Summary:
With scalar stores, M0 is clobbered and therefore marked as implicitly
defined. However, it is also dead.
This fixes an assertion when the Greedy Register Allocator decides to
optimize a spill/restore pair away again (via tryHintsRecoloring).
Reviewers: arsenm
Subscribers: qcolombet, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D33319
llvm-svn: 306375
Summary:
Added support based on merged SDWA pseudo instructions. Now peephole allow one scalar operand, omod and clamp modifiers.
Added several subtarget features for GFX9 SDWA.
This diff also contains changes from D34026.
Depends D34026
Reviewers: vpykhtin, rampitec, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D34241
llvm-svn: 305986
Summary: Previously there were two separate pseudo instruction for SDWA on VI and on GFX9. Created one pseudo instruction that is union of both of them. Added verifier to check that operands conform either VI or GFX9.
Reviewers: dp, arsenm, vpykhtin
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, artem.tamazov
Differential Revision: https://reviews.llvm.org/D34026
llvm-svn: 305886
I did this a long time ago with a janky python script, but now
clang-format has built-in support for this. I fed clang-format every
line with a #include and let it re-sort things according to the precise
LLVM rules for include ordering baked into clang-format these days.
I've reverted a number of files where the results of sorting includes
isn't healthy. Either places where we have legacy code relying on
particular include ordering (where possible, I'll fix these separately)
or where we have particular formatting around #include lines that
I didn't want to disturb in this patch.
This patch is *entirely* mechanical. If you get merge conflicts or
anything, just ignore the changes in this patch and run clang-format
over your #include lines in the files.
Sorry for any noise here, but it is important to keep these things
stable. I was seeing an increasing number of patches with irrelevant
re-ordering of #include lines because clang-format was used. This patch
at least isolates that churn, makes it easy to skip when resolving
conflicts, and gets us to a clean baseline (again).
llvm-svn: 304787
1. RegisterClass::getSize() is split into two functions:
- TargetRegisterInfo::getRegSizeInBits(const TargetRegisterClass &RC) const;
- TargetRegisterInfo::getSpillSize(const TargetRegisterClass &RC) const;
2. RegisterClass::getAlignment() is replaced by:
- TargetRegisterInfo::getSpillAlignment(const TargetRegisterClass &RC) const;
This will allow making those values depend on subtarget features in the
future.
Differential Revision: https://reviews.llvm.org/D31783
llvm-svn: 301221
Summary:
Fix a compiler bug when the lane select happens to end up in a VGPR.
Clarify the semantic of the corresponding intrinsic to be that of
the corresponding GLSL: the lane select must be uniform across a
wave front, otherwise results are undefined.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D32343
llvm-svn: 301197
This is possible in ways that are not compiler bugs,
so stop asserting on them.
This emits an extra error when emitting objects when it
can't encode the new pseudo, but I'm not sure that matters.
llvm-svn: 299712
Previously compiler often extracted common immediates into specific register, e.g.:
```
%vreg0 = S_MOV_B32 0xff;
%vreg2 = V_AND_B32_e32 %vreg0, %vreg1
%vreg4 = V_AND_B32_e32 %vreg0, %vreg3
```
Because of this SDWA peephole failed to find SDWA convertible pattern. E.g. in previous example this could be converted into 2 SDWA src operands:
```
SDWA src: %vreg2 src_sel:BYTE_0
SDWA src: %vreg4 src_sel:BYTE_0
```
With this change peephole check if operand is either immediate or register that is copy of immediate.
llvm-svn: 299202
As we introduced target triple environment amdgiz and amdgizcl, the address
space values are no longer enums. We have to decide the value by target triple.
The basic idea is to use struct AMDGPUAS to represent address space values.
For address space values which are not depend on target triple, use static
const members, so that they don't occupy extra memory space and is equivalent
to a compile time constant.
Since the struct is lightweight and cheap, it can be created on the fly at
the point of usage. Or it can be added as member to a pass and created at
the beginning of the run* function.
Differential Revision: https://reviews.llvm.org/D31284
llvm-svn: 298846
StructurizeCFG can't handle cases with multiple
returns creating regions with multiple exits.
Create a copy of UnifyFunctionExitNodes that only
unifies exit nodes that skips exit nodes
with uniform branch sources.
llvm-svn: 298729
Since v_max_f32_e64/v_max_f16_e64 can be folded if the target
instruction supports the clamp bit, we also need to maintain
modifiers when converting v_mac to v_mad.
This fixes a rendering issue with Dirt Rally because a v_mac
instruction with the clamp bit set was converted to a v_mad
but that bit was lost during the conversion.
Fixes: e184e01dd79 ("AMDGPU: Fold FP clamp as modifier bit")
Patch by Samuel Pitoiset <samuel.pitoiset@gmail.com>
llvm-svn: 297556
Before frame offsets are calculated, try to eliminate the
frame indexes used by SGPR spills. Then we can delete them
after.
I think for now we can be sure that no other instruction
will be re-using the same frame indexes. It should be easy
to notice if this assumption ever breaks since everything
asserts if it tries to use a dead frame index later.
The unused emergency stack slot seems to still be left behind,
so an additional 4 bytes is still wasted.
llvm-svn: 295753
The operand types were defined to fit the fp16_to_fp node, which
has the half as an integer type. v_cvt_f32_f16 does support
source modifiers, so change this to have an FP type and modifiers.
