Legal if we have hardware support for floating point, libcalls
otherwise.
Also add the necessary support for libcalls in the legalizer helper.
llvm-svn: 323726
Patch by: Bas Nieuwenhuizen
Just use the _e64 variant if needed. This should be possible as per
def : Pat <
(int_amdgcn_kill (i1 (setcc f32:$src, InlineFPImm<f32>:$imm, cond:$cond))),
(SI_KILL_F32_COND_IMM_PSEUDO $src, (bitcast_fpimm_to_i32 $imm), (cond_as_i32imm $cond))
> ;
I don't think we can get an immediate for the other operand for which we
need the second 32-bit word.
https://reviews.llvm.org/D42302
llvm-svn: 323706
We currently emit up to 15-byte NOPs on all targets (apart from Silvermont), which stalls performance on some targets with decoders that struggle with 2 or 3 more '66' prefixes.
This patch flags recent AMD targets (btver1/znver1) to still emit 15-byte NOPs and bdver* targets to emit 11-byte NOPs. All other targets now emit 10-byte NOPs apart from SilverMont CPUs which still emit 7-byte NOPS.
Differential Revision: https://reviews.llvm.org/D42616
llvm-svn: 323693
Summary:
Apparently, we missed on constraining register classes of VReg-operands of all the instructions
built from a destination pattern but the root (top-level) one. The issue exposed itself
while selecting G_FPTOSI for armv7: the corresponding pattern generates VTOSIZS wrapped
into COPY_TO_REGCLASS, so top-level COPY_TO_REGCLASS gets properly constrained,
while nested VTOSIZS (or rather its destination virtual register to be exact) does not.
Fixing this by issuing GIR_ConstrainSelectedInstOperands for every nested GIR_BuildMI.
https://bugs.llvm.org/show_bug.cgi?id=35965
rdar://problem/36886530
Patch by Roman Tereshin
Reviewers: dsanders, qcolombet, rovka, bogner, aditya_nandakumar, volkan
Reviewed By: dsanders, qcolombet, rovka
Subscribers: aemerson, javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D42565
llvm-svn: 323692
Summary:
It seems it's main effect is to create addition copies when values are inr register that do not support this trick, which increase register pressure and makes the code bigger.
The main noteworthy regression I was able to observe was pattern of the type (setcc (trunc (and X, C)), 0) where C is such as it would benefit from the hi register trick. To prevent this, a new pattern is added to materialize such pattern using a 32 bits test. This has the added benefit of working with any constant that is materializable as a 32bits immediate, not just the ones that can leverage the high register trick, as demonstrated by the test case in test-shrink.ll using the constant 2049 .
Reviewers: craig.topper, niravd, spatel, hfinkel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42646
llvm-svn: 323690
This reverts commit r322917 due to multiple performance regressions in spec2006
and spec2017. XFAILed llvm/test/CodeGen/AArch64/big-callframe.ll which initially
motivated this change.
llvm-svn: 323683
Summary:
As discussed in D42244, we have difficulty describing the legality of some
operations. We're not able to specify relationships between types.
For example, declaring the following
setAction({..., 0, s32}, Legal)
setAction({..., 0, s64}, Legal)
setAction({..., 1, s32}, Legal)
setAction({..., 1, s64}, Legal)
currently declares these type combinations as legal:
{s32, s32}
{s64, s32}
{s32, s64}
{s64, s64}
but we currently have no means to say that, for example, {s64, s32} is
not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/
G_UNMERGE_VALUES have relationships between the types that are currently
described incorrectly.
Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics
differently to atomics. The necessary information is in the MMO but we have no
way to use this in the legalizer. Similarly, there is currently no way for the
register type and the memory type to differ so there is no way to cleanly
represent extending-load/truncating-store in a way that can't be broken by
optimizers (resulting in illegal MIR).
It's also difficult to control the legalization strategy. We've added support
for legalizing non-power of 2 types but there's still some hardcoded assumptions
about the strategy. The main one I've noticed is that type0 is always legalized
before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if
you need to widen the container. It will converge on the same result eventually
but it will take a much longer route when legalizing type0 than if you legalize
type1 first.
Lastly, the definition of legality and the legalization strategy is kept
separate which is not ideal. It's helpful to be able to look at a one piece of
code and see both what is legal and the method the legalizer will use to make
illegal MIR more legal.
