Currently runtime metadata is emitted as an ELF section with name .AMDGPU.runtime_metadata.
However there is a standard way to convey vendor specific information about how to run an ELF binary, which is called vendor-specific note element (http://www.netbsd.org/docs/kernel/elf-notes.html).
This patch lets AMDGPU backend emits runtime metadata as a note element in .note section.
Differential Revision: https://reviews.llvm.org/D25781
llvm-svn: 286502
Codegen prepare sinks comparisons close to a user is we have only one register
for conditions. For AMDGPU we have many SGPRs capable to hold vector conditions.
Changed BE to report we have many condition registers. That way IR LICM pass
would hoist an invariant comparison out of a loop and codegen prepare will not
sink it.
With that done a condition is calculated in one block and used in another.
Current behavior is to store workitem's condition in a VGPR using v_cndmask
and then restore it with yet another v_cmp instruction from that v_cndmask's
result. To mitigate the issue a forward propagation of a v_cmp 64 bit result
to an user is implemented. Additional side effect of this is that we may
consume less VGPRs in a cost of more SGPRs in case if holding of multiple
conditions is needed, and that is a clear win in most cases.
llvm-svn: 286171
hange explores the fact that LDS reads may be reordered even if access
the same location.
Prior the change, algorithm immediately stops as soon as any memory
access encountered between loads that are expected to be merged
together. Although, Read-After-Read conflict cannot affect execution
correctness.
Improves hcBLAS CGEMM manually loop-unrolled kernels performance by 44%.
Also improvement expected on any massive sequences of reads from LDS.
Differential Revision: https://reviews.llvm.org/D25944
llvm-svn: 285919
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
This will prevent following regression when enabling i16 support (D18049):
test/CodeGen/AMDGPU/ctlz.ll
test/CodeGen/AMDGPU/ctlz_zero_undef.ll
Differential Revision: https://reviews.llvm.org/D25802
llvm-svn: 285716
I wanted to implement this as a target independent expansion, however when
targets say they want to expand FP_TO_FP16 what they actually want is
the unsafe math expansion when possible and expansion to a libcall in all
other cases.
The only way to make this work as a target independent would be to add logic
to target's TargetLowering construction to mark theses nodes as Expand when
LegalizeDAG can use the unsafe expansion and mark them as LibCall when it
cannot. I think this would be possible, but I think it would be too fragile
and complex as it would require targets to keep their expansion logic up
to date with the code in LegalizeDAG.
Reviewers: bogner, ab, t.p.northover, arsenm
Subscribers: wdng, llvm-commits, nhaehnle
Differential Revision: https://reviews.llvm.org/D25999
llvm-svn: 285704
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:
When finding a match for a merge and collecting the instructions that must
be moved, keep in mind that the instruction we merge might actually use one
of the defs that are being moved.
Fixes piglit spec/arb_enhanced_layouts/execution/component-layout/vs-tcs-load-output[-indirect].
The fact that the ds_read in the test case is not eliminated suggests that
there might be another problem related to alias analysis, but that's a
separate problem: this pass should still work correctly even when earlier
optimization passes missed something or were disabled.
Reviewers: tstellarAMD, arsenm
Subscribers: kzhuravl, wdng, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D25829
llvm-svn: 285273
Add missing ISA versions 7.0.2/8.0.4/8.1.0. to backend.
Refactor processor definition to use ISA version features.
Fixed ISA version for stoney.
Based on Laurent Morichetti's patch.
Differential Revision: https://reviews.llvm.org/D25919
llvm-svn: 285210
Summary:
This test had run lines disabling/enabling the promote alloca pass, but
enabling/disabling promote alloca had no impact on the output.
Reviewers: arsenm
Subscribers: mgrang, kzhuravl, wdng, nhaehnle, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D25787
llvm-svn: 285197
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
Summary:
If we are loading an i16 value from a 32-bit memory location, then
we need to be able to truncate the loaded value to i16.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D25198
llvm-svn: 284397
Summary:
We are using this helper for our 24-bit arithmetic combines, so we are now able to eliminate multi-use operations that mask the high-bits of 24-bit inputs (e.g. and x, 0xffffff)
Reviewers: arsenm, nhaehnle
Subscribers: tony-tye, arsenm, kzhuravl, wdng, nhaehnle, llvm-commits, yaxunl
Differential Revision: https://reviews.llvm.org/D24672
llvm-svn: 284267
Summary:
This will be used for 64-bit MULHU, which is in turn used for the 64-bit
divide-by-constant optimization (see D24822).
Reviewers: arsenm, tstellarAMD
Subscribers: kzhuravl, wdng, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D25289
llvm-svn: 284224
Retrying after upstream changes.
