The original motivation for this was to implement moreElementsVector of shuffles
on AArch64, which resulted in complex sequences of artifacts like unmerge(unmerge(concat...))
which the combiner couldn't handle. It seemed here that the better option,
instead of writing ever-more-complex combines, was to have a way to find
the original "non-artifact" source registers for a given definition, walking
through arbitrary expressions of unmerge/concat/insert. As long as the bits
aren't extended or truncated, this is a pretty simple algorithm that avoids
the need for lots of combines and instead jumps straight to the final result
we want.
I've only used this new technique in 2 places within tryCombineUnmerge, using it
in more general situations resulted in infinite loops in AMDGPU. So for now
it's used when we would otherwise fail to combine and that seems to work.
In order to support looking through G_INSERTs, I also had to add it as an
artifact in isArtifact(), which caused a whole lot of issues in tests. AMDGPU
started infinite looping since full legalization of G_INSERT doensn't seem to
be there. To work around this, I've temporarily added a CLI option to use the
old behaviour so that the MIR tests will still run and terminate.
Other minor changes include no longer making >128b G_MERGE/UNMERGE legal.
We never had isel support for that anyway and it was a remnant of the legacy
legalizer rules. However being legal prevented the combiner from checking if it
was dead and deleting them.
Differential Revision: https://reviews.llvm.org/D104355
This enables proper lowering of non-byte sized loads. We still aren't
faithfully preserving memory types everywhere, so the legality checks
still only consider the size.
Previously we didn't preserve the memory type and had to blindly
interpret a number of bytes. Now that non-byte memory accesses are
representable, we can handle these correctly.
Ported from DAG version (minus some weird special case i1 legality
checking which I don't fully understand, and we don't have a way to
query for)
For now, this is NFC and the test changes are placeholders. Since the
legality queries are still relying on byte-flattened memory sizes, the
legalizer can't actually see these non-byte accesses. This keeps this
change self contained without merging it with the larger patch to
switch to LLT memory queries.
This will currently accept the old number of bytes syntax, and convert
it to a scalar. This should be removed in the near future (I think I
converted all of the tests already, but likely missed a few).
Not sure what the exact syntax and policy should be. We can continue
printing the number of bytes for non-generic instructions to avoid
test churn and only allow non-scalar types for generic instructions.
This will currently print the LLT in parentheses, but accept parsing
the existing integers and implicitly converting to scalar. The
parentheses are a bit ugly, but the parser logic seems unable to deal
without either parentheses or some keyword to indicate the start of a
type.
This is to allow 64 bit constant rematerialization. If a constant
is split into two separate moves initializing sub0 and sub1 like
now RA cannot rematerizalize a 64 bit register.
This gives 10-20% uplift in a set of huge apps heavily using double
precession math.
Fixes: SWDEV-292645
Differential Revision: https://reviews.llvm.org/D104874
GlobalISel is relying on regular MachineMemOperands to track all of
the memory properties of accesses. Just the raw byte size is
insufficent to disambiguate all situations. For example, if we need to
split an unaligned extending load, we need to know the number of bits
in the original source value and can't infer it from the result
type. This is also a problem for extending vector loads.
This does decrease the maximum representable size from the full
uint64_t bytes to a maximum of 16-bits. No in tree testcases hit this,
other than places using UINT64_MAX for unknown sizes. This may be an
issue for G_MEMCPY and co., although they can just use unknown size
for large static sizes. This also has potential for backend abuse by
relying on the type when it really shouldn't be relevant after
selection.
This does not include the necessary MIR printer/parser changes to
represent this.
Adds legalizer, register bank select, and instruction
select support for G_SBFX and G_UBFX. These opcodes generate
scalar or vector ALU bitfield extract instructions for
AMDGPU. The instructions allow both constant or register
values for the offset and width operands.
The 32-bit scalar version is expanded to a sequence that
combines the offset and width into a single register.
There are no 64-bit vgpr bitfield extract instructions, so the
operations are expanded to a sequence of instructions that
implement the operation. If the width is a constant,
then the 32-bit bitfield extract instructions are used.
Moved the AArch64 specific code for creating G_SBFX to
CombinerHelper.cpp so that it can be used by other targets.
Only bitfield extracts with constant offset and width values
are handled currently.
Differential Revision: https://reviews.llvm.org/D100149
Add SReg_224, VReg_224, AReg_224, etc.
Link 224-bit types with v7i32/v7f32.
