Summary:
For these loads that write to the HI part of a register, we should chain them to the op that writes to the LO part
of the register to maintain the appropriate order.
Reviewers:
rampitec, arsenm
Differential Revision:
https://reviews.llvm.org/D56454
llvm-svn: 351379
Summary:
This allows moving the condition from the intrinsic to the standard ICmp
opcode, so that LLVM can do simplifications on it. The icmp.i1 intrinsic
is an identity for retrieving the SGPR mask.
And we can also get the mask from and i1, or i1, xor i1.
Reviewers: arsenm, nhaehnle
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D52060
llvm-svn: 351150
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
This removes check for single use from general ShrinkDemandedConstant
to the BE because of the AArch64 regression after D56289/rL350475.
After several hours of experiments I did not come up with a testcase
failing on any other targets if check is not performed.
Moreover, direct call to ShrinkDemandedConstant is not really needed
and superceed by SimplifyDemandedBits.
Differential Revision: https://reviews.llvm.org/D56406
llvm-svn: 350684
Summary:
Moving SMRD to VMEM in SIFixSGPRCopies is rather bad for performance if
the load is really uniform. So select the scalar load intrinsics directly
to either VMEM or SMRD buffer loads based on divergence analysis.
If an offset happens to end up in a VGPR -- either because a floating
point calculation was involved, or due to other remaining deficiencies
in SIFixSGPRCopies -- we use v_readfirstlane.
There is some unrelated churn in tests since we now select MUBUF offsets
in a unified way with non-scalar buffer loads.
Change-Id: I170e6816323beb1348677b358c9d380865cd1a19
Reviewers: arsenm, alex-t, rampitec, tpr
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D53283
llvm-svn: 348050
Summary:
The VirtReg2Value mapping is crucial for getting consistently
reliable divergence information into the SelectionDAG. This
patch fixes a bunch of issues that lead to incorrect divergence
info and introduces tight assertions to ensure we don't regress:
1. VirtReg2Value is generated lazily; there were some cases where
a lookup was performed before all relevant virtual registers were
created, leading to an out-of-sync mapping. Those cases were:
- Complex code to lower formal arguments that generated CopyFromReg
nodes from live-in registers (fixed by never querying the mapping
for live-in registers).
- Code that generates CopyToReg for formal arguments that are used
outside the entry basic block (fixed by never querying the
mapping for Register nodes, which don't need the divergence info
anyway).
2. For complex values that are lowered to a sequence of registers,
all registers must be reflected in the VirtReg2Value mapping.
I am not adding any new tests, since I'm not actually aware of any
bugs that these problems are causing with trunk as-is. However,
I recently added a test case (in r346423) which fails when D53283 is
applied without this change. Also, the new assertions should provide
most of the effective test coverage.
There is one test change in sdwa-peephole.ll. The underlying issue
is that since the divergence info is now correct, the DAGISel will
select V_OR_B32 directly instead of S_OR_B32. This leads to an extra
COPY which affects the behavior of MachineLICM in a way that ends up
with the S_MOV_B32 with the constant in a different basic block than
the V_OR_B32, which is presumably what defeats the peephole.
Reviewers: alex-t, arsenm, rampitec
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D54340
llvm-svn: 348049
Also revert fix r347876
One of the buildbots was reporting a failure in some relevant tests that I can't
repro or explain at present, so reverting until I can isolate.
llvm-svn: 347911
My change svn-id: 347871 caused a buildbot failure due to an unused
variable def (used in an assert).
Change-Id: Ia882d18bb6fa79b4d7bbfda422b9ea5d23eab336
llvm-svn: 347876
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
llvm-svn: 347871
This allows to avoid scratch use or indirect VGPR addressing for
small vectors.
Differential Revision: https://reviews.llvm.org/D54606
llvm-svn: 347231
An extractelement with non-constant index will be lowered either to
scratch or movrel loop in most cases. This patch converts such
instruction into a set of selects if vector size is not too big.
