Previously, if a floating-point type was legal, but FNEG wasn't legal,
we would use FSUB. Instead, we should use integer ops, to preserve the
semantics. (Alternatively, there's a compiler-rt call we could use, but
there isn't much reason to use that.)
It turns out we actually are still using this obscure codepath in a few
cases: on some targets, we have "legal" floating-point types that don't
actually support any floating-point operations. In particular, ARM and
AArch64 are using this path.
The implementation for SelectionDAG is pretty simple because we can
reuse the infrastructure from FCOPYSIGN.
See also 9a3dc3e, the corresponding change to type legalization.
Also includes a "bonus" change to STRICT_FSUB legalization, so we can
lower a STRICT_FSUB to a float libcall.
Includes the changes to both LegalizeDAG and GlobalISel so we don't have
inconsistent results in the future.
Fixes https://bugs.llvm.org/show_bug.cgi?id=46792 .
Differential Revision: https://reviews.llvm.org/D84287
For <8 x s32> = fptrunc <8 x s64> the fewerElementsVector action tries to break
down the source vector into the final source vectors of <2 x s64> using unmerge.
This fixes a crash due to using the wrong number of elements for the breakdown
type.
Also add some legalizer tests for explicitly G_FPTRUNC which we didn't have.
Differential Revision: https://reviews.llvm.org/D87814
AArch64, X86 and Mips currently directly consumes these and custom
lowering to produce a libcall, but really these should follow the
normal legalization process through the libcall/lower action.
Some of the lower implementations were relying on this, however the
type was not set depending on which form .lower* helper form you were
using. For instance, if you used an unconditonal lower(), the type was
never set. Most of the lower actions do not benefit from a type
parameter, and just expand in terms of the original operation's types.
However, some lowerings could benefit from an additional type hint to
combine a promotion and an expansion. An example of this is for
add/sub sat. The DAG integer legalization tries to use smarter
expansions directly when promoting the integer type, and doesn't
always produce the same instruction with a wider type.
Treat this as an optional hint argument, that only means something for
specific lower actions. It may be useful to generalize this mechanism
to pass a full list of type indexes and desired types, but I haven't
run into a case like that yet.
This mirrors the support for the equivalent extracts. This also
creates a huge mess that would be greatly improved if we had any bit
operation combines.
Use the same basic strategy as LegalizeVectorTypes. Try to index into
smaller pieces if there's a constant index, and otherwise fall back to
a stack temporary.
Use pad with undef and unmerge with unused results. This is annoyingly
similar to several other places in LegalizerHelper, but they're all
slightly different.
For AMDGPU, vectors with elements < 32 bits should be indexed in
32-bit elements and the desired bits extracted from there. For
elements > 64-bits, these should be reduce to 64/32 elements to enable
the normal dynamic indexing paths.
In the dynamic index cases, this produces shorter code most of the
time. This does immediately regress the constant index cases, but this
should be fixed once we have the most basic of shift combines.
The element size > 64 case is pretty much ported from the exisiting
DAG implementation for extract element promote. The increasing element
size case is new.
I still think it's highly questionable that we have two intrinsics
with identical behavior and only vary by the name of the libcall used
if it happens to be lowered that way, but try to reduce the feature
delta between SDAG and GlobalISel for recently added intrinsics. I'm
not sure which opcode should be considered the canonical one, but
lower roundeven back to round.
Add support in LegalizerHelper for lowering G_SADDSAT etc. either
using add/subtract-with-overflow or using max/min instructions.
Enable this lowering for AMDGPU so it can be tested. The legalization
rules are still approximate and skips out on using the clamp bit to
treat these as legal, which has never been used before. This also
doesn't yet try to deal with expanding SALU cases.
Try harder to find a canonical unmerge type when trying to cover the
desired target type. Handle finding a compatible unmerge type for two
vectors with different element types. This will return the largest
multiple of the source vector element that will evenly divide the
target vector type.
Also make the handling mixing scalars and vectors, and prefer the
source element type as the unmerge target type.
Add narrowScalarFor action.
Add narrow scalar for typeIndex == 0 for G_FPTOSI/G_FPTOUI.
