The pre-RA scheduler does load/store clustering, but post-RA
scheduler undoes it. Add mutation to prevent it.
Differential Revision: https://reviews.llvm.org/D38014
llvm-svn: 313670
Because the stack growth direction and addressing is done
in the same direction, modifying SP at the beginning of the
call sequence was incorrect. If we had a stack passed argument,
we would end up skipping that number of bytes before pushing
arguments, leaving unused/inconsistent space.
The callee creates fixed stack objects in its frame, so
the space necessary for these is already logically allocated
in the callee, so we just let the callee increment SP if
it really requires it.
llvm-svn: 313279
Using SplitCSR for the frame register was very broken. Often
the copies in the prolog and epilog were optimized out, in addition
to them being inserted after the true prolog where the FP
was clobbered.
I have a hacky solution which works that continues to use
split CSR, but for now this is simpler and will get to working
programs.
llvm-svn: 313274
We already have a combine for this pattern when the input to shl is add, so we just need to enable the transformation when the input is or.
Original patch by @tstellar
Differential Revision: https://reviews.llvm.org/D19325
llvm-svn: 313251
MachineScheduler when clustering loads or stores checks if base
pointers point to the same memory. This check is done through
comparison of base registers of two memory instructions. This
works fine when instructions have separate offset operand. If
they require a full calculated pointer such instructions can
never be clustered according to such logic.
Changed shouldClusterMemOps to accept base registers as well and
let it decide what to do about it.
Differential Revision: https://reviews.llvm.org/D37698
llvm-svn: 313208
These two instructions are normally selected, but when the
two address pass converts mac into mad we end up with the
mad where we could have one of these.
Differential Revision: https://reviews.llvm.org/D37389
llvm-svn: 312928
A mrt exp with vm=1 must be in exact (non-WQM) mode, as it also exports
the exec mask as the valid mask to determine which pixels to render.
This commit marks any exp as needing to be in exact mode.
Actually, if there are multiple mrt exps, only one needs to have vm=1,
and only that one needs to be in exact mode. But that is an optimization
for another day.
Differential Revision: https://reviews.llvm.org/D36305
llvm-svn: 312915
The various scalar bit operations set SCC,
so one is erased or moved it needs to be recomputed.
Not sure why the existing tests don't fail on this.
llvm-svn: 312819
Summary:
Mesa still uses a hack where empty inline assembly is used as a kind of
optimization barrier. This exposed a problem where not enough wait states
were inserted, because the hazard recognizer implicitly assumed that each
inline assembly "instruction" has at least one wait state.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D37205
llvm-svn: 312635
When packet size equals packet align and is power of 2, transform
__read_pipe* and __write_pipe* to specialized library function.
Differential Revision: https://reviews.llvm.org/D36831
llvm-svn: 312598
Summary:
This fixes a bug that was exposed on gfx9 in various
GL45-CTS.shaders.loops.*_iterations.select_iteration_count_fragment tests,
e.g. GL45-CTS.shaders.loops.do_while_uniform_iterations.select_iteration_count_fragment
Reviewers: arsenm
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D36193
llvm-svn: 312337
Issues addressed since original review:
- Moved removal of dead instructions found by
LiveIntervals::shrinkToUses() outside of loop iterating over
instructions to avoid instructions being deleted while pointed to by
iterator.
- Fixed ARMLoadStoreOptimizer bug exposed by this change in r311907.
- The pass no longer forwards COPYs to physical register uses, since
doing so can break code that implicitly relies on the physical
register number of the use.
- The pass no longer forwards COPYs to undef uses, since doing so
can break the machine verifier by creating LiveRanges that don't
end on a use (since the undef operand is not considered a use).
[MachineCopyPropagation] Extend pass to do COPY source forwarding
This change extends MachineCopyPropagation to do COPY source forwarding.
This change also extends the MachineCopyPropagation pass to be able to
be run during register allocation, after physical registers have been
assigned, but before the virtual registers have been re-written, which
allows it to remove virtual register COPY LiveIntervals that become dead
through the forwarding of all of their uses.
llvm-svn: 312328
It caused PR34387: Assertion failed: (RegNo < NumRegs && "Attempting to access record for invalid register number!")
> Issues identified by buildbots addressed since original review:
> - Fixed ARMLoadStoreOptimizer bug exposed by this change in r311907.
> - The pass no longer forwards COPYs to physical register uses, since
> doing so can break code that implicitly relies on the physical
> register number of the use.
> - The pass no longer forwards COPYs to undef uses, since doing so
> can break the machine verifier by creating LiveRanges that don't
> end on a use (since the undef operand is not considered a use).
>
> [MachineCopyPropagation] Extend pass to do COPY source forwarding
>
> This change extends MachineCopyPropagation to do COPY source forwarding.