Differential Revision: https://reviews.llvm.org/D54351
llvm-svn: 346800
The main caller of this already has an MVT and several targets called getSimpleVT inside without checking isSimple. This makes the simpleness explicit.
llvm-svn: 346180
UBSan detected an error in our ISelLowering that is exposed only when
you have a dmask == 0x1. Fix this by adding in an explicit check to
ensure we don't do the UBSan detected shl << 32.
llvm-svn: 345962
This patch should not introduce any behavior changes. It consists of
mostly one of two changes:
1. Replacing fall through comments with the LLVM_FALLTHROUGH macro
2. Inserting 'break' before falling through into a case block consisting
of only 'break'.
We were already using this warning with GCC, but its warning behaves
slightly differently. In this patch, the following differences are
relevant:
1. GCC recognizes comments that say "fall through" as annotations, clang
doesn't
2. GCC doesn't warn on "case N: foo(); default: break;", clang does
3. GCC doesn't warn when the case contains a switch, but falls through
the outer case.
I will enable the warning separately in a follow-up patch so that it can
be cleanly reverted if necessary.
Reviewers: alexfh, rsmith, lattner, rtrieu, EricWF, bollu
Differential Revision: https://reviews.llvm.org/D53950
llvm-svn: 345882
Our a16 support was only enabled for sample/gather and buffer
load/store, but not for image load/store operations (which take an i16
as the pixel index rather than a half).
Fix our isel lowering and add test cases to prove it out.
Differential Revision: https://reviews.llvm.org/D53750
llvm-svn: 345710
Introduce new versions that follow the IEEE semantics
to help with legalization that may need quieted inputs.
There are some regressions from inserting unnecessary
canonicalizes when these are matched from fast math
fcmp + select which should be fixed in a future commit.
llvm-svn: 344914
Summary:
To workaround a hardware issue in the (base + offset) calculation
when base is negative. The impact on code quality should be limited
since SILoadStoreOptimizer still runs afterwards and is able to
combine loads/stores based on known sign information.
This fixes visible corruption in Hitman on SI (easily reproducible
by running benchmark mode).
Change-Id: Ia178d207a5e2ac38ae7cd98b532ea2ae74704e5f
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=99923
Reviewers: arsenm, mareko
Subscribers: jholewinski, kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D53160
llvm-svn: 344698
Summary:
Moving SMRD to VMEM in SIFixSGPRCopies is rather bad for performance if
the load is really uniform. So select the scalar load intrinsics directly
to either VMEM or SMRD buffer loads based on divergence analysis.
If an offset happens to end up in a VGPR -- either because a floating
point calculation was involved, or due to other remaining deficiencies
in SIFixSGPRCopies -- we use v_readfirstlane.
There is some unrelated churn in tests since we now select MUBUF offsets
in a unified way with non-scalar buffer loads.
Change-Id: I170e6816323beb1348677b358c9d380865cd1a19
Reviewers: arsenm, alex-t, rampitec, tpr
Subscribers: kzhuravl, jvesely, wdng, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D53283
llvm-svn: 344696
The isAmdCodeObjectV2 is a misleading name which actually checks whether the os
is amdhsa or mesa.
Also add a test to make sure we do not generate old kernel header for code
object v3.
Differential Revision: https://reviews.llvm.org/D52897
llvm-svn: 343813
Summary:
The new buffer/tbuffer intrinsics handle an out-of-range immediate
offset by moving/adding offset&-4096 to a vgpr, leaving an in-range
immediate offset, with a chance of the move/add being CSEd for similar
loads/stores.
However it turns out that a negative offset in a vgpr is illegal, even
if adding the immediate offset makes it legal again.
Therefore, this commit disables the offset&-4096 thing if the offset is
negative.
Differential Revision: https://reviews.llvm.org/D52683
Change-Id: Ie02f0a74f240a138dc2a29d17cfbd9e350e4ed13
llvm-svn: 343672
If the alignment is at least 4, this should report true.
Something still seems off with how < 4-byte types are
handled here though.
Fixing this seems to change how some combines get
to where they get, but somehow isn't changing the net
result.
llvm-svn: 342879
Summary:
GFX9 and above support sin/cos instructions with a greater range and thus don't
require a fract instruction prior to invocation.
Added a subtarget feature to reflect this and added code to take advantage of
expanded range on GFX9+
Also updated the tests to check correct behaviour
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D51933
Change-Id: I1c1f1d3726a5ae32116646ca5cfa1ab4ef69e5b0
llvm-svn: 342222
If an argument was passed on the stack, this
was using the default alignment.
