Implement AArch64 variant of shouldCoalesce() to detect a known failing case
and prevent the coalescing of a 32-bit copy into a 64-bit sign-extending load.
Do not coalesce in the following case:
COPY where source is bottom 32 bits of a 64-register,
and destination is a 32-bit subregister of a 64-bit register,
ie it causes the rest of the register to be implicitly set to zero.
A mir test has been added.
In the test case, the 32-bit copy implements a 32 to 64 bit zero extension
and relies on the upper 32 bits being zeroed.
Coalescing to the result of the 64-bit load meant overwriting
the upper 32 bits incorrectly when the loaded byte was negative.
Reviewed By: john.brawn
Differential Revision: https://reviews.llvm.org/D85956
This was reverted in 503deec218
because it caused gigantic increase (3x) in branch mispredictions
in certain benchmarks on certain CPU's,
see https://reviews.llvm.org/D84108#2227365.
It has since been investigated and here are the results:
https://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20200907/827578.html
> It's an amazingly severe regression, but it's also all due to branch
> mispredicts (about 3x without this). The code layout looks ok so there's
> probably something else to deal with. I'm not sure there's anything we can
> reasonably do so we'll just have to take the hit for now and wait for
> another code reorganization to make the branch predictor a bit more happy :)
>
> Thanks for giving us some time to investigate and feel free to recommit
> whenever you'd like.
>
> -eric
So let's just reland this.
Original commit message:
I've been looking at missed vectorizations in one codebase.
One particular thing that stands out is that some of the loops
reach vectorizer in a rather mangled form, with weird PHI's,
and some of the loops aren't even in a rotated form.
After taking a more detailed look, that happened because
the loop's headers were too big by then. It is evident that
SimplifyCFG's common code hoisting transform is at fault there,
because the pattern it handles is precisely the unrotated
loop basic block structure.
Surprizingly, `SimplifyCFGOpt::HoistThenElseCodeToIf()` is enabled
by default, and is always run, unlike it's friend, common code sinking
transform, `SinkCommonCodeFromPredecessors()`, which is not enabled
by default and is only run once very late in the pipeline.
I'm proposing to harmonize this, and disable common code hoisting
until //late// in pipeline. Definition of //late// may vary,
here currently i've picked the same one as for code sinking,
but i suppose we could enable it as soon as right after
loop rotation happens.
Experimentation shows that this does indeed unsurprizingly help,
more loops got rotated, although other issues remain elsewhere.
Now, this undoubtedly seriously shakes phase ordering.
This will undoubtedly be a mixed bag in terms of both compile- and
run- time performance, codesize. Since we no longer aggressively
hoist+deduplicate common code, we don't pay the price of said hoisting
(which wasn't big). That may allow more loops to be rotated,
so we pay that price. That, in turn, that may enable all the transforms
that require canonical (rotated) loop form, including but not limited to
vectorization, so we pay that too. And in general, no deduplication means
more [duplicate] instructions going through the optimizations. But there's still
late hoisting, some of them will be caught late.
As per benchmarks i've run {F12360204}, this is mostly within the noise,
there are some small improvements, some small regressions.
One big regression i saw i fixed in rG8d487668d09fb0e4e54f36207f07c1480ffabbfd, but i'm sure
this will expose many more pre-existing missed optimizations, as usual :S
llvm-compile-time-tracker.com thoughts on this:
http://llvm-compile-time-tracker.com/compare.php?from=e40315d2b4ed1e38962a8f33ff151693ed4ada63&to=c8289c0ecbf235da9fb0e3bc052e3c0d6bff5cf9&stat=instructions
* this does regress compile-time by +0.5% geomean (unsurprizingly)
* size impact varies; for ThinLTO it's actually an improvement
The largest fallout appears to be in GVN's load partial redundancy
elimination, it spends *much* more time in
`MemoryDependenceResults::getNonLocalPointerDependency()`.
Non-local `MemoryDependenceResults` is widely-known to be, uh, costly.
