-amdgpu-inline-max-bb option could lead to a suboptimal
codegen preventing inlining of really simple functions
including pure wrapper calls. Relax the cutoff by allowing
to call a function with a single block on the grounds
that it will not increase total number of blocks after
inlining.
Differential Revision: https://reviews.llvm.org/D97744
The expected use case is for frontends to insert this into
shaders that are to be run under a debugger. The shader can
then be resumed or single stepped from the point of the call
under debugger control.
Differential Revision: https://reviews.llvm.org/D97670
I copied the nearly identical function from AArch64 into AMDGPU, so
fix this duplication.
Mips and X86 have their own more exotic versions which should be
removed. However replacing those is better left for a separate patch
since it requires other changes to avoid regressions.
Previously we would use a bundle to hint the register allocator to not
overwrite the pointers in a sequence of loads to avoid breaking soft
clauses. This bundling was based on a fuzzy register pressure
heuristic, so we could not guarantee using more registers than are
really available. This would result in register allocator failing on
unsatisfiable bundles. Use a kill to artificially extend the live
ranges, so we can always succeed at register allocation even if it
means extra spills in the worst case.
This seems to capture most of the benefit of the bundle while avoiding
most of the risk presented by the bundle. However the lit tests do
show a handful of regressions. In some cases with sequences of
volatile loads, unused load components end up getting reallocated to
the next load which forces a wait between. There are also a few small
scheduling regressions where a hazard used to be avoided, and one
spill torture test which for some reason nearly doubles the stack
usage. There is also a bit of noise from leftover kills (it may make
sense for post-RA pseudos to strip all of these out).
If a region was not constrained by a high register pressure
and was not rescheduled without clustering we can skip
rescheduling it ClusteredLowOccupancyReschedule stage.
This improves scheduling speed by 25% on some kernels.
Differential Revision: https://reviews.llvm.org/D97506
We are attempting rescheduling without load store clustering
if occupancy limits were not met with clustering. Skip this
for regions which do not have any loads or stores at all.
In a set of kernels I am experimenting with this improves
scheduling time by ~30%.
Differential Revision: https://reviews.llvm.org/D97342
This allows GlobalISel to use this instruction where available. I assume
SelectionDAG always selects s_xnor_b32 so it isn't affected by this
change.
Differential Revision: https://reviews.llvm.org/D97560
gfx90a operations require even aligned registers, but this was
previously achieved by reserving registers inside the full class.
Ideally this would be captured in the static instruction definitions
for the operands, and we would have different instructions per
subtarget. The hackiest part of this is we need to manually reassign
AGPR register classes after instruction selection (we get away without
this for VGPRs since those types are actually registered for legal
types).
Update the list of s_sendmsg messages known to the assembler and
disassembler and validate the ones that were added or removed in gfx9
and gfx10.
Differential Revision: https://reviews.llvm.org/D97295
This is to limit compile time. I did experiments with some
inputs and found that compile time keeps reasonable for this
pass if we have less than 100000 virtual registers and then
starts to explode somewhere between 100000 and 150000.
Differential Revision: https://reviews.llvm.org/D97218
Prefer to keep uniform (non-divergent) multiplies on the scalar ALU when
possible. This significantly improves some game cases by eliminating
v_readfirstlane instructions when the result feeds into a scalar
operation, like the address calculation for a scalar load or store.
Since isDivergent is only an approximation of whether a value is in
SGPRs, it can potentially regress some situations where a uniform value
ends up in a VGPR. These should be rare in real code, although the test
changes do contain a number of examples.
Most of the test changes are just using s_mul instead of v_mul/mad which
is generally better for both register pressure and latency (at least on
GFX10 where sgpr pressure doesn't affect occupancy and vector ALU
instructions have significantly longer latency than scalar ALU). Some
R600 tests now use MULLO_INT instead of MUL_UINT24.
GlobalISel appears to handle more scenarios in the desirable way,
although it can also be thrown off and fails to select the 24-bit
multiplies in some cases.
