This transform was added with rL351346, and we had
an escape for shufps, but we also want one for
unpckps vs. vpermps because vpermps doesn't take
an immediate shuffle index operand.
llvm-svn: 352333
While i have no intention of actually commiting regeneration
of the check lines in these test files with update_llc_test_checks,
lack of that whitespace breaks that util, which is mildly inconvenient.
llvm-svn: 352318
This fixes loads like 's1 = load %p (load 1 from %p)' being combined with an
extend into an illegal 's8 = g_extload %p (load 1 from %p)' which doesn't do any
extension, by avoiding touching those < s8 size loads.
This bug was uncovered by a verifier update r351584, which I reverted it to keep
the bots green.
llvm-svn: 352311
The add+and sequence followed by a branch can
happen e.g. when looping over the set bits of an integer:
```
while (x != 0) {
func(x & ~x);
x &= x - 1;
}
```
Reviewed By: ctopper
Differential Revision: https://reviews.llvm.org/D57296
llvm-svn: 352306
def32 here means the producing instruction zeroed bits 63:32. We already do this for zext, but it looks like we can get an and+anyext sometimes.
Spotted in the diffs from D33587.
llvm-svn: 352303
As discussed on PR24545, we should try to commute X86::COND_A 'icmp ugt' cases to X86::COND_B 'icmp ult' to more optimally bind the carry flag output to a SBB instruction.
Differential Revision: https://reviews.llvm.org/D57281
llvm-svn: 352289
We often generate X86ISD::SBB(X, 0) for carry flag arithmetic.
I had tried to create test cases for the ADC equivalent (which often uses the same pattern) but haven't managed to find anything yet.
Differential Revision: https://reviews.llvm.org/D57169
llvm-svn: 352288
The IR enforced limit for the address space is 24-bits, but LLT was
only using 23-bits. Additionally, the argument to the constructor was
truncating to 16-bits.
A similar problem still exists for the number of vector elements. The
IR enforces no limit, so if you try to use a vector with > 65535
elements the IRTranslator asserts in the LLT constructor.
llvm-svn: 352264
Summary: We have isel patterns for this, but we're missing some load patterns and all broadcast patterns. A DAG combine seems like a better fit for this.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D56971
llvm-svn: 352260
These intrinsics may return different values every time they are called
and should not be CSE'd. IntrInaccessibleMemOnly appears to be the right
attribute to model this behavior.
Differential Revision: https://reviews.llvm.org/D57259
llvm-svn: 352256
Summary:
I'm not sure why we were using SEXTLOAD. EXTLOAD seems more appropriate since we don't care about the upper bits.
This patch changes this and then modifies the X86 post legalization combine to emit a extending shuffle instead of a sign_extend_vector_inreg. Could maybe use an any_extend_vector_inreg, but I just did what we already do in LowerLoad. I think we can actually get rid of this code entirely if we switch to -x86-experimental-vector-widening-legalization.
On AVX512 targets I think we might be able to use a masked vpmovzx and not have to expand this at all.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D57186
llvm-svn: 352255
DAGCombiner::visitBITCAST will perform:
fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
As shown in double-bitmanip-dagcombines.ll, this can be advantageous. But
RV32FD doesn't use bitcast directly (as i64 isn't a legal type), and instead
uses RISCVISD::SplitF64. This patch adds an equivalent DAG combine for
SplitF64.
llvm-svn: 352247
Summary:
Currently, if an instruction with a memory operand has no debug information,
X86DiscriminateMemOps will generate one based on the first line of the
enclosing function, or the last seen debug info.
This may cause confusion in certain debugging scenarios. The long term
approach would be to use the line number '0' in such cases, however, that
brings in challenges: the base discriminator value range is limited
(4096 values).
For the short term, adding an opt-in flag for this feature.
See bug 40319 (https://bugs.llvm.org/show_bug.cgi?id=40319)
Reviewers: dblaikie, jmorse, gbedwell
Reviewed By: dblaikie
Subscribers: aprantl, eraman, hiraditya
Differential Revision: https://reviews.llvm.org/D57257
llvm-svn: 352246
(fcopysign a, (fneg b)) will be expanded to bitwise operations by
DAGTypeLegalizer::SoftenFloatRes_FCOPYSIGN if the floating point type isn't
legal. Arguably it might be worth doing a combine even if it is legal.
llvm-svn: 352240
If bottom of block BB has only one successor OldTop, in most cases it is profitable to move it before OldTop, except the following case:
-->OldTop<-
| . |
| . |
| . |
---Pred |
| |
BB-----
Move BB before OldTop can't reduce the number of taken branches, this patch detects this case and prevent the moving.