>
> This change also extends the MachineCopyPropagation pass to be able to
> be run during register allocation, after physical registers have been
> assigned, but before the virtual registers have been re-written, which
> allows it to remove virtual register COPY LiveIntervals that become dead
> through the forwarding of all of their uses.
llvm-svn: 312178
Issues identified by buildbots addressed since original review:
- Fixed ARMLoadStoreOptimizer bug exposed by this change in r311907.
- The pass no longer forwards COPYs to physical register uses, since
doing so can break code that implicitly relies on the physical
register number of the use.
- The pass no longer forwards COPYs to undef uses, since doing so
can break the machine verifier by creating LiveRanges that don't
end on a use (since the undef operand is not considered a use).
[MachineCopyPropagation] Extend pass to do COPY source forwarding
This change extends MachineCopyPropagation to do COPY source forwarding.
This change also extends the MachineCopyPropagation pass to be able to
be run during register allocation, after physical registers have been
assigned, but before the virtual registers have been re-written, which
allows it to remove virtual register COPY LiveIntervals that become dead
through the forwarding of all of their uses.
llvm-svn: 312154
Summary:
Reverts r311008 to reinstate r310825 with a fix.
Refine alias checking for pseudo vs value to be conservative.
This fixes the original failure in builtbot unittest SingleSource/UnitTests/2003-07-09-SignedArgs.
Reviewers: hfinkel, nemanjai, efriedma
Reviewed By: efriedma
Subscribers: bjope, mcrosier, nhaehnle, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D36900
llvm-svn: 312126
If denorms are not flushed we can use max instead of multiplication
by 1. For double that is simply faster, while for float and half
it is shorter, because mul uses constant bus and VOP3.
Differential Revision: https://reviews.llvm.org/D36856
llvm-svn: 312095
Under -cl-fast-relaxed-math we could use native_sqrt, but f64 was
allowed to produce HSAIL's nsqrt instruction. HSAIL is not here
and we stick with non-existing native_sqrt(double) as a result.
Add check for f64 to not return native functions and also remove
handling of f64 case for fold_sqrt.
Differential Revision: https://reviews.llvm.org/D37223
llvm-svn: 311900
Two issues identified by buildbots were addressed:
- The pass no longer forwards COPYs to physical register uses, since
doing so can break code that implicitly relies on the physical
register number of the use.
- The pass no longer forwards COPYs to undef uses, since doing so
can break the machine verifier by creating LiveRanges that don't
end on a use (since the undef operand is not considered a use).
[MachineCopyPropagation] Extend pass to do COPY source forwarding
This change extends MachineCopyPropagation to do COPY source forwarding.
This change also extends the MachineCopyPropagation pass to be able to
be run during register allocation, after physical registers have been
assigned, but before the virtual registers have been re-written, which
allows it to remove virtual register COPY LiveIntervals that become dead
through the forwarding of all of their uses.
Reviewers: qcolombet, javed.absar, MatzeB, jonpa
Subscribers: jyknight, nemanjai, llvm-commits, nhaehnle, mcrosier, mgorny
Differential Revision: https://reviews.llvm.org/D30751
llvm-svn: 311135
This reverts commit r311038.
Several buildbots are breaking, and at least one appears to be due to
the forwarding of physical regs enabled by this change. Reverting while
I investigate further.
llvm-svn: 311062
This change extends MachineCopyPropagation to do COPY source forwarding.
This change also extends the MachineCopyPropagation pass to be able to
be run during register allocation, after physical registers have been
assigned, but before the virtual registers have been re-written, which
allows it to remove virtual register COPY LiveIntervals that become dead
through the forwarding of all of their uses.
Reviewers: qcolombet, javed.absar, MatzeB, jonpa
Subscribers: jyknight, nemanjai, llvm-commits, nhaehnle, mcrosier, mgorny
Differential Revision: https://reviews.llvm.org/D30751
llvm-svn: 311038
r310825 caused the clang-ppc64le-linux-lnt bot to go red
(http://lab.llvm.org:8011/builders/clang-ppc64le-linux-lnt/builds/5712)
because of a test-suite failure of
SingleSource/UnitTests/2003-07-09-SignedArgs
This reverts commit 0028f6a87224fb595a1c19c544cde9b003035996.
llvm-svn: 311008
Summary:
This patch teaches PostDominatorTree about infinite loops. It is built on top of D29705 by @dberlin which includes a very detailed motivation for this change.
What's new is that the patch also teaches the incremental updater how to deal with reverse-unreachable regions and how to properly maintain and verify tree roots. Before that, the incremental algorithm sometimes ended up preserving reverse-unreachable regions after updates that wouldn't appear in the tree if it was constructed from scratch on the same CFG.
This patch makes the following assumptions:
- A sequence of updates should produce the same tree as a recalculating it.
- Any sequence of the same updates should lead to the same tree.
- Siblings and roots are unordered.
The last two properties are essential to efficiently perform batch updates in the future.
