Previously we emitted an fmadd and a fmadd+fneg and combined them with a shufflevector. But this doesn't follow the correct exception behavior for unselected elements so the backend can't merge them into the fmaddsub/fmsubadd instructions.
This patch restores the the fmaddsub intrinsics so we don't have two arithmetic operations. We lose out on optimization opportunity in the non-strict FP case, but I don't think this is a big loss. If someone gives us a test case we can look into adding instcombine/dagcombine improvements. I'd rather not have the frontend do completely different things for strict and non-strict.
This still has problems because target specific intrinsics don't support strict semantics yet. We also still have all of the problems with masking. But we at least generate the right instruction in constrained mode now.
Differential Revision: https://reviews.llvm.org/D74268
This commit removes the artificial types <512 x i1> and <1024 x i1>
from HVX intrinsics, and makes v512i1 and v1024i1 no longer legal on
Hexagon.
It may cause existing bitcode files to become invalid.
* Converting between vector predicates and vector registers must be
done explicitly via vandvrt/vandqrt instructions (their intrinsics),
i.e. (for 64-byte mode):
%Q = call <64 x i1> @llvm.hexagon.V6.vandvrt(<16 x i32> %V, i32 -1)
%V = call <16 x i32> @llvm.hexagon.V6.vandqrt(<64 x i1> %Q, i32 -1)
The conversion intrinsics are:
declare <64 x i1> @llvm.hexagon.V6.vandvrt(<16 x i32>, i32)
declare <128 x i1> @llvm.hexagon.V6.vandvrt.128B(<32 x i32>, i32)
declare <16 x i32> @llvm.hexagon.V6.vandqrt(<64 x i1>, i32)
declare <32 x i32> @llvm.hexagon.V6.vandqrt.128B(<128 x i1>, i32)
They are all pure.
* Vector predicate values cannot be loaded/stored directly. This directly
reflects the architecture restriction. Loading and storing or vector
predicates must be done indirectly via vector registers and explicit
conversions via vandvrt/vandqrt instructions.
Summary:
These are in some sense the inverse of vmovl[bt]q: they take a vector
of n wide elements and truncate each to half its width. So they only
write half a vector's worth of output data, and therefore they also
take an 'inactive' parameter to provide the other half of the data in
the output vector. So vmovnb overwrites the even lanes of 'inactive'
with the narrowed values from the main input, and vmovnt overwrites
the odd lanes.
LLVM had existing codegen which generates these MVE instructions in
response to IR that takes two vectors of wide elements, or two vectors
of narrow ones. But in this case, we have one vector of each. So my
clang codegen strategy is to narrow the input vector of wide elements
by simply reinterpreting it as the output type, and then we have two
narrow vectors and can represent the operation as a vector shuffle
that interleaves lanes from both of them.
Even so, not all the cases I needed ended up being selected as a
single MVE instruction, so I've added a couple more patterns that spot
combinations of the 'MVEvmovn' and 'ARMvrev32' SDNodes which can be
generated as a VMOVN instruction with operands swapped.
This commit adds the unpredicated forms only.
Reviewers: dmgreen, miyuki, MarkMurrayARM, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D74337
Summary:
These intrinsics take a vector of 2n elements, and return a vector of
n wider elements obtained by sign- or zero-extending every other
element of the input vector. They're represented in IR as a
shufflevector that extracts the odd or even elements of the input,
followed by a sext or zext.
Existing LLVM codegen already matches this pattern and generates the
VMOVLB instruction (which widens the even-index input lanes). But no
existing isel rule was generating VMOVLT, so I've added some. However,
the new rules currently only work in little-endian MVE, because the
pattern they expect from isel lowering includes a bitconvert which
doesn't have the right semantics in big-endian.
The output of one existing codegen test is improved by those new
rules.
This commit adds the unpredicated forms only.
Reviewers: dmgreen, miyuki, MarkMurrayARM, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D74336
Summary:
These intrinsics just reorder the lanes of a vector, so the natural IR
representation is as a shufflevector operation. Existing LLVM codegen
already recognizes those particular shufflevectors and generates the
MVE VREV instruction.
This commit adds the unpredicated forms only.
Reviewers: dmgreen, miyuki, MarkMurrayARM, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D74334
Summary:
This commit adds the unpredicated intrinsics for the unary operations
vabsq (absolute value), vnegq (arithmetic negation), vmvnq (bitwise
complement), vqabsq and vqnegq (saturating versions of abs and neg for
signed integers, in the sense that they give INT_MAX if an input lane
is INT_MIN).