There does not appear to be a proper solution to this issue,
other than silencing the compile-time performance regression
by tuning cut-off thresholds in `MemoryDependenceResults`,
at the cost of potentially regressing run-time performance.
D84609 attempts to move in that direction, but the path is unclear
and is going to take some time.
If we look at stats before/after diffs, some excerpts:
* RawSpeed (the target) {F12360200}
* -14 (-73.68%) loops not rotated due to the header size (yay)
* -272 (-0.67%) `"Number of live out of a loop variables"` - good for vectorizer
* -3937 (-64.19%) common instructions hoisted
* +561 (+0.06%) x86 asm instructions
* -2 basic blocks
* +2418 (+0.11%) IR instructions
* vanilla test-suite + RawSpeed + darktable {F12360201}
* -36396 (-65.29%) common instructions hoisted
* +1676 (+0.02%) x86 asm instructions
* +662 (+0.06%) basic blocks
* +4395 (+0.04%) IR instructions
It is likely to be sub-optimal for when optimizing for code size,
so one might want to change tune pipeline by enabling sinking/hoisting
when optimizing for size.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D84108
This reverts commit 503deec218.
Add the functionality to lower SVE rounding operations for passthru variant.
Created a new test case file for all rounding operations.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D86793
Add support in llvm-readobj for displaying them and support in the
asm parsser, AArch64TargetStreamer and MCWin64EH for emitting them.
The directives for the remaining basic opcodes have names that
match the opcode in the documentation.
The directives for custom stack cases, that are named
MSFT_OP_TRAP_FRAME, MSFT_OP_MACHINE_FRAME, MSFT_OP_CONTEXT
and MSFT_OP_CLEAR_UNWOUND_TO_CALL, are given matching assembler
directive names that fit into the rest of the opcode naming;
.seh_trap_frame, .seh_context, .seh_clear_unwound_to_call
The opcode MSFT_OP_MACHINE_FRAME is mapped to the existing
opecode enum UOP_PushMachFrame that is used on x86_64, and also
uses the corresponding existing x86_64 directive name
.seh_pushframe.
Differential Revision: https://reviews.llvm.org/D86889
Unwinders may only preserve the lower 64bits of Neon and SVE registers,
as only the registers in the base ABI are guaranteed to be preserved
over the exception edge. The caller will need to preserve additional
registers for when the call throws an exception and the unwinder has
tried to recover state.
For e.g.
svint32_t bar(svint32_t);
svint32_t foo(svint32_t x, bool *err) {
try { bar(x); } catch (...) { *err = true; }
return x;
}
`z0` needs to be spilled before the call to `bar(x)` and reloaded before
returning from foo, as the exception handler may have clobbered z0.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D84737
This reverts commit bc9a29b9ee.
The reasoning that this patch was wrong was itself incorrect
(see discussion on llvm-commits). This patch does seem to be exposing
a latent SVE code generation bug on non-public tests, which should
not block a correctness fix for public, non-SVE use cases.
This reverts commit e9d9a61208.
This patch was previously revert by 04879086b4
with the reapplication being done after breaking the assert used to
ensure SP is always 16-byte aligned, which is a requirement of the AAPCS.
For extra context the latest patch caused runtime failures when
building with "-march=armv8-a+sve -mllvm -aarch64-sve-vector-bits-min=256".
This requires adding a missing 'const' to the definition because
the callers are using const args, but there should be no change
in behavior.
The intrinsic method was added with D86798 / rG096527214033
This ensures that you get the same output regardless if generating
code directly to an object file or if generating assembly and
assembling that.
Add implementations of the EmitARM64WinCFI*() methods in
AArch64TargetAsmStreamer, and fill in one blank in MCAsmStreamer.
Add corresponding directive handlers in AArch64AsmParser and
COFFAsmParser.
Some SEH directive names have been picked to match the prior art
for SEH assembly directives for x86_64, e.g. the spelling of
".seh_startepilogue" matching the preexisting ".seh_endprologue".