This patch adds a layer onto the LegalizerInfo (to be removed when all targets
have been migrated) which resolves all these issues.
Here are the rules for shift and division:
for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV})
getActionDefinitions(BinOp)
.legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal
.clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32
// If type0 is >s64 then NarrowScalar to s64
.widenScalarToPow2(0) // Round type0 scalars up to powers of 2
.unsupported(); // Otherwise, it's unsupported
This describes everything needed to both define legality and describe how to
make illegal things legal.
Here's an example of a complex rule:
getActionDefinitions(G_INSERT)
.unsupportedIf([=](const LegalityQuery &Query) {
// If type0 is smaller than type1 then it's unsupported
return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits();
})
.legalIf([=](const LegalityQuery &Query) {
// If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal
// We don't need to worry about large type1's because unsupportedIf caught that.
const LLT &Ty0 = Query.Types[0];
const LLT &Ty1 = Query.Types[1];
if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0)
return false;
return isPowerOf2_32(Ty1.getSizeInBits()) &&
(Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8);
})
.clampScalar(0, s32, s64)
.widenScalarToPow2(0)
.maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16
.maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32
.widenScalarToPow2(1) // Round type1 scalars up to powers of 2
.unsupported();
This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than
type1 (in practice, this would be a default rule for G_INSERT). It also uses one
to describe the legal cases. This particular predicate is equivalent to:
.legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}})
In terms of performance, I saw a slight (~6%) performance improvement when
AArch64 was around 30% ported but it's pretty much break even right now.
I'm going to take a look at constexpr as a means to reduce the initialization
cost.
Future work:
* Make it possible for opcodes to share rulesets. There's no need for
G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's
no technical barrier to this, it just hasn't been done yet.
* Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64)
* Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules.
* Improve initialization cost using constexpr
Possible future work:
* It's possible to make these rulesets change the MIR directly instead of
returning a description of how to change the MIR. This should remove a little
overhead caused by parsing the description and routing to the right code, but
the real motivation is that it removes the need for LegalizeAction::Custom.
With Custom removed, there's no longer a requirement that Custom legalization
change the opcode to something that's considered legal.
Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner
Reviewed By: bogner
Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D42251
llvm-svn: 323681
Summary:
Fix a few places that were modifying code after register
allocation to set the renamable bit correctly to avoid failing the
validation added in D42449.
llvm-svn: 323675
Summary:
The improvements to the LegalizerInfo discussed in D42244 require that
LegalizerInfo::LegalizeAction be available for use in other classes. As such,
it needs to be moved out of LegalizerInfo. This has been done separately to the
next patch to minimize the noise in that patch.
llvm-svn: 323669
Summary:
All variants of isLogicalImm[Not](32|64) can be combined into a single templated function, same for printLogicalImm(32|64).
By making it use a template instead, further SVE patches can use it for other data types as well (e.g. 8, 16 bits).
Reviewers: fhahn, rengolin, aadg, echristo, kristof.beyls, samparker
Reviewed By: samparker
Subscribers: aemerson, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D42294
llvm-svn: 323646
Summary: This was broken long ago in D12208, which failed to account for
the fact that 64-bit SPARC uses a stack bias of 2047, and it is the
*unbiased* value which should be aligned, not the biased one. This was
seen to be an issue with Rust.
Patch by: jrtc27 (James Clarke)
Reviewers: jyknight, venkatra
Reviewed By: jyknight
Subscribers: jacob_hansen, JDevlieghere, fhahn, fedor.sergeev, llvm-commits
Differential Revision: https://reviews.llvm.org/D39425
llvm-svn: 323643
Create and use FP16Pat FullFP16Pat helper patterns to make the difference
explicit.
Differential Revision: https://reviews.llvm.org/D42634
llvm-svn: 323640
The Large System Extension added an atomic compare-and-swap instruction
that operates on a pair of 64-bit registers, which we can use to
implement a 128-bit cmpxchg.
Because i128 is not a legal type for AArch64 we have to do all of the
instruction selection in C++, and the instruction requires even/odd
register pairs, so we have to wrap it in REG_SEQUENCE and EXTRACT_SUBREG
nodes. This is very similar to what we do for 64-bit cmpxchg in the ARM
backend.
Differential revision: https://reviews.llvm.org/D42104
llvm-svn: 323634
We can use the same input for both operands to get a free compare with zero.
