Add a new llvm.amdgcn.ballot intrinsic modeled on the ballot function
in GLSL and other shader languages. It returns a bitfield containing the
result of its boolean argument in all active lanes, and zero in all
inactive lanes.
This is intended to replace the existing llvm.amdgcn.icmp and
llvm.amdgcn.fcmp intrinsics after a suitable transition period.
Use the new intrinsic in the atomic optimizer pass.
Differential Revision: https://reviews.llvm.org/D65088
Two one-use checks were added with rGfdcb27105537,
but only the first one is necessary to limit an
increase in instruction count. The second transform
only creates one instruction, so it is always a
reasonable canonicalization/optimization.
The sll/srl/sra scalar vector shifts can be replaced with generic shifts if the shift amount is known to be in range.
This also required public DemandedElts variants of llvm::computeKnownBits to be exposed (PR36319).
If a call argument has the "returned" attribute, we can simplify
the call to the value of that argument. This was already partially
handled by InstSimplify/InstCombine for the case where the argument
is an integer constant, and the result is thus known via known bits.
The non-constant (or non-int) argument cases weren't handled though.
This previously landed as an InstSimplify transform, but was reverted
due to assertion failures when compiling the Linux kernel. The reason
is that simplifying a call to another call breaks assumptions in
call graph updating during inlining. As the code is not easy to fix,
and there is no particularly strong motivation for having this in
InstSimplify, the transform is only performed in InstCombine instead.
Differential Revision: https://reviews.llvm.org/D75815
If we know the SSE shift amount is out of range then we can simplify to zero value (logical) or a 'signsplat' bitwidth-1 shift (arithmetic). This allows us to remove the equivalent ConstantInt constant folding path from simplifyX86immShift.
The slli/srli/srai 'immediate' vector shifts (although its not immediate anymore to match gcc) can be replaced with generic shifts if the shift amount is known to be in range.
Summary: Prevent InstCombine from removing llvm.assume for which the arguement is true when they have operand bundles with usefull information.
Reviewers: jdoerfert, nikic, lebedev.ri
Reviewed By: jdoerfert
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76147
According to LangRef:
If len is not a positive integer multiple of element_size, then the behaviour of the intrinsic is undefined.
Add InstCombine rule to transform intrinsic to undef operation.
This is a follow-up for D76116.
Reviewers: reames
Reviewed By: reames
Subscribers: hiraditya, jfb, dantrushin, llvm-commits
Differential Revision: https://reviews.llvm.org/D76215
When simplifying a call without uses, replaceInstUsesWith() is
going to do nothing, but we'll skip all following folds. We can
only run into this problem with calls that both simplify and are
not trivially dead if unused, which currently seems to happen only
with calls to undef, as the test diff shows. When extending
SimplifyCall() to handle "returned" attributes, this becomes a much
bigger problem, so I'm fixing this first.
Differential Revision: https://reviews.llvm.org/D75814
This removes everything but int_x86_avx512_mask_vcvtph2ps_512 which provides the SAE variant, but even this can use the fpext generic if the rounding control is the default.
Differential Revision: https://reviews.llvm.org/D75162
Use UnaryOperator::CreateFNeg instead.
Summary:
With the introduction of the native fneg instruction, the
fsub -0.0, %x idiom is obsolete. This patch makes LLVM
emit fneg instead of the idiom in all places.
Reviewed By: cameron.mcinally
Differential Revision: https://reviews.llvm.org/D75130
InstCombine removes pairs of start+end intrinsics that don't
have anything in between them. Currently this is done by starting
at the start intrinsic and scanning forwards. This patch changes
it to start at the end intrinsic and scan backwards.
The motivation here is as follows: When we process the start
intrinsic, we have not yet looked at the following instructions,
which may still get folded/removed. If they do, we will only be
able to remove the start/end pair on the next iteration. When we
process the end intrinsic, all the instructions before it have
already been visited, and we don't run into this problem.
Differential Revision: https://reviews.llvm.org/D75011
Followup to D73919 with another batch of replacements of
setOperand() -> replaceOperand(), to make sure the old
operand gets DCEd right away.
Differential Revision: https://reviews.llvm.org/D74932
This changes the SimplifyLibCalls utility to accept an IRBuilderBase,
which allows us to pass through the IRBuilder used by InstCombine.
This will ensure that new instructions get added to the worklist.
The annotated test-case drops from 4 to 2 InstCombine iterations thanks
to this.