Differential Revision: https://reviews.llvm.org/D57067
llvm-svn: 352236
We also need to combine to masked truncating with saturation stores, but I'm leaving that for a future patch.
This does regress some tests that used truncate wtih saturation followed by a masked store. Those now use a truncating store and use min/max to saturate.
Differential Revision: https://reviews.llvm.org/D57218
llvm-svn: 352230
The intrinsic names erroneously used the .f32 variant. As the return and
argument types were still double the intrinsics calls worked properly.
llvm-svn: 352211
Simplify to the generic ISD::ADD/SUB if we don't make use of the result flag.
This mainly helps with ADDCARRY/SUBBORROW intrinsics which get expanded to X86ISD::ADD/SUB but could be simplified further.
Noticed in some of the test cases in PR31754
Differential Revision: https://reviews.llvm.org/D57234
llvm-svn: 352210
This isn't the final fix for our reduction/horizontal codegen, but it takes care
of a lot of the problems. After we narrow the shuffle, existing combines for
insert/extract and binops kick in, and we end up with cheaper 128-bit ops.
The avg and mul reduction tests show an existing shuffle lowering hole for
AVX2/AVX512. I think in its most minimal form this is:
https://bugs.llvm.org/show_bug.cgi?id=40434
...but we might need multiple fixes to get it right.
Differential Revision: https://reviews.llvm.org/D57156
llvm-svn: 352209
This target-independent code won't trigger for cases such as RV32FD where
custom SelectionDAG nodes are generated. These new tests demonstrate such
cases. Additionally, float-arith.ll was updated so that fneg.s, fsgnjn.s, and
fabs.s selection patterns are actually exercised.
llvm-svn: 352199
Same as ARM.
On this occasion we split some of the instruction select tests for more
complicated instructions into their own files, so we can reuse them for
ARM and Thumb mode. Likewise for the legalizer tests.
llvm-svn: 352188
Fast selection of llvm icmp and fcmp instructions is not handled well about VSX instruction support.
We'd use VSX float comparison instruction instead of non-vsx float comparison instruction
if the operand register class is VSSRC or VSFRC because i32 and i64 are mapped to VSSRC and
VSFRC correspondingly if VSX feature is opened.
If the target does not have corresponding VSX instruction comparison for some type,
just copy VSX-related register to common float register class and use non-vsx comparison instruction.
Differential Revision: https://reviews.llvm.org/D57078
llvm-svn: 352174
Follow the same custom legalisation strategy as used in D57085 for
variable-length shifts (see that patch summary for more discussion). Although
we may lose out on some late-stage DAG combines, I think this custom
legalisation strategy is ultimately easier to reason about.
There are some codegen changes in rv64m-exhaustive-w-insts.ll but they are all
neutral in terms of the number of instructions.
Differential Revision: https://reviews.llvm.org/D57096
llvm-svn: 352171
The previous DAG combiner-based approach had an issue with infinite loops
between the target-dependent and target-independent combiner logic (see
PR40333). Although this was worked around in rL351806, the combiner-based
approach is still potentially brittle and can fail to select the 32-bit shift
variant when profitable to do so, as demonstrated in the pr40333.ll test case.
This patch instead introduces target-specific SelectionDAG nodes for
SHLW/SRLW/SRAW and custom-lowers variable i32 shifts to them. pr40333.ll is a
good example of how this approach can improve codegen.
This adds DAG combine that does SimplifyDemandedBits on the operands (only
lower 32-bits of first operand and lower 5 bits of second operand are read).
This seems better than implementing SimplifyDemandedBitsForTargetNode as there
is no guarantee that would be called (and it's not for e.g. the anyext return
test cases). Also implements ComputeNumSignBitsForTargetNode.
There are codegen changes in atomic-rmw.ll and atomic-cmpxchg.ll but the new
instruction sequences are semantically equivalent.