For targets without legal f16, this requires recognizing the
bit operations and trying to produce them.
llvm-svn: 293857
Leave early ifcvt disabled for now since there are some
shader-db regressions.
This causes some immediate improvements, but could be better.
The cost checking that the pass does is based on critical path
length for out of order CPUs which we do not want so it skips out
on many cases we want.
llvm-svn: 293016
Inline spiller can decide to move a spill as early as possible in the basic block.
It will skip phis and label, but we also need to make sure it skips instructions
in the basic block prologue which restore exec mask.
Added isPositionLike callback in TargetInstrInfo to detect instructions which
shall be skipped in addition to common phis, labels etc.
Differential Revision: https://reviews.llvm.org/D27997
llvm-svn: 292554
Rename from addOperand to just add, to match the other method that has been
added to MachineInstrBuilder for adding more than just 1 operand.
See https://reviews.llvm.org/D28057 for the whole discussion.
Differential Revision: https://reviews.llvm.org/D28556
llvm-svn: 291891
When the instruction is processed the first time, it may be
deleted resulting in crashes. While the new test adds the same
user to the worklist twice, this particular case doesn't crash
but I'm not sure why.
llvm-svn: 290191
Since 32-bit instructions with 32-bit input immediate behavior
are used to materialize 16-bit constants in 32-bit registers
for 16-bit instructions, determining the legality based
on the size is incorrect. Change operands to have the size
specified in the type.
Also adds a workaround for a disassembler bug that
produces an immediate MCOperand for an operand that
is supposed to be OPERAND_REGISTER.
The assembler appears to accept out of bounds immediates and
truncates them, but this seems to be an issue for 32-bit
already.
llvm-svn: 289306
Use vaddr/vdst for the same purposes.
This also fixes a beg in SIInsertWaits for the
operand check. The stored value operand is currently called
data0 in the single offset case, not data.
llvm-svn: 288188
m0 may need to be written for spill code, so
we don't want general code uses relying on the
value stored in it.
This introduces a few code quality regressions where copies
from m0 are not coalesced into copies of a copy of m0.
llvm-svn: 287841
Summary:
The addr64-based legalization is incorrect for MUBUF instructions with idxen
set as well as for BUFFER_LOAD/STORE_FORMAT_* instructions. This affects
e.g. shaders that access buffer textures.
Since we never actually need the addr64-legalization in shaders, this patch
takes the easy route and keys off the calling convention. If this ever
affects (non-OpenGL) compute, the type of legalization needs to be chosen
based on some TSFlag.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98664
Reviewers: arsenm, tstellarAMD
Subscribers: kzhuravl, wdng, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D26747
llvm-svn: 287339
Summary:
1. Don't try to copy values to and from the same register class.
2. Replace copies with of registers with immediate values with v_mov/s_mov
instructions.
The main purpose of this change is to make MachineSink do a better job of
determining when it is beneficial to split a critical edge, since the pass
assumes that copies will become move instructions.
This prevents a regression in uniform-cfg.ll if we enable critical edge
splitting for AMDGPU.
Reviewers: arsenm
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: https://reviews.llvm.org/D23408
llvm-svn: 287131
The wave barrier represents the discardable barrier. Its main purpose is to
carry convergent attribute, thus preventing illegal CFG optimizations. All lanes
in a wave come to convergence point simultaneously with SIMT, thus no special
instruction is needed in the ISA. The barrier is discarded during code generation.
Differential Revision: https://reviews.llvm.org/D26585
llvm-svn: 287007
nThis avoids the nasty problems caused by using
memory instructions that read the exec mask while
spilling / restoring registers used for control flow
masking, but only for VI when these were added.
This always uses the scalar stores when enabled currently,
but it may be better to still try to spill to a VGPR
and use this on the fallback memory path.
The cache also needs to be flushed before wave termination
if a scalar store is used.
llvm-svn: 286766
If the branch was on a read-undef of vcc, passes that used
analyzeBranch to invert the branch condition wouldn't preserve
the undef flag resulting in a verifier error.
Fixes verifier failures in a future commit.
Also fix verifier error when inserting copy for vccz
corruption bug.
llvm-svn: 286133
Summary:
The post-RA scheduler occasionally uses additional implicit operands when
the vector implicit operand as a whole is killed, but some subregisters
are still live because they are directly referenced later. Unfortunately,
this seems incredibly subtle to reproduce.
Fixes piglit spec/glsl-110/execution/variable-indexing/vs-temp-array-mat2-index-wr.shader_test
and others.
Reviewers: arsenm, tstellarAMD
Subscribers: kzhuravl, wdng, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D25656
llvm-svn: 285835
This is the conservatively correct way because it's easy to
move or replace a scalar immediate. This was incorrect in the case
when the register class wasn't known from the static instruction
definition, but still needed to be an SGPR. The main example of this
is inlineasm has an SGPR constraint.
Also start verifying the register classes of inlineasm operands.
llvm-svn: 285762
Instructions with a 32-bit base encoding with an optional
32-bit literal encoded after them report their size as 4
for the disassembler. Consider these when computing the
MachineInstr size. This fixes problems caused by size estimate
consistency in BranchRelaxation.
llvm-svn: 285743
Summary:
Flat instruction can return out of order, so we need always need to wait
for all the outstanding flat operations.