For the directives for saving registers, the exact spelling
from the arm64 documentation is picked, e.g. ".seh_save_reg" (to follow
that naming for all the other ones, e.g. ".seh_save_fregp_x"), while
the corresponding one for x86_64 is plain ".seh_savereg" without the
second underscore.
Directives in the epilogues have the same names as in prologues,
e.g. .seh_savereg, even though the registers are restored, not
saved, at that point.
Differential Revision: https://reviews.llvm.org/D86529
A couple of AArch64 tests were failing on Solaris, both sparc and x86:
LLVM :: MC/AArch64/SVE/add-diagnostics.s
LLVM :: MC/AArch64/SVE/cpy-diagnostics.s
LLVM :: MC/AArch64/SVE/cpy.s
LLVM :: MC/AArch64/SVE/dup-diagnostics.s
LLVM :: MC/AArch64/SVE/dup.s
LLVM :: MC/AArch64/SVE/mov-diagnostics.s
LLVM :: MC/AArch64/SVE/mov.s
LLVM :: MC/AArch64/SVE/sqadd-diagnostics.s
LLVM :: MC/AArch64/SVE/sqsub-diagnostics.s
LLVM :: MC/AArch64/SVE/sub-diagnostics.s
LLVM :: MC/AArch64/SVE/subr-diagnostics.s
LLVM :: MC/AArch64/SVE/uqadd-diagnostics.s
LLVM :: MC/AArch64/SVE/uqsub-diagnostics.s
For example, reduced from `MC/AArch64/SVE/add-diagnostics.s`:
add z0.b, z0.b, #0, lsl #8
missed the expected diagnostics
$ ./bin/llvm-mc -triple=aarch64 -show-encoding -mattr=+sve add.s
add.s:1:21: error: immediate must be an integer in range [0, 255] with a shift amount of 0
add z0.b, z0.b, #0, lsl #8
^
The message is `Match_InvalidSVEAddSubImm8`, emitted in the generated
`lib/Target/AArch64/AArch64GenAsmMatcher.inc` for `MCK_SVEAddSubImm8`.
When comparing the call to `::AArch64Operand::isSVEAddSubImm<char>` on both
Linux/x86_64 and Solaris, I find
875 bool IsByte = std::is_same<int8_t, std::make_signed_t<T>>::value;
is `false` on Solaris, unlike Linux.
The problem boils down to the fact that `int8_t` is plain `char` on
Solaris: both the sparc and i386 psABIs have `char` as signed. However,
with
9887 DiagnosticPredicate DP(Operand.isSVEAddSubImm<int8_t>());
in `lib/Target/AArch64/AArch64GenAsmMatcher.inc`, `std::make_signed_t<int8_t>`
above yieds `signed char`, so `std::is_same<int8_t, signed char>` is `false`.
This can easily be fixed by also allowing for `int8_t` here and in a few
similar places.
Tested on `amd64-pc-solaris2.11`, `sparcv9-sun-solaris2.11`, and
`x86_64-pc-linux-gnu`.
Differential Revision: https://reviews.llvm.org/D85225
There's a special case in hasAttribute for None when pImpl is null. If pImpl is not null we dispatch to pImpl->hasAttribute which will always return false for Attribute::None.
So if we just want to check for None its sufficient to just check that pImpl is null. Which can even be done inline.
This patch adds a helper for that case which I hope will speed up our getSubtargetImpl implementations.
Differential Revision: https://reviews.llvm.org/D86744
This patch changes ElementCount so that the Min and Scalable
members are now private and can only be accessed via the get
functions getKnownMinValue() and isScalable(). In addition I've
added some other member functions for more commonly used operations.
Hopefully this makes the class more useful and will reduce the
need for calling getKnownMinValue().