Differential Revision: https://reviews.llvm.org/D57085
llvm-svn: 352169
This patch exploits the instructions that store a single element from a vector
to preform a (store (extract_elt)). We already have code that does this with
ISA 3.0 instructions that were added to handle i8/i16 types. However, we had
never exploited the existing ones that handle f32/f64/i32/i64 types.
Differential revision: https://reviews.llvm.org/D56175
llvm-svn: 352131
https://reviews.llvm.org/D57178
Now add a hook in TargetPassConfig to query if CSE needs to be
enabled. By default this hook returns false only for O0 opt level but
this can be overridden by the target.
As a consequence of the default of enabled for non O0, a few tests
needed to be updated to not use CSE (by passing in -O0) to the run
line.
reviewed by: arsenm
llvm-svn: 352126
This patch adds support for vector @llvm.ceil intrinsics when full 16 bit
floating point support isn't available.
To do this, this patch...
- Implements basic isel for G_UNMERGE_VALUES
- Teaches the legalizer about 16 bit floats
- Teaches AArch64RegisterBankInfo to respect floating point registers on
G_BUILD_VECTOR and G_UNMERGE_VALUES
- Teaches selectCopy about 16-bit floating point vectors
It also adds
- A legalizer test for the 16-bit vector ceil which verifies that we create a
G_UNMERGE_VALUES and G_BUILD_VECTOR when full fp16 isn't supported
- An instruction selection test which makes sure we lower to G_FCEIL when
full fp16 is supported
- A test for selecting G_UNMERGE_VALUES
And also updates arm64-vfloatintrinsics.ll to show that the new ceiling types
work as expected.
https://reviews.llvm.org/D56682
llvm-svn: 352113
It should be emitted when any floating-point operations (including
calls) are present in the object, not just when calls to printf/scanf
with floating point args are made.
The difference caused by this is very subtle: in static (/MT) builds,
on x86-32, in a program that uses floating point but doesn't print it,
the default x87 rounding mode may not be set properly upon
initialization.
This commit also removes the walk of the types pointed to by pointer
arguments in calls. (To assist in opaque pointer types migration --
eventually the pointee type won't be available.)
That latter implies that it will no longer consider a call like
`scanf("%f", &floatvar)` as sufficient to emit _fltused on its
own. And without _fltused, `scanf("%f")` will abort with error R6002. This
new behavior is unlikely to bite anyone in practice (you'd have to
read a float, and do nothing with it!), and also, is consistent with
MSVC.
Differential Revision: https://reviews.llvm.org/D56548
llvm-svn: 352076
Select zero extending and sign extending load for MIPS32.
Use size from MachineMemOperand to determine number of bytes to load.
Differential Revision: https://reviews.llvm.org/D57099
llvm-svn: 352038
Use CombinerHelper to combine extending load instructions.
G_LOAD combined with G_ZEXT, G_SEXT or G_ANYEXT gives G_ZEXTLOAD,
G_SEXTLOAD or G_LOAD with same type as def of extending instruction
respectively.
Similarly G_ZEXTLOAD combined with G_ZEXT gives G_ZEXTLOAD and
G_SEXTLOAD combined with G_SEXT gives G_SEXTLOAD with same type
as def of extending instruction.
Differential Revision: https://reviews.llvm.org/D56914
llvm-svn: 352037
After creating new PHI instructions during isel pseudo expansion, the NoPHIs
property of MF should be reset in case it was previously set.
Review: Ulrich Weigand
llvm-svn: 352030
This patch adds a new ReadAdvance definition named ReadInt2Fpu.
ReadInt2Fpu allows x86 scheduling models to accurately describe delays caused by
data transfers from the integer unit to the floating point unit.
ReadInt2Fpu currently defaults to a delay of zero cycles (i.e. no delay) for all
x86 models excluding BtVer2. That means, this patch is only a functional change
for the Jaguar cpu model only.
Tablegen definitions for instructions (V)PINSR* have been updated to account for
the new ReadInt2Fpu. That read is mapped to the the GPR input operand.
On Jaguar, int-to-fpu transfers are modeled as a +6cy delay. Before this patch,
that extra delay was added to the opcode latency. In practice, the insert opcode
only executes for 1cy. Most of the actual latency is actually contributed by the
so-called operand-latency. According to the AMD SOG for family 16h, (V)PINSR*
latency is defined by expression f+1, where f is defined as a forwarding delay
from the integer unit to the fpu.