To achieve this, I'm adding an IRBuilderBase::OperandBundlesGuard,
which is basically the same as the existing InsertPointGuard and
FastMathFlagsGuard, but for operand bundles. Also add a
setDefaultOperandBundles() method so these can be set outside the
constructor.
Differential Revision: https://reviews.llvm.org/D74792
AMDGPU and x86 at least both have separate controls for whether
denormal results are flushed on output, and for whether denormals are
implicitly treated as 0 as an input. The current DAGCombiner use only
really cares about the input treatment of denormals.
Adds a replaceOperand() helper, which is like Instruction.setOperand()
but adds the old operand to the worklist. This reduces the amount of
missing or incorrect worklist management.
This only applies the helper to a relatively small subset of
setOperand() calls in InstCombine, namely those of the pattern
`I.setOperand(); return &I;`, where it is most obviously applicable.
Differential Revision: https://reviews.llvm.org/D73803
This renames Worklist.AddDeferred() to Worklist.add() and
Worklist.Add() to Worklist.push(). The intention here is that
Worklist.add() should be the go-to method for explicit worklist
management, while the raw Worklist.push() is mostly for
InstCombine internals. I will then migrate uses of Worklist.push()
to Worklist.add() in followup changes.
As suggested by spatel on D73411 I'm also changing the remaining
method names to lowercase first character, in line with current
coding standards.
Differential Revision: https://reviews.llvm.org/D73745
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
Currently there are 4 different mechanisms for controlling denormal
flushing behavior, and about as many equivalent frontend controls.
- AMDGPU uses the fp32-denormals and fp64-f16-denormals subtarget features
- NVPTX uses the nvptx-f32ftz attribute
- ARM directly uses the denormal-fp-math attribute
- Other targets indirectly use denormal-fp-math in one DAGCombine
- cl-denorms-are-zero has a corresponding denorms-are-zero attribute
AMDGPU wants a distinct control for f32 flushing from f16/f64, and as
far as I can tell the same is true for NVPTX (based on the attribute
name).
Work on consolidating these into the denormal-fp-math attribute, and a
new type specific denormal-fp-math-f32 variant. Only ARM seems to
support the two different flush modes, so this is overkill for the
other use cases. Ideally we would error on the unsupported
positive-zero mode on other targets from somewhere.
Move the logic for selecting the flush mode into the compiler driver,
instead of handling it in cc1. denormal-fp-math/denormal-fp-math-f32
are now both cc1 flags, but denormal-fp-math-f32 is not yet exposed as
a user flag.
-cl-denorms-are-zero, -fcuda-flush-denormals-to-zero and
-fno-cuda-flush-denormals-to-zero will be mapped to
-fp-denormal-math-f32=ieee or preserve-sign rather than the old
attributes.
Stop emitting the denorms-are-zero attribute for the OpenCL flag. It
has no in-tree users. The meaning would also be target dependent, such
as the AMDGPU choice to treat this as only meaning allow flushing of
f32 and not f16 or f64. The naming is also potentially confusing,
since DAZ in other contexts refers to instructions implicitly treating
input denormals as zero, not necessarily flushing output denormals to
zero.
This also does not attempt to change the behavior for the current
attribute. The LangRef now states that the default is ieee behavior,
but this is inaccurate for the current implementation. The clang
handling is slightly hacky to avoid touching the existing
denormal-fp-math uses. Fixing this will be left for a future patch.
AMDGPU is still using the subtarget feature to control the denormal
mode, but the new attribute are now emitted. A future change will
switch this and remove the subtarget features.
When multiple guard intrinsics are merged into one, currently the
result of eraseInstFromFunction() is returned -- however, this
should only be done if the current instruction is being removed.
In this case we're removing a different instruction and should
instead report that the current one has been modified by returning it.
For this test case, this reduces the number of instcombine iterations
from 5 to 2 (the minimum possible).
Differential Revision: https://reviews.llvm.org/D72558
The instructions use a mask to either pack disjoint bits together(pext) or spread bits to disjoint locations(pdep). If the mask is all 0s then no bits are extracted or deposited. If the mask is all ones, then the source value is written to the result since no compression or expansion happens. Otherwise if both the source and mask are constant we can walk the bits in the source/mask and calculate the result.
There other crazier things we could do like computeKnownBits or turning pext into shift/and if only a single contiguous range of bits is extracted.
Fixes PR44389
Differential Revision: https://reviews.llvm.org/D71952
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
Constructor invocations such as `APFloat(APFloat::IEEEdouble(), 0.0)`
may seem like they accept a FP (floating point) value, but the overload
they reach is actually the `integerPart` one, not a `float` or `double`
overload (which only exists when `fltSemantics` isn't passed).