We already use this trick in a couple places where we explicitly create PTESTM with the same input twice. This generalizes it.
I'm hoping to remove the ISD opcodes and move this to isel patterns like we do for scalar cmp/test.
llvm-svn: 323605
Legalization is still biased to turn LT compares in to GT by swapping operands to avoid needing extra isel patterns to commute.
I'm hoping to remove TESTM/TESTNM next and this should simplify that by making EQ/NE more similar.
llvm-svn: 323604
If broadcasting from another shuffle, attempt to simplify it.
We can probably generalize this a lot more (embedding in combineX86ShufflesRecursively), but BROADCAST is one of the more troublesome as it accepts inputs of different sizes to the result.
llvm-svn: 323602
Previously we had to materialize all 1s in a register using vpternlog or pcmpeq and then xor with that. By using vpternlog directly we can do it in one operation.
This is implemented using isel patterns, but we should maybe consider creating a generalized vpternlog combiner.
llvm-svn: 323572
A correctly aligned address may happen to be separated into a variable
part and a constant part, where the constant part does not match the
alignment needed in a load/store that uses this address. Such a constant
cannot be used as an immediate offset in an indexed instruction.
When lowering a global address, make sure that if there is an offset
folded into the global, the offset is valid for all uses in load/store
instructions.
llvm-svn: 323562
X86ISelLowering.cpp:34130:5: error: return type 'llvm::SDValue' must
match previous return type 'const llvm::SDValue' when lambda expression
has unspecified explicit return type
llvm-svn: 323557
Previously some targets printed their own message at the start of Select to indicate what they were selecting. For the targets that didn't, it means there was no print of the root node before any custom handling in the target executed. So if the target did something custom and never called SelectNodeCommon, no print would be made. For the targets that did print a message in Select, if they didn't custom handle a node SelectNodeCommon would reprint the root node before walking the isel table.
It seems better to just print the message before the call to Select so all targets behave the same. And then remove the root node printing from SelectNodeCommon and just leave a message that says we're starting the table search.
There were also some oddities in blank line behavior. Usually due to a \n after a call to SelectionDAGNode::dump which already inserted a new line.
llvm-svn: 323551
We currently coalesce v4i32 extracts from all 4 elements to 2 v2i64 extracts + shifts/sign-extends.
This seems to have been added back in the days when we tended to spill vectors and reload scalars, or ended up with repeated shuffles moving everything down to 0'th index. I don't think either of these are likely these days as we have better EXTRACT_VECTOR_ELT and VECTOR_SHUFFLE handling, and the existing code tends to make it very difficult for various vector and load combines.
Differential Revision: https://reviews.llvm.org/D42308
llvm-svn: 323541
- using qualified pointer addrspace in intrinsics class to avoid .f32 mangling
- changed too common atomic mangling to ds
- added missing intrinsics to AMDGPUTTIImpl::getTgtMemIntrinsic
Reviewed by: b-sumner
Differential Revision: https://reviews.llvm.org/D42383
llvm-svn: 323516
load instruction
The function `Thumb1InstrInfo::loadRegFromStackSlot` accepts only the `tGPR`
register class. The function serves to emit a `tLDRspi` instruction and
certainly any subset of the `tGPR` register class is a valid destination of the
load.
Differential revision: https://reviews.llvm.org/D42535
llvm-svn: 323514
This is the groundwork for Armv8.2-A FP16 code generation .
Clang passes and returns _Float16 values as floats, together with the required
bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318.
We will implement half-precision argument passing/returning lowering in the ARM
backend soon, but for now this means that this:
_Float16 sub(_Float16 a, _Float16 b) {
return a + b;
}
gets lowered to this:
define float @sub(float %a.coerce, float %b.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
<SNIP>
%add = fadd half %1, %3
<SNIP>
}
When FullFP16 is *not* supported, we don't make f16 a legal type, and we get
legalization for "free", i.e. nothing changes and everything works as before.
And also f16 argument passing/returning is handled.
When FullFP16 is supported, we do make f16 a legal type, and have 2 places that
we need to patch up: f16 argument passing and returning, which involves minor
tweaks to avoid unnecessary code generation for some bitcasts.
As a "demonstrator" that this works for the different FP16, FullFP16, softfp
modes, etc., I've added match rules to the VSUB instruction description showing
that we can codegen this instruction from IR, but more importantly, also to
some conversion instructions. These conversions were causing issue before in
the FP16 and FullFP16 cases.