Alternative solution considered and rejected was to allow selecting
MUL_[UI]24 to S_MUL_I32. I've rejected this because the definition of
those SD operations works is don't-care on the most significant 8 bits,
and this fact is used in some combines via SimplifyDemandedBits.
Based on a patch by Nicolai Hähnle.
Differential Revision: https://reviews.llvm.org/D97063
As a followup to D95291, getOperandsScalarizationOverhead was still
using a VF as a vector factor if the arguments were scalar, and would
assert on certain matrix intrinsics with differently sized vector
arguments. This patch removes the VF arg, instead passing the Types
through directly. This should allow it to more accurately compute the
cost without having to guess at which operands will be vectorized,
something difficult with more complex intrinsics.
This adjusts one SVE test as it is now calling the wrong intrinsic vs
veccall. Without invalid InstructCosts the cost of the scalarized
intrinsic is too low. This should get fixed when the cost of
scalarization is accounted for with scalable types.
Differential Revision: https://reviews.llvm.org/D96287
getIntrinsicInstrCost takes a IntrinsicCostAttributes holding various
parameters of the intrinsic being costed. It can either be called with a
scalar intrinsic (RetTy==Scalar, VF==1), with a vector instruction
(RetTy==Vector, VF==1) or from the vectorizer with a scalar type and
vector width (RetTy==Scalar, VF>1). A RetTy==Vector, VF>1 is considered
an error. Both of the vector modes are expected to be treated the same,
but because this is confusing many backends end up getting it wrong.
Instead of trying work with those two values separately this removes the
VF parameter, widening the RetTy/ArgTys by VF used called from the
vectorizer. This keeps things simpler, but does require some other
modifications to keep things consistent.
Most backends look like this will be an improvement (or were not using
getIntrinsicInstrCost). AMDGPU needed the most changes to keep the code
from c230965ccf working. ARM removed the fix in
dfac521da1, webassembly happens to get a fixup for an SLP cost
issue and both X86 and AArch64 seem to now be using better costs from
the vectorizer.
Differential Revision: https://reviews.llvm.org/D95291
Enabled "bound_ctrl:1" and disabled "bound_ctrl:-1" syntax.
Corrected printer to output "bound_ctrl:1" instead of "bound_ctrl:0".
See bug 35397 for detailed issue description.
Differential Revision: https://reviews.llvm.org/D97048
fixed-abi uses pre-defined and predictable
SGPR/VGPRs for passing arguments. This patch makes
this scheme default when HSA OS is specified in triple.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D96340
Track lanes when processing definitions for marking WQM/WWM.
If all lanes have been defined then marking can stop.
This prevents marking unnecessary instructions as WQM/WWM.
In particular this fixes a bug where values passing through
V_SET_INACTIVE would me marked as requiring WWM.
Reviewed By: piotr
Differential Revision: https://reviews.llvm.org/D95503
AMDGPU currently has a lot of pre-processing code to pre-split
argument types into 32-bit pieces before passing it to the generic
code in handleAssignments. This is a bit sloppy and also requires some
overly fancy iterator work when building the calls. It's better if all
argument marshalling code is handled directly in
handleAssignments. This handles more situations like decomposing large
element vectors into sub-element sized pieces.
This should mostly be NFC, but does change the generated code by
shifting where the initial argument packing instructions are placed. I
think this is nicer looking, since it now emits the packing code
directly after the relevant copies, rather than after the copies for
the remaining arguments.
This doubles down on gfx6/gfx7 using the gfx8+ ABI for 16-bit
types. This is ultimately the better option, but incompatible with the
DAG. Fixing this requires more work, especially for f16.
We can always look through single-argument (LCSSA) phi nodes when
performing alias analysis. getUnderlyingObject() already does this,
but stripPointerCastsAndInvariantGroups() does not. We still look
through these phi nodes with the usual aliasPhi() logic, but
sometimes get sub-optimal results due to the restrictions on value
equivalence when looking through arbitrary phi nodes. I think it's
generally beneficial to keep the underlying object logic and the
pointer cast stripping logic in sync, insofar as it is possible.