When it comes to the first one, we can decide later that the consistency between freshly built tree and an updated one doesn't matter match, as there are many correct ways to pick roots in infinite loops, and to relax this assumption. That should enable us to recalculate postdominators less frequently.
This patch is pretty conservative when it comes to incremental updates on reverse-unreachable regions and ends up recalculating the whole tree in many cases. It should be possible to improve the performance in many cases, if we decide that it's important enough.
That being said, my experiments showed that reverse-unreachable are very rare in the IR emitted by clang when bootstrapping clang. Here are the statistics I collected by analyzing IR between passes and after each removePredecessor call:
```
# functions: 52283
# samples: 337609
# reverse unreachable BBs: 216022
# BBs: 247840796
Percent reverse-unreachable: 0.08716159869015269 %
Max(PercRevUnreachable) in a function: 87.58620689655172 %
# > 25 % samples: 471 ( 0.1395104988314885 % samples )
... in 145 ( 0.27733680163724345 % functions )
```
Most of the reverse-unreachable regions come from invalid IR where it wouldn't be possible to construct a PostDomTree anyway.
I would like to commit this patch in the next week in order to be able to complete the work that depends on it before the end of my internship, so please don't wait long to voice your concerns :).
Reviewers: dberlin, sanjoy, grosser, brzycki, davide, chandlerc, hfinkel
Reviewed By: dberlin
Subscribers: nhaehnle, javed.absar, kparzysz, uabelho, jlebar, hiraditya, llvm-commits, dberlin, david2050
Differential Revision: https://reviews.llvm.org/D35851
llvm-svn: 310940
The pass does simplifications of well known AMD library calls.
If given -amdgpu-prelink option it works in a pre-link mode which
allows to reference new library functions which will be linked in
later.
In addition it also used to process traditional AMD option
-fuse-native which allows to replace some of the functions with
their fast native implementations from the library.
The necessary glue to pass the prelink option and translate
-fuse-native is to be added to the driver.
Differential Revision: https://reviews.llvm.org/D36436
llvm-svn: 310731
Summary:
Now that we've made all the necessary backend changes, we can add a new
intrinsic which exposes the new capabilities to IR producers. Since
llvm.amdgpu.update.dpp is a strict superset of llvm.amdgpu.mov.dpp, we
should deprecate the former. We also add tests for all the functionality
that was added in previous changes, now that we can access it via an IR
construct.
Reviewers: tstellar, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D34718
llvm-svn: 310399
Summary:
All instructions with the DPP modifier may not write to certain lanes of
the output if bound_ctrl=1 is set or any bits in bank_mask or row_mask
aren't set, so the destination register may be both defined and modified.
The right way to handle this is to add a constraint that the destination
register is the same as one of the inputs. We could tie the destination
to the first source, but that would be too restrictive for some use-cases
where we want the destination to be some other value before the
instruction executes. Instead, add a fake "old" source and tie it to the
destination. Effectively, the "old" source defines what value unwritten
lanes will get. We'll expose this functionality to users with a new
intrinsic later.
Also, we want to use DPP instructions for computing derivatives, which
means we need to set WQM for them. We also need to enable the entire
wavefront when using DPP intrinsics to implement nonuniform subgroup
reductions, since otherwise we'll get incorrect results in some cases.
To accomodate this, add a new operand to all DPP instructions which will
be interpreted by the SI WQM pass. This will be exposed with a new
intrinsic later. We'll also add support for Whole Wavefront Mode later.
I also fixed llvm.amdgcn.mov.dpp to overwrite the source and fixed up
the test. However, I could also keep the old behavior (where lanes that
aren't written are undefined) if people want it.
Reviewers: tstellar, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D34716
llvm-svn: 310283
Try to avoid mutually exclusive features. Don't use
a real default GPU, and use a fake "generic". The goal
is to make it easier to see which set of features are
incompatible between feature strings.
Most of the test changes are due to random scheduling changes
from not having a default fullspeed model.
llvm-svn: 310258
Summary:
This intrinsic lets us set inactive lanes to an identity value when
implementing wavefront reductions. In combination with Whole Wavefront
Mode, it lets inactive lanes be skipped over as required by GLSL/Vulkan.
Lowering the intrinsic needs to happen post-RA so that RA knows that the
destination isn't completely overwritten due to the EXEC shenanigans, so
we need another pseudo-instruction to represent the un-lowered
intrinsic.
Reviewers: tstellar, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D34719
llvm-svn: 310088
Summary:
Whole Wavefront Wode (WWM) is similar to WQM, except that all of the
lanes are always enabled, regardless of control flow. This is required
for implementing wavefront reductions in non-uniform control flow, where
we need to use the inactive lanes to propagate intermediate results, so
they need to be enabled. We need to propagate WWM to uses (unless
they're explicitly marked as exact) so that they also propagate
intermediate results correctly. We do the analysis and exec mask munging
during the WQM pass, since there are interactions with WQM for things
that require both WQM and WWM. For simplicity, WWM is entirely
block-local -- blocks are never WWM on entry or exit of a block, and WWM
is not propagated to the block level. This means that computations
involving WWM cannot involve control flow, but we only ever plan to use
WWM for a few limited purposes (none of which involve control flow)
anyways.