Legalize using narrowScalarFor as s16->s32 G_FPTOSI/G_FPTOUI
followed by s32->s64 G_SEXT/G_ZEXT.
Differential Revision: https://reviews.llvm.org/D84010
Add widenScalar for TypeIdx == 0 for G_SITOFP/G_UITOFP.
Legailize, using widenScalar, as s64->s32 G_SITOFP/G_UITOFP
followed by s32->s16 G_FPTRUNC.
Differential Revision: https://reviews.llvm.org/D83880
This was passing in all the parameters needed to construct a
LegalizerHelper in the custom legalization, when it's simpler to just
pass in the existing helper.
This is slightly more annoying to use in the common case where you
don't need the legalizer helper, but we could add back the common
parameters back in addition to the helper.
I didn't propagate this to all the internal target changes that this
logically implies, but did update a sample one for
legalizeMinNumMaxNum.
This is in preparation for moving AMDGPU load/store legalization
entirely into custom lowering. The current set of legalization actions
is really constraining and not really capable of expressing all the
actions needed to legalize loads/stores. In particular there's no way
to express when the memory access itself needs to change size vs. the
result type. There's also a lot of redundancy since the same
split/widen actions need to be applied in both vector and scalar
cases. All of the sub-cases logically belong as steps in the legalizer
helper, but it will be easier to consider everything at once in custom
lowering.
Put AND before ADD in LegalizerHelper::lowerFPTRUNC_F64_TO_F16
in order to match algorithm from AMDGPUTargetLowering::LowerFP_TO_FP16.
Differential Revision: https://reviews.llvm.org/D81666
It was annoying enough that every custom lowering needed to set the
insert point, but this was made worse since now these all needed to be
updated to setInstrAndDebugLoc. Consolidate these so every
legalization action has the right insert position by default.
This should fix dropping debug info in every custom AMDGPU
legalization.
The current relationship between LegalizerHelper and MachineIRBuilder
confuses me, because the LegalizerHelper modifies the MachineIRBuilder
which it does not own. Constructing a LegalizerHelper destroys the
insert point, since the constructor calls setMF, which clears all the
fields. Try to separate these functions, so it's possible to construct
a LegalizerHelper from an existing MachineIRBuilder without losing the
insert point/debug loc.
Currently this code exists in widenScalar for G_MERGE_VALUE
sources. I'm not sure if the existing expansion in widenScalar should
be removed or not. The widenScalar variant tries to extend to the
requested size, but this just uses the original bitwidth.
Summary:
While lowering memory intrinsics, GIsel attempts to form a tail call to
a library routine.
There might be a DBG_LABEL or something after the intrinsic call,
though: in that case, GIsel should still be able to form the tail call,
and should also delete the debug insts after the tail call as the
transform makes them invalid.
Reviewers: dsanders, aemerson
Subscribers: hiraditya, aprantl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78335
Summary:
When narrowing G_IMPLICIT_DEF where the original size is not a multiple
of the narrow size, emit a smaller G_IMPLICIT_DEF and use G_ANYEXT.
To prevent a potential endless loop in the legalizer, the condition
to combine G_ANYEXT(G_IMPLICIT_DEF) is changed from isInstUnsupported
to !isInstLegal, since in this case the combine is only valid if
consequent legalization of the newly combined G_IMPLICIT_DEF does not
introduce G_ANYEXT due to narrowing.
Although this legalization for G_IMPLICIT_DEF would also be valid for
the general case, it actually caused a lot of code regressions when
tried due to superfluous COPYs and combines not getting hit anymore.
Reviewers: dsanders, aemerson, volkan, arsenm, aditya_nandakumar
Reviewed By: arsenm
Subscribers: jvesely, nhaehnle, kerbowa, wdng, rovka, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76598
Summary: I think it would be better to require the alignment to be >= 1. It is currently confusing to allow both values.
Reviewers: courbet
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77372
The legalizer has a tendency to lose DebugLoc's when expanding or
combining instructions. The verifier that detected these isn't ready
for upstreaming yet but this patch fixes the cases that came up when
applying it to our out-of-tree backend's CodeGen tests.
This pattern comes up a few more times in this file and probably in
the backends too but I'd prefer to fix the others separately (and
preferably when the lost-location verifier detects them).