Reviewers: tony-tye, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, llvm-commits, yaxunl
Differential Revision: https://reviews.llvm.org/D25998
llvm-svn: 285479
Also add glc bit to the scalar loads since they exist on VI
and change the caching behavior.
This currently has an assembler bug where the glc bit is incorrectly
accepted on SI/CI which do not have it.
llvm-svn: 285463
It's possible to have a use of the private resource descriptor or
scratch wave offset registers even though there are no allocated
stack objects. This would result in continuing to use the maximum
number reserved registers. This could go over the number of SGPRs
available on VI, or violate the SGPR limit requested by
the function attributes.
llvm-svn: 285435
Summary:
The v_movreld machine instruction is used with three operands that are
in a sense tied to each other (the explicit VGPR_32 def and the implicit
VGPR_NN def and use). There is no way to express that using the currently
available operand bits, and indeed there are cases where the Two Address
instructions pass does the wrong thing.
This patch introduces a new set of pseudo instructions that are identical
in intended semantics as v_movreld, but they only have two tied operands.
Having to add a new set of pseudo instructions is admittedly annoying, but
it's a fairly straightforward and solid approach. The only alternative I
see is to try to teach the Two Address instructions pass about Three Address
instructions, and I'm afraid that's trickier and is going to end up more
fragile.
Note that v_movrels does not suffer from this problem, and so this patch
does not touch it.
This fixes several GL45-CTS.shaders.indexing.* tests.
Reviewers: tstellarAMD, arsenm
Subscribers: kzhuravl, wdng, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D25633
llvm-svn: 284980
This is the most basic handling of the indirect access
pseudos using GPR indexing mode. This currently only enables
the mode for a single v_mov_b32 and then disables it.
This is much more complicated to use than the movrel instructions,
so a new optimization pass is probably needed to fold the access
into the uses and keep the mode enabled for them.
llvm-svn: 284031
For some reason there are both of these available, except
for scalar 64-bit compares which only has u64. I'm not sure
why there are both (I'm guessing it's for the one bit inputs we
don't use), but for consistency always using the
unsigned one.
llvm-svn: 282832
Fixes to allow spilling all registers at the end of the block
work with exec modifications. Don't emit s_and_saveexec_b64 for
if lowering, and instead emit copies. Mark control flow mask
instructions as terminators to get correct spill code placement
with fast regalloc, and then have a separate optimization pass
form the saveexec.
This should work if SGPRs are spilled to VGPRs, but
will likely fail in the case that an SGPR spills to memory
and no workitem takes a divergent branch.
llvm-svn: 282667
If the literal is being folded into src0, it doesn't matter
if it's an SGPR because it's being replaced with the literal.
Also fixes initially selecting 32-bit versions of some instructions
which also confused commuting.
llvm-svn: 281117
Summary:
Prevously assembler parsed all literals as either 32-bit integers or 32-bit floating-point values. Because of this we couldn't support f64 literals.
E.g. in instruction "v_fract_f64 v[0:1], 0.5", literal 0.5 was encoded as 32-bit literal 0x3f000000, which is incorrect and will be interpreted as 3.0517578125E-5 instead of 0.5. Correct encoding is inline constant 240 (optimal) or 32-bit literal 0x3FE00000 at least.
With this change the way immediate literals are parsed is changed. All literals are always parsed as 64-bit values either integer or floating-point. Then we convert parsed literals to correct form based on information about type of operand parsed (was literal floating or binary) and type of expected instruction operands (is this f32/64 or b32/64 instruction).
Here are rules how we convert literals:
- We parsed fp literal:
- Instruction expects 64-bit operand:
- If parsed literal is inlinable (e.g. v_fract_f64_e32 v[0:1], 0.5)
- then we do nothing this literal
- Else if literal is not-inlinable but instruction requires to inline it (e.g. this is e64 encoding, v_fract_f64_e64 v[0:1], 1.5)
- report error
- Else literal is not-inlinable but we can encode it as additional 32-bit literal constant
- If instruction expect fp operand type (f64)
- Check if low 32 bits of literal are zeroes (e.g. v_fract_f64 v[0:1], 1.5)
- If so then do nothing
- Else (e.g. v_fract_f64 v[0:1], 3.1415)
- report warning that low 32 bits will be set to zeroes and precision will be lost
- set low 32 bits of literal to zeroes
- Instruction expects integer operand type (e.g. s_mov_b64_e32 s[0:1], 1.5)
- report error as it is unclear how to encode this literal
- Instruction expects 32-bit operand:
- Convert parsed 64 bit fp literal to 32 bit fp. Allow lose of precision but not overflow or underflow
- Is this literal inlinable and are we required to inline literal (e.g. v_trunc_f32_e64 v0, 0.5)
- do nothing
- Else report error
- Do nothing. We can encode any other 32-bit fp literal (e.g. v_trunc_f32 v0, 10000000.0)
- Parsed binary literal:
- Is this literal inlinable (e.g. v_trunc_f32_e32 v0, 35)
- do nothing
- Else, are we required to inline this literal (e.g. v_trunc_f32_e64 v0, 35)
- report error
- Else, literal is not-inlinable and we are not required to inline it
- Are high 32 bit of literal zeroes or same as sign bit (32 bit)
- do nothing (e.g. v_trunc_f32 v0, 0xdeadbeef)
- Else
- report error (e.g. v_trunc_f32 v0, 0x123456789abcdef0)
For this change it is required that we know operand types of instruction (are they f32/64 or b32/64). I added several new register operands (they extend previous register operands) and set operand types to corresponding types:
'''
enum OperandType {
OPERAND_REG_IMM32_INT,
OPERAND_REG_IMM32_FP,
OPERAND_REG_INLINE_C_INT,
OPERAND_REG_INLINE_C_FP,
}
'''
This is not working yet:
- Several tests are failing
- Problems with predicate methods for inline immediates
- LLVM generated assembler parts try to select e64 encoding before e32.