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search which only checks for parallel stores
through the chain subgraph. This is cleaner as the separation of
non-interfering loads/stores from the store-merging logic.
Whem merging stores, search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited. This improves the quality of the
output SelectionDAG and generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the the chain aggregation in the merged stores across
code paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seemed sufficient to not cause regressions in
tests.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and
merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll -
This test appears to work but no longer exhibits the spill behavior.
Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, dsanders, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 284151
Because everything live is spilled at the end of a
block by fast regalloc, assume this will happen and
avoid the copies of the resource descriptor.
llvm-svn: 284119
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
Summary:
There was a bug with sequences like
s_mov_b64 s[0:1], exec
s_and_b64 s[2:3]<def>, s[0:1], s[2:3]<kill>
...
s_mov_b64_term exec, s[2:3]
because s[2:3] was defined and used in the same instruction, ending up with
SaveExecInst inside OtherUseInsts.
Note that the test case also exposes an unrelated bug.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98028
Reviewers: tstellarAMD, arsenm
Subscribers: kzhuravl, wdng, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D25306
llvm-svn: 283528
Summary: Both computeKnownBits and ComputeNumSignBits can now do a simple
look-through of EXTRACT_VECTOR_ELT. It will compute the result based
on the known bits (or known sign bits) for the vector that the element
is extracted from.
Reviewers: bogner, tstellarAMD, mkuper
Subscribers: wdng, RKSimon, jyknight, llvm-commits, nhaehnle
Differential Revision: https://reviews.llvm.org/D25007
llvm-svn: 283347
This avoids llc using the hosts OS/vendor as defaults and triggering
unwanted behaviour in the tests. This should deal with the buildbot
breakages on windows after r283140.
llvm-svn: 283149
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
Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search which only checks for parallel stores
through the chain subgraph. This is cleaner as the separation of
non-interfering loads/stores from the store-merging logic.
Whem merging stores, search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited. This improves the quality of the
output SelectionDAG and generally the output CodeGen (with some
exceptions).
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the the chain aggregation in the merged stores across
code paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seemed sufficient to not cause regressions in
tests.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable. Some tests relying on the order were changed to use
volatile memory operations
Noteworthy tests:
CodeGen/AArch64/argument-blocks.ll -
It's not entirely clear what the test_varargs_stackalign test is
supposed to be asserting, but the new code looks right.
CodeGen/AArch64/arm64-memset-inline.lli -
CodeGen/AArch64/arm64-stur.ll -
CodeGen/ARM/memset-inline.ll -
The backend now generates *worse* code due to store merging
succeeding, as we do do a 16-byte constant-zero store efficiently.
CodeGen/AArch64/merge-store.ll -
Improved, but there still seems to be an extraneous vector insert
from an element to itself?
CodeGen/PowerPC/ppc64-align-long-double.ll -
Worse code emitted in this case, due to the improved store->load
forwarding.
CodeGen/X86/dag-merge-fast-accesses.ll -
CodeGen/X86/MergeConsecutiveStores.ll -
CodeGen/X86/stores-merging.ll -
CodeGen/Mips/load-store-left-right.ll -
Restored correct merging of non-aligned stores
CodeGen/AMDGPU/promote-alloca-stored-pointer-value.ll -
Improved. Correctly merges buffer_store_dword calls
CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll -
Improved. Sidesteps loading a stored value and merges two stores
CodeGen/X86/pr18023.ll -
This test has been removed, as it was asserting incorrect
behavior. Non-volatile stores *CAN* be moved past volatile loads,
and now are.
CodeGen/X86/vector-idiv.ll -
CodeGen/X86/vector-lzcnt-128.ll -
It's basically impossible to tell what these tests are actually
testing. But, looks like the code got better due to the memory
operations being recognized as non-aliasing.
CodeGen/X86/win32-eh.ll -
Both loads of the securitycookie are now merged.
CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll -
This test appears to work but no longer exhibits the spill
behavior.
Reviewers: arsenm, hfinkel, tstellarAMD, nhaehnle, jyknight
Subscribers: wdng, nhaehnle, nemanjai, arsenm, weimingz, niravd, RKSimon, aemerson, qcolombet, resistor, tstellarAMD, t.p.northover, spatel
Differential Revision: https://reviews.llvm.org/D14834
llvm-svn: 282600
This check currently doesn't seem to do anything useful on any in-tree target:
On non-x86, it always evaluates to false, so we never hit the code path that
creates the shuffle with zero.
On x86, it just forwards to isShuffleMaskLegal(), which is a reasonable thing to
query in general, but doesn't make sense if only restricted to zero blends.
Differential Revision: https://reviews.llvm.org/D24625
llvm-svn: 282567
We were trying to avoid using a FrameIndex operand in non-pointer
operands in a convoluted way, and would break because of
using TargetFrameIndex. The TargetFrameIndex should only be used
in the case where it makes sense to fold it as part of the addressing
mode, otherwise it requires materialization like a normal constant.