With this patch we get marginally better results:
aa.NumMayAlias | 5010069 | 5009861
aa.NumMustAlias | 347518 | 347674
aa.NumNoAlias | 27201336 | 27201528
...
licm.NumPromoted | 1293 | 1296
I've renamed the relevant strip method to stripPointerCastsForAliasAnalysis(),
as we're past the point where we can explicitly spell out everything
that's getting stripped.
Differential Revision: https://reviews.llvm.org/D96668
We did not have atomic flags on SMRD, did not copy TSFlags
to real instructions, and did not have ret/noret atomic map.
At the moment it is NFC, but needed for D96469.
Differential Revision: https://reviews.llvm.org/D96823
* Update skip-if-dead.ll with tests for wave32.
* Fix the crash in verifier in one newly enabled test by adding
missing fixImplicitOperands in branch insertion code.
```
*** Bad machine code: Using an undefined physical register ***
- function: test_kill_divergent_loop
- basic block: %bb.2 bb (0xad96308)
- instruction: S_CBRANCH_VCCNZ %bb.1, implicit $vcc_lo
- operand 1: implicit $vcc_lo
LLVM ERROR: Found 1 machine code errors.
```
* Simplify "cbranch_kill" to not use interp instructions.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D96793
The helper function isBoolSGPR is too aggressive when determining
when a v_cndmask can be skipped on a boolean value because the
function does not check the operands of and/or/xor.
This can be problematic for the Add/Sub combines that can leave
bits set even for inactive lanes leading to wrong results.
Fix this by inspecting the operands of and/or/xor recursively.
Differential Revision: https://reviews.llvm.org/D86878
The AMD GPU SIMemoryLegalizer was using the ordering address space
rather than the instruction address space when determining the
s_waitcnt to generate to ensure that a read-modify-write atomic has
completed. This resulted in additional unnecessary counters being
waited on.
Differential Revision: https://reviews.llvm.org/D96743
In a future commit, soft clauses will be hinted with kill instructions
rather than forced together with bundles. Look for kills that look
like this, and erase them. I'm not sure if the check for specific uses
is worthwhile, or if it would be better to just unconditionally erase
kills.
This reduces test churn in a future patch.
This was allowing debug instructions to break the bundling, which
would change scheduling behavior. Bundle debug info / kills inside
the bundle. This seems to work OK, although the asm printer doesn't
understand these in a bundle. This implicitly expects the memory
legalizer to unbundle. It would probably be slightly nicer to move
these after.
Rewrite the loop to be clearer and make sure we don't end a bundle on
a meta instruction, only allow them in between other valid bundle
instructions.
The API is a bit awkward since you need to index into an array in the
passed struct. I guess an alternative would be to pass all of the
individual fields.
This commit fixes how metadata is handled in CloneModule to be sound,
and improves how it's handled in CloneFunctionInto (although the latter
is still awkward when called within a module).
Ruiling Song pointed out in PR48841 that CloneModule was changed to
unsoundly use the RF_ReuseAndMutateDistinctMDs flag (renamed in
fa35c1f80f for clarity). This flag papered
over a crash caused by other various changes made to CloneFunctionInto
over the past few years that made it unsound to use cloning between
different modules.
(This commit partially addresses PR48841, fixing the repro from
preprocessed source but not textual IR. MDNodeMapper::mapDistinctNode
became unsound in df763188c9 and this
commit does not address that regression.)
RF_ReuseAndMutateDistinctMDs is designed for the IRMover to use,
avoiding unnecessary clones of all referenced metadata when linking
between modules (with IRMover, the source module is discarded after
linking). It never makes sense to use when you're not discarding the
source. This commit drops its incorrect use in CloneModule.
Sadly, the right thing to do with metadata when cloning a function is
complicated, and this patch doesn't totally fix it.
The first problem is that there are two different types of referenceable
metadata and it's not obvious what to with one of them when remapping.