More changes are required for several AsmOperands.
Reviewers: vpykhtin, tstellarAMD
Subscribers: arsenm, kzhuravl, artem.tamazov
Differential Revision: https://reviews.llvm.org/D22922
llvm-svn: 281050
- Implemented amdgpu-flat-work-group-size attribute
- Implemented amdgpu-num-active-waves-per-eu attribute
- Implemented amdgpu-num-sgpr attribute
- Implemented amdgpu-num-vgpr attribute
- Dynamic LDS constraints are in a separate patch
Patch by Tom Stellard and Konstantin Zhuravlyov
Differential Revision: https://reviews.llvm.org/D21562
llvm-svn: 280747
Summary:
I put this code here, because I want to re-use it in a few other places.
This supersedes some of the immediate folding code we have in SIFoldOperands.
I think the peephole optimizers is probably a better place for folding
immediates into copies, since it does some register coalescing in the same time.
This will also make it easier to transition SIFoldOperands into a smarter pass,
where it looks at all uses of instruction at once to determine the optimal way to
fold operands. Right now, the pass just considers one operand at a time.
Reviewers: arsenm
Subscribers: wdng, nhaehnle, arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D23402
llvm-svn: 280744
readlane/writelane do not support using m0 as the output/input.
Constrain the register class of spill vregs to try to avoid this,
but also handle spilling of the physreg when necessary by inserting
an additional copy to a normal SGPR.
llvm-svn: 280584
Summary:
The SILoadStoreOptimizer will need to use AliasAnalysis here in order to
move it before scheduling.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D23813
llvm-svn: 279963
There's only one use of this for the convenience
of a pattern. I think v_mov_b64_pseudo should also be
moved, but SIFoldOperands does currently make use of it.
llvm-svn: 279901
This is a mechanical change of comments in switches like fallthrough,
fall-through, or fall-thru to use the LLVM_FALLTHROUGH macro instead.
llvm-svn: 278902
Summary:
We were using reserved VGPRs for SGPR spilling and this was causing
some programs with a workgroup size of 1024 to use more than 64
registers, which is illegal.
Reviewers: arsenm, mareko, nhaehnle
Subscribers: nhaehnle, arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D22032
llvm-svn: 276980
ABIArgOffset is a problem because properly fsetting the
KernArgSize requires that the reserved area before the
real kernel arguments be correctly aligned, which requires
fixing clover.
llvm-svn: 276766
This is to help moveSILowerControlFlow to before regalloc.
There are a couple of tradeoffs with this. The complete CFG
is visible to more passes, the loop body avoids an extra copy of m0,
vcc isn't required, and immediate offsets can be shrunk into s_movk_i32.
The disadvantage is the register allocator doesn't understand that
the single lane's vector is dead within the loop body, so an extra
register is used to outlive the loop block when expanding the
VGPR -> m0 loop. This also now results in worse waitcnt insertion
before the loop instead of after for pending operations at the point
of the indexing, but that should be fixed by future improvements to
cross block waitcnt insertion.
v_movreld_b32's operands are now modeled more correctly since vdst
is not a true output. This is kind of a hack to treat vdst as a
use operand. Extra checking is required in the verifier since
I can't seem to get tablegen to emit an implicit operand for a
virtual register.
llvm-svn: 275934
In this situation:
%VGPR2<def> = BUFFER_LOAD_DWORD_OFFSET %SGPR8_SGPR9_SGPR10_SGPR11,
%VGPR7<def,tied3> = V_MAC_F32_e32 %VGPR0<undef>, %VGPR1<kill>, %VGPR7<kill,tied0>, %EXEC<imp-use>
%VGPR3_VGPR4_VGPR5_VGPR6<def> = COPY %VGPR0_VGPR1_VGPR2_VGPR3
%VGPR4<def> = COPY %VGPR2
The copy for VGPR1 -> VGPR4 was an error from reading undefined VGPR1,
but VGPR4 is defined immediately after this copy.
llvm-svn: 275635
Remove remaining implicit conversions from MachineInstrBundleIterator to
MachineInstr* from the AMDGPU backend. In most cases, I made them less
attractive by preferring MachineInstr& or using a ranged-based for loop.