This wasn't working reliably and failed in the added testcase, hitting
the assert when processing the frame index.
The TargetFrameIndex was coming from trying to produce an AssertZext
limiting the maximum stack size. I'm not sure this was correct to begin
with, because it is apparently possible to have a single workitem
dispatch that requires all 4G of private memory.
llvm-svn: 281824
This reduces the number of copies and reg_sequences
when using fp constant vectors. This significantly
reduces the code size in local-stack-alloc-bug.ll
llvm-svn: 281822
Summary: i8, i16, and f16 values are not extended to 32-bit in the HSA kernel ABI.
Reviewers: arsenm
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, llvm-commits, yaxunl
Differential Revision: https://reviews.llvm.org/D24621
llvm-svn: 281789
These clean up some unnecessary or instructions in
cases with complex loops.
In the original testcase I noticed this, the same
or with exec was repeated 5 or 6 times in a row. With
this only one is emitted or sometimes a copy.
llvm-svn: 281786
Summary:
The main challenge in lowering kernel arguments for AMDGPU is determing the
memory type of the argument. The generic calling convention code assumes
that only legal register types can be stored in memory, but this is not the
case for AMDGPU.
This consolidates all the logic AMDGPU uses for deducing memory types into a single
function. This will make it much easier to support different ABIs in the future.
Reviewers: arsenm
Subscribers: arsenm, wdng, nhaehnle, llvm-commits, yaxunl
Differential Revision: https://reviews.llvm.org/D24614
llvm-svn: 281781
Summary:
mesa3d will use the same kernel calling convention as amdhsa, but it will
handle everything else like the default 'unknown' OS type.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D22783
llvm-svn: 281779
This addresses a TODO to handle operations besides and. This
also starts eliminating no-op operations with a constant that
can emerge later.
llvm-svn: 281488
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
We want each register to have a canonical type, which means the best place to
store this is in MachineRegisterInfo rather than on every MachineInstr that
happens to use or define that register.
Most changes following from this are pretty simple (you need an MRI anyway if
you're going to be doing any transformations, so just check the type there).
But legalization doesn't really want to check redundant operands (when, for
example, a G_ADD only ever has one type) so I've made use of MCInstrDesc's
operand type field to encode these constraints and limit legalization's work.
As an added bonus, more validation is possible, both in MachineVerifier and
MachineIRBuilder (coming soon).
llvm-svn: 281035
Summary:
Also removed duplicate code from AMDGPUTargetAsmStreamer.
This change only change how amd_kernel_code_t is parsed and printed. No variable names are changed.
Reviewers: vpykhtin, tstellarAMD
Subscribers: arsenm, wdng, nhaehnle
Differential Revision: https://reviews.llvm.org/D24296
llvm-svn: 281028
Add the ability to computeKnownBits and SimplifyDemandedBits to extract the known zero/one bits from BUILD_VECTOR, returning the known bits that are shared by every vector element.
This is an initial step towards determining the sign bits of a vector (PR29079).
Differential Revision: https://reviews.llvm.org/D24253
llvm-svn: 280927
OpenCL kernels have hidden kernel arguments for global offset and printf buffer. For consistency, these hidden argument should be included in the runtime metadata. Also updated kernel argument kind metadata.
Differential Revision: https://reviews.llvm.org/D23424
llvm-svn: 280829
- 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:
This contains two changes that reduce the time spent in WQM, with the
intention of reducing bandwidth required by VMEM loads:
1. Sampling instructions by themselves don't need to run in WQM, only their
coordinate inputs need it (unless of course there is a dependent sampling
instruction). The initial scanInstructions step is modified accordingly.
2. When switching back from WQM to Exact, switch back as soon as possible.
This affects the logic in processBlock.
This should always be a win or at best neutral.
There are also some cleanups (e.g. remove unused ExecExports) and some new
debugging output.
Reviewers: arsenm, tstellarAMD, mareko
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: http://reviews.llvm.org/D22092
llvm-svn: 280590
Summary:
This fixes a rare bug in polygon stippling with non-monolithic pixel shaders.
The underlying problem is as follows: the prolog part contains the polygon
stippling sequence, i.e. a kill. The main part then enables WQM based on the
_reduced_ exec mask, effectively undoing most of the polygon stippling.
Since we cannot know whether polygon stippling will be used, the main part
of a non-monolithic shader must always return to exact mode to fix this
problem.
Reviewers: arsenm, tstellarAMD, mareko
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D23131
llvm-svn: 280589
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
Subregister definitions are considered uses for the purpose of tracking
liveness of the whole register. At the same time, when calculating live
interval subranges, subregister defs should not be treated as uses.