Link existing 192-bit types to newly added v3i64/v3f64/v6i32/v6f32.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D104622
This pass aims to optimize VGPR live-range in a typical divergent if-else
control flow. For example:
def(a)
if(cond)
use(a)
... // A
else
use(a)
As AMDGPU access vgpr with respect to active-mask, we can mark `a` as
dead in region A. For details, please refer to the comments in
implementation file.
The pass is enabled by default, the frontend can disable it through
"-amdgpu-opt-vgpr-liverange=false".
Differential Revision: https://reviews.llvm.org/D102212
- Take the same principle as the conversion from f64 to i64 with extra
necessary pre- and post-processing. It helps to reduce that conversion
sequence by half compared to legacy one.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D104427
This can be seen as a follow up to commit 0ee439b705,
that changed the second argument of __powidf2, __powisf2 and
__powitf2 in compiler-rt from si_int to int. That was to align with
how those runtimes are defined in libgcc.
One thing that seem to have been missing in that patch was to make
sure that the rest of LLVM also handle that the argument now depends
on the size of int (not using the si_int machine mode for 32-bit).
When using __builtin_powi for a target with 16-bit int clang crashed.
And when emitting libcalls to those rtlib functions, typically when
lowering @llvm.powi), the backend would always prepare the exponent
argument as an i32 which caused miscompiles when the rtlib was
compiled with 16-bit int.
The solution used here is to use an overloaded type for the second
argument in @llvm.powi. This way clang can use the "correct" type
when lowering __builtin_powi, and then later when emitting the libcall
it is assumed that the type used in @llvm.powi matches the rtlib
function.
One thing that needed some extra attention was that when vectorizing
calls several passes did not support that several arguments could
be overloaded in the intrinsics. This patch allows overload of a
scalar operand by adding hasVectorInstrinsicOverloadedScalarOpd, with
an entry for powi.
Differential Revision: https://reviews.llvm.org/D99439
G_INSERT legalization is incomplete and doesn't work very
well. Instead try to use sequences of G_MERGE_VALUES/G_UNMERGE_VALUES
padding with undef values (although this can get pretty large).
For the case of load/store narrowing, this is still performing the
load/stores in irregularly sized pieces. It might be cleaner to split
this down into equal sized pieces, and rely on load/store merging to
optimize it.
Avoid having to round up to v8f32/VReg_256 when only 5 VGPRs are
required for a MIMG address operand.
Maintain _V8 instruction variants of pseudo instructions allowing
assembly prior to GFX10 to work as-is. Currently the validator
can tell for GFX10 what the correct size is, so will disallow
oversize address registers.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D103672
This allows to convert the add instruction to s_addk_i32 and
v_add_nc_u32 instead of needing v_add_co_u32 when converting to a VALU
instruction.
Differential Revision: https://reviews.llvm.org/D103322
Before packing LDS globals into a sorted structure, make sure that
their alignment is properly updated based on their size. This will make
sure that the members of sorted structure are properly aligned, and
hence it will further reduce the probability of unaligned LDS access.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D103261
Before packing LDS globals into a sorted structure, make sure that
their alignment is properly updated based on their size. This will make
sure that the members of sorted structure are properly aligned, and
hence it will further reduce the probability of unaligned LDS access.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D103261
Adjusting the load register type is a widenScalar type action, not a
lowering. lowerLoad should be reserved for operations that change the
memory access size, such as unaligned load decomposition. With this
trying to adjust the register type, it was hard to avoid infinite
loops in the legalizer. Adds a bandaid to avoid regressing a few
AArch64 tests, but I'm not sure what the exact condition is and
there's probably a cleaner way to do this.
For AMDGPU this regresses handling of some cases for unaligned loads,
but the way this is currently working is a pretty ugly hack.
This function can change regbank for registers which already have a selected
bank. Depending on the instruction where these registers were used it can
cause instruction selection to fail.
Differential Revision: https://reviews.llvm.org/D98515
This function can change regbank for registers which already have a selected
bank. Depending on the instruction where these registers were used it can
cause instruction selection to fail.
Use existing KnownBits helpers from KnownBits.h to simplify G_ICMPs.
E.g.
x == x -> true
x != x -> false
load(x) > 1 -> true (when the load is known to be greater than 1)
And so on.
Differential Revision: https://reviews.llvm.org/D102542
Large loads on target that does not useFlatForGlobal have to be split
in regbankselect. This did not happen in case when destination had vgpr
bank and address had sgpr bank.
Instead of checking if address bank is sgpr check bank of the destination.
Differential Revision: https://reviews.llvm.org/D101992
First clean up the strange API of tryConstantFoldOp where it took an
immediate operand value, but no indication of which operand it was the
value for.
Second clean up the loop that calls tryConstantFoldOp so that it does
not have to restart from the beginning every time it folds an
instruction.