Summary:
The existing helper function can only create a libcall to functions available in
RTLIB. Add a helper function that can create a libcall to a given function name
using the provided calling convention.
Reviewers: aditya_nandakumar, t.p.northover, rovka, arsenm, dsanders
Reviewed By: arsenm
Subscribers: wdng, hiraditya, volkan, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76845
For some operations, the type is unimportant and only the number of
bits matters. For example I don't want to treat <4 x s8> as a legal
type, but I also don't want to decompose loads of this into smaller
pieces to get legal register types.
On AMDGPU in SelectionDAG, we legalize a number of operations (most
notably load and store) by coercing all types to vectors of i32. For
GlobalISel, I'm trying very hard to avoid doing this for every type,
but I don't think this strategy can be completely avoided. I'm trying
to avoid bitcasts for any legitimately legal type we can operate on,
since the intervening bitcasts have proven to be a hassle.
For loads, I think I can get away without ever casting the result
type, and handling any arbitrary bitwidth during selection (I will
eventually want new tablegen support to help with this, rather than
having to add every possible type as legal). The unmerge required to
do anything with the value should expand to the expected shifts. This
is trickier for stores, since it would now require handling a wide
array of truncates during selection which I don't want.
Future potentially interesting case are for vector indexing, where
sub-dword type should be indexed in s32 pieces.
Summary:
Widening G_UNMERGE_VALUES to a type which is larger than the
original source type is the same as widening it to the same
type as the source type: in both cases, G_UNMERGE_VALUES has
to be replaced with bit arithmetic which. Although the arithmetic
itself is independent of whether the source type is smaller
or equal to the widen type, widening the source type to the
widen type should result in less artifacts being emitted,
since this is the type that the user explicitly requested.
Reviewers: arsenm, dsanders, aemerson, aditya_nandakumar
Reviewed By: arsenm, dsanders
Subscribers: jvesely, wdng, nhaehnle, rovka, hiraditya, volkan, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76494
I used the implementation for floor instead of round. It also turns
out the OpenCL builtin library wasn't using the round builtin, but
implemented the expanded form.
Summary: When narrowing a scalar G_EXTRACT where the destination lines up perfectly with a single result of the emitted G_UNMERGE_VALUES a COPY should be emitted instead of unconditionally trying to emit a G_MERGE_VALUES.
Reviewers: arsenm, dsanders
Reviewed By: arsenm
Subscribers: wdng, rovka, hiraditya, volkan, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75743
If the loaded memory size was smaller than the result size, this would
produce out of bounds memory accesses. I'm wondering if we need a
distinct narrow memory legalize action type, since a case I care about
is decomposing a 4-byte unaligned access into 4 extending loads, which
would leave the original result register type. I'm currently awkwardly
using narrowScalar to handle unaligned accesses that need to be split.
Changes the handling of odd breakdowns, and avoids using
G_EXTRACT/G_INSERT. Pad with undef to a wider size, and unmerge. Also
avoid introducing instructions for the fully undef components.
This is more or less directly ported from the AMDGPU custom lowering
for FP_TO_FP16. I made a few minor fixups (using G_UNMERGE_VALUES
instead of creating shift/trunc to extract the two halves, and zexting
an inverted compare instead of select_cc).
This also does not include the fast math expansion the DAG which
converts to f32 and then to f16. I think that belongs in a
pre-legalize combine instead.
Allows more flexible use of buildMerge in places where
use operands are available as SrcOp since it does not
require explicit conversion to Register.
Simplify code with new buildMerge.
Differential Revision: https://reviews.llvm.org/D74223
The type passed to lower was invalid, so I'm not sure how this was
even working before. The source and destination type also do not have
to match, so make sure to use the right ones.
Start using a new strategy with a combination of merge and unmerges.
This allows scalarizing before lowering, which in cases like
<2 x s128> avoids producing giant illegal shifts.