This may lead to possible loss of data, by the conversion from `float`
or `double` to `integerPart`.
To prevent future mistakes, a new constructor overload, which accepts
any FP value and marked with `delete`, to prevent its usage.
Fixes PR34095.
Differential Revision: https://reviews.llvm.org/D70425
Summary:
If a user writing C code using the ACLE MVE intrinsics generates a
predicate and then complements it, then the resulting IR will use the
`pred_v2i` IR intrinsic to turn some `<n x i1>` vector into a 16-bit
integer; complement that integer; and convert back. This will generate
machine code that moves the predicate out of the `P0` register,
complements it in an integer GPR, and moves it back in again.
This InstCombine rule replaces `i2v(~v2i(x))` with a direct complement
of the original predicate vector, which we can already instruction-
select as the VPNOT instruction which complements P0 in place.
Reviewers: ostannard, MarkMurrayARM, dmgreen
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70484
If the sign of the sign argument is known (this could be extended to use ValueTracking),
then we can use fneg+fabs to clear/set the sign bit of the magnitude argument.
http://llvm.org/docs/LangRef.html#llvm-copysign-intrinsic
This transform is already done in DAGCombiner, but we can do it sooner in IR as
suggested in PR44153:
https://bugs.llvm.org/show_bug.cgi?id=44153
We have effectively no analysis for copysign in IR, so we are taking the unusual step
of increasing the number of IR instructions for the negative constant case.
Differential Revision: https://reviews.llvm.org/D70792
If you're writing C code using the ACLE MVE intrinsics that passes the
result of a vcmp as input to a predicated intrinsic, e.g.
mve_pred16_t pred = vcmpeqq(v1, v2);
v_out = vaddq_m(v_inactive, v3, v4, pred);
then clang's codegen for the compare intrinsic will create calls to
`@llvm.arm.mve.pred.v2i` to convert the output of `icmp` into an
`mve_pred16_t` integer representation, and then the next intrinsic
will call `@llvm.arm.mve.pred.i2v` to convert it straight back again.
This will be visible in the generated code as a `vmrs`/`vmsr` pair
that move the predicate value pointlessly out of `p0` and back into it again.
To prevent that, I've added InstCombine rules to remove round trips of
the form `v2i(i2v(x))` and `i2v(v2i(x))`. Also I've taught InstCombine
about the known and demanded bits of those intrinsics. As a result,
you now get just the generated code you wanted:
vpt.u16 eq, q1, q2
vaddt.u16 q0, q3, q4
Reviewers: ostannard, MarkMurrayARM, dmgreen
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70313
The MVE VADC instruction reads and writes the carry bit at bit 29 of
the FPSCR register. The corresponding ACLE intrinsic is specified to
work with an integer in which the carry bit is stored at bit 0. So if
a user writes a code sequence in C that passes the carry from one VADC
to the next, like this,
s0 = vadcq_u32(a0, b0, &carry);
s1 = vadcq_u32(a1, b1, &carry);
then clang will generate IR for each of those operations that shifts
the carry bit up into bit 29 before the VADC, and after it, shifts it
back down and masks off all but the low bit. But in this situation
what you really wanted was two consecutive VADC instructions, so that
the second one directly reads the value left in FPSCR by the first,
without wasting several instructions on pointlessly clearing the other
flag bits in between.
This commit explains to InstCombine that the other bits of the flags
operand don't matter, and adds a test that demonstrates that all the
code between the two VADC instructions can be optimized away as a
result.
Reviewers: dmgreen, miyuki, ostannard
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67162
As @reames pointed out post-commit, rL371518 adds additional rounding
in some cases, when doing constant folding of the multiplication.
This breaks a guarantee llvm.fma makes and must be avoided.
This patch reapplies rL371518, but splits off the simplifications not
requiring rounding from SimplifFMulInst as SimplifyFMAFMul.
Reviewers: spatel, lebedev.ri, reames, scanon
Reviewed By: reames
Differential Revision: https://reviews.llvm.org/D67434
llvm-svn: 372899
If we generate the gc.relocate, and then later prove two arguments to the statepoint are equivalent, we should canonicalize the gc.relocate to the form we would have produced if this had been known before rewriting.
llvm-svn: 372771
"Implementations are free to malloc() a buffer containing either (size + 1) bytes or (strnlen(s, size) + 1) bytes. Applications should not assume that strndup() will allocate (size + 1) bytes when strlen(s) is smaller than size."
llvm-svn: 372647
Summary:
Motivation:
- If we can fold it to strdup, we should (strndup does more things than strdup).
