Implement an intrinsic for use lowering LDS variables to different
addresses from different kernels. This will allow kernels that cannot
reach an LDS variable to avoid wasting space for it.
There are a number of implicit arguments accessed by intrinsic already
so this implementation closely follows the existing handling. It is slightly
novel in that this SGPR is written by the kernel prologue.
It is necessary in the general case to put variables at different addresses
such that they can be compactly allocated and thus necessary for an
indirect function call to have some means of determining where a
given variable was allocated. Claiming an arbitrary SGPR into which
an integer can be written by the kernel, in this implementation based
on metadata associated with that kernel, which is then passed on to
indirect call sites is sufficient to determine the variable address.
The intent is to emit a __const array of LDS addresses and index into it.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D125060
We form VOPD instructions in the GCNCreateVOPD pass by combining
back-to-back component instructions. There are strict register
constraints for creating a legal VOPD, namely that the matching operands
(e.g. src0x and src0y, src1x and src1y) must be in different register
banks. We add a PostRA scheduler
mutation to put possible VOPD components back-to-back.
Depends on D128442, D128270
Reviewed By: #amdgpu, rampitec
Differential Revision: https://reviews.llvm.org/D128656
GFX11 has a new message type MSG_DEALLOC_VGPRS which can be used to
release a shader's VGPRs. Sending this at the end of a shader (just
before the s_endpgm) can help overall system performance in cases where
the s_endpgm would have to wait for outstanding VMEM stores to complete
before releasing the VGPRs.
Differential Revision: https://reviews.llvm.org/D128442
The sched_barrier builtin allow the scheduler's behavior to be shaped by users
when very specific codegen is needed in order to create highly optimized code.
This patch adds more granular control over the types of instructions that are
allowed to be reordered with respect to one or multiple sched_barriers. A mask
is used to specify groups of instructions that should be allowed to be scheduled
around a sched_barrier. The details about this mask may be used can be found in
llvm/include/llvm/IR/IntrinsicsAMDGPU.td.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D127123
Avoid the dependency on TargetInstrInfo, which depends on the subtarget
and therefore the individual function.
Currently AMDGPU is constructing PseudoSourceValue instances in MachineFunctionInfo.
In order to facilitate copying MachineFunctionInfo, we need to stop allocating these
there. Alternatively we could allow targets to subclass PseudoSourceValueManager,
and allocate them similarly to MachineFunctionInfo.
This patch implements a DAG mutation which adds edges between different groups of instructions. The purpose is to try to generate code that conforms to a pipeline (groupA instructions occur before groupB, groupB -> groupC, and so on). Currently the pipeline order is hardcoded as VMEM->DSRead->MFMA->DSWrite, but the patch was designed to be easily extensible. Alias analysis is problematic for pipelining as memory instructions will usually not be able to be reordered w.r.t one another.
Differential Revision: https://reviews.llvm.org/D125997
This patch adds cluster edges between independent MFMA instructions. Additionally, it propogates all predecessors of cluster insts to the root of the cluster(s), and all successors to the leaf(ves) of the cluster(s) -- this is done to remove the possibility that those insts will be interspersed within the cluster.
Reviewed By: kerbowa
Differential Revision: https://reviews.llvm.org/D124678
As older waves execute long sequences of VALU instructions, this may
prevent younger waves from address calculation and then issuing their
VMEM loads, which in turn leads the VALU unit to idle. This patch tries
to prevent this by temporarily raising the wave's priority.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D124246
static_cast is a little safer here since the compiler will
ensure we're casting to a class derived from
yaml::MachineFunctionInfo.
I believe this first appeared on AMDGPU and was copied to the
other two targets.
Spotted when it was being copied to RISCV in D123178.
Differential Revision: https://reviews.llvm.org/D123260
This was a workaround for not supporting indirect calls when
instcombine didn't eliminate constant expression casts of the callee
at -O0. Indirect calls are supposed to work now, so drop the hack.
Code using indirect calls is broken without this, and there isn't
really much value in supporting the old attempt to vary the argument
placement based on uses. This resulted in more argument shuffling code
anyway.
Also have the option stop implying all inputs need to be passed. This
will no rely on the amdgpu-no-* attributes to avoid passing
unnecessary values.