Once all the backends are fixed I'll make the operator explicit so that
this doesn't bitrot back.
llvm-svn: 274906
Because of the special immediate operand, the constant
bus is already used so SGPRs are never useful.
r263212 changed the name of the immediate operand, which
broke the verifier check for the restriction.
llvm-svn: 274564
This is mostly a mechanical change to make TargetInstrInfo API take
MachineInstr& (instead of MachineInstr* or MachineBasicBlock::iterator)
when the argument is expected to be a valid MachineInstr. This is a
general API improvement.
Although it would be possible to do this one function at a time, that
would demand a quadratic amount of churn since many of these functions
call each other. Instead I've done everything as a block and just
updated what was necessary.
This is mostly mechanical fixes: adding and removing `*` and `&`
operators. The only non-mechanical change is to split
ARMBaseInstrInfo::getOperandLatencyImpl out from
ARMBaseInstrInfo::getOperandLatency. Previously, the latter took a
`MachineInstr*` which it updated to the instruction bundle leader; now,
the latter calls the former either with the same `MachineInstr&` or the
bundle leader.
As a side effect, this removes a bunch of MachineInstr* to
MachineBasicBlock::iterator implicit conversions, a necessary step
toward fixing PR26753.
Note: I updated WebAssembly, Lanai, and AVR (despite being
off-by-default) since it turned out to be easy. I couldn't run tests
for AVR since llc doesn't link with it turned on.
llvm-svn: 274189
Split AMDGPUSubtarget into amdgcn/r600 specific subclasses.
This removes most of the static_casting of the basic codegen
classes everywhere, and tries to restrict the features
visible on the wrong target.
llvm-svn: 273652
Summary:
We now use a standard fixup type applying the pc-relative address of
constant address space variables, and we have the GlobalAddress lowering
code add the required 4 byte offset to the global address rather than
doing it as part of the fixup.
This refactoring will make it easier to use the same code for global
address space variables and also simplifies the code.
Re-commit this after fixing a bug where we were trying to use a
reference to a Triple object that had already been destroyed.
Reviewers: arsenm, kzhuravl
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: http://reviews.llvm.org/D21154
llvm-svn: 272705
Summary:
We now use a standard fixup type applying the pc-relative address of
constant address space variables, and we have the GlobalAddress lowering
code add the required 4 byte offset to the global address rather than
doing it as part of the fixup.
This refactoring will make it easier to use the same code for global
address space variables and also simplifies the code.
Reviewers: arsenm, kzhuravl
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: http://reviews.llvm.org/D21154
llvm-svn: 272675
Summary:
Mesa and other users must set this to enable coalescing:
- STRIDE = 0
- SWIZZLE_ENABLE = 1
This makes one particular compute shader 8x faster.
Reviewers: tstellarAMD, arsenm
Subscribers: arsenm, kzhuravl
Differential Revision: http://reviews.llvm.org/D21136
llvm-svn: 272556
The condition reg of the cndmask_b64 expansion can't be killed by
the first one, and the implicit super register implicit def is needed.
llvm-svn: 272554
This used to be free, copying and moving DebugLocs became expensive
after the metadata rewrite. Passing by reference eliminates a ton of
track/untrack operations. No functionality change intended.
llvm-svn: 272512
The current SGPR spilling test does not stress this
because it is using s_buffer_load instructions to
increase SGPR pressure and spill, but their output
operands have the same SReg_32_XM0 constraint. This fixes
an error when the SReg_32 output from most instructions
is spilled.
llvm-svn: 270301
We only really need this to be true for SIFixSGPRCopies.
I'm not sure there's any way this could happen before that point.
Fixes a case where MachineCSE could introduce a cross block
scc use.
llvm-svn: 269391
Summary:
This includes a hazard recognizer implementation to replace some of
the hazard handling we had during frame index elimination.
Reviewers: arsenm
Subscribers: qcolombet, arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D18602
llvm-svn: 268143
Summary:
These instructions can add an immediate offset to the address, like other
ds instructions.
Reviewers: arsenm
Subscribers: arsenm, scchan
Differential Revision: http://reviews.llvm.org/D19233
llvm-svn: 268043
Summary:
The expression is detected as a redundant expression.
Turn out, this is probably a bug.
```
/home/etienneb/llvm/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp:306:26: warning: both side of operator are equivalent [misc-redundant-expression]
if (isSMRD(*FirstLdSt) && isSMRD(*FirstLdSt)) {
```
Reviewers: rnk, tstellarAMD
Subscribers: arsenm, cfe-commits
Differential Revision: http://reviews.llvm.org/D19460
llvm-svn: 267415
Summary:
This intrinsic returns true if the current thread belongs to a live pixel
and false if it belongs to a pixel that we are executing only for derivative
computation. It will be used by Mesa to implement gl_HelperInvocation.
Note that for pixels that are killed during the shader, this implementation
also returns true, but it doesn't matter because those pixels are always
disabled in the EXEC mask.
This unearthed a corner case in the instruction verifier, which complained
about a v_cndmask 0, 1, exec, exec<imp-use> instruction. That's stupid but
correct code, so make the verifier accept it as such.
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D19191
llvm-svn: 267102
Summary:
The added testcase, which triggered this, was derived from a shader-db case
via bugpoint. A separate question is why scalar branching wasn't used.