This is NFCI but there are some minor changes caused by the order in
which things are folded.
Differential Revision: https://reviews.llvm.org/D100031
Preexisting waitcnt may not update the scoreboard if the instruction
being examined needed to wait on fewer counters than what was encoded in
the old waitcnt instruction. Fixing this results in the elimination of
some redudnat waitcnt.
These changes also enable combining consecutive waitcnt into a single
S_WAITCNT or S_WAITCNT_VSCNT instruction.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D100281
Widen 1 and 2 byte scalar loads to 4 bytes when sufficiently
aligned to avoid using a global load.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D100430
An address can be a uniform sum of two i64 bit values.
That regularly happens in a loop where index is an induction
variable promoted to 64 bit by the LSR. We can materialize
zero in a VGPR and still use SADDR form of the load.
Differential Revision: https://reviews.llvm.org/D101591
Unfortunately the current call lowering code is built on top of the
legacy MVT/DAG based code. However, GlobalISel was not using it the
same way. In short, the DAG passes legalized types to the assignment
function, and GlobalISel was passing the original raw type if it was
simple.
I do believe the DAG lowering is conceptually broken since it requires
picking a type up front before knowing how/where the value will be
passed. This ends up being a problem for AArch64, which wants to pass
i1/i8/i16 values as a different size if passed on the stack or in
registers.
The argument type decision is split across 3 different places which is
hard to follow. SelectionDAG builder uses
getRegisterTypeForCallingConv to pick a legal type, tablegen gives the
illusion of controlling the type, and the target may have additional
hacks in the C++ part of the call lowering. AArch64 hacks around this
by not using the standard AnalyzeFormalArguments and special casing
i1/i8/i16 by looking at the underlying type of the original IR
argument.
I believe people have generally assumed the calling convention code is
processing the original types, and I've discovered a number of dead
paths in several targets.
x86 actually relies on the opposite behavior from AArch64, and relies
on x86_32 and x86_64 sharing calling convention code where the 64-bit
cases implicitly do not work on x86_32 due to using the pre-legalized
types.
AMDGPU targets without legal i16/f16 have always used a broken ABI
that promotes to i32/f32. GlobalISel accidentally fixed this to be the
ABI we should have, but this fixes it so we're using the worse ABI
that is compatible with the DAG. Ideally we would fix the DAG to match
the old GlobalISel behavior, but I don't wish to fight that battle.
A new native GlobalISel call lowering framework should let the target
process the incoming types directly.
CCValAssigns select a "ValVT" and "LocVT" but the meanings of these
aren't entirely clear. Different targets don't use them consistently,
even within their own call lowering code. My current belief is the
intent was "ValVT" is supposed to be the legalized value type to use
in the end, and and LocVT was supposed to be the ABI passed type
(which is also legalized).
With the default CCState::Analyze functions always passing the same
type for these arguments, these only differ when the TableGen part of
the lowering decide to promote the type from one legal type to
another. AArch64's i1/i8/i16 hack ends up inverting the meanings of
these values, so I had to add an additional hack to let the target
interpret how large the argument memory is.
Since targets don't consistently interpret ValVT and LocVT, this
doesn't produce quite equivalent code to the initial DAG
lowerings. I've opted to consistently interpret LocVT as the in-memory
size for stack passed values, and ValVT as the register type to assign
from that memory. We therefore produce extending loads directly out of
the IRTranslator, whereas the DAG would emit regular loads of smaller
values. This will also produce loads/stores that are wider than the
argument value if the allocated stack slot is larger (and there will
be undef padding bytes). If we had the optimizations to reduce
load/stores based on truncated values, this wouldn't produce a
different end result.
Since ValVT/LocVT are more consistently interpreted, we now will emit
more G_BITCASTS as requested by the CCAssignFn. For example AArch64
was directly assigning types to some physical vector registers which
according to the tablegen spec should have been casted to a vector
with a different element type.
This also moves the responsibility for inserting
G_ASSERT_SEXT/G_ASSERT_ZEXT from the target ValueHandlers into the
generic code, which is closer to how SelectionDAGBuilder works.
I had to xfail an x86 test since I don't see a quick way to fix it
right now (I filed bug 50035 for this). It's broken independently of
this change, and only triggers since now we end up with more ands
which hit the improperly handled selection pattern.
I also observed that FP arguments that need promotion (e.g. f16 passed
as f32) are broken, and use regular G_TRUNC and G_ANYEXT.
TLDR; the current call lowering infrastructure is bad and nobody has
ever understood how it chooses types.
Amara Emerson