- Annotation mechanism. (Works for strdup well).
strdup and strndup are part of C 20 (currently posix fns), so we should optimize them.
Reviewers: efriedma, jdoerfert
Reviewed By: jdoerfert
Subscribers: lebedev.ri, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67679
llvm-svn: 372636
This allows us to fold fma's that multiply with 0.0. Also, the
multiply by 1.0 case is handled there as well. The fneg/fabs cases
are not handled by SimplifyFMulInst, so we need to keep them.
Reviewers: spatel, anemet, lebedev.ri
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D67351
llvm-svn: 371518
This makes the functions in Loads.h require a type to be specified
independently of the pointer Value so that when pointers have no structure
other than address-space, it can still do its job.
Most callers had an obvious memory operation handy to provide this type, but a
SROA and ArgumentPromotion were doing more complicated analysis. They get
updated to merge the properties of the various instructions they were
considering.
llvm-svn: 365468
Prefer the more exact intrinsic to remove a use of the input value
and possibly make further transforms easier (we will still need
to match patterns with funnel-shift of wider types as pieces of
bswap, especially if we want to canonicalize to funnel-shift with
constant shift amount). Discussed in D46760.
llvm-svn: 364187
I'm not 100% sure about this, since I'm worried about IR transforms
that might end up introducing divergence downstream once replaced with
a constant, but I haven't come up with an example yet.
llvm-svn: 363406
When the byval attribute has a type, it must match the pointee type of
any parameter; but InstCombine was not updating the attribute when
folding casts of various kinds away.
llvm-svn: 362643
Based on the overflow direction information added in D62463, we can
now fold always overflowing signed saturating add/sub to signed min/max.
Differential Revision: https://reviews.llvm.org/D62544
llvm-svn: 362006
In order to fold an always overflowing signed saturating add/sub,
we need to know in which direction the always overflow occurs.
This patch splits up AlwaysOverflows into AlwaysOverflowsLow and
AlwaysOverflowsHigh to pass through this information (but it is
not used yet).
Differential Revision: https://reviews.llvm.org/D62463
llvm-svn: 361858
We were turning roundss/sd/ps/pd intrinsics with immediates of 1 or 2 into
llvm.floor/ceil. The llvm.ceil/floor intrinsics are supposed to correspond
to the libm functions. For the libm functions we need to disable the
precision exception so the llvm.floor/ceil functions should always map to
encodings 0x9 and 0xA.
We had a mix of isel patterns where some used 0x9 and 0xA and others used
0x1 and 0x2. We need to be consistent and always use 0x9 and 0xA.
Since we have no way in isel of knowing where the llvm.ceil/floor came
from, we can't map X86 specific intrinsics with encodings 1 or 2 to it.
We could map 0x9 and 0xA to llvm.ceil/floor instead, but I'd really like
to see a use case and optimization advantage first.
I've left the backend test cases to show the blend we now emit without
the extra isel patterns. But I've removed the InstCombine tests completely.
llvm-svn: 361425
This patch rewrites the existing PACKSS/PACKUS constant folding code to expand as a generic expansion.
This is a first NFCI step toward expanding PACKSS/PACKUS intrinsics which are acting as non-saturating truncations (although technically the expansion could be used in all cases - but we'll probably want to be conservative).
llvm-svn: 359111
If we have a masked.load from a location we know to be dereferenceable, we can simply issue a speculative unconditional load against that address. The key advantage is that it produces IR which is well understood by the optimizer. The select (cnd, load, passthrough) form produced should be pattern matchable back to hardware predication if profitable.
Differential Revision: https://reviews.llvm.org/D59703
llvm-svn: 359000
If we have a store to a piece of memory which is known constant, then we know the store must be storing back the same value. As a result, the store (or memset, or memmove) must either be down a dead path, or a noop. In either case, it is valid to simply remove the store.
The motivating case for this involves a memmove to a buffer which is constant down a path which is dynamically dead.
Note that I'm choosing to implement the less aggressive of two possible semantics here. We could simply say that the store *is undefined*, and prune the path. Consensus in the review was that the more aggressive form might be a good follow on change at a later date.