Shaders can ask for WWM using the @llvm.amdgcn.wwm intrinsic. There
isn't yet a way to turn WWM off -- that will be added in a future
change.
Finally, it turns out that turning on inactive lanes causes a number of
problems with register allocation. While the best long-term solution
seems like teaching LLVM's register allocator about predication, for now
we need to add some hacks to prevent ourselves from getting into trouble
due to constraints that aren't currently expressed in LLVM. For the gory
details, see the comments at the top of SIFixWWMLiveness.cpp.
Reviewers: arsenm, nhaehnle, tpr
Subscribers: kzhuravl, wdng, mgorny, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D35524
llvm-svn: 310087
Summary:
Previously, we assumed that certain types of instructions needed WQM in
pixel shaders, particularly DS instructions and image sampling
instructions. This was ok because with OpenGL, the assumption was
correct. But we want to start using DPP instructions for derivatives as
well as other things, so the assumption that we can infer whether to use
WQM based on the instruction won't continue to hold. This intrinsic lets
frontends like Mesa indicate what things need WQM based on their
knowledge of the API, rather than second-guessing them in the backend.
We need to keep around the old method of enabling WQM, but eventually we
should remove it once Mesa catches up. For now, this will let us use DPP
instructions for computing derivatives correctly.
Reviewers: arsenm, tpr, nhaehnle
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D35167
llvm-svn: 310085
In case if SI_KILL is in between of the SI_IF and SI_END_CF we need
to preserve the bits actually flipped by if rather then restoring
the original mask.
Differential Revision: https://reviews.llvm.org/D36299
llvm-svn: 310031
Summary:
Following the docs, we need at least 5 wait states between an EXEC write
and an instruction that uses DPP.
Reviewers: tstellar, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D34849
llvm-svn: 310013
Summary:
If a PHI has at lease one VGPR operand, we have to fix the PHI
in SIFixSGPRCopies.
Reviewer:
Matt
Differential Revision:
http://reviews.llvm.org/D34727
llvm-svn: 309959
This was failing on out of bounds access to the extra operands
on the s_swappc_b64 beyond those in the instruction definition.
This was working, but somehow regressed within the past few weeks,
although I don't see any obvious commit.
llvm-svn: 309782
This pattern shows up when lowering byval copies on AMDGPU.
The byval object access is split into 4-byte chunks, adding a
constant offset to the FixedStack base. When some of the offsets
turn into ors, this prevents combining the constant offsets.
This makes it not apparent that the object is there when matching
addressing modes, so it ends up using a scratch wave offset
relative access and the lengthy frame index expansion for that.
llvm-svn: 309775
With SI_END_CF elimination for some nested control flow we can now
eliminate saved exec register completely by turning a saveexec version
of instruction into just a logical instruction.
Differential Revision: https://reviews.llvm.org/D36007
llvm-svn: 309766
Add a pass to remove redundant S_OR_B64 instructions enabling lanes in
the exec. If two SI_END_CF (lowered as S_OR_B64) come together without any
vector instructions between them we can only keep outer SI_END_CF, given
that CFG is structured and exec bits of the outer end statement are always
not less than exec bit of the inner one.
This needs to be done before the RA to eliminate saved exec bits registers
but after register coalescer to have no vector registers copies in between
of different end cf statements.
Differential Revision: https://reviews.llvm.org/D35967
llvm-svn: 309762
The src0 register must match src1 or src2, but if these
were undefined they could end up using different implicit_defed
virtual registers. Force these to use one undef vreg or pick the
defined other register.
Also fixes producing invalid nodes without the right number of
inputs when src2 is undef.
llvm-svn: 309743
Includes a hack to fix the type selected for
the GlobalAddress of the function, which will be
fixed by changing the default datalayout to use
generic pointers for 0.
llvm-svn: 309732
Also refine the flat check to respect flat-for-global feature,
and constant fallback should check global handling, not
specifically MUBUF.
llvm-svn: 309471
This allows handling of a lot more of the interesting
cases in Blender. Most of the large functions unlikely
to be inlined have this pattern.
This is a special case for what clang emits for OpenCL 3
element vectors. Annoyingly, these are emitted as
<3 x elt>* pointers, but accessed as <4 x elt>* operations.
This also needs to handle cases where a struct containing
a single vector is used.
llvm-svn: 309419
It is better to return arguments directly in registers
if we are making a call rather than introducing expensive
stack usage. In one of sample compile from one of
Blender's many kernel variants, this fires on about
~20 different functions. Future improvements may be to
recognize simple cases where the pointer is indexing a small
array. This also fails when the store to the out argument
is in a separate block from the return, which happens in
a few of the Blender functions. This should also probably
be using MemorySSA which might help with that.