Differential Revision: https://reviews.llvm.org/D24190
llvm-svn: 280532
Add runtime metdata for pointee alignment of pointer type kernel argument. The key is KeyArgPointeeAlign and the value is a 32 bit unsigned integer.
Differential Revision: https://reviews.llvm.org/D24145
llvm-svn: 280399
Summary: This fixes some OpenCV tests that were broken by libclc commit r276443.
Reviewers: arsenm, jvesely
Subscribers: arsenm, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D24051
llvm-svn: 280274
Summary:
Simply replace usage of aliases to functions with aliasee.
This came up when bitcode linking to builtin library and
calls to aliases not being resolved.
Also made minor improvements to existing test.
Reviewers: tstellarAMD, alex-t, vpykhtin
Subscribers: arsenm, wdng, rampitec
Differential Revision: https://reviews.llvm.org/D24023
llvm-svn: 280221
Summary:
The SILoadStoreOptimizer can now look ahead more then one instruction when
looking for instructions to merge, which greatly improves the number of
loads/stores that we are able to merge.
Moving the pass before scheduling avoids increasing register pressure after
the scheduler, so that the scheduler's register pressure estimates will be
more accurate. It also gives more consistent results, since it is no longer
affected by minor scheduling changes.
Reviewers: arsenm
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: https://reviews.llvm.org/D23814
llvm-svn: 279991
Summary:
For shrinking SOPK instructions, we were creating a hint to tell the
register allocator to use the register allocated for src0 for the dst
operand as well. However, this seems to not work sometimes depending
on the order virtual registers are assigned physical registers.
To fix this, I've added a second allocation hint which does the reverse,
asks that the register allocated for dst is used for src0.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D23862
llvm-svn: 279968
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
Summary:
If the scheduler clusters the loads, then the offsets will be sorted,
but it is possible for the scheduler to scheduler loads together
without out explicitly clustering them, which would give us non-sorted
offsets.
Also, we will want to do this if we move the load/store optimizer before
the scheduler.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D23776
llvm-svn: 279870
This was for some reason skipping operands that are subregisters
instead of keeping the same subregister index.
v_movreld_b32 expects src0 to be the subregister of the tied
super register use/def.
e.g.
v_movreld_b32 v0, v9, <imp-def, tied3> v[0:3], <imp-use, tied2> v[0:3]
was being replaced with
v[4:7] = copy v[0:3]
v_movreld_b32 v0, v9, <imp-def, tied3> v[4:7], <imp-use, tied2> v[4:7],
which really writes to v[0:3]
llvm-svn: 279804
Summary:
This patch implements readlane/readfirstlane intrinsics.
TODO: need to define a new register class to consider the case
that the source could be a vector register or M0.
Reviewed by:
arsenm and tstellarAMD
Differential Revision:
http://reviews.llvm.org/D22489
llvm-svn: 279660
The register allocator can split a live interval of a register into a set
of smaller intervals. After the allocation of registers is complete, the
rewriter will modify the IR to replace virtual registers with the corres-
ponding physical registers. At this stage, if a register corresponding
to a subregister of a virtual register is used, the rewriter will check
if that subregister is undefined, and if so, it will add the <undef> flag
to the machine operand. The function verifying liveness of the subregis-
ter would assume that it is undefined, unless any of the subranges of the
live interval proves otherwise.
The problem is that the live intervals created during splitting do not
have any subranges, even if the original parent interval did. This could
result in the <undef> flag placed on a register that is actually defined.
Differential Revision: http://reviews.llvm.org/D21189
llvm-svn: 279625
Specifying isSSA is an extra line at best and results in invalid MI at
worst. Compute the value instead.
Differential Revision: http://reviews.llvm.org/D22722
llvm-svn: 279600
Do most of the lowering in a pre-RA pass. Keep the skip jump
insertion late, plus a few other things that require more
work to move out.
One concern I have is now there may be COPY instructions
which do not have the necessary implicit exec uses
if they will be lowered to v_mov_b32.
This has a positive effect on SGPR usage in shader-db.
llvm-svn: 279464
The ppc64 multistage bot fails on this.
This reverts commit r279124.
Also Revert "CodeGen: Add/Factor out LiveRegUnits class; NFCI" because it depends on the previous change
This reverts commit r279171.
llvm-svn: 279199
Summary:
We need to use floating-point compares to ensure that s_cbranch_vcc*
instructions are always generated. With integer compares, future
optimizations could cause s_cbranch_scc* to be generated instead.
Reviewers: arsenm, nhaehnle
Subscribers: llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D23401
llvm-svn: 279148
Re-apply r276044 with off-by-1 instruction fix for the reload placement.