Differential Revision: https://reviews.llvm.org/D86065
Previously in addTypeForNeon, we would set the operations for bfloat vectors
like other generic types. But as bfloat is a storage-only type a number of
operations shouldn't be set. This patch fixes that.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D85101
This patch adjusts the following ARM/AArch64 LLVM IR intrinsics:
- neon_bfmmla
- neon_bfmlalb
- neon_bfmlalt
so that they take and return bf16 and float types. Previously these
intrinsics used <8 x i8> and <4 x i8> vectors (a rudiment from
implementation lacking bf16 IR type).
The neon_vbfdot[q] intrinsics are adjusted similarly. This change
required some additional selection patterns for vbfdot itself and
also for vector shuffles (in a previous patch) because of SelectionDAG
transformations kicking in and mangling the original code.
This patch makes the generated IR cleaner (less useless bitcasts are
produced), but it does not affect the final assembly.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D86146
Original Commit Message:
After the commit r368987 (rG643adb55769e) was landed, the frame record (FP and LR register)
may be placed in the middle of a stack frame if a function has both callee-saved
general-purpose registers and floating point registers. This will break the stack unwinders
that simply walk through the frame records (based on the guarantee from AAPCS64
"The Frame Pointer" section). This commit fixes the problem by adding the frame record offset.
Patch By: logan
Differential Revision: D70800
This patch adds code to recognize vector shuffles which can be
represented as VDUP (splat) of a vector lane with of a different
(wider) type than the original vector lane type.
For example:
shufflevector <4 x i16> %v, <4 x i16> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
is essentially:
shufflevector <2 x i32> %v, <2 x i32> undef, <2 x i32> <i32 0, i32 0>
Such patterns are generated by the SelectionDAG machinery in some cases
(see DAGCombiner::visitBITCAST in DAGCombiner.cpp, the "Remove double
bitcasts from shuffles" part).
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D86225
AArch64, X86 and Mips currently directly consumes these and custom
lowering to produce a libcall, but really these should follow the
normal legalization process through the libcall/lower action.
Original Commit Message:
After the commit r368987 (rG643adb55769e) was landed, the frame record (FP and LR register)
may be placed in the middle of a stack frame if a function has both callee-saved
general-purpose registers and floating point registers. This will break the stack unwinders
that simply walk through the frame records (based on the guarantee from AAPCS64
"The Frame Pointer" section). This commit fixes the problem by adding the frame record offset.
Patch By: logan
The version of `st1d` that operates with vector plus immediate
addressing mode uses the alias `st1d { <Zn>.d }, <Pg>, [<Za>.d]` for
rendering `st1d { <Zn>.d }, <Pg>, [<Za>.d, #0]`. The disassembler was
generating `<Zn>.s` instead of `<Zn>.d>`.
Differential Revision: https://reviews.llvm.org/D86633
When floating point callee-saved registers were used, the frame pointer would
incorrectly point to the bottom of the CSR space (containing saved floating-point
registers), rather than to the frame record.
While all frame offsets were calculated consistently, resulting in working code,
this prevented stack walkers from being about to traverse the frame list.
The switch in AArch64Operand::print was changed in D45688 so the shift
can be printed after printing the register. This is implemented with
LLVM_FALLTHROUGH and was broken in D52485 when BTIHint was put between
the register and shift operands.
Reviewed By: ostannard
Differential Revision: https://reviews.llvm.org/D86535
This fixes an issue where the restore point of callee-saves in the
function epilogues was incorrectly calculated when the basic block
consisted of only a RET instruction. This caused dealloc instructions
to be inserted in between the block of callee-save restore instructions,
rather than before it.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D86099
Also updates isConstOrConstSplatFP to allow the mul(A,-1) -> neg(A)
transformation when -1 is expressed as an ISD::SPLAT_VECTOR.
Differential Revision: https://reviews.llvm.org/D86415
As disscussed in post-commit review starting with
https://reviews.llvm.org/D84108#2227365
while this appears to be mostly a win overall, especially code-size-wise,
this appears to shake //certain// code pattens in a way that is extremely
unfavorable for performance (+30% runtime regression)
on certain CPU's (i personally can't reproduce).