I'm not sure this is correct as a FunctionPass, but
MemoryDependenceAnalysis seems to not work with
a ModulePass.
I'm also not sure where it should run.I think it should
run before DeadArgumentElimination, so maybe either
EP_CGSCCOptimizerLate or EP_ScalarOptimizerLate.
llvm-svn: 309416
We need to pass something to functions for this to work.
It isn't derivable just from the kernarg segment pointer
because the implicit arguments are placed after the
kernel arguments.
Also fixes missing test for the intrinsic.
llvm-svn: 309398
Currently SI_IF results in a s_and_saveexec_b64 followed by s_xor_b64.
The xor is used to extract only the changed bits. In case of a simple
if region where the only use of that value is in the SI_END_CF to
restore the old exec mask, we can omit the xor and perform an or of
the exec mask with the original exec value saved by the
s_and_saveexec_b64.
Differential Revision: https://reviews.llvm.org/D35861
llvm-svn: 309185
Check the actual memory type stored and not the extended value size
when considering if truncated store merge is worthwhile.
Reviewers: efriedma, RKSimon, spatel, jyknight
Reviewed By: efriedma
Subscribers: llvm-commits, nhaehnle
Differential Revision: https://reviews.llvm.org/D35623
llvm-svn: 308833
On AMDGPU SGPR spills are really spilled to another register.
The spiller creates the spills to new frame index objects,
which is used as a placeholder.
This will eventually be replaced with a reference to a position
in a VGPR to write to and the frame index deleted. It is
most likely not a real stack location that can be shared
with another stack object.
This is a problem when StackSlotColoring decides it should
combine a frame index used for a normal VGPR spill with
a real stack location and a frame index used for an SGPR.
Add an ID field so that StackSlotColoring has a way
of knowing the different frame index types are
incompatible.
llvm-svn: 308673
As an approximation of the existing handling to avoid
regressions. Fixes using too many registers with calls
on subtargets with the SGPR allocation bug.
llvm-svn: 308326
Introduce pseudo-registers for registers needed for stack
access, which are replaced during finalizeLowering.
Note these pseudo-registers are currently only used for the
used register location, and not for determining their
input argument register.
This is better because it avoids the need to try to predict
whether a call will be emitted from the IR, and also
detects stack objects introduced by legalization.
Test changes are from the HasStackObjects check being more
accurate since stack objects introduced during legalization
are now known.
llvm-svn: 308325
This wasn't necessary before since they are always enabled
for kernels, but this is necessary if they need to be
forwarded to a callable function.
llvm-svn: 308226
Summary:
Previously, CodeGen checked first src operand type to determine if omod is supported by instruction. This isn't correct for some instructions: e.g. V_CMP_EQ_F32 has floating-point src operands but desn't support omod.
Changed .td files to check if dst operand instead of src operand.
Reviewers: arsenm, vpykhtin
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D35350
llvm-svn: 308179
The type needs to be casted back to the original argument type.
Fixes an assert that for some reason is only run when
using -debug.
Includes an additional combine to avoid test regressions
from having conversions mixed with multiple Assert[SZ]ext
nodes. On subtargets where i16 is legal, this was producing an i32
register with an i16 AssertZExt, truncated to i16 with another i8
AssertZExt.
t2: i32,ch = CopyFromReg t0, Register:i32 %vreg0
t3: i16 = truncate t2
t5: i16 = AssertZext t3, ValueType:ch:i8
t6: i8 = truncate t5
t7: i32 = zero_extend t6
llvm-svn: 308082
In moveToVALU(), move to vector ALU is performed, all instrs in
the use chain will be visited. We do not want the same node to be
pushed to the visit worklist more than once.
Differential Revision: https://reviews.llvm.org/D34726
llvm-svn: 308039
Since GFX9 supports denorm modes for v_min_f32/v_max_f32 that
is possible to further optimize fcanonicalize and remove it
if applied to min/max given their operands are known not to be
an sNaN or that sNaNs are not supported.
Additionally we can remove fcanonicalize if denorms are supported
for the VT and we know that its argument is never a NaN.
Differential Revision: https://reviews.llvm.org/D35335
llvm-svn: 307976
We are using multiplication by 1.0 to flush denormals and quiet sNaNs.
That is possible to omit this multiplication if source of the
fcanonicalize instruction is known to be flushed/quieted, i.e.
if it comes from another instruction known to do the normalization
and we are using IEEE mode to quiet sNaNs.
Differential Revision: https://reviews.llvm.org/D35218
llvm-svn: 307848
OpenCL 2.0 introduces the notion of memory scopes in atomic operations to
global and local memory. These scopes restrict how synchronization is
achieved, which can result in improved performance.
