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
Summary:
Without a MachineMemOperand, the scheduler was assuming MIMG instructions
were ordered memory references, so no loads or stores could be reordered
across them.
Reviewers: arsenm
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tony-tye
Differential Revision: https://reviews.llvm.org/D27536
llvm-svn: 290179
Retrying after fixing after removing load-store factoring through
token factors in favor of improved token factor operand pruning
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.
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: 289659
Summary:
This change adds some verification in the IR verifier around struct path
TBAA metadata.
Other than some basic sanity checks (e.g. we get constant integers where
we expect constant integers), this checks:
- That by the time an struct access tuple `(base-type, offset)` is
"reduced" to a scalar base type, the offset is `0`. For instance, in
C++ you can't start from, say `("struct-a", 16)`, and end up with
`("int", 4)` -- by the time the base type is `"int"`, the offset
better be zero. In particular, a variant of this invariant is needed
for `llvm::getMostGenericTBAA` to be correct.
- That there are no cycles in a struct path.
- That struct type nodes have their offsets listed in an ascending
order.
- That when generating the struct access path, you eventually reach the
access type listed in the tbaa tag node.
Reviewers: dexonsmith, chandlerc, reames, mehdi_amini, manmanren
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D26438
llvm-svn: 289402
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
Summary:
There is no point in setting SGPRS=104, because VI allocates SGPRs
in multiples of 16, so 104 -> 112. That enables us to use all 102 SGPRs
for general purposes.
Reviewers: tstellarAMD
Subscribers: qcolombet, arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tony-tye
Differential Revision: https://reviews.llvm.org/D27149
llvm-svn: 289260
test/CodeGen/MIR should contain tests that intent to test the MIR
printing or parsing. Tests that test something else should be in
test/CodeGen/TargetName even when they are written in .mir.
As a rule of thumb, only tests using "llc -run-pass none" should be in
test/CodeGen/MIR.
llvm-svn: 289254
Retrying after fixing overly aggressive load-store forwarding optimization.
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.
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: 289221
These were selecting directly to the VOP2 form instead
of VOP3 like the i32 instructions. Fixes regressions in
future commits where an immediate isn't folded because it was
initially used for the second operand.
Because uniform 16-bit operations are promoted to i32, it's
difficult to get a simple testcase where this matters. Fold
failures in SIFoldOperands here tend to be hidden by commute
and fold in SIShrinkInstructions.
llvm-svn: 289189
Not having this legal led to combine failures, resulting
in dumb things like bitcasts of constants not being folded
away.
The only reason I'm leaving the v_mov_b32 hack that f32
already uses is to avoid madak formation test regressions.
PeepholeOptimizer has an ordering issue where the immediate
fold attempt is into the sgpr->vgpr copy instead of the actual
use. Running it twice avoids that problem.
llvm-svn: 289096
The correct commutable opcode was set to itself, so this
was simply swapping the operands to commute instead of also
changing the opcode to v_subrev_u16.
llvm-svn: 289093
Multiple metadata values for records such as opencl.ocl.version, llvm.ident
and similar are created after linking several modules. For some of them, notably
opencl.ocl.version, this creates semantic problem because we cannot tell which
version of OpenCL the composite module conforms.
Moreover, such repetitions of identical values often create a huge list of
unneeded metadata, which grows bitcode size both in memory and stored on disk.
It can go up to several Mb when linked against our OpenCL library. Lastly, such
long lists obscure reading of dumped IR.
The pass unifies metadata after linking.
Differential Revision: https://reviews.llvm.org/D25381
llvm-svn: 289092
Summary:
LC can currently select scalar load for uniform memory access
basing on readonly memory address space only. This restriction
originated from the fact that in HW prior to VI vector and scalar caches
are not coherent. With MemoryDependenceAnalysis we can check that the
memory location corresponding to the memory operand of the LOAD is not
clobbered along the all paths from the function entry.
Reviewers: rampitec, tstellarAMD, arsenm
Subscribers: wdng, arsenm, nhaehnle
Differential Revision: https://reviews.llvm.org/D26917
llvm-svn: 289076
Summary:
Without the fix to isFrameOffsetLegal to consider the instruction's
immediate offset, the new test case hits the corresponding assertion in
resolveFrameIndex, because the LocalStackSlotAllocation pass re-uses a
different base register.
With only the fix to isFrameOffsetLegal, code quality reduces in a bunch of
places because frame base registers are added where they're not needed.
This is addressed by properly implementing needsFrameBaseReg, which also
helps to avoid unnecessary zero frame indices in a bunch of other places.