I've also added match rules to the VLDRH and VSTRH desriptions, so that we can
actually compile the entire half-precision sub code example above. This showed
that these loads and stores had the wrong addressing mode specified: AddrMode5
instead of AddrMode5FP16, which turned out not be implemented at all, so that
has also been added.
This is the minimal patch that shows all the different moving parts. In patch
2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the
remaining Armv8.2-A FP16 instruction descriptions.
Thanks to Sam Parker and Oliver Stannard for their help and reviews!
Differential Revision: https://reviews.llvm.org/D38315
llvm-svn: 323512
Type legalization would prevent any i64 operands to the build_vector from existing before we get here. The coverage bots show this code as uncovered.
llvm-svn: 323506
The original autoupgrade for kunpck intrinsics used a bitcasted scalar shift, or, and. This combine would turn this into a concat_vectors. Now the kunpck intrinsics are autoupgraded to a vector shuffle that will become a concat_vectors.
llvm-svn: 323504
This listed all legal 128-bit integer types individually, but since we already know we have a legal type and its integer, we can just check is128BitVector.
llvm-svn: 323502
When pass creates a MOV instruction for
lea (%base,%index,1), %dst => mov %base,%dst; add %index,%dst
modification it should clean the killed flag for base
if base is equal to index.
Otherwise verifier complains about usage of killed register in add instruction.
Reviewers: lsaba, zvi, zansari, aaboud
Reviewed By: lsaba
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42522
llvm-svn: 323497
This patch enables aggressive FMA by default on T99, and provides a -mllvm
option to enable the same on other AArch64 micro-arch's (-mllvm
-aarch64-enable-aggressive-fma).
Test case demonstrating the effects on T99 is included.
Patch by: steleman (Stefan Teleman)
Differential Revision: https://reviews.llvm.org/D40696
llvm-svn: 323474
The asm parser puts the lock prefix in the MCInst flags so we need to check that in addition to TSFlags. This matches what the ATT printer does.
llvm-svn: 323469
The regular expressions and the imul names caused some instructions to be matched by multiple regexs creating unpredictable results.
This changes them all to use explicit instrs instead.
While doing this I also found that some instructions in Skylake were missing load latency so I fixed that too.
llvm-svn: 323406
The IMUL instruction names mixed with the prefix matching of the instregex lead to some strange matches. The worst being that several memory instructions are using the register form latency.
I don't know what the right answer is, so I've left TODOs and will try to work with the AMD folks to get this cleaned up.
llvm-svn: 323405
MMX instrutions all start with MMX_ so the 64 isn't needed for disambigutation.
SSE/AVX1 instructions are assumed 128-bit so we don't need to say 128.
AVX2 instructions should use a Y to indicate 256-bits.
llvm-svn: 323402
These were treated as optional suffixes, but the regular expressions are already prefix matches so this is unnecessary. It breaks the binary search optimization in tablegen due to the top level question mark.
llvm-svn: 323401
The code in EmitFunctionEntryCode needs to know the maximum stack
alignment, but it runs very early in the selection process (before
lowering). The final stack alignment may change during lowering, so
the code needs to be moved to where the alignment is known.
llvm-svn: 323374
The tablegen imported patterns for sext(load(a)) don't check for single uses
of the load or delete the original after matching. As a result two loads are
left in the generated code. This particular issue will be fixed by adding
support for a G_SEXTLOAD opcode in future.
There are however other potential issues around this that wouldn't be fixed by
a G_SEXTLOAD, so until we have a proper solution we don't try to handle volatile
loads at all in the AArch64 selector.
Fixes/works around PR36018.
llvm-svn: 323371
As discussed in D41484, PMADDWD for 'zero extended' vXi32 is nearly always a better option than PMULLD:
On SNB it will result in code that isn't any faster, but not any slower so we may as well keep it.
On KNL it only has half the throughput, so I've disabled it on there - ideally there'd be a better way than this.
Differential Revision: https://reviews.llvm.org/D42258
llvm-svn: 323367
Summary:
Move reserveRegisterTuples into AMDGPURegisterInfo and use it in
R600RegisterInfo::getReservedRegs and
R600InstrInfo::reserveIndirectRegisters to ensure that all super
registers of reserved registers are also marked as reserved.