- `!0 = !{!1}` is metadata's version of a constant. Programatically it's
called "uniqued" (probably a better term would be "constant") because,
like `ConstantArray`, it's stored in uniquing tables. Once it's
constructed, it's illegal to change its arguments.
- `!0 = distinct !{!1}` is a bit closer to a global variable. It's legal
to change the operands after construction.
What should be done with distinct metadata when cloning functions within
the same module?
- Should new, cloned nodes be created?
- Should all references point to the same, old nodes?
The answer depends on whether that metadata is effectively owned by a
function.
And that's the second problem. Referenceable metadata's ownership model
is not clear or explicit. Technically, it's all stored on an
LLVMContext. However, any metadata that is `distinct`, that transitively
references a `distinct` node, or that transitively references a
GlobalValue is specific to a Module and is effectively owned by it. More
specifically, some metadata is effectively owned by a specific Function
within a module.
Effectively function-local metadata was introduced somewhere around
c10d0e5ccd, which made it illegal for two
functions to share a DISubprogram attachment.
When cloning a function within a module, you need to clone the
function-local debug info and suppress cloning of global debug info (the
status quo suppresses cloning some global debug info but not all). When
cloning a function to a new/different module, you need to clone all of
the debug info.
Here's what I think we should do (eventually? soon? not this patch
though):
- Distinguish explicitly (somehow) between pure constant metadata owned
by the LLVMContext, global metadata owned by the Module, and local
metadata owned by a GlobalValue (such as a function).
- Update CloneFunctionInto to trigger cloning of all "local" metadata
(only), perhaps by adding a bit to RemapFlag. Alternatively, split
out a separate function CloneFunctionMetadataInto to prime the
metadata map that callers are updated to call ahead of time as
appropriate.
Here's the somewhat more isolated fix in this patch:
- Converted the `ModuleLevelChanges` parameter to `CloneFunctionInto` to
an enum called `CloneFunctionChangeType` that is one of
LocalChangesOnly, GlobalChanges, DifferentModule, and ClonedModule.
- The code maintaining the "functions uniquely own subprograms"
invariant is now only active in the first two cases, where a function
is being cloned within a single module. That's necessary because this
code inhibits cloning of (some) "global" metadata that's effectively
owned by the module.
- The code maintaining the "all compile units must be explicitly
referenced by !llvm.dbg.cu" invariant is now only active in the
DifferentModule case, where a function is being cloned into a new
module in isolation.
- CoroSplit.cpp's call to CloneFunctionInto in CoroCloner::create
uses LocalChangeOnly, since fa635d730f
only set `ModuleLevelChanges` to trigger cloning of local metadata.
- CloneModule drops its unsound use of RF_ReuseAndMutateDistinctMDs
and special handling of !llvm.dbg.cu.
- Fixed some outdated header docs and left a couple of FIXMEs.
Differential Revision: https://reviews.llvm.org/D96531
We are using AtomicNoRet map in multiple places to determine
if an instruction atomic, rtn or nortn atomic. This method
does not work always since we have some instructions which
only has rtn or nortn version.
One such instruction is ds_wrxchg_rtn_b32 which does not have
nortn version. This has caused changes in memory legalizer
tests.
Differential Revision: https://reviews.llvm.org/D96639
Add intrinsic which demotes all active lanes to helper lanes.
This is used to implement demote to helper Vulkan extension.
In practice demoting a lane to helper simply means removing it
from the mask of live lanes used for WQM/WWM/Exact mode.
Where the shader does not use WQM, demotes just become kills.
Additionally add llvm.amdgcn.live.mask intrinsic to complement
demote operations. In theory llvm.amdgcn.ps.live can be used
to detect helper lanes; however, ps.live can be moved by LICM.
The movement of ps.live cannot be remedied without changing
its type signature and such a change would require ps.live
users to update as well.
Reviewed By: piotr
Differential Revision: https://reviews.llvm.org/D94747
These two instructions are VOP3P and have op_sel_hi bits,
however do not use op_sel_hi. That is recommended to set
unused op_sel_hi bits to 1. However, we cannot decode
both representations with 1 and 0 if bits are set to
default value 1. If bits are set to be ignored with '?'
initializer then encoding defaults them to 0.