This is a variant of scavengeRegister() that works for
enterBasicBlockEnd()/backward(). The benefit of the backward mode is
that it is not affected by incomplete kill flags.
This patch also changes
PrologEpilogInserter::doScavengeFrameVirtualRegs() to use the register
scavenger in backwards mode.
Differential Revision: http://reviews.llvm.org/D21885
llvm-svn: 279124
If a loop is not rotated (for example when optimizing for size), the latch is not the backedge. If we promote an expression to post-inc form, we not only increase register pressure and add a COPY for that IV expression but for all IVs!
Motivating testcase:
void f(float *a, float *b, float *c, int n) {
while (n-- > 0)
*c++ = *a++ + *b++;
}
It's imperative that the pointer increments be located in the latch block and not the header block; if not, we cannot use post-increment loads and stores and we have to keep both the post-inc and pre-inc values around until the end of the latch which bloats register usage.
llvm-svn: 278658
Summary:
This patch define and implement amdgcn image intrinsics with sampler.
1. define vdata type to be llvm_anyfloat_ty, address type to be llvm_anyfloat_ty,
and rsrc type to be llvm_anyint_ty. As a result, we expect the intrinsics name
to have three suffixes to overload each of these three types;
2. D128 as well as two other flags are implied in the three types, for example,
if you use v8i32 as resource type, then r128 is 0!
3. don't expose TFE flag, and other flags are exposed in the instruction order:
unrm, glc, slc, lwe and da.
Differential Revision: http://reviews.llvm.org/D22838
Reviewed by:
arsenm and tstellarAMD
llvm-svn: 278291
Insert before the skip branch if one is created.
This is a somewhat more natural placement relative
to the skip branches, and makes it possible to implement
analyzeBranch for skip blocks.
The test changes are mostly due to a quirk where
the block label is not emitted if there is a terminator
that is not also a branch.
llvm-svn: 278273
Summary:
See the new test case for one that was (non-deterministically) crashing
on trunk and deterministically hit the assertion that I added in D23302.
Basically, the machine function contains a sequence
DS_WRITE_B32 %vreg4, %vreg14:sub0, ...
DS_WRITE_B32 %vreg4, %vreg14:sub0, ...
%vreg14:sub1<def> = COPY %vreg14:sub0
and SILoadStoreOptimizer::mergeWrite2Pair merges the two DS_WRITE_B32
instructions into one before calling repairIntervalsInRange.
Now repairIntervalsInRange wants to repair %vreg14, in particular, and
ends up trying to repair %vreg14:sub1 as well, but that only becomes
active _after_ the range that is to be repaired, hence the crash due
to LR.find(...) == LR.begin() at the start of repairOldRegInRange.
I believe that just skipping those subrange is fine, but again, not too
familiar with that code.
Reviewers: MatzeB, kparzysz, tstellarAMD
Subscribers: llvm-commits, MatzeB
Differential Revision: https://reviews.llvm.org/D23303
llvm-svn: 278268
Summary:
This is the setting of the Vulkan closed source driver.
It decreases the max wave count from 10 to 8.
26010 shaders in 14650 tests
Totals:
VGPRS: 829593 -> 808440 (-2.55 %)
Spilled SGPRs: 81878 -> 42226 (-48.43 %)
Spilled VGPRs: 367 -> 358 (-2.45 %)
Scratch VGPRs: 1764 -> 1748 (-0.91 %) dwords per thread
Code Size: 36677864 -> 35923932 (-2.06 %) bytes
There is a massive decrease in SGPR spilling in general and -7.4% spilled
VGPRs for DiRT Showdown (= SGPRs spilled to scratch?)
Reviewers: arsenm, tstellarAMD, nhaehnle
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D23034
llvm-svn: 277867
Adding missing tests for OCL type names for half, float, double, char, short, long, and unknown.
Patch by Aaron En Ye Shi.
Differential Revision: https://reviews.llvm.org/D22964
llvm-svn: 277759
This just tests that the register limit isn't exceeded,
so the regisetr allocation doesn't need to be great.'
The critically slow part is all in greedy RA, so
switch to basic.
llvm-svn: 277700
rewriteOperands() always performed liveness queries at the base index
rather than the RegSlot/Base as apropriate for the machine operand. This
could lead to illegal rewriting in some cases.
llvm-svn: 277661
Summary:
Two types of stores are possible in pixel shaders: stores to memory that are
explicitly requested at the API level, and stores that are an implementation
detail of register spilling or lowering of arrays.
For the first kind of store, we must ensure that helper pixels have no effect
and hence WQM must be disabled. The second kind of store must always be
executed, because the written value may be loaded again in a way that is
relevant for helper pixels as well -- and there are no externally visible
effects anyway.
This is a candidate for the 3.9 release branch.