I'm not sure there's an observable change from this. This
was observable due to bugs in expansion of unaligned
loads and stores, but since that is fixed I don't think
this matters much.
llvm-svn: 342133
This already worked if only one register piece was used,
but didn't if a type was split into multiple, unequal
sized pieces.
Fixes not splitting 3i16/v3f16 into two registers for
AMDGPU.
This will also allow fixing the ABI for 16-bit vectors
in a future commit so that it's the same for all subtargets.
llvm-svn: 341801
The intention is to enable the extract_vector_elt load combine,
and doing this for other operations interferes with more
useful optimizations on vectors.
Handle any type of load since in principle we should do the
same combine for the various load intrinsics.
llvm-svn: 341219
Summary:
This is patch 1 of the new DivergenceAnalysis (https://reviews.llvm.org/D50433).
The purpose of this patch is to free up the name DivergenceAnalysis for the new generic
implementation. The generic implementation class will be shared by specialized
divergence analysis classes.
Patch by: Simon Moll
Reviewed By: nhaehnle
Subscribers: jvesely, jholewinski, arsenm, nhaehnle, mgorny, jfb, llvm-commits
Differential Revision: https://reviews.llvm.org/D50434
Change-Id: Ie8146b11be2c50d5312f30e11c7a3036a15b48cb
llvm-svn: 341071
Summary:
Patch by Marek Olsak and David Stuttard, both of AMD.
This adds a new amdgcn intrinsic supporting s.buffer.load, in particular
multiple dword variants. These are convenient to use from some front-end
implementations.
Also modified the existing llvm.SI.load.const intrinsic to common up the
underlying implementation.
This modification also requires that we can lower to non-uniform loads correctly
by splitting larger dword variants into sizes supported by the non-uniform
versions of the load.
V2: Addressed minor review comments.
V3: i1 glc is now i32 cachepolicy for consistency with buffer and
tbuffer intrinsics, plus fixed formatting issue.
V4: Added glc test.
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D51098
Change-Id: I83a6e00681158bb243591a94a51c7baa445f169b
llvm-svn: 340684
This was hackily adding in the 4-bytes reserved for the callee's
emergency stack slot. Treat it like a normal stack allocation
so we get the correct alignment padding behavior. This fixes
an inconsistency between the caller and callee.
llvm-svn: 340396
Summary:
This commit adds new intrinsics
llvm.amdgcn.raw.buffer.load
llvm.amdgcn.raw.buffer.load.format
llvm.amdgcn.raw.buffer.load.format.d16
llvm.amdgcn.struct.buffer.load
llvm.amdgcn.struct.buffer.load.format
llvm.amdgcn.struct.buffer.load.format.d16
llvm.amdgcn.raw.buffer.store
llvm.amdgcn.raw.buffer.store.format
llvm.amdgcn.raw.buffer.store.format.d16
llvm.amdgcn.struct.buffer.store
llvm.amdgcn.struct.buffer.store.format
llvm.amdgcn.struct.buffer.store.format.d16
llvm.amdgcn.raw.buffer.atomic.*
llvm.amdgcn.struct.buffer.atomic.*
with the following changes from the llvm.amdgcn.buffer.*
intrinsics:
* there are separate raw and struct versions: raw does not have an
index arg and sets idxen=0 in the instruction, and struct always sets
idxen=1 in the instruction even if the index is 0, to allow for the
fact that gfx9 does bounds checking differently depending on whether
idxen is set;
* there is a combined cachepolicy arg (glc+slc)
* there are now only two offset args: one for the offset that is
included in bounds checking and swizzling, to be split between the
instruction's voffset and immoffset fields, and one for the offset
that is excluded from bounds checking and swizzling, to go into the
instruction's soffset field.
The AMDISD::BUFFER_* SD nodes always have an index operand, all three
offset operands, combined cachepolicy operand, and an extra idxen
operand.
The obsolescent llvm.amdgcn.buffer.* intrinsics continue to work.