When printing instruction latency from MCA (see InstructionInfoView.cpp) and LLC
(only when flag -print-schedule is speified), we now need to account for any
extra forwarding delays. We do this by checking if scheduling classes declare
any negative ReadAdvance entries. Quoting a code comment in TargetSchedule.td:
"A negative advance effectively increases latency, which may be used for
cross-domain stalls". When computing the instruction latency for the purpose of
our scheduling tests, we now add any extra delay to the formula. This avoids
regressing existing codegen and mca schedule tests. It comes with the cost of an
extra (but very simple) hook in MCSchedModel.
Differential Revision: https://reviews.llvm.org/D57056
llvm-svn: 351965
Summary:
With XNACK, an smem load whose result is coalesced with an operand (thus
it overwrites its own operand) cannot appear in a clause, because some
other instruction might XNACK and restart the whole clause.
The clause breaker already realized that an smem that overwrites an
operand cannot appear in a clause, and broke the clause. The problem
that this commit fixes is that the SIFormMemoryClauses optimization
formed a bundle with early clobber, which caused the earlier code that
set up the coalesced operand to be removed as dead.
Differential Revision: https://reviews.llvm.org/D57008
Change-Id: I703c4d5b0bf7d6060222bec491f45c18bb3c0016
llvm-svn: 351950
Currently in Arm code, we allocate LR first, under the assumption that
it needs to be saved anyway. Unfortunately this has the disadvantage
that it will require any instructions using it to be the longer thumb2
instructions, not the shorter thumb1 ones.
This switches the order when we are optimising for minsize, returning to
the default order so that more lower registers can be used. It can end
up requiring more pushed registers, but on average produces smaller code.
Differential Revision: https://reviews.llvm.org/D56008
llvm-svn: 351938
In the last stage of type promotion, we replace any zext that uses a
new trunc with the operand of the trunc. This is okay when we only
allowed one type to be optimised, but now its the case that the trunc
maybe needed to produce a more narrow type than the one we were
optimising for. So we need to check this before doing the replacement.
Differential Revision: https://reviews.llvm.org/D57041
llvm-svn: 351935
The current check in CombineToPreIndexedLoadStore is too
conversative, preventing a pre-indexed store when the base pointer
is a predecessor of the value being stored. Instead, we should check
the pointer operand of the store.
Differential Revision: https://reviews.llvm.org/D56719
llvm-svn: 351933
As part of speculation hardening, the stack pointer gets masked with the
taint register (X16) before a function call or before a function return.
Since there are no instructions that can directly mask writing to the
stack pointer, the stack pointer must first be transferred to another
register, where it can be masked, before that value is transferred back
to the stack pointer.
Before, that temporary register was always picked to be x17, since the
ABI allows clobbering x17 on any function call, resulting in the
following instruction pattern being inserted before function calls and
returns/tail calls:
mov x17, sp
and x17, x17, x16
mov sp, x17
However, x17 can be live in those locations, for example when the call
is an indirect call, using x17 as the target address (blr x17).
To fix this, this patch looks for an available register just before the
call or terminator instruction and uses that.
In the rare case when no register turns out to be available (this
situation is only encountered twice across the whole test-suite), just
insert a full speculation barrier at the start of the basic block where
this occurs.
Differential Revision: https://reviews.llvm.org/D56717
llvm-svn: 351930
Two backend optimizations failed to handle cases when compiled with -g, due
to failing to consider DBG_VALUE instructions. This was in
SystemZTargetLowering::emitSelect() and
SystemZElimCompare::getRegReferences().
This patch makes sure that DBG_VALUEs are recognized so that they do not
affect these optimizations.
Tests for branch-on-count, load-and-trap and consecutive selects.
Review: Ulrich Weigand
https://reviews.llvm.org/D57048
llvm-svn: 351928
#pragma clang loop pipeline(disable)
Disable SWP optimization for the next loop.
“disable” is the only possible value.
#pragma clang loop pipeline_initiation_interval(number)
Set value of initiation interval for SWP
optimization to specified number value for
the next loop. Number is the positive value
greater than 0.