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D19208
llvm-svn: 266825
Perform store clustering just like load clustering. This change add
StoreClusterMutation in machine-scheduler. To control StoreClusterMutation,
added enableClusterStores() in TargetInstrInfo.h. This is enabled only on
AArch64 for now.
This change also add support for unscaled stores which were not handled in
getMemOpBaseRegImmOfs().
llvm-svn: 266437
Summary:
It seems like this was broken in r252327. I thought we had test cases
for this, but it's really hard to tirgger spills of this exact register
size since they aren't used very much.
Reviewers: arsenm, nhaehnle
Subscribers: nhaehnle, arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D19021
llvm-svn: 266152
Summary: This makes it possible to insert nops at the end of blocks.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D18549
llvm-svn: 265678
This makes it possible to distinguish between mesa shaders
and other kernels even in the presence of compute shaders.
Patch By: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Differential Revision: http://reviews.llvm.org/D18559
llvm-svn: 265589
Summary:
This helps prevent load clustering from drastically increasing register
pressure by trying to cluster 4 SMRDx8 loads together. The limit of 16
bytes was chosen, because it seems like that was the original intent
of setting the limit to 4 instructions, but more analysis could show
that a different limit is better.
This fixes yields small decreases in register usage with shader-db, but
also helps avoid a large increase in register usage when lane mask
tracking is enabled in the machine scheduler, because lane mask tracking
enables more opportunities for load clustering.
shader-db stats:
2379 shaders in 477 tests
Totals:
SGPRS: 49744 -> 48600 (-2.30 %)
VGPRS: 34120 -> 34076 (-0.13 %)
Code Size: 1282888 -> 1283184 (0.02 %) bytes
LDS: 28 -> 28 (0.00 %) blocks
Scratch: 495616 -> 492544 (-0.62 %) bytes per wave
Max Waves: 6843 -> 6853 (0.15 %)
Wait states: 0 -> 0 (0.00 %)
Reviewers: nhaehnle, arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D18451
llvm-svn: 264589
Summary:
Whole quad mode is already enabled for pixel shaders that compute
derivatives, but it must be suspended for instructions that cause a
shader to have side effects (i.e. stores and atomics).
This pass addresses the issue by storing the real (initial) live mask
in a register, masking EXEC before instructions that require exact
execution and (re-)enabling WQM where required.
This pass is run before register coalescing so that we can use
machine SSA for analysis.
The changes in this patch expose a problem with the second machine
scheduling pass: target independent instructions like COPY implicitly
use EXEC when they operate on VGPRs, but this fact is not encoded in
the MIR. This can lead to miscompilation because instructions are
moved past changes to EXEC.
This patch fixes the problem by adding use-implicit operands to
target independent instructions. Some general codegen passes are
relaxed to work with such implicit use operands.
Reviewers: arsenm, tstellarAMD, mareko
Subscribers: MatzeB, arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D18162
llvm-svn: 263982
The constant is now at source operand 1 (previously at 2).
This is also how it is in legacy AMD sp3 assembler.
Update tests.
Differential Revision: http://reviews.llvm.org/D17984
llvm-svn: 263212
Summary:
This is necessary for when we run out of VGPRs and can no
longer use v_{read,write}_lane for spilling SGPRs.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17592
llvm-svn: 262732
Summary:
This patch impleemnts DS_PERMUTE/DS_BPERMUTE instruction definitions and intrinsics,
which are new since VI.
Reviewers: tstellarAMD, arsenm
Subscribers: llvm-commits, arsenm
Differential Revision: http://reviews.llvm.org/D17614
llvm-svn: 262356
Summary:
Instead of trying to replace SMRD instructions with a VGPR base pointer
with an equivalent MUBUF instruction, we now copy the base pointer to
SGPRs using v_readfirstlane.
This is safe to do, because any load selected as an SMRD instruction
has been proven to have a uniform base pointer, so each thread in the
wave will have the same pointer value in VGPRs.
This will fix some errors on VI from trying to replace SMRD instructions
with addr64-enabled MUBUF instructions that don't exist.
Reviewers: arsenm, cfang, nhaehnle
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17305
llvm-svn: 261385
Summary: This change renames output operand for VOP instructions from dst to vdst. This is needed to enable decoding named operands for disassembler.
Reviewers: vpykhtin, tstellarAMD, arsenm
Subscribers: arsenm, llvm-commits, nhaustov
Projects: #llvm-amdgpu-spb
Differential Revision: http://reviews.llvm.org/D16920
llvm-svn: 260986
Introduce a subtarget feature for this, and leave the default with
the current behavior which assumes up to 16-byte loads/stores can
be used. The field also seems to have the ability to be set to 2 bytes,
but I'm not sure what that would be used for.
llvm-svn: 260651
Summary:
It's possible to have resource descriptors and samplers stored in
VGPRs, either by a VMEM instruction or in the case of samplers,
floating-point calculations. When this happens, we need to use
v_readfirstlane to copy these values back to sgprs.
Reviewers: mareko, arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17102
llvm-svn: 260599
Summary:
When we split SMRD instructions into two MUBUFs we were adding the users
of the newly created MUBUFs to the VALU worklist. However, the only
users these instructions had was the REG_SEQUENCE that was inserted
by splitSMRD when the original SMRD instruction was split.
