Use IR analysis to infer when an addrspacecast operand is nonnull, then
lower it to an intrinsic that the DAG can use to skip the null check.
I did this using an intrinsic as it's non-intrusive. An alternative
would have been to allow something like `!nonnull` on `addrspacecast`
then lower that to a custom opcode (or add an operand to the
addrspacecast MIR/DAG opcodes), but it's a lot of boilerplate for just
one target's use case IMO.
I'm hoping that when we switch to GISel that we can move all this logic
to the MIR level without losing info, but currently the DAG doesn't see
enough so we need to act in CGP.
Fixes: SWDEV-316445
Remove LSH transform and restore previous lowering.
Fixes conformance issue in
[77615](https://github.com/llvm/llvm-project/pull/77615) where OpenCL
integer_ops tests fail for integer_clz.
Co-authored-by: Leon Clark <leoclark@amd.com>
AMDGPUInstructionSelector should no longer attempt to select S1 G_PHIs.
Remove MIR test that attempts to inst-select divergent vcc(S1) G_PHI.
Lane mask merging algorithm for GlobalISel is now responsible for
selecting divergent S1 G_PHIs in AMDGPUGlobalISelDivergenceLowering.
Uniform S1 G_PHIs should be lowered to S32 G_PHIs in reg bank select
pass. In summary S1 G_PHIs should not reach AMDGPUInstructionSelector.
Basic implementation of lane mask merging for GlobalISel.
Lane masks on GlobalISel are registers with sgpr register class
and S1 LLT - required by machine uniformity analysis.
Implements equivalent of lowerPhis from SILowerI1Copies.cpp in:
patch 1: https://github.com/llvm/llvm-project/pull/75340
patch 2: https://github.com/llvm/llvm-project/pull/75349
patch 3: https://github.com/llvm/llvm-project/pull/80003
patch 4: https://github.com/llvm/llvm-project/pull/78431
patch 5: is in this commit:
AMDGPU/GlobalISelDivergenceLowering: constrain incoming registers
Previously, in PHIs that represent lane masks, incoming registers
taken as-is were not selected as lane masks. Such registers are not
being merged with another lane mask and most often only have S1 LLT.
Implement constrainAsLaneMask by constraining incoming registers
taken as-is with lane mask attributes, essentially transforming them
to lane masks. This is final step in having PHI instructions created
in this pass to be fully instruction-selected.
Follow the 2019 rules and order -0 as less than +0 and +0 as greater
than -0. As currently defined this isn't required for the intrinsics,
but is a better QoI.
This will avoid the workaround in libc added by #83158
Insert waitcnts for loads and atomics before stores with system scope.
Scope is field in instruction encoding and corresponds to desired
coherence level in cache hierarchy.
Intrinsic stores can set scope in cache policy operand.
If volatile keyword is used on generic stores memory legalizer will set
scope to system. Generic stores, by default, get lowest scope level.
Waitcnts are not required if it is guaranteed that memory is cached.
For example vulkan shaders can guarantee this.
TODO: implement flag for frontends to give us a hint not to insert
waits.
Expecting vulkan flag to be implemented as vulkan:private MMRA.
IGLP itself will be in SavedMutations via mutations added during
Scheduler creation, thus falling back results in reapplying IGLP.
In PostRA scheduling, if we have multiple regions with IGLP
instructions, then we may have infinite loop.
Disable the feature for now.
foldOperands() for REG_SEQUENCE has recursion that can trigger an infinite loop
as the method can modify the operand order, which messes up the range-based
for loop. This patch fixes the issue by caching the uses for processing beforehand,
and then iterating over the cache rather using the instruction iterator.
befa925acac8fd6a9266e introduced preliminary hoisting of COPY
instructions when the user of the COPY is inside the same loop. That
optimization appeared to be too aggressive and hoisted too many COPY's
greatly increasing register pressure causing performance regressions for
AMDGPU target.
This is intended to fix the regression by hoisting COPY instruction only
if either:
- User of COPY can be hoisted (other args are invariant)
or
- Hoisting COPY doesn't bring high register pressure
When initializing MachineSSAUpdater save all attributes of current
virtual register and create new virtual registers with same attributes.
Now new virtual registers have same both register class or bank and LLT.
Previously new virtual registers had same register class but LLT was not
set (LLT was set to default/empty LLT).
Required by GlobalISel for AMDGPU, new 'lane mask' virtual registers
created by MachineSSAUpdater need to have both register class and LLT.
patch 4 from: https://github.com/llvm/llvm-project/pull/73337
When using Greedy Register Allocation, there are times where
early-clobber values are ignored, and assigned the same register. This
is illeagal behaviour for these intructions. To get around this, using
Pseudo instructions for early-clobber registers gives them a definition
and allows Greedy to assign them to a different register. This then
meets the ARM Architecture Reference Manual and matches the defined
behaviour.
This patch takes the existing RISC-V patch and makes it target
independent, then adds support for the ARM Architecture. Doing this will
ensure early-clobber restraints are followed when using the ARM
Architecture. Making the pass target independent will also open up
possibility that support other architectures can be added in the future.