These pragmas could be used for debugging or reducing
compile time purposes. It is possible to disable SWP for
concrete loops to save compilation time or to find bugs
by not doing SWP to certain loops. It is possible to set
value of initiation interval to concrete number to save
compilation time by not doing extra pipeliner passes or
to check created schedule for specific initiation interval.
That is llvm part of the fix
Clang part of fix: https://reviews.llvm.org/D55710
Patch by Alexey Lapshin!
Differential Revision: https://reviews.llvm.org/D56403
llvm-svn: 351923
Each hwasan check requires emitting a small piece of code like this:
https://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html#memory-accesses
The problem with this is that these code blocks typically bloat code
size significantly.
An obvious solution is to outline these blocks of code. In fact, this
has already been implemented under the -hwasan-instrument-with-calls
flag. However, as currently implemented this has a number of problems:
- The functions use the same calling convention as regular C functions.
This means that the backend must spill all temporary registers as
required by the platform's C calling convention, even though the
check only needs two registers on the hot path.
- The functions take the address to be checked in a fixed register,
which increases register pressure.
Both of these factors can diminish the code size effect and increase
the performance hit of -hwasan-instrument-with-calls.
The solution that this patch implements is to involve the aarch64
backend in outlining the checks. An intrinsic and pseudo-instruction
are created to represent a hwasan check. The pseudo-instruction
is register allocated like any other instruction, and we allow the
register allocator to select almost any register for the address to
check. A particular combination of (register selection, type of check)
triggers the creation in the backend of a function to handle the check
for specifically that pair. The resulting functions are deduplicated by
the linker. The pseudo-instruction (really the function) is specified
to preserve all registers except for the registers that the AAPCS
specifies may be clobbered by a call.
To measure the code size and performance effect of this change, I
took a number of measurements using Chromium for Android on aarch64,
comparing a browser with inlined checks (the baseline) against a
browser with outlined checks.
Code size: Size of .text decreases from 243897420 to 171619972 bytes,
or a 30% decrease.
Performance: Using Chromium's blink_perf.layout microbenchmarks I
measured a median performance regression of 6.24%.
The fact that a perf/size tradeoff is evident here suggests that
we might want to make the new behaviour conditional on -Os/-Oz.
But for now I've enabled it unconditionally, my reasoning being that
hwasan users typically expect a relatively large perf hit, and ~6%
isn't really adding much. We may want to revisit this decision in
the future, though.
I also tried experimenting with varying the number of registers
selectable by the hwasan check pseudo-instruction (which would result
in fewer variants being created), on the hypothesis that creating
fewer variants of the function would expose another perf/size tradeoff
by reducing icache pressure from the check functions at the cost of
register pressure. Although I did observe a code size increase with
fewer registers, I did not observe a strong correlation between the
number of registers and the performance of the resulting browser on the
microbenchmarks, so I conclude that we might as well use ~all registers
to get the maximum code size improvement. My results are below:
Regs | .text size | Perf hit
-----+------------+---------
~all | 171619972 | 6.24%
16 | 171765192 | 7.03%
8 | 172917788 | 5.82%
4 | 177054016 | 6.89%
Differential Revision: https://reviews.llvm.org/D56954
llvm-svn: 351920
It might be a bit nicer to use the fancy .legalIf and co. predicates,
but this was requiring more boilerplate and disables the coverage
assertions.
llvm-svn: 351886
For AMDGPU the shift amount is never 64-bit, and
this needs to use a 32-bit shift.
X86 uses i8, but seemed to be hacking around this before.
llvm-svn: 351882
The existing test unintentionally shows that we have prematurely
optimized the shuffle into a vector concat and lost the undef info,
so it is not affected by a basic improvement to
SimplifyDemandedVectorElts.
llvm-svn: 351834
vecbo (insertsubv undef, X, Z), (insertsubv undef, Y, Z) --> insertsubv VecC, (vecbo X, Y), Z
This is another step in generic vector narrowing. It's also a step towards more horizontal op
formation specifically for x86 (although we still failed to match those in the affected tests).
The scalarization cases are also not optimal (we should be scalarizing those), but it's still
an improvement to use a narrower vector op when we know part of the result must be constant
because both inputs are undef in some vector lanes.
I think a similar match but checking for a constant operand might help some of the cases in
D51553.
Differential Revision: https://reviews.llvm.org/D56875
llvm-svn: 351825
Amara Emerson