Fixes piglit glsl-1.50/execution/variable-indexing/gs-output-array-vec4-index-wr.shader_test
Reviewers: arsenm, tstellarAMD
Subscribers: qcolombet, kzhuravl, wdng, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D27344
llvm-svn: 289048
Summary:
If we write an immediate to a VGPR and then copy the VGPR to an
SGPR, we can replace the copy with a S_MOV_B32 sgpr, imm, rather than
moving the copy to the SALU.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D27272
llvm-svn: 288849
The structured CFG is just an aid to inserting exec
mask modification instructions, once that is done
we don't really need it anymore. We also
do not analyze blocks with terminators that
modify exec, so this should only be impacting
true branches.
llvm-svn: 288744
Structure the definitions a bit more like the other classes.
The main change here is to split EXP with the done bit set
to a separate opcode, so we can set mayLoad = 1 so that it won't
be reordered before the other exp stores, since this has the special
constraint that if the done bit is set then this should be the last
exp in she shader.
Previously all exp instructions were inferred to have unmodeled
side effects.
llvm-svn: 288695
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
Since the spill is for the whole wave, these
don't have the swizzling problems that vector stores do
and a single 4-byte allocation is enough to spill a 64 element
register. This should reduce the number of spill instructions and
put all the spills for a register in the same cacheline.
This should save allocated private size, but for now it doesn't.
The extra slots are allocated for each component, but never used
because the frame layout is essentially finalized before frame
indices are replaced. For always using the scalar store path,
this should probably be moved into processFunctionBeforeFrameFinalized.
llvm-svn: 288445
The coalescer eliminates copies from reserved registers of the form:
%vregX = COPY %rY
in the case where %rY is a reserved register. However this turns out to
be invalid if only some of the subregisters are reserved (see also
https://reviews.llvm.org/D26648).
Differential Revision: https://reviews.llvm.org/D26687
llvm-svn: 288428
This is not in the list of valid inputs for the encoding.
When spilling, copies from exec can be folded directly
into the spill instruction which results in broken
stores.
This only fixes the operand constraints, more codegen
work is required to avoid emitting the invalid
spills.
This sort of breaks the dbg.value test. Because the
register class of the s_load_dwordx2 changes, there
is a copy to SReg_64, and the copy is the operand
of dbg_value. The copy is later dead, and removed
from the dbg_value.
llvm-svn: 288191
It isn't generally safe to fold the frame index
directly into the operand since it will possibly
not be an inline immediate after it is expanded.
This surprisingly seems to produce better code, since
the FI doesn't prevent folding other immediate operands.
llvm-svn: 288185
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_b32
and then restore it with yet another v_cmp instruction from that v_cndmask's
result. To mitigate the issue a propagation of source SGPR pair in place of v_cmp
is implemented. Additional side effect of this is that we may consume less VGPRs
at a cost of more SGPRs in case if holding of multiple conditions is needed, and
that is a clear win in most cases.
Differential Revision: https://reviews.llvm.org/D26114
llvm-svn: 288053
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
The size and offset were wrong. The size of the object was
being used for the size of the access, when here it is really
being split into 4-byte accesses. The underlying object size
is set in the MachinePointerInfo, which also didn't have the
offset set.
llvm-svn: 287806
A target intrinsic may be defined as possibly reading memory,
but the call site may have additional knowledge that it doesn't read
memory. The intrinsic lowering will expect the pessimistic
assumption of the intrinsic definition, so the chain should
still be used.
llvm-svn: 287593
Summary:
The 32-bit instructions don't zero the high 16-bits like the 16-bit
instructions do.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, llvm-commits, tony-tye
Differential Revision: https://reviews.llvm.org/D26828
llvm-svn: 287342
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
wbinvl.* are vector instruction that do not sue vector registers.
v2: check only M?BUF instructions
Differential Revision: https://reviews.llvm.org/D26633
llvm-svn: 287056
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
Summary:
Extend image intrinsics to support data types of V1F32 and V2F32.
TODO: we should define a mapping table to change the opcode for data type of V2F32 but just one channel is active,
even though such case should be very rare.
Reviewers:
tstellarAMD
Differential Revision:
http://reviews.llvm.org/D26472
llvm-svn: 286860
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
Summary:
This pass was assuming that when a PHI instruction defined a register
used by another PHI instruction that the defining insstruction would
be legalized before the using instruction.
This assumption was causing the pass to not legalize some PHI nodes
within divergent flow-control.
This fixes a bug that was uncovered by r285762.
Reviewers: nhaehnle, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, tony-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D26303
llvm-svn: 286676
addSchedBarrierDeps() is supposed to add use operands to the ExitSU
node. The current implementation adds uses for calls/barrier instruction
and the MBB live-outs in all other cases. The use
operands of conditional jump instructions were missed.
Also added code to macrofusion to set the latencies between nodes to
zero to avoid problems with the fusing nodes lingering around in the
pending list now.
Differential Revision: https://reviews.llvm.org/D25140
llvm-svn: 286544
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
Matt Arsenault