Reviewers: arsenm, tstellarAMD, mareko
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: https://reviews.llvm.org/D22675
llvm-svn: 277504
Summary:
There are cases where uniform branch conditions are computed in VGPRs, and
we didn't correctly mark those as WQM.
The stray change in basic-branch.ll is because invoking the LiveIntervals
analysis leads to the detection of a dead register that would otherwise not
be seen at -O0.
This is a candidate for the 3.9 branch, as it fixes a possible hang.
Reviewers: arsenm, tstellarAMD, mareko
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D22673
llvm-svn: 277500
Summary:
SI_ELSE is lowered into two parts:
s_or_saveexec_b64 dst, src (at the start of the basic block)
s_xor_b64 exec, exec, dst (at the end of the basic block)
The idea is that dst contains the exec mask of the preceding IF block. It can
happen that SIWholeQuadMode decides to switch from WQM to Exact mode inside
the basic block that contains SI_ELSE, in which case it introduces an instruction
s_and_b64 exec, exec, s[...]
which masks out bits that can correspond to both the IF and the ELSE paths.
So the resulting sequence must be:
s_or_savexec_b64 dst, src
s_and_b64 exec, exec, s[...] <-- added by SIWholeQuadMode
s_and_b64 dst, dst, exec <-- added by SILowerControlFlow
s_xor_b64 exec, exec, dst
Whether to add the additional s_and_b64 dst, dst, exec is currently determined
via the ExecModified tracking. With this change, it is instead determined by
an additional flag on SI_ELSE which is set by SIWholeQuadMode.
Finally: It also occured to me that an alternative approach for the long run
is for SILowerControlFlow to unconditionally emit
s_or_saveexec_b64 dst, src
...
s_and_b64 dst, dst, exec
s_xor_b64 exec, exec, dst
and have a pass that detects and cleans up the "redundant AND with exec"
pattern where possible. This could be useful anyway, because we also add
instructions
s_and_b64 vcc, exec, vcc
before s_cbranch_scc (in moveToALU), and those are often redundant. I have
some pending changes to how KILL is lowered that could also benefit from
such a cleanup pass.
In any case, this current patch could help in the short term with the whole
ExecModified business.
Reviewers: tstellarAMD, arsenm
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D22846
llvm-svn: 276972
Change the syntax to use `%0.sub8` to denote a subregister.
This seems like a more natural fit to denote subregisters; I also plan
to introduce a new ":classname" syntax in upcoming patches to denote the
register class of a vreg.
Note that this commit disallows plain identifiers to start with a '.'
character. This shouldn't affect anything as external names/IR
references are all prefixed with '$'/'%', plain identifiers are only
used for instruction names, register mask names and subreg indexes.
Differential Revision: https://reviews.llvm.org/D22390
llvm-svn: 276815
This reverts commit r276298.
Data stored in .rodata can have a negative offset from .text, but we
don't support negative values in relocations yet.
This caused a regression in one of the amp conformance tests:
5_Data_Cont/5_2_a_v/5_2_3_m/Assignment/Test.02.01
llvm-svn: 276498
Summary:
The llvm.invariant.start and llvm.invariant.end intrinsics currently
support specifying invariant memory objects only in the default address
space.
With this change, these intrinsics are overloaded for any adddress space
for memory objects
and we can use these llvm invariant intrinsics in non-default address
spaces.
Example: llvm.invariant.start.p1i8(i64 4, i8 addrspace(1)* %ptr)
This overloaded intrinsic is needed for representing final or invariant
memory in managed languages.
Reviewers: apilipenko, reames
Subscribers: llvm-commits
llvm-svn: 276447
The size can exceed s_movk_i32's limit, and we don't
want to use it this early since it inhibits optimizations.
This should probably be merged to the release branch.
llvm-svn: 276438
Summary:
The llvm.invariant.start and llvm.invariant.end intrinsics currently
support specifying invariant memory objects only in the default address space.
With this change, these intrinsics are overloaded for any adddress space for memory objects
and we can use these llvm invariant intrinsics in non-default address spaces.
Example: llvm.invariant.start.p1i8(i64 4, i8 addrspace(1)* %ptr)
This overloaded intrinsic is needed for representing final or invariant memory in managed languages.
Reviewers: tstellarAMD, reames, apilipenko
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D22519
llvm-svn: 276316
This should be all the low-level instruction selection needs to determine how
to implement an operation, with the remaining context taken from the opcode
(e.g. G_ADD vs G_FADD) or other flags not based on type (e.g. fast-math).
llvm-svn: 276158
Reverting this commit for now as it seems to be causing failures on
test-suite tests on the clang-ppc64le-linux-lnt bot.
This reverts commit r276044.
llvm-svn: 276068
If 2.5 ulp is acceptable, denormals are not required, and
isn't a reciprocal which will already be handled, replace
with a faster fdiv.