This is done entirely in clang: all of these operations have existing
isel patterns and existing tests for them on the LLVM side, so I've
just made clang emit the same IR that those patterns already match.
Reviewers: dmgreen, miyuki, MarkMurrayARM, ostannard
Reviewed By: MarkMurrayARM
Subscribers: kristof.beyls, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D74331
With REQUIRES: x86-register-target added to the tests.
Also remove some unneeded FIXMEs
But add a FIXME for bad IR generation for FMADDSUB/FMSUBADD with
constrained FP.
Original patch by Kevin P. Neal
This reverts commit 208470dd5d.
Tests fail:
error: unable to create target: 'No available targets are compatible with triple "x86_64-apple-darwin"'
This happens on clang-hexagon-elf, clang-cmake-armv7-quick, and
clang-cmake-armv7-quick bots.
If anyone has any suggestions on why then I'm all ears.
Differential Revision: https://reviews.llvm.org/D73570
Revert "[FPEnv][X86] Speculative fix for failures introduced by eda495426."
This reverts commit 80e17e5fcc.
The speculative fix didn't solve the test failures on Hexagon, ARMv6, and
MSVC AArch64.
When constrained floating point is enabled the X86-specific builtins don't
use constrained intrinsics in some cases. Fix that.
Differential Revision: https://reviews.llvm.org/D73570
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
The constrained fcmp intrinsics don't allow the TRUE/FALSE predicates.
Using them will assert. To workaround this I'm emitting the old X86 specific intrinsics that were never removed from the backend when we switched to using fcmp in IR. We have no way to mark them as being strict, but that's true of all target specific intrinsics so doesn't seem like we need to solve that here.
I've also added support for selecting between signaling and quiet.
Still need to support SAE which will require using a target specific
intrinsic. Also need to fix masking to not use an AND instruction
after the compare.
Differential Revision: https://reviews.llvm.org/D72906
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
Summary:
This is a follow up on https://reviews.llvm.org/D71473#inline-647262.
There's a caveat here that `Align(1)` relies on the compiler understanding of `Log2_64` implementation to produce good code. One could use `Align()` as a replacement but I believe it is less clear that the alignment is one in that case.
Reviewers: xbolva00, courbet, bollu
Subscribers: arsenm, dylanmckay, sdardis, nemanjai, jvesely, nhaehnle, hiraditya, kbarton, jrtc27, atanasyan, jsji, Jim, kerbowa, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D73099
When constrained floating point is enabled the SystemZ-specific builtins
don't use constrained intrinsics in some cases. Fix that.
Differential Revision: https://reviews.llvm.org/D72722
Summary:
This change implements the expansion in two parts:
- Add a utility function emitAMDGPUPrintfCall() in LLVM.
- Invoke the above function from Clang CodeGen, when processing a HIP
program for the AMDGPU target.
The printf expansion has undefined behaviour if the format string is
not a compile-time constant. As a sufficient condition, the HIP
ToolChain now emits -Werror=format-nonliteral.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D71365
This change introduces three new builtins (which work on both pointers
and integers) that can be used instead of common bitwise arithmetic:
__builtin_align_up(x, alignment), __builtin_align_down(x, alignment) and
__builtin_is_aligned(x, alignment).
I originally added these builtins to the CHERI fork of LLVM a few years ago
to handle the slightly different C semantics that we use for CHERI [1].
Until recently these builtins (or sequences of other builtins) were
required to generate correct code. I have since made changes to the default
C semantics so that they are no longer strictly necessary (but using them
does generate slightly more efficient code). However, based on our experience
using them in various projects over the past few years, I believe that adding
these builtins to clang would be useful.
These builtins have the following benefit over bit-manipulation and casts
via uintptr_t:
- The named builtins clearly convey the semantics of the operation. While
checking alignment using __builtin_is_aligned(x, 16) versus
((x & 15) == 0) is probably not a huge win in readably, I personally find
__builtin_align_up(x, N) a lot easier to read than (x+(N-1))&~(N-1).
- They preserve the type of the argument (including const qualifiers). When
using casts via uintptr_t, it is easy to cast to the wrong type or strip
qualifiers such as const.