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, t-tye, jfb, llvm-commits
Differential Revision: https://reviews.llvm.org/D50306
Change-Id: If897ea7dc34fcbf4d5496e98cc99a934f62fc205
llvm-svn: 340269
Summary:
This commit adds new intrinsics
llvm.amdgcn.raw.tbuffer.load
llvm.amdgcn.struct.tbuffer.load
llvm.amdgcn.raw.tbuffer.store
llvm.amdgcn.struct.tbuffer.store
with the following changes from the llvm.amdgcn.tbuffer.* intrinsics:
* there are separate raw and struct versions: raw does not have an index
arg and sets idxen=0 in the instruction, and struct always sets
idxen=1 in the instruction even if the index is 0, to allow for the
fact that gfx9 does bounds checking differently depending on whether
idxen is set;
* there is a combined format arg (dfmt+nfmt)
* there is a combined cachepolicy arg (glc+slc)
* there are now only two offset args: one for the offset that is
included in bounds checking and swizzling, to be split between the
instruction's voffset and immoffset fields, and one for the offset
that is excluded from bounds checking and swizzling, to go into the
instruction's soffset field.
The AMDISD::TBUFFER_* SD nodes always have an index operand, all three
offset operands, combined format operand, combined cachepolicy operand,
and an extra idxen operand.
The tbuffer pseudo- and real instructions now also have a combined
format operand.
The obsolescent llvm.amdgcn.tbuffer.* and llvm.SI.tbuffer.store
intrinsics continue to work.
V2: Separate raw and struct intrinsics.
V3: Moved extract_glc and extract_slc defs to a more sensible place.
V4: Rebased on D49995.
V5: Only two separate offset args instead of three.
V6: Pseudo- and real instructions have joint format operand.
V7: Restored optionality of dfmt and nfmt in assembler.
V8: Addressed minor review comments.
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D49026
Change-Id: If22ad77e349fac3a5d2f72dda53c010377d470d4
llvm-svn: 340268
a generically extensible collection of extra info attached to
a `MachineInstr`.
The primary change here is cleaning up the APIs used for setting and
manipulating the `MachineMemOperand` pointer arrays so chat we can
change how they are allocated.
Then we introduce an extra info object that using the trailing object
pattern to attach some number of MMOs but also other extra info. The
design of this is specifically so that this extra info has a fixed
necessary cost (the header tracking what extra info is included) and
everything else can be tail allocated. This pattern works especially
well with a `BumpPtrAllocator` which we use here.
I've also added the basic scaffolding for putting interesting pointers
into this, namely pre- and post-instruction symbols. These aren't used
anywhere yet, they're just there to ensure I've actually gotten the data
structure types correct. I'll flesh out support for these in
a subsequent patch (MIR dumping, parsing, the works).
Finally, I've included an optimization where we store any single pointer
inline in the `MachineInstr` to avoid the allocation overhead. This is
expected to be the overwhelmingly most common case and so should avoid
any memory usage growth due to slightly less clever / dense allocation
when dealing with >1 MMO. This did require several ergonomic
improvements to the `PointerSumType` to reasonably support the various
usage models.
This also has a side effect of freeing up 8 bits within the
`MachineInstr` which could be repurposed for something else.
The suggested direction here came largely from Hal Finkel. I hope it was
worth it. ;] It does hopefully clear a path for subsequent extensions
w/o nearly as much leg work. Lots of thanks to Reid and Justin for
careful reviews and ideas about how to do all of this.
Differential Revision: https://reviews.llvm.org/D50701
llvm-svn: 339940
This will allow the library to just use __builtin_expf directly
without expanding this itself. Note f64 still won't work because
there is no exp instruction for it.
llvm-svn: 339902
Handle fmul, fsub and preserve flags.
Also really test minnum/maxnum reductions.
The existing tests were only checking from
minnum/maxnum matched from a fast math compare
and select which is not the same.
llvm-svn: 339820
`MachineMemOperand` pointers attached to `MachineSDNodes` and instead
have the `SelectionDAG` fully manage the memory for this array.
Prior to this change, the memory management was deeply confusing here --
The way the MI was built relied on the `SelectionDAG` allocating memory
for these arrays of pointers using the `MachineFunction`'s allocator so
that the raw pointer to the array could be blindly copied into an
eventual `MachineInstr`. This creates a hard coupling between how
`MachineInstr`s allocate their array of `MachineMemOperand` pointers and
how the `MachineSDNode` does.