This change extends existing notion of synchronization scopes in LLVM to
support arbitrary scopes expressed as target-specific strings, in addition to
the already defined scopes (single thread, system).
The LLVM IR and MIR syntax for expressing synchronization scopes has changed
to use *syncscope("<scope>")*, where <scope> can be "singlethread" (this
replaces *singlethread* keyword), or a target-specific name. As before, if
the scope is not specified, it defaults to CrossThread/System scope.
Implementation details:
- Mapping from synchronization scope name/string to synchronization scope id
is stored in LLVM context;
- CrossThread/System and SingleThread scopes are pre-defined to efficiently
check for known scopes without comparing strings;
- Synchronization scope names are stored in SYNC_SCOPE_NAMES_BLOCK in
the bitcode.
Differential Revision: https://reviews.llvm.org/D21723
llvm-svn: 307722
Adds loop expansions for known-size and unknown-sized memcpy calls, allowing the
target to provide the operand types through TTI callbacks. The default values
for the TTI callbacks use int8 operand types and matches the existing behaviour
if they aren't overridden by the target.
Differential revision: https://reviews.llvm.org/D32536
llvm-svn: 307346
Regardless of relaxation options such as -cl-fast-relaxed-math
we are producing rather long code for fdiv via amdgcn_fdiv_fast
intrinsic. This intrinsic is used to replace fdiv with 2.5ulp
metadata and does not handle denormals, thus believed to be fast.
An fdiv instruction can also have fast math flag either by itself
or together with fpmath metadata. Clang used with a relaxation flag
always produces both metadata and fast flag:
%div = fdiv fast float %v, %0, !fpmath !12!12 = !{float 2.500000e+00}
Current implementation ignores fast flag and favors metadata. An
instruction with just fast flag would be lowered to a fastest rcp +
mul, but that never happen on practice because of described mutual
clang and BE behavior.
This change allows an "fdiv fast" to be always lowered as rcp + mul.
Differential Revision: https://reviews.llvm.org/D34844
llvm-svn: 307308
Summary:
During remat, some subranges might end up having invalid segments which caused problems for later
coalescing.
Added in a check to remove segments that are invalidated as part of the remat.
See http://llvm.org/PR33524
Subscribers: MatzeB, qcolombet
Differential Revision: https://reviews.llvm.org/D34391
llvm-svn: 307247
This is a short-term fix for PR33650 aimed to get the modules build bots green again.
Remove all the places where we use the LLVM_YAML_IS_(FLOW_)?SEQUENCE_VECTOR
macros to try to locally specialize a global template for a global type. That's
not how C++ works.
Instead, we now centrally define how to format vectors of fundamental types and
of string (std::string and StringRef). We use flow formatting for the former
cases, since that's the obvious right thing to do; in the latter case, it's
less clear what the right choice is, but flow formatting is really bad for some
cases (due to very long strings), so we pick block formatting. (Many of the
cases that were using flow formatting for strings are improved by this change.)
Other than the flow -> block formatting change for some vectors of strings,
this should result in no functionality change.
Differential Revision: https://reviews.llvm.org/D34907
Corresponding updates to clang, clang-tools-extra, and lld to follow.
llvm-svn: 306878
Given no NaNs and no signed zeroes it folds:
(fmul X, (select (fcmp X > 0.0), -1.0, 1.0)) -> (fneg (fabs X))
(fmul X, (select (fcmp X > 0.0), 1.0, -1.0)) -> (fabs X)
Differential Revision: https://reviews.llvm.org/D34579
llvm-svn: 306592
That is pretty common for clang to produce code like
(shl %x, (and %amt, 31)). In this situation we can still perform
trunc (shl) into shl (trunc) conversion given the known value
range of shift amount.
Differential Revision: https://reviews.llvm.org/D34723
llvm-svn: 306499
Depending on the compare code that can be either an argument of
sext or negate of it. This helps to avoid v_cndmask_b64 instruction
for sext. A reversed value can be further simplified and folded into
its parent comparison if possible.
Differential Revision: https://reviews.llvm.org/D34545
llvm-svn: 306446
Apparently this replacement can really be substituting the
same as the original register. Avoid restarting the loop
when there's been no change in the register uses.
llvm-svn: 306441
Also factored out function to check if a boolean is an already
deserialized value which does not require v_cndmask_b32 to be
loaded. Added binary logical operators to its check.
Differential Revision: https://reviews.llvm.org/D34500
llvm-svn: 306439
Summary:
1. Instruction V_CVT_U32_F32 allow omod operand (see SIInstrInfo.td:1435). In fact this operand shouldn't be allowed here. This fix checks if SDWA pseudo instruction has OMod operand and then copy it.