This implements the basic pipelining structure of exp/mfma interleaving
for better extensibility. While it does have improved extensibility,
there are controls which only enable it for DAGs with certain
characteristics (matching the DAGs it has been designed against).
The logic was supposed to be choosing between {0, 1, -1} as an
adjustment to the FP bit pattern. However, the adjustment itself was
used as the bit pattern instead which result in garbage results.
We did something pretty naive:
- round FP64 -> BF16 by first rounding to FP32
- skip FP32 -> BF16 rounding entirely
- taking the top 16 bits of a FP32 which will turn some NaNs into
infinities
Let's do this in a more principled way by rounding types with more
precision than FP32 to FP32 using round-inexact-to-odd which will negate
double rounding issues.
LLVM function calls carry convergence control tokens as operand bundles, where
the tokens themselves are produced by convergence control intrinsics. This patch
implements convergence control tokens in MIR as follows:
1. Introduce target-independent ISD opcodes and MIR opcodes for convergence
control intrinsics.
2. Model token values as untyped virtual registers in MIR.
The change also introduces an additional ISD opcode CONVERGENCECTRL_GLUE and a
corresponding machine opcode with the same spelling. This glues the convergence
control token to SDNodes that represent calls to intrinsics. The glued token is
later translated to an implicit argument in the MIR.
The lowering of calls to user-defined functions is target-specific. On AMDGPU,
the convergence control operand bundle at a non-intrinsic call is translated to
an explicit argument to the SI_CALL_ISEL instruction. Post-selection adjustment
converts this explicit argument to an implicit argument on the SI_CALL
instruction.
This combine transforms an unmerge where only the first element is used
into a truncate. That works OK for scalar but for vector needs to insert
a bitcast to integers, perform the truncate then bitcast back to
vectors. This generates more awkward code than using an Unmerge.
gfx11 chips may, in some conditions, behave incorrectly with S_CLAUSE
instructions (hard clauses) containing more than 32 operations (that is,
whose arguments exceed 0x1f). However, gfx10 targets will work
successfully with clauses of up to length 63.
Therefore, define the MaxHardClauseLength property on GCNSubtarget and
make it a subtarget feature via tablegen, thus allowing us to specify,
both now and in the future, the maximum viable size of clauses on
various hardware from the tablegen definition. If MaxHardClauseLength is
0, which is the default, the hardware does not support hard clauses.
When a whole register is added a basic block's liveins, use
LaneBitmask::getAll for the live lanes instead of trying to calculate an
accurate mask of the lanes that comprise the register.
This simplifies the code and matches other places where a whole register
is marked as livein.
This also avoids problems when regunits that are synthesized by TableGen
to represent ad hoc aliasing have a lane mask of 0.
Fixes#78942
The dot is too confusing for tools. Output temporaries would have
'10.3-generic' so tools could parse it as an extension, device libs &
the associated clang driver logic are also confused by the dot.
After discussions, we decided it's better to just remove the '.' from
the target name than fix each issue one by one.
"ninja check-llvm" is failing on tip of tree.
This reverts commit ec0aa1646e9953d1a8d0d15dc381d3250c854572.
This reverts commit 1b65742f8c71f576381fe85d5e34579b24f2d874.
In this case, a trivial GEP chain has the form:
```
%ptr = getelementptr sameType, %base, constant
%val = getelementptr sameType, %ptr, %variable
```
That is, a one-index GEP consumes another (of the same basis and result
type) one-index GEP, where the inner GEP uses a constant index and the
outer GEP uses a variable index. For chains of this type, it is trivial
to reorder them (by simply swapping the indexes). The result of doing so
is better AddrMode matching for users of the ultimate ptr produced by
GEP chain.
Future patches can extend this to support non-trivial GEP chains (e.g.
those with different basis types and/or multiple indices).
Summary:
This patch adds a new intrinsic and builtin function mirroring the
existing `__builtin_readcyclecounter`. The difference is that this
implementation targets a separate counter that some targets have which
returns a fixed frequency clock that can be used to determine elapsed
time, this is different compared to the cycle counter which often has
variable frequency.
This patch only adds support for the NVPTX and AMDGPU targets.
This is done as a new and separate builtin rather than an argument to
`readcyclecounter` to avoid needing to change existing code and to make
the separation more explicit.
These generic targets include multiple GPUs and will, in the future,
provide a way to build once and run on multiple GPU, at the cost of less
optimization opportunities.
Note that this is just doing the compiler side of things, device libs an
runtimes/loader/etc. don't know about these targets yet, so none of them
actually work in practice right now. This is just the initial commit to
make LLVM aware of them.
This contains the documentation changes for both this change and #76954
as well.
Reverts llvm/llvm-project#79586
This broke the AMDGPU OpenMP Offload buildbot.
The typical error message was that the GPU attempted to read beyong the
largest legal address.
Error message:
AMDGPU fatal error 1: Received error in queue 0x7f8363f22000:
HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION: The agent attempted to
access memory beyond the largest legal address.
At the moment, the emergency spill slot is a fixed object for entry
functions and chain functions, and a regular stack object otherwise.
This patch adopts the latter behaviour for entry/chain functions too. It
seems this was always the intention [1] and it will also save us a bit
of stack space in cases where the first stack object has a large
alignment.
[1]
34c8b835b1