We need to make sure to add the users of the original SMRD to the VALU
worklist before it is split.
I have a test case, but it requires one other bug fix, so it will be
added in a later commt.
Reviewers: mareko, arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17101
llvm-svn: 260588
If the two operands to an instruction were both
subregisters of the same super register, it would incorrectly
think this counted as the same constant bus use.
This fixes the verifier error in fmin_legacy.ll which
was missing -verify-machineinstrs.
llvm-svn: 260495
Summary:
Also delete all the stub functions that are identical to the
implementations in TargetInstrInfo.cpp.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D16609
llvm-svn: 259054
Summary:
It is off by default, but can be used
with --misched=si
Patch by: Axel Davy
Reviewers: arsenm, tstellarAMD, nhaehnle
Subscribers: nhaehnle, solenskiner, arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D11885
llvm-svn: 257609
Summary:
Multi-dword constant loads generated unnecessary moves from SGPRs into VGPRs,
increasing the code size and VGPR pressure. These moves are now folded away.
Note that this lack of operand folding was not a problem for VMEM loads,
because COPY nodes from VReg_Nnn to VGPR32 are eliminated by the register
coalescer.
Some tests are updated, note that the fsub.ll test explicitly checks that
the move is elided.
With the IR generated by current Mesa, the changes are obviously relatively
minor:
7063 shaders in 3531 tests
Totals:
SGPRS: 351872 -> 352560 (0.20 %)
VGPRS: 199984 -> 200732 (0.37 %)
Code Size: 9876968 -> 9881112 (0.04 %) bytes
LDS: 91 -> 91 (0.00 %) blocks
Scratch: 1779712 -> 1767424 (-0.69 %) bytes per wave
Wait states: 295164 -> 295337 (0.06 %)
Totals from affected shaders:
SGPRS: 65784 -> 66472 (1.05 %)
VGPRS: 38064 -> 38812 (1.97 %)
Code Size: 1993828 -> 1997972 (0.21 %) bytes
LDS: 42 -> 42 (0.00 %) blocks
Scratch: 795648 -> 783360 (-1.54 %) bytes per wave
Wait states: 54026 -> 54199 (0.32 %)
Reviewers: tstellarAMD, arsenm, mareko
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15875
llvm-svn: 257074
Summary:
When copying aggregate registers within the same register class, there may
be an overlap between source and destination that forces us to do the copy
backwards.
Do the simplest possible thing that guarantees the correct order of moves
when there are overlaps, and does whatever when there is no overlap. (The
last part forces some trivial adjustments to test cases.)
Together with r255906, this fixes a VM fault in Unreal Elemental Demo.
While at it, change the generation of kill and def flags to something that
looks more reasonable. This method is used very late during compilation, so
it probably doesn't matter in practice, and to be honest, I don't know if
this change is actually correct because the semantics in connection with
aggregate registers vs. sub-registers are not clear to me.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93264
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15622
llvm-svn: 256072
Summary:
The method insertNOPs expected the number of wait states to be passed as
parameter, while eliminateFrameIndex passed the immediate argument for the
S_NOP, leading to an off-by-one error. Rename the method to make the
meaning of its parameter clearer. The number of 4 / 5 wait states (which
is what the method has always _tried_ to do according to the comment) is
correct according to the hardware docs.
I stumbled upon this while trying to track down the cause of
https://bugs.freedesktop.org/show_bug.cgi?id=93264. While clearly needed,
this patch unfortunately does not fix that bug...
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15542
llvm-svn: 255906
Summary:
This allows us to remove the END_OF_TEXT_LABEL hack we had been using
and simplifies the fixups used to compute the address of constant
arrays.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D15257
llvm-svn: 255204
Don't use commuteInstruction, and don't commute if
doing so will not improve legality. Skip the more
complex checks for literal operands and constant bus restrictions,
which are not a concern for VOP2 instructions because src1
does not accept SGPRs or constants and few implicitly
read vcc.
This gets called quite a few times and the
attempts at commuting are a significant fraction
of the time spent in SIFixSGPRCopies, so it's
somewhat worthwhile to optimize. With this patch and others
leading up to it, this reduces the compile time of SIFixSGPRCopies
on some of the LuxMark 2 kernels from ~8ms to ~5ms on my system.
llvm-svn: 254452
If we know we have stack objects, we reserve the registers
that the private buffer resource and wave offset are passed
and use them directly.
If not, reserve the last 5 SGPRs just in case we need to spill.
After register allocation, try to pick the next available registers
instead of the last SGPRs, and then insert copies from the inputs
to the reserved registers in the progloue.
This also only selectively enables all of the input registers
which are really required instead of always enabling them.
llvm-svn: 254331
It does not work because of emergency stack slots.
This pass was supposed to eliminate dummy registers for the
spill instructions, but the register scavenger can introduce
more during PrologEpilogInserter, so some would end up
left behind if they were needed.
The potential for spilling the scratch resource descriptor
and offset register makes doing something like this
overly complicated. Reserve registers to use for the resource
descriptor and use them directly in eliminateFrameIndex.
Also removes creating another scratch resource descriptor
when directly selecting scratch MUBUF instructions.
The choice of which registers are reserved is temporary.