Summary:
Applying this cleanup:
- MIRBuilder.buildInstr(TargetOpcode::G_ASHR)
- .addDef(Shifted)
- .addUse(Res)
- .addUse(ShiftAmt);
+ MIRBuilder.buildAShr(Shifted, Res, ShiftAmt);
caused an assertion failure here:
llc: /home/jayfoad2/git/llvm-project/llvm/lib/CodeGen/MachineRegisterInfo.cpp:404: llvm::MachineInstr *llvm::MachineRegisterInfo::getVRegDef(unsigned int) const: Assertion `(I.atEnd() || std::next(I) == def_instr_end()) && "getVRegDef assumes a single definition or no definition"' failed.
#4 0x00000000050a6d96 in llvm::MachineRegisterInfo::getVRegDef (this=0x74606a0, Reg=2147483650) at /home/jayfoad2/git/llvm-project/llvm/lib/CodeGen/MachineRegisterInfo.cpp:403
#5 0x00000000066148f6 in llvm::getConstantVRegValWithLookThrough (VReg=2147483650, MRI=..., LookThroughInstrs=false, HandleFConstant=true) at /home/jayfoad2/git/llvm-project/llvm/lib/CodeGen/GlobalISel/Utils.cpp:244
#6 0x00000000066147da in llvm::getConstantVRegVal (VReg=2147483650, MRI=...) at /home/jayfoad2/git/llvm-project/llvm/lib/CodeGen/GlobalISel/Utils.cpp:210
#7 0x0000000006615367 in llvm::ConstantFoldBinOp (Opcode=101, Op1=2147483650, Op2=2147483656, MRI=...) at /home/jayfoad2/git/llvm-project/llvm/lib/CodeGen/GlobalISel/Utils.cpp:341
#8 0x000000000657eee0 in llvm::CSEMIRBuilder::buildInstr (this=0x7465010, Opc=101, DstOps=..., SrcOps=..., Flag=...) at /home/jayfoad2/git/llvm-project/llvm/lib/CodeGen/GlobalISel/CSEMIRBuilder.cpp:160
#9 0x0000000003645958 in llvm::MachineIRBuilder::buildAShr (this=0x7465010, Dst=..., Src0=..., Src1=..., Flags=...) at /home/jayfoad2/git/llvm-project/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h:1298
#10 0x00000000065c35b1 in llvm::LegalizerHelper::lower (this=0x7fffffffb5f8, MI=..., TypeIdx=0, Ty=...) at /home/jayfoad2/git/llvm-project/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp:2020
because at this point there are two instructions defining Res: the
original G_SMULO/G_UMULO and the new G_MUL that we built. The fix is
to modify the original mul in place, so that there is only ever one
definition of Res.
Reviewers: arsenm, aditya_nandakumar
Subscribers: wdng, rovka, hiraditya, volkan, Petar.Avramovic, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72842
Commit 9965b12fd1 was supposed to change the offset constant when
lowering load/stores, but only introduced this change for loads. This
patch adds the same fix for stores.
This is passed to legalizeCustom, but not intrinsic. Also remove the
MRI argument, since you can get that from the MachineIRBuilder.
I'm not sure why MachineIRBuilder has a private observer member, and
this is passed separately.
G_CTPOP is generated from llvm.ctpop.<type> intrinsics, clang generates
these intrinsics from __builtin_popcount and __builtin_popcountll.
Add lower and narrow scalar for G_CTPOP.
Lower G_CTPOP for MIPS32.
Differential Revision: https://reviews.llvm.org/D73216
llvm.cttz.<type> intrinsic has additional i1 argument is_zero_undef,
it tells whether zero as the first argument produces a defined result.
G_CTTZ is generated from llvm.cttz.<type> (<type> <src>, i1 false)
intrinsics, clang generates these intrinsics from __builtin_ctz and
__builtin_ctzll.
G_CTTZ_ZERO_UNDEF comes from llvm.cttz.<type> (<type> <src>, i1 true).
Clang generates such intrinsics as parts of expansion of builtin_ffs
and builtin_ffsll. It is also traditionally part of and many
algorithms that are now predicated on avoiding zero-value inputs.
Add narrow scalar (algorithm uses G_CTTZ_ZERO_UNDEF) for G_CTTZ.
Lower G_CTTZ and G_CTTZ_ZERO_UNDEF for MIPS32.
Differential Revision: https://reviews.llvm.org/D73215
Amara Emerson