Simplify the lowering tests by using per function
subtarget features.
llvm-svn: 276051
This is a variant of scavengeRegister() that works for
enterBasicBlockEnd()/backward(). The benefit of the backward mode is
that it is not affected by incomplete kill flags.
This patch also changes
PrologEpilogInserter::doScavengeFrameVirtualRegs() to use the register
scavenger in backwards mode.
Differential Revision: http://reviews.llvm.org/D21885
llvm-svn: 276044
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
This is currently only called with GEP users. A direct
alloca would only happen with current typed pointers
for arrays which are a perverse case.
Also fix crashes on 0 x and 1 x arrays.
llvm-svn: 275869
Summary:
The work item intrinsics are not available for the shader
calling conventions. And even if we did hook them up most
shader stages haves some extra restrictions on the amount
of available LDS.
Reviewers: tstellarAMD, arsenm
Subscribers: nhaehnle, arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D20728
llvm-svn: 275779
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
Added emitting metadata to elf for runtime.
Runtime requires certain information (metadata) about kernels to be able to execute and query them. Such information is emitted to an elf section as a key-value pair stream.
Differential Revision: https://reviews.llvm.org/D21849
llvm-svn: 275566
Also stop trying to insert skip blocks at end_cf. This
was inserting them at the end of the block which doesn't make
sense. The skip should be inserted at the beginning of the block
right after the end cf. Just remove this for now since no tests
seem to stress this and I think this can be handled more generally
later.
Fixes bug 28550
llvm-svn: 275510
It wasn't actually running the pass, and since it is
missing the llvm prefix, the eh intrinsic was not
really an IntrinsicInst.
Also add missing test for lifetime markers.
llvm-svn: 275370
Currently the MIR framework prints all its outputs (errors and actual
representation) on stderr.
This patch fixes that by printing the regular output in the output
specified with -o.
Differential Revision: http://reviews.llvm.org/D22251
llvm-svn: 275314
Summary:
Previously, constant index insertelements would be turned into SI_INDIRECT_DST,
which is bound to prevent some optimization opportunities. Worse, it mislead
the heuristic that decides whether immediates should be lowered to S_MOV_B32
or V_MOV_B32 in a way that resulted in unnecessary v_readfirstlanes.
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: http://reviews.llvm.org/D22217
llvm-svn: 275160
Summary:
Setting MIMG to 0 has a bunch of unexpected side effects, including that
isVMEM returns false which leads to incorrect treatment in the hazard
recognizer. The reason I noticed it is that it also leads to incorrect
treatment in VGPR-to-SGPR copies, which is one cause of the referenced bug.
The only reason why MIMG was set to 0 is to signal the special handling of
dmasks, but that can be checked differently.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96877
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: http://reviews.llvm.org/D22210
llvm-svn: 275113
Summary:
The main bug fix here is using the 32-bit encoding of V_ADD_I32 in
materializeFrameBaseRegister and resolveFrameIndex, so that arbitrary
immediates work.
The second part is that we may now require the SegmentWaveByteOffset
even when there are initially no stack objects and VGPR spilling isn't
enabled, for stack slots that are allocated later. This means that some
bits become effectively dead and can be cleaned up.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96602
Tested-by: Kai Wasserbäch <kai@dev.carbon-project.org>
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: http://reviews.llvm.org/D21551
llvm-svn: 275108
Summary:
Since "AMDGPU: Fix verifier errors in SILowerControlFlow", the logic that
ensures that a non-void-returning shader falls off the end of the last
basic block was effectively disabled, since SI_RETURN is now used.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96731
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: http://reviews.llvm.org/D21975
llvm-svn: 274612
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
Due to visit order problems, in the case of an unaligned copy
the legalized DAG fails to eliminate extra instructions introduced
by the expansion of both unaligned parts.
llvm-svn: 274397
There was a combine before to handle the simple copy case.
Split this into handling loads and stores separately.
We might want to change how this handles some of the vector
extloads, since this can result in large code size increases.
llvm-svn: 274394
This was reverted in r268740 because of problems with corresponding Clang change.
Clang change was updated and resubmitted in r274220.
Check calling convention in AMDGPUMachineFunction::isKernel
This will be used for AMDGPU_HSA_KERNEL symbol type in output ELF.
Also, in the future unused non-kernels may be optimized.
Reviewers: tstellarAMD, arsenm
Subscribers: arsenm, joker.eph, llvm-commits
Differential Revision: http://reviews.llvm.org/D19917
llvm-svn: 274341
Don't cast GV expression to MCSymbolRefExpr. r272705 changed GV to binary
expressions by including offset even if the offset it 0
(we haven't hit this sooner since tested workloads don't include static offsets)
We don't really care about the type of expression, so set it directly.