For now it attempts to pick the next available registers
after the user and system SGPRs.
llvm-svn: 254329
v2: added more tests, moved the SALU->VALU conversion to a separate function
It looks like it's not possible to get subregisters in the S_ABS lowering
code, and I don't feel like guessing without testing what the correct code
would look like.
llvm-svn: 254095
The operand layout is slightly different for the atomic
opcodes from the usual MUBUF loads and stores.
This should only fix it on SI/CI. VI is still broken
because it still emits the addr64 replacement.
llvm-svn: 252140
This was checking for a variety of situations that should
never happen. This saves a tiny bit of compile time.
We should not be selecting instructions with invalid operands in the
first place. Most of the time for registers copys are inserted
to the correct operand register class.
For VOP3, since all operand types are supported and literal
constants never are, we just need to verify the constant bus
requirements (all immediates should be legal inline ones).
The only possibly tricky case to maybe worry about is if when
legalizing operands in moveToVALU with s_add_i32 and similar
instructions. If the original s_add_i32 had a literal constant
and we need to replace it with v_add_i32_e64 we would have an
unsupported literal operand. However, I don't think we should worry
about that because SIFoldOperands should handle folding literal
constant operands into the SALU instructions based on the uses.
At SIFoldOperands time, the legality and profitability of
operand types is a bit different.
llvm-svn: 250951
This stops using an unknown reg class operand.
Currently build_vector selection has a broken looking check
where it tries to use a VGPR reg class and an SGPR one if it
sees an SGPR use.
With the source operand has an explicit VGPR class,
illegal copies will be inserted that SIFixSGPRCopies will take care
of normally later, which will allow removing the weird check
of build_vector users. Without this, when removed v_movrels_b32 would
still be emitted even though all of the values were only stored in
SGPRs.
llvm-svn: 249494
to prevent setting a huge stride, because DATA_FORMAT has a different
meaning if ADD_TID_ENABLE is set.
This is a candidate for stable llvm 3.7.
Tested-and-Reviewed-by: Christian König <christian.koenig@amd.com>
llvm-svn: 248858
The splitting of > 4 dword SMRD instructions
if using an offset in an SGPR instead of an immediate
was not setting the destination register,
resulting an an instruction missing an operand
which would assert later.
Test will be included in a following commit
which fixes a related issue.
llvm-svn: 248739
It's easier to understand creating a full instruction
than the current situation where sometimes a new
instruction is created and sometimes it is awkwardly
mutated in place.
llvm-svn: 248627
When buffer resource descriptors were built, the upper two components
of the descriptor were first composed into a 64-bit register because
legalizeOperands assumed all operands had the same register class.
Fix that problem, but keep the workaround. I'm not sure anything
actually is actually emitting such a REG_SEQUENCE now.
If multiple resource descriptors are set up with different base
pointers, this is copied with a single s_mov_b64. We probably
should fix this better by recognizing a pair of s_mov_b32 later,
but for now delete the dead code.
llvm-svn: 248585
Instead of always inserting a copy in case
the super register is itself a subregister,
only extract to the super reg class if this is
actually the case.
This shouldn't really change codegen, but
makes looking at the output of SIFixSGPRCopies
easier to read.
llvm-svn: 248467
Instead of extracting both 32-bit components from the 128-bit
register. This produces fewer copies and is easier for
the copy peephole optimizer to understand and see the actual uses
as extracts from a reg_sequence.
This avoids needing to handle subregister composing in the
PeepholeOptimizer's ValueTracker for this case.
llvm-svn: 247162
Without a memory operand, mayLoad or mayStore instructions
are treated as hasUnorderedMemRef, which results in much worse
scheduling.
We really should have a verifier check that any
non-side effecting mayLoad or mayStore has a memory operand.
There are a few instructions (interp and images) which I'm
not sure what / where to add these.
llvm-svn: 246356
There is no context where s_mov_b64 is emitted
and could potentially be moved to the VALU.
It is currently only emitted for materializing
immediates, which can't be dependent on vector sources.
The immediate splitting is already done when selecting
constants. I'm not sure what contexts if any the register
splitting would have been used before.
Also clean up using s_mov_b64 in place of v_mov_b64_pseudo,
although this isn't required and just skips the extra step
of eliminating the copy from the SReg_64.
llvm-svn: 246080
When splitting 64-bit operations, create the correct
VALU instructions immediately.
This was splitting things like s_or_b64 into the two
s_or_b32s and then pushing the new instructions
onto the worklist. There's no reason we need
to do this intermediate step.
llvm-svn: 246077
This commit fixes a bug in the class 'SIInstrInfo' where the implicit register
machine operands were added to a machine instruction in an incorrect order -
the implicit uses were added before the implicit defs.
I found this bug while working on moving the implicit register operand
verification code from the MIR parser to the machine verifier.
This commit also makes the method 'addImplicitDefUseOperands' in the machine
instruction class public so that it can be reused in the 'SIInstrInfo' class.
Reviewers: Matt Arsenault
Differential Revision: http://reviews.llvm.org/D11689
llvm-svn: 243799
Summary:
Replace the switch on instruction opcode with a switch on register size.
This way we don't need to update the switch statement when we add new
SMRD variants.
Reviewers: arsenm
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D11601
llvm-svn: 243652
Nicolai Haehnle