Differential Revision: https://reviews.llvm.org/D60659
llvm-svn: 358919
In the process, use the existing masked.load combine which is slightly stronger, and handles a mix of zero and undef elements in the mask.
llvm-svn: 358913
In InstCombine, we use an idiom of "store i1 true, i1 undef" to indicate we've found a path which we've proven unreachable. We can't actually insert the unreachable instruction since that would require changing the CFG. We leave that to simplifycfg later.
This just factors out that idiom creation so we don't duplicate the same mostly undocument idiom creation in multiple places.
llvm-svn: 358600
If a constant shift amount is used, then only some of the LHS/RHS
operand bits are demanded and we may be able to simplify based on
that. InstCombineSimplifyDemanded already had the necessary support
for that, we just weren't calling it with fshl/fshr as root.
In particular, this allows us to relax some masked funnel shifts
into simple shifts, as shown in the tests.
Patch by Shawn Landden.
Differential Revision: https://reviews.llvm.org/D60660
llvm-svn: 358515
Summary:
Enable some of the existing size optimizations for cold code under PGO.
A ~5% code size saving in big internal app under PGO.
The way it gets BFI/PSI is discussed in the RFC thread
http://lists.llvm.org/pipermail/llvm-dev/2019-March/130894.html
Note it doesn't currently touch loop passes.
Reviewers: davidxl, eraman
Reviewed By: eraman
Subscribers: mgorny, javed.absar, smeenai, mehdi_amini, eraman, zzheng, steven_wu, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59514
llvm-svn: 358422
ssubo X, C is equivalent to saddo X, -C. Make the transformation in
InstCombine and allow the logic implemented for saddo to fold prior
usages of add nsw or sub nsw with constants.
Patch by Dan Robertson.
Differential Revision: https://reviews.llvm.org/D60061
llvm-svn: 358099
First step towards removing the MOVMSK intrinsics completely - this patch expands MOVMSK to the pattern:
e.g. PMOVMSKB(v16i8 x):
%cmp = icmp slt <16 x i8> %x, zeroinitializer
%int = bitcast <16 x i8> %cmp to i16
%res = zext i16 %int to i32
Which is correctly handled by ISel and FastIsel (give or take an annoying movzx move....): https://godbolt.org/z/rkrSFW
Differential Revision: https://reviews.llvm.org/D60256
llvm-svn: 357909
If we know we're not storing a lane, we don't need to compute the lane. This could be improved by using the undef element result to further prune the mask, but I want to separate that into its own change since it's relatively likely to expose other problems.
Differential Revision: https://reviews.llvm.org/D57247
llvm-svn: 356590
Follow-up to:
rL356338
rL356369
We can calculate an arbitrary vector constant minus the bitwidth, so there's
no need to limit this transform to scalars and splats.
llvm-svn: 356372
Follow-up to:
rL356338
Rotates are a special case of funnel shift where the 2 input operands
are the same value, but that does not need to be a restriction for the
canonicalization when the shift amount is a constant.
llvm-svn: 356369
This was noted as a backend problem:
https://bugs.llvm.org/show_bug.cgi?id=41057
...and subsequently fixed for x86:
rL356121
But we should canonicalize these in IR for the benefit of all targets
and improve IR analysis such as CSE.
llvm-svn: 356338
A change of two parts:
1) A generic enhancement for all callers of SDVE to exploit the fact that if all lanes are undef, the result is undef.
2) A GEP specific piece to strengthen/fix the vector index undef element handling, and call into the generic infrastructure when visiting the GEP.
The result is that we replace a vector gep with at least one undef in each lane with a undef. We can also do the same for vector intrinsics. Once the masked.load patch (D57372) has landed, I'll update to include call tests as well.
Differential Revision: https://reviews.llvm.org/D57468
llvm-svn: 356293
The shift argument is defined to be modulo the bitwidth, so if that argument
is a constant, we can always reduce the constant to its minimal form to allow
better CSE and other follow-on transforms.
We need to be careful to ignore constant expressions here, or we will likely
infinite loop. I'm adding a general vector constant query for that case.
Differential Revision: https://reviews.llvm.org/D59374
llvm-svn: 356192
This indicates an intrinsic parameter is required to be a constant,
and should not be replaced with a non-constant value.
Add the attribute to all AMDGPU and generic intrinsics that comments
indicate it should apply to. I scanned other target intrinsics, but I
don't see any obvious comments indicating which arguments are intended
to be only immediates.
This breaks one questionable testcase for the autoupgrade. I'm unclear
on whether the autoupgrade is supposed to really handle declarations
which were never valid. The verifier fails because the attributes now
refer to a parameter past the end of the argument list.
llvm-svn: 355981
Sebastian Neubauer