2. There were several problems with support of VOPC instructions in SDWA peephole pass.
Reviewers: tstellar, arsenm, vpykhtin, airlied, kzhuravl
Subscribers: wdng, nhaehnle, yaxunl, dstuttard, tpr, sarnex, t-tye
Differential Revision: https://reviews.llvm.org/D34626
llvm-svn: 306413
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
Remove invalid shortcut in fixupKills(): A register needs to be marked
live even when we are not adding a kill flag. This is because a
partially live register must not get a kill flags, but it still needs to
be fully marked live when walking backwards.
llvm-svn: 306352
Fixes bug 33597.
Use of substituteRegister in the tied operand case messes
up the register use iterator, causing some uses to be left
unprocessed.
llvm-svn: 306333
This should not be treated as a different version of
private_segment_buffer. These are distinct things with
different uses and register classes, and requires the
function argument info to have more context about the
function's type and environment.
Also add missing test coverage for the intrinsic, and
emit an error for HSA. This also encovers that the intrinsic
is broken unless there happen to be stack objects.
llvm-svn: 306264
Intrinsic already existed for llvm.SI.tbuffer.store
Needed tbuffer.load and also re-implementing the intrinsic as llvm.amdgcn.tbuffer.*
Added CodeGen tests for the 2 new variants added.
Left the original llvm.SI.tbuffer.store implementation to avoid issues with existing code
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tony-tye, tpr
Differential Revision: https://reviews.llvm.org/D30687
llvm-svn: 306031
Summary:
Despite that this instructions are listed in VOP2, they are treated as VOP3 in specs. They should not support SDWA.
There are no real instructions for them, but there are pseudo instructions.
Reviewers: arsenm, vpykhtin, cfang
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D34403
llvm-svn: 305999
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
If one of the arguments of adde/sube is zero we can fold another
add/sub into it.
Differential Revision: https://reviews.llvm.org/D34374
llvm-svn: 305964
This simplification allows to avoid generating v_cndmask_b32
to serialize condition code between compare and use.
Differential Revision: https://reviews.llvm.org/D34300
llvm-svn: 305962
If there is an immediate operand we shall not shrink V_SUBB_U32
and V_ADDC_U32, it does not fit e32 encoding.
Differential Revison: https://reviews.llvm.org/D34291
llvm-svn: 305840
Before it was possible to partially fold use instructions
before the defs. After the xor is folded into a copy, the same
mov can end up in the fold list twice, so on the second attempt
it will fail expecting to see a register to fold.
llvm-svn: 305821
This does some improvements/cleanup to the recently introduced
scavengeRegisterBackwards() functionality:
- Rewrite findSurvivorBackwards algorithm to use the existing
LiveRegUnit::accumulateBackward() code. This also avoids the Available
and Candidates bitset and just need 1 LiveRegUnit instance
(= 1 bitset).
- Pick registers in allocation order instead of register number order.
llvm-svn: 305817
The offset may not be an inline immediate, so this needs
to be materialized into a register. The post-RA run of
SIShrinkInstructions is able to fold it later if it can.
llvm-svn: 305761
It adds it for the target after inlining but before SROA where
we can get most out of it.
Differential Revision: https://reviews.llvm.org/D34366
llvm-svn: 305759
Re-apply r276044/r279124/r305516. Fixed a problem where we would refuse
to place spills as the very first instruciton of a basic block and thus
artifically increase pressure (test in
test/CodeGen/PowerPC/scavenging.mir:spill_at_begin)
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: 305625
Revert because of reports of some PPC input starting to spill when it
was predicted that it wouldn't and no spillslot was reserved.
This reverts commit r305516.
llvm-svn: 305566
Re-apply r276044/r279124. Trying to reproduce or disprove the ppc64
problems reported in the stage2 build last time, which I cannot
reproduce right now.
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: 305516
Summary:
Alloca promotion pass not dealing with non-canonical input
Added some additional checks so the pass simply backs-off forms it can't deal with (non-canonical)
Also added some test cases in non-canonical form to check that it no longer crashes
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D31710
llvm-svn: 305079
The V_MQSAD_PK_U16_U8, V_QSAD_PK_U16_U8, and V_MQSAD_U32_U8 take more than 1 pass in hardware. For these three instructions, the destination registers must be different than all sources, so that the first pass does not overwrite sources for the following passes.
Differential Revision: https://reviews.llvm.org/D33783
llvm-svn: 304998
If -simplify-mir option is passed then MIRPrinter will not print such fields.
This change also required some lit test cases in CodeGen directory to be changed.
Reviewed By: MatzeB
Differential Revision: https://reviews.llvm.org/D32304
llvm-svn: 304779
If a tied source operand was undef, it would be replaced but not
update the other tied operand, which would end up using different
virtual registers.
llvm-svn: 304747
Fixes bug #33302. Pass did not account that Src1 of max instruction
can be an immediate.
Differential Revision: https://reviews.llvm.org/D33884
llvm-svn: 304696
Remove dependency of SDWA pass on SIShrinkInstructions.