Before this change, under certain circumstances, the registers %t1_x and
%t1_xyzw would be marked as reserved, but %t1_xy and %t1_xyz would not
be, leading to the register allocator sometimes assigning a register to
%t1_xy, which is invalid since %t1_x is reserved.
Reviewers: arsenm, tstellar, MatzeB, qcolombet
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D42448
llvm-svn: 323356
Summary: For long shifts, the inlined version takes about 20 instructions on Thumb1. To avoid the code bloat, expand to __aeabi_ calls if target is Thumb1.
Reviewers: samparker
Reviewed By: samparker
Subscribers: samparker, aemerson, javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D42401
llvm-svn: 323354
The regexs are treated as a prefix match already so the checking for optional text at the end provides no value. Instead it prevents the binary search optimization in tablegen from kicking in due to the top level question mark.
llvm-svn: 323351
Summary:
Loads/stores of some NEON vector types are promoted to other vector
types with different lane sizes but same vector size. This is not a
problem in little-endian but, when in big-endian, it requires
additional byte reversals required to preserve the lane ordering
while keeping the right endianness of the data inside each lane.
For example:
%1 = load <4 x half>, <4 x half>* %p
results in the following assembly:
ld1 { v0.2s }, [x1]
rev32 v0.4h, v0.4h
This patch changes the promotion of these loads/stores so that the
actual vector load/store (LD1/ST1) takes care of the endianness
correctly and there is no need for further byte reversals. The
previous code now results in the following assembly:
ld1 { v0.4h }, [x1]
Reviewers: olista01, SjoerdMeijer, efriedma
Reviewed By: efriedma
Subscribers: aemerson, rengolin, javed.absar, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D42235
llvm-svn: 323325
This matches what MSVC does for alloca() function calls on ARM.
Even if MSVC doesn't support VLAs at the language level, it does
support the alloca function.
On the clang level, both the _alloca() (when emulating MSVC, which is
what the alloca() function expands to) and __builtin_alloca() builtin
functions, and VLAs, map to the same LLVM IR "alloca" function - so
within LLVM they're not distinguishable from each other.
Differential Revision: https://reviews.llvm.org/D42292
llvm-svn: 323308
I think these instructions used to be named differently and the regular expression reflected that. I guess we must have correct itinerary information that made this not matter for the scheduler test?
llvm-svn: 323305
There are a couple tricky things with this patch.
I had to add an override of isVectorLoadExtDesirable to stop DAG combine from combining sign_extend with loads after legalization since we legalize sextload using a load+sign_extend. Overriding this hook actually prevents a lot sextloads from being created in the first place.
I also had to add isel patterns because DAG combine blindly combines sign_extend+truncate to a smaller sign_extend which defeats what legalization was trying to do.
Differential Revision: https://reviews.llvm.org/D42407
llvm-svn: 323301
This is already a simplification, and should help with avoiding a plt
reference when calling an intrinsic with -fno-plt.
With this change we return false for null GVs, so the caller only
needs to check the new metadata to decide if it should use foo@plt or
*foo@got.
llvm-svn: 323297
Remove FeatureSlowMisaligned128Store from cyclone flags.
This flag causes splitting of 16 byte wide stores into 2 stored of 8
bytes. This was useful on older apple CPUs which were slow for 16byte
stores that were not aligned on 16byte. As the compiler often cannot
predict the actual alignment, the splitting was choosen.
This has been a topic for a lot of debate as the splitting also
decreases performance for some benchmarks. Measuring the effects on
newer apple chips (rdar://35525421) shows that it harms more cases than
it helps. So it is time to retire this workaround.
llvm-svn: 323289
Summary:
Fix an issue that's similar to what D41411 fixed:
float(__int128(float_var)) shouldn't be optimized to xscvdpsxds +
xscvsxdsp, as they mean (float)(int64_t)float_var.
Reviewers: jtony, hfinkel, echristo
Subscribers: sanjoy, nemanjai, hiraditya, llvm-commits, kbarton
Differential Revision: https://reviews.llvm.org/D42400
llvm-svn: 323270
I don't know if the unused classes were intended to be used and that the VEX version is really different than the legacy SSE version. Agner's tables don't show any differences. I'm just cleaning up assuming the current behavior is correct.
llvm-svn: 323263
Eric Liu