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71469
Simon Moll