- If the alignment argument is a constant value, clang can check that it is
a power-of-two and within the range of the type. Since the semantics of
these builtins is well defined compared to arbitrary bit-manipulation,
it is possible to add a UBSAN checker that the run-time value is a valid
power-of-two. I intend to add this as a follow-up to this change.
- The builtins avoids int-to-pointer casts both in C and LLVM IR.
In the future (i.e. once most optimizations handle it), we could use the new
llvm.ptrmask intrinsic to avoid the ptrtoint instruction that would normally
be generated.
- They can be used to round up/down to the next aligned value for both
integers and pointers without requiring two separate macros.
- In many projects the alignment operations are already wrapped in macros (e.g.
roundup2 and rounddown2 in FreeBSD), so by replacing the macro implementation
with a builtin call, we get improved diagnostics for many call-sites while
only having to change a few lines.
- Finally, the builtins also emit assume_aligned metadata when used on pointers.
This can improve code generation compared to the uintptr_t casts.
[1] In our CHERI compiler we have compilation mode where all pointers are
implemented as capabilities (essentially unforgeable 128-bit fat pointers).
In our original model, casts from uintptr_t (which is a 128-bit capability)
to an integer value returned the "offset" of the capability (i.e. the
difference between the virtual address and the base of the allocation).
This causes problems for cases such as checking the alignment: for example, the
expression `if ((uintptr_t)ptr & 63) == 0` is generally used to check if the
pointer is aligned to a multiple of 64 bytes. The problem with offsets is that
any pointer to the beginning of an allocation will have an offset of zero, so
this check always succeeds in that case (even if the address is not correctly
aligned). The same issues also exist when aligning up or down. Using the
alignment builtins ensures that the address is used instead of the offset. While
I have since changed the default C semantics to return the address instead of
the offset when casting, this offset compilation mode can still be used by
passing a command-line flag.
Reviewers: rsmith, aaron.ballman, theraven, fhahn, lebedev.ri, nlopes, aqjune
Reviewed By: aaron.ballman, lebedev.ri
Differential Revision: https://reviews.llvm.org/D71499
Summary:
A few of the ARM MVE builtins directly return a structure type. This
causes an assertion failure at code-gen time if you try to assign the
result of the builtin to a variable, because the `RValue` created in
`EmitBuiltinExpr` from the `llvm::Value` produced by codegen is always
made by `RValue::get()`, which creates a non-aggregate `RValue` that
will fail an assertion when `AggExprEmitter::withReturnValueSlot` calls
`Src.getAggregatePointer()`. A similar failure occurs if you try to use
the struct return value directly to extract one field, e.g.
`vld2q(address).val[0]`.
The existing code-gen tests for those MVE builtins pass the returned
structure type directly to the C `return` statement, which apparently
managed to avoid that particular code path, so we didn't notice the
crash.
Now `EmitBuiltinExpr` checks the evaluation kind of the builtin's return
value, and does the necessary handling for aggregate returns. I've added
two extra test cases, both of which crashed before this change.
Reviewers: dmgreen, rjmccall
Reviewed By: rjmccall
Subscribers: kristof.beyls, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D72271
We already recognize the __builtin versions of these, might as well
recognize the libcall version.
Differential Revision: https://reviews.llvm.org/D72028
This replaces the fsub -0.0 idiom with an fneg instruction. We didn't see to have a test that showed the current codegen. Just some tests for constant folding and a test that was only checking the declare lines for libcalls. The latter just checked that we did not have a declare for @conj when using __builtin_conj.
Differential Revision: https://reviews.llvm.org/D72012
This just updates an IRBuilder interface to take Functions instead of
Values so the type can be derived, and fixes some callsites in Clang to
call the updated API.
Summary:
The vector pattern `(a + b + 1) / 2` was previously selected to an
avgr_u instruction regardless of nuw flags, but this is incorrect in
the case where either addition may have an unsigned wrap. This CL
changes the existing pattern to require both adds to have nuw flags
and adds builtin functions and intrinsics for the avgr_u instructions
because the corrected pattern is not representable in C.
Reviewers: aheejin
Subscribers: dschuff, sbc100, jgravelle-google, hiraditya, sunfish, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71648
Summary:
The instructions were originally implemented via builtins and
intrinsics so users would have to explicitly opt-in to using
them. This was useful while were validating whether these instructions
should have been merged into the spec proposal. Now that they have
been, we can use normal codegen patterns, so the intrinsics and
builtins are no longer useful.