So until the behaviour is better understood, and a path forward is mapped,
let's back this out for now.
This reverts commit 1d51dc38d8.
There are no nxv16i8/nxv8i16 SDIV instructions, so these fixed width operations must be promoted to nxv4i32.
Differential Revision: https://reviews.llvm.org/D86114
The getSrcFromCopy helper nowadays return a MachineOperand pointer,
so talking about zero_reg was incorrect as it nowadays return
a nullptr when not finding a copy like instruction.
For scalable vector shifts the prediacte is typically all active,
which gets selected to an unpredicated shift by immediate. When
code generating for fixed length vectors the predicate is based
on the vector length and so additional patterns are required to
make use of SVE's predicated shift by immediate instructions.
Differential Revision: https://reviews.llvm.org/D86204
TargetRegisterInfo::getMinimalPhysRegClass() returns rtcGPR64RegClassID for X16
and X17, as it's the last matching class. This in turn gets passed to
AArch64RegisterBankInfo::getRegBankFromRegClass(), which hits an unreachable.
It seems sensible to handle this case, so copies from X16 and X17 work.
Copying from X17 is used in inline assembly in libunwind for pointer
authentication.
Differential Revision: https://reviews.llvm.org/D85720
If we have a mask, and a value x, where (x & mask) == x, we can drop the AND
and just use x.
This is about a 0.4% geomean code size improvement on CTMark at -O3 for AArch64.
In AArch64, this is most useful post-legalization. Patterns like this often
show up when legalizing s1s, which must be extended to larger types.
e.g.
```
%cmp:_(s32) = G_ICMP ...
%and:_(s32) = G_AND %cmp, 1
```
Since G_ICMP only produces a single bit, there's no reason to mask it with the
G_AND.
Differential Revision: https://reviews.llvm.org/D85463
We probably want to introduce pseudo-instructions at some point, like
we have for binary operations, but this seems okay for now.
One thing I'm not sure about is whether we should be doing this as a
DAGCombine instead of directly pattern-matching it. I don't see any big
downside to doing it this way, though.
Differential Revision: https://reviews.llvm.org/D85681
This isn't necessaary for ACLE, but could be useful in other situations.
And the change is simple.
Differential Revision: https://reviews.llvm.org/D85251
Right shift patterns will no longer incorrectly accept a shift
amount of zero. At the same time they will allow larger shift
amounts that are now saturated to their upper bound.
Patterns have been extended to enable immediate forms for shifts
taking an arbitrary predicate.
This patch also unifies the code path for immediate parsing so the
i64 based shifts are no longer treated specially.
Differential Revision: https://reviews.llvm.org/D86084
This cleans up copies that the legalizer or other combines leave around. They
can occasionally end up escaping as moves.
Differential Revision: https://reviews.llvm.org/D85964
This patch implements initial backend support for a -mtune CPU controlled by a "tune-cpu" function attribute. If the attribute is not present X86 will use the resolved CPU from target-cpu attribute or command line.
This patch adds MC layer support a tune CPU. Each CPU now has two sets of features stored in their GenSubtargetInfo.inc tables . These features lists are passed separately to the Processor and ProcessorModel classes in tablegen. The tune list defaults to an empty list to avoid changes to non-X86. This annoyingly increases the size of static tables on all target as we now store 24 more bytes per CPU. I haven't quantified the overall impact, but I can if we're concerned.
One new test is added to X86 to show a few tuning features with mismatched tune-cpu and target-cpu/target-feature attributes to demonstrate independent control. Another new test is added to demonstrate that the scheduler model follows the tune CPU.
I have not added a -mtune to llc/opt or MC layer command line yet. With no attributes we'll just use the -mcpu for both. MC layer tools will always follow the normal CPU for tuning.
Differential Revision: https://reviews.llvm.org/D85165
The code wasn't taking into account that the two operands
passed to ptest could be identical and was trying to erase
them twice.
Differential Revision: https://reviews.llvm.org/D85892
Simon Wallis