The patch is a hack to force ignored '?' bits to 1 on
encoding for these instructions.
There is still canonicalization happens on disasm print
if incoming values are non-default, so that disasm output
does not match binary input, but this is pre-existing
problem for all instructions with '?' bits.
Fixes: SWDEV-272540
Differential Revision: https://reviews.llvm.org/D96543
Allow different GICustomOperandRenderers to use the same RendererFn.
This avoids the need for targets to define a bunch of identical C++
renderer functions with different names.
Without this fix TableGen would have emitted code that tried to define
the GICR enumeration with duplicate enumerators.
Differential Revision: https://reviews.llvm.org/D96587
If we have an instruction where more than one pointer operands
are derived from the same promoted alloca, we are fixing it for
one argument and do not fix a second use considering this user
done.
Fix this by deferring processing of memory intrinsics until all
potential operands are replaced.
Fixes: SWDEV-271358
Differential Revision: https://reviews.llvm.org/D96386
Fixes a testcase that was overcommitting large register tuples to a
bundle, which the register allocator could not possibly satisfy. This
was producing a bundle which used nearly all of the available SGPRs
with a series of 16-dword loads (not all of which are freely available
to use).
This is a quick hack for some deeper issues with how the clause
bundler tracks register pressure.
Overall the pressure tracking used here doesn't make sense and is too
imprecise for what it needs to avoid the allocator failing. The
pressure estimate does not account for the alignment requirements of
large SGPR tuples, so this was really underestimating the pressure
impact. This also ignores the impact of the extended live range of the
use registers after the bundle is introduced. Additionally, it didn't
account for some wide tuples not being available due to reserved
registers.
This regresses a few cases. These end up introducing more
spilling. This is also a function of the global pressure being used in
the decision to bundle, not the local pressure impact of the bundle
itself.
When merging a pair of DS reads or writes needs to materialize the base
offset in a vgpr, choose a value that is aligned to as high a power of
two as possible. This maximises the chance that different pairs can use
the same base offset, in which case the base offset registers can be
commoned up by MachineCSE.
Differential Revision: https://reviews.llvm.org/D96421
Move implementation of kill intrinsics to WQM pass. Add live lane
tracking by updating a stored exec mask when lanes are killed.
Use live lane tracking to enable early termination of shader
at any point in control flow.
Reviewed By: piotr
Differential Revision: https://reviews.llvm.org/D94746
Add mimgopc object to represent the opcode allowing different
opcodes for different hardware variants.
This enables image_atomic_fcmpswap, image_atomic_fmin, and
image_atomic_fmax on GFX10
Reviewed By: foad, rampitec
Differential Revision: https://reviews.llvm.org/D96309
We need to avoid setting the kill flag on the CSR spill if there's an
additional use of the register after the spill.
This does rely on consistency between the entry block liveins and the
MRI's function live ins, which is not something the verifier checks
now.
This was taking the calling convention from the parent function,
instead of the callee. Avoids regressions in a future patch when the
caller and callee have different type breakdowns.
For some reason AArch64's lowerFormalArguments seems to intentionally
ignore the parent isVarArg.
This reverts commit 502a67dd7f.
This expose a failure in test-suite build on PowerPC,
revert to unblock buildbot first,
Dave will re-commit in https://reviews.llvm.org/D96287.
Thanks Dave.
The generated calling convention code shouldn't see these types since
we split large types into 32-bit chunks before the calling convention
code is triggered.
GlobalISel ends up directly calls the generated CC code before
checking for the register count breakdown. Arguably this difference is
a bug, but this was dead code for the DAG anyway.
When running the tests on PowerPC and x86, the lit test GlobalISel/trunc.ll fails at the memory sanitize step. This seems to be due to wrong invalid logic (which matches even if it shouldn't) and likely missing variable initialisation."