- CUDA cannot associate memory space with pointer types. Even though Clang could add extra attributes to specify the address space explicitly on a pointer type, it breaks the portability between Clang and NVCC.
- This change proposes to assume the address space from a pointer from the assumption built upon target-specific address space predicates, such as `__isGlobal` from CUDA. E.g.,
```
foo(float *p) {
__builtin_assume(__isGlobal(p));
// From there, we could assume p is a global pointer instead of a
// generic one.
}
```
This makes the code portable without introducing the implementation-specific features.
Note that NVCC starts to support __builtin_assume from version 11.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D112041
Run post-RA SIShrinkInstructions just before post-RA scheduling, instead
of afterwards. After the fixes in D112305 and D112317 this seems to make
no difference, but it paves the way for scheduler tweaks that are
sensitive to the e32 vs e64 encoding of VALU instructions.
Differential Revision: https://reviews.llvm.org/D112341
This has a couple of benefits:
1. It can sometimes fix clusters that got broken apart when the register
allocator inserted a copy.
2. Post-RA scheduling does not have to worry about increasing register
pressure, which in some cases gives it more freedom to reorder
instructions.
Testing on a collection of 10,000 graphics shaders compiled for gfx1010
showed:
- The average length of each run of one or more load instructions
increased by about 1%.
- The number of runs of two or more load instructions increased by
about 4%.
Differential Revision: https://reviews.llvm.org/D111646
The new pass walks kernel's pointer arguments, then loads from them.
If a loaded value is a pointer and loaded pointer is unmodified in
the kernel before the load, then promote loaded pointer to global.
Then recursively continue.
Differential Revision: https://reviews.llvm.org/D111464
This has a couple of benefits:
1. It can sometimes fix clusters that got broken apart when the register
allocator inserted a copy.
2. Post-RA scheduling does not have to worry about increasing register
pressure, which in some cases gives it more freedom to reorder
instructions.
Testing on a collection of 10,000 graphics shaders compiled for gfx1010
showed:
- The average length of each run of one or more load instructions
increased by about 1%.
- The number of runs of two or more load instructions increased by
about 4%.
This moves the registry higher in the LLVM library dependency stack.
Every client of the target registry needs to link against MC anyway to
actually use the target, so we might as well move this out of Support.
This allows us to ensure that Support doesn't have includes from MC/*.
Differential Revision: https://reviews.llvm.org/D111454
This was introduced in D32628 but it does not seem to be required any
more. At least it does not show any problems in check-llvm in an
LLVM_ENABLE_EXPENSIVE_CHECKS build.
Differential Revision: https://reviews.llvm.org/D110692
ASan device library functions (those starts with the prefix __asan_)
are at the moment undergoing through undesired optimizations due to
internalization. Hence, in order to avoid such undesired optimizations
on ASan device library functions, do not internalize them in the first
place.
Reviewed By: yaxunl
Differential Revision: https://reviews.llvm.org/D110468
Use GCNHazardRecognizer in postra sched.
Updated tests for the new schedules.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D109536
Change-Id: Ia86ba2ae168f12fb34b4d8efdab491f84d936cde
Previously we assumed all callable functions did not need any
implicitly passed inputs, and added attributes to functions to
indicate when they were necessary. Requiring attributes for
correctness is pretty ugly, and it makes supporting indirect and
external calls more complicated.
This inverts the direction of the attributes, so an undecorated
function is assumed to need all implicit imputs. This enables
AMDGPUAttributor by default to mark when functions are proven to not
need a given input. This strips the equivalent functionality from the
legacy AMDGPUAnnotateKernelFeatures pass.
However, AMDGPUAnnotateKernelFeatures is not fully removed at this
point although it should be in the future. It is still necessary for
the two hacky amdgpu-calls and amdgpu-stack-objects attributes, which
would be better served by a trivial analysis on the IR during
selection. Additionally, AMDGPUAnnotateKernelFeatures still
redundantly handles the uniform-work-group-size attribute to be
removed in a future commit.
At this point when not using -amdgpu-fixed-function-abi, we are still
modifying the ABI based on these newly negated attributes. In the
future, this option will be removed and the locations for implicit
inputs will always be fixed. We will then use the new attributes to
avoid passing the values when unnecessary.