Reviewers: aheejin
Subscribers: dschuff, sbc100, jgravelle-google, hiraditya, sunfish, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71500
This has two main effects:
- Optimizes debug info size by saving 221.86 MB of obj file size in a
Windows optimized+debug build of 'all'. This is 3.03% of 7,332.7MB of
object file size.
- Incremental step towards decoupling target intrinsics.
The enums are still compact, so adding and removing a single
target-specific intrinsic will trigger a rebuild of all of LLVM.
Assigning distinct target id spaces is potential future work.
Part of PR34259
Reviewers: efriedma, echristo, MaskRay
Reviewed By: echristo, MaskRay
Differential Revision: https://reviews.llvm.org/D71320
This adds the family of `vshlq_n` and `vshrq_n` ACLE intrinsics, which
shift every lane of a vector left or right by a compile-time
immediate. They mostly work by expanding to the IR `shl`, `lshr` and
`ashr` operations, with their second operand being a vector splat of
the immediate.
There's a fiddly special case, though. ACLE specifies that the
immediate in `vshrq_n` can take values up to //and including// the bit
size of the vector lane. But LLVM IR thinks that shifting right by the
full size of the lane is UB, and feels free to replace the `lshr` with
an `undef` half way through the optimization pipeline. Hence, to keep
this legal in source code, I have to detect it at codegen time.
Logical (unsigned) right shifts by the element size are handled by
simply emitting the zero vector; arithmetic ones are converted into a
shift of one bit less, which will always give the same output.
In order to do that check, I also had to enhance the tablegen
MveEmitter so that it can cope with converting a builtin function's
operand into a bare integer to pass to a code-generating subfunction.
Previously the only bare integers it knew how to handle were flags
generated from within `arm_mve.td`.
Reviewers: dmgreen, miyuki, MarkMurrayARM, ostannard
Reviewed By: dmgreen, MarkMurrayARM
Subscribers: echristo, hokein, rdhindsa, kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71065
Change the IRBuilder and clang so that constrained FP intrinsics will be
emitted for builtins when appropriate. Only non-target-specific builtins
are affected in this patch.
Differential Revision: https://reviews.llvm.org/D70256
Summary:
This adds the family of `vshlq_n` and `vshrq_n` ACLE intrinsics, which
shift every lane of a vector left or right by a compile-time
immediate. They mostly work by expanding to the IR `shl`, `lshr` and
`ashr` operations, with their second operand being a vector splat of
the immediate.
There's a fiddly special case, though. ACLE specifies that the
immediate in `vshrq_n` can take values up to //and including// the bit
size of the vector lane. But LLVM IR thinks that shifting right by the
full size of the lane is UB, and feels free to replace the `lshr` with
an `undef` half way through the optimization pipeline. Hence, to keep
this legal in source code, I have to detect it at codegen time.
Logical (unsigned) right shifts by the element size are handled by
simply emitting the zero vector; arithmetic ones are converted into a
shift of one bit less, which will always give the same output.
In order to do that check, I also had to enhance the tablegen
MveEmitter so that it can cope with converting a builtin function's
operand into a bare integer to pass to a code-generating subfunction.
Previously the only bare integers it knew how to handle were flags
generated from within `arm_mve.td`.
Reviewers: dmgreen, miyuki, MarkMurrayARM, ostannard
Reviewed By: MarkMurrayARM
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71065
AggValueSlot
This reapplies 8a5b7c3570 after a null
dereference bug in CGOpenMPRuntime::emitUserDefinedMapper.
Original commit message:
This is needed for the pointer authentication work we plan to do in the
near future.
a63a81bd99/clang/docs/PointerAuthentication.rst
Summary:
Add support for vcadd_* family of intrinsics. This set of intrinsics is
available in Armv8.3-A.
The fp16 versions require the FP16 extension, which has been available
(opt-in) since Armv8.2-A.
Reviewers: t.p.northover
Reviewed By: t.p.northover
Subscribers: t.p.northover, kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D70862
This replaces the A32 NEON vqadds, vqaddu, vqsubs and vqsubu intrinsics
with the target independent sadd_sat, uadd_sat, ssub_sat and usub_sat.
This helps generate vqadds from standard IR nodes, which might be
produced from the vectoriser. The old variants are removed in the
process.
Differential Revision: https://reviews.llvm.org/D69350
Yaxun (Sam) Liu