The goal is to move SDWA even higher in the stack to avoid second run
of MachineLICM, MachineCSE and SIFoldOperands.
Also added handling to preserve original src modifiers.
Differential Revision: https://reviews.llvm.org/D33860
llvm-svn: 304665
Summary:
These are mostly legal, but will probably need special lowering for some
cases.
Reviewers: arsenm
Reviewed By: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, igorb, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D33791
llvm-svn: 304628
SIFoldOperands can commute operands even if no folding was done.
This change is to preserve IR is no folding was done.
Differential Revision: https://reviews.llvm.org/D33802
llvm-svn: 304625
-enable-si-insert-waitcnts=1 becomes the default
-enable-si-insert-waitcnts=0 to use old pass
Differential Revision: https://reviews.llvm.org/D33730
llvm-svn: 304551
- new waitcnt pass remains off by default; -enable-si-insert-waitcnts=1 to enable it
- fix handling of PERMUTE ops
- fix insertion of waitcnt instrs at function begin/end ( port of analogous code that was added to old waitcnt pass )
- add new test
Differential Revision: https://reviews.llvm.org/D33114
llvm-svn: 304311
An encoding does not allow to use SDWA in an instruction with
scalar operands, either literals or SGPRs. That is however possible
to copy these operands into a VGPR first.
Several copies of the value are produced if multiple SDWA conversions
were done. To cleanup MachineLICM (to hoist copies out of loops),
MachineCSE (to remove duplicate copies) and SIFoldOperands (to replace
SGPR to VGPR copy with immediate copy right to the VGPR) runs are added
after the SDWA pass.
Differential Revision: https://reviews.llvm.org/D33583
llvm-svn: 304219
[AMDGPU] add intrinsic for s_getpc
Summary: The s_getpc instruction is exposed as intrinsic llvm.amdgcn.s.getpc.
Patch by Tim Corringham
llvm-svn: 304031
Rename the DEBUG_TYPE to match the names of corresponding passes where
it makes sense. Also establish the pattern of simply referencing
DEBUG_TYPE instead of repeating the passname where possible.
llvm-svn: 303921
This fixes 17 of the 41 -verify-machineinstrs test failures identified in PR33045
Differential Revision: https://reviews.llvm.org/D33451
llvm-svn: 303691
Summary:
Promoting Alloca to Vector and Promoting Alloca to LDS are two independent handling of Alloca and should not affect each other.
As a result, we should not give up promoting to vector if there is not enough LDS. This patch factors out the local memory usage
related checking out and replace it after the calling convention checking.
Reviewer:
arsenm
Differential Revision:
http://reviews.llvm.org/D33139
llvm-svn: 303684
Perform DAG combine:
and (srl x, c), mask => shl (bfe x, nb + c, mask >> nb), nb
Where nb is a number of trailing zeroes in mask.
It replaces two instructions with two and BFE is generally a more
expensive one. However this is only done if we are selecting a byte
or word at an aligned boundary which results in a proper SDWA
operand pattern. It is only done if SDWA is supported.
TODO: improve SDWA pass to actually convert this pattern. It is not
done now because we have an immediate in the instruction, which has
be moved into a VGPR.
Differential Revision: https://reviews.llvm.org/D33455
llvm-svn: 303681
shl (or|add x, c2), c1 => or|add (shl x, c1), (c2 << c1)
This allows to fold a constant into an address in some cases as
well as to eliminate second shift if the expression is used as
an address and second shift is a result of a GEP.
Differential Revision: https://reviews.llvm.org/D33432
llvm-svn: 303641
Turn expensive 64 bit shift into 32 bit if shift does not overflow int:
shl (ext x) => zext (shl x)
Differential Revision: https://reviews.llvm.org/D33367
llvm-svn: 303569
pruneSubRegValues() needs to remove subregister ranges starting at
instructions that later get removed by eraseInstrs(). It missed to check
one case in which eraseInstrs() would remove an instruction.
Fixes http://llvm.org/PR32688
llvm-svn: 303396
Summary:
There should be no intesection between SDWA operands and potential MIs. E.g.:
```
v_and_b32 v0, 0xff, v1 -> src:v1 sel:BYTE_0
v_and_b32 v2, 0xff, v0 -> src:v0 sel:BYTE_0
v_add_u32 v3, v4, v2
```
In that example it is possible that we would fold 2nd instruction into 3rd (v_add_u32_sdwa) and then try to fold 1st instruction into 2nd (that was already destroyed). So if SDWAOperand is also a potential MI then do not apply it.
Reviewers: vpykhtin, arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye
Differential Revision: https://reviews.llvm.org/D32804
llvm-svn: 303347
Partially implement callee-side for arguments and return values.
byval doesn't work properly, and most likely sret or other on-stack
return values most as well.
llvm-svn: 303308
Matt Arsenault