Fixes: r272705
Consider section relative relocations. Since all const as data is in one boffer section relative is equivalent to abs32.
Fixes: r273166
Differential Revision: http://reviews.llvm.org/D21633
llvm-svn: 273785
Debugger prologue is emitted if -mattr=+amdgpu-debugger-emit-prologue.
Debugger prologue writes work group IDs and work item IDs to scratch memory at fixed location in the following format:
- offset 0: work group ID x
- offset 4: work group ID y
- offset 8: work group ID z
- offset 16: work item ID x
- offset 20: work item ID y
- offset 24: work item ID z
Set
- amd_kernel_code_t::debug_wavefront_private_segment_offset_sgpr to scratch wave offset reg
- amd_kernel_code_t::debug_private_segment_buffer_sgpr to scratch rsrc reg
- amd_kernel_code_t::is_debug_supported to true if all debugger features are enabled
Differential Revision: http://reviews.llvm.org/D20335
llvm-svn: 273769
Summary:
Offset folding only works if you are emitting relocations, and we don't
emit relocations for local address space globals.
Reviewers: arsenm, nhaustov
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: http://reviews.llvm.org/D21647
llvm-svn: 273765
In bidirectional scheduling this gives more stable results than just
comparing the "reason" fields of the top/bottom node because the reason
field may be higher depending on what other nodes are in the queue.
Differential Revision: http://reviews.llvm.org/D19401
llvm-svn: 273755
COPY was lacking a scheduling class, define it to avoid regressions in
the upcoming change to the bidirectional MachineScheduler. Approved by
tstellar on IRC.
Differential Revision: http://reviews.llvm.org/D21540
llvm-svn: 273751
Un XFAIL a few tests plus a few more I had lying around
in my tree, which seem to all work now but I don't see tests
that quite test the same things.
llvm-svn: 273655
The only real reason to use it is for testing, so replace
it with a command line option instead of a potentially function
dependent feature.
llvm-svn: 273653
The exit-on-error flag was necessary in order to avoid an assertion when
handling DYNAMIC_STACKALLOC nodes in SelectionDAGLegalize.
We can avoid the assertion by creating some dummy nodes. This enables us to
remove the exit-on-error flag on the first 2 run lines (SI), but on the third
run line (R600) we would run into another assertion when trying to reserve
indirect registers. This patch also replaces that assertion with an early exit
from the function.
Fixes PR27761.
Differential Revision: http://reviews.llvm.org/D20852
llvm-svn: 273550
The main sin this was committing was using terminator
instructions in the middle of the block, and then
not updating the block successors / predecessors.
Split the blocks up to avoid this and introduce new
pseudo instructions for branches taken with exec masking.
Also use a pseudo instead of emitting s_endpgm and erasing
it in the special case of a non-void return.
llvm-svn: 273467
The implicit operand is added by the initial instruction construction,
so this was adding an additional vcc use. The original one
was missing the undef flag the original condition had,
so the verifier would complain.
llvm-svn: 273182
This will help sneak undefs past GVN into the DAG for
some tests.
Also add missing intrinsic for rsq_legacy, even though the node
was already selected to the instruction. Also start passing
the debug location to intrinsic errors.
llvm-svn: 273181
Don't use AllocateStack because kernel arguments have nothing
to do with the stack. The ensureMaxAlignment call was still
changing the stack alignment.
llvm-svn: 273080
This should select to s_trap, but that requires
additonal work to setup and enable the trap handler.
For now emit s_endpgm so bugpoint stops getting stuck
on the unsupported call to abort.
Emit a warning that this will only terminate the wave and
not really trap.
llvm-svn: 273062
Summary:
This fixes two related bugs. First, the generic optimization passes
unfortunately generate negative constant offsets but the hardware treats
SOffset as an unsigned value.
Second, there is a hardware bug on SI and CI, where address clamping in MUBUF
instructions does not work correctly when SOffset is larger than the buffer
size. This patch works around this bug by never using SOffset.
An alternative workaround would be to do the clamping manually when SOffset
is too large, but generating the required code sequence during instruction
selection would be rather involved, and in any case the resulting code would
probably be worse.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96360
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits, kzhuravl
Differential Revision: http://reviews.llvm.org/D21326
llvm-svn: 272761
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
Summary:
We need to set the fixup type to FK_Data_4 for the
SCRATCH_RSRC_DWORD[01] symbols, since these require absolute
relocations, and fixup_si_rodata is for relative relocations.
Reviewers: arsenm, kzhuravl
Subscribers: arsenm, kzhuravl, llvm-commits
Differential Revision: http://reviews.llvm.org/D21153
llvm-svn: 272417
Stanislav Mekhanoshin