1. Splitted out some parts of R600 target to separate modules/headers.
2. Reduced some include lists in headers.
3. Found and fixed issue with override `GCNTargetMachine::getSubtargetImpl()`
and `R600TargetMachine::getSubtargetImpl()` had different return value type
than base class.
4. Minor forward declarations cleanup.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D108596
This patch introduces a new code object metadata field, ".kind"
which is used to add support for init and fini kernels.
HSAStreamer will use function attributes, "device-init" and
"device-fini" to distinguish between init and fini kernels from
the regular kernels and will emit metadata with ".kind" set to
"init" and "fini" respectively.
To reduce the number of init and fini kernels, the ctors and
dtors present in the llvm's global.ctors and global.dtors lists
are called from a single init and fini kernel respectively.
Reviewed by: yaxunl
Differential Revision: https://reviews.llvm.org/D105682
This patch introduces a new code object metadata field, ".kind"
which is used to add support for init and fini kernels.
HSAStreamer will use function attributes, "device-init" and
"device-fini" to distinguish between init and fini kernels from
the regular kernels and will emit metadata with ".kind" set to
"init" and "fini" respectively.
To reduce the number of init and fini kernels, the ctors and
dtors present in the llvm's global.ctors and global.dtors lists
are called from a single init and fini kernel respectively.
Reviewed by: yaxunl
Differential Revision: https://reviews.llvm.org/D105682
Pulled out the OptimizationLevel class from PassBuilder in order to be able to access it from within the PassManager and avoid include conflicts.
Reviewed By: mtrofin
Differential Revision: https://reviews.llvm.org/D107025
This patch introduces a pass that uses the Attributor to deduce AMDGPU specific attributes.
Reviewed By: jdoerfert, arsenm
Differential Revision: https://reviews.llvm.org/D104997
This is SCC pass, moving it to the end of SCC PM saves one
Function PM. This needs the analysis to take into account
memory access width since it is now places after the
load/store optimizer (D105651).
Differential Revision: https://reviews.llvm.org/D105652
First, collect the register usage in each function, then apply the
maximum register usage of all functions to functions with indirect
calls.
This is more accurate than guessing the maximum register usage without
looking at the actual usage.
As before, assume that indirect calls will hit a function in the
current module.
Differential Revision: https://reviews.llvm.org/D105839
AMDGPU normally spills SGPRs to VGPRs. Previously, since all register
classes are handled at the same time, this was problematic. We don't
know ahead of time how many registers will be needed to be reserved to
handle the spilling. If no VGPRs were left for spilling, we would have
to try to spill to memory. If the spilled SGPRs were required for exec
mask manipulation, it is highly problematic because the lanes active
at the point of spill are not necessarily the same as at the restore
point.
Avoid this problem by fully allocating SGPRs in a separate regalloc
run from VGPRs. This way we know the exact number of VGPRs needed, and
can reserve them for a second run. This fixes the most serious
issues, but it is still possible using inline asm to make all VGPRs
unavailable. Start erroring in the case where we ever would require
memory for an SGPR spill.
This is implemented by giving each regalloc pass a callback which
reports if a register class should be handled or not. A few passes
need some small changes to deal with leftover virtual registers.
In the AMDGPU implementation, a new pass is introduced to take the
place of PrologEpilogInserter for SGPR spills emitted during the first
run.
One disadvantage of this is currently StackSlotColoring is no longer
used for SGPR spills. It would need to be run again, which will
require more work.
Error if the standard -regalloc option is used. Introduce new separate
-sgpr-regalloc and -vgpr-regalloc flags, so the two runs can be
controlled individually. PBQB is not currently supported, so this also
prevents using the unhandled allocator.
- In [D98783](https://reviews.llvm.org/D98783), an extra GlobalDCE pass
is inserted before the internalization pass to ensure a global
variable without users could be internalized even if there are dead
users. Instead of inserting a dedicated optimization pass, the
dead user checking, i.e. 'use_empty()', should be preceeded with
constant dead user removal to ensure an accurate result.
Differential Revision: https://reviews.llvm.org/D105590
Jon Chesterfield