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D95878
getIntrinsicInstrCost takes a IntrinsicCostAttributes holding various
parameters of the intrinsic being costed. It can either be called with a
scalar intrinsic (RetTy==Scalar, VF==1), with a vector instruction
(RetTy==Vector, VF==1) or from the vectorizer with a scalar type and
vector width (RetTy==Scalar, VF>1). A RetTy==Vector, VF>1 is considered
an error. Both of the vector modes are expected to be treated the same,
but because this is confusing many backends end up getting it wrong.
Instead of trying work with those two values separately this removes the
VF parameter, widening the RetTy/ArgTys by VF used called from the
vectorizer. This keeps things simpler, but does require some other
modifications to keep things consistent.
Most backends look like this will be an improvement (or were not using
getIntrinsicInstrCost). AMDGPU needed the most changes to keep the code
from c230965ccf working. ARM removed the fix in
dfac521da1, webassembly happens to get a fixup for an SLP cost
issue and both X86 and AArch64 seem to now be using better costs from
the vectorizer.
Differential Revision: https://reviews.llvm.org/D95291
As mentioned in TODO comment, casting double to float causes NaNs to change bits.
To avoid the change, this patch adds support for single-floating-point immediate value on MachineCode.
Patch by Yuta Saito.
Differential Revision: https://reviews.llvm.org/D77384
When widening, each half of the v2s16 operands needs to be sign extended
for G_ASHR or zero extended for G_LSHR.
Differential Revision: https://reviews.llvm.org/D96048
SALU min/max s32 instructions exist so use them. This means that
regbankselect can handle min/max much like add/sub/mul/shifts.
Differential Revision: https://reviews.llvm.org/D96047
If amdgpu-unsafe-fp-atomics is specified, allow {flat|global}_atomic_add_f32 even if atomic modes don't match.
Differential Revision: https://reviews.llvm.org/D95391
When SGPRs are spilled to VGPRs, they can overwrite any lane. We need
to preserve the value of inactive lanes in function calls, so we save
the register even if it is marked as caller saved.
Also, teach buildPrologSpill to work when no registers are free like in
CodeGen/AMDGPU/pei-scavenge-vgpr-spill.mir and update the comment on
findScratchNonCalleeSaveRegister as it is not used anymore to realign
the stack pointer since D95865.
Differential Revision: https://reviews.llvm.org/D95946
For the fixed ABI, set this in the initial argument constructor,
rather than relying on the allocation logic to set the values. Also
stop passing them for amdgpu_gfx, since the DAG path seems to skip
these. I'm unclear on what amdgpu_gfx's expectations are. This will
allow moving the special input registers out of the normal argument
range.
This reverts commits 62af0305b7cc..677a3529d3e6 from D93708.
They cause failures in the sanitizer builds because of uninitialized
values.
A fix is in D95878, but it might take some time until this is pushed,
so reverting the changes for now.
We don't register i128 as a legal type with addRegisterClass, but it
appears in the list of legal register types. This inconsistency
resulted in the asm constraint lowering trying to use 2 128-bit
registers for these operands. This would leave behind a dead def that
would waste registers.
Regresses GlobalISel tests for i128 load/store, but these aren't very
important right now. Ideally these would not depend on the list of
register types.
This should only consider whether the pressure impact of the bundle at
the given point in the program will decrease the occupancy. High VGPR
pressure was incorrectly blocking the formation of scalar bundles, and
vice versa. This was also blocking bundling from high pressure
situations at other points in the program.
The temporary register is only used to compute the frame pointer.
The frame pointer is overwritten and not used in between, so we
can reuse the frame pointer for the computation, saving one register.
Differential Revision: https://reviews.llvm.org/D95865
Saving callee-save registers happens in whole wave mode. Exec is saved
to a free register, which can be reused to save the frame pointer.
Therefore, saving the fp needs to happen after saving csrs.
Differential Revision: https://reviews.llvm.org/D95861
I guess instructions were marked as frame-setup by accident, they are
restores as part of the epilog.
Differential Revision: https://reviews.llvm.org/D95783
Simon Pilgrim