This change is motivated in large part by a change I am making to how
`MachineFunction` allocates these pointers, but it seems like a layering
improvement as well.
This would run the risk of increasing allocations overall, but I've
implemented an optimization that should avoid that by storing a single
`MachineMemOperand` pointer directly instead of allocating anything.
This is expected to be a net win because the vast majority of uses of
these only need a single pointer.
As a side-effect, this makes the API for updating a `MachineSDNode` and
a `MachineInstr` reasonably different which seems nice to avoid
unexpected coupling of these two layers. We can map between them, but we
shouldn't be *surprised* at where that occurs. =]
Differential Revision: https://reviews.llvm.org/D50680
llvm-svn: 339740
If one of the elements is undef, use the canonicalized constant
from the other element instead of 0.
Splat vectors are more useful for other optimizations, such
as matching vector clamps. This was breaking on clamps
of half3 from the undef 4th component.
llvm-svn: 339512
Everything should quiet, and I think everything should
flush.
I assume the min3/med3/max3 follow the same rules
as regular min/max for flushing, which should at
least be conservatively correct.
There are still more operations that need to
be handled.
llvm-svn: 339065
Not sure why this was checking for denormals for f16.
My interpretation of the IEEE standard is conversions
should produce a canonical result, and the ISA manual
says denormals are created when appropriate.
llvm-svn: 339064
If denormals are enabled, denormals are canonical.
Also fix a few other issues. minnum/maxnum are supposed
to canonicalize. Temporarily improve workaround for the
instruction behavior change in gfx9.
Handle selects and fcopysign.
The tests were also largely broken, since they were
checking for a flush used on some targets after the
store of the result.
llvm-svn: 339061
Add a parameter for testing specifically for
sNaNs - at least one instruction pattern on AMDGPU
needs to check specifically for this.
Also handle more cases, and add a target hook
for custom nodes, similar to the hooks for known
bits.
llvm-svn: 338910
Summary:
By not reconstructing the operand list of the SDNode, this change makes
it easier to add the forthcoming new tbuffer and buffer intrinsics.
Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D49995
Change-Id: I0cb79ef0801532645d7dd954a6d7355139db7b38
llvm-svn: 338784
Mutate the node type during selection when it
doesn't matter. This avoids an intermediate bitcast
node on targets with legal i16/f16.
Also fixes missing output modifiers on v_cvt_pkrtz_f32_f16,
which I assume are OK.
llvm-svn: 338619
Summary:
Add _L to _LZ image intrinsic table mapping to table gen.
In ISelLowering check if image intrinsic has lod and if it's equal
to zero, if so remove lod and change opcode to equivalent mapped _LZ.
Change-Id: Ie24cd7e788e2195d846c7bd256151178cbb9ec71
Subscribers: arsenm, mehdi_amini, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, steven_wu, dexonsmith, llvm-commits
Differential Revision: https://reviews.llvm.org/D49483
llvm-svn: 338523
When lowering calling conventions, prefer to decompose vectors
into the constitute register types. This avoids artifical constraints
to satisfy a wide super-register.
This improves code quality because now optimizations don't need to
deal with the super-register constraint. For example the immediate
folding code doesn't deal with 4 component reg_sequences, so by
breaking the register down earlier the existing immediate folding
code is able to work.
This also avoids the need for the shader input processing code
to manually split vector types.
llvm-svn: 338416
We could choose a free 0 for this, but this
matches the behavior for fmul undef, 1.0. Also,
the NaN use is more useful for folding use operations
although if it's not eliminated it is more expensive
in terms of code size.
llvm-svn: 338376
SelectionDAGBuilder widens v3i32/v3f32 arguments to
to v4i32/v4f32 which consume an additional register.
In addition to wasting argument space, this produces extra
instructions since now it appears the 4th vector component has
a meaningful value to most combines.
llvm-svn: 338197
This reverts commit r337021.
WARNING: MemorySanitizer: use-of-uninitialized-value
#0 0x1415cd65 in void write_signed<long>(llvm::raw_ostream&, long, unsigned long, llvm::IntegerStyle) /code/llvm-project/llvm/lib/Support/NativeFormatting.cpp:95:7
#1 0x1415c900 in llvm::write_integer(llvm::raw_ostream&, long, unsigned long, llvm::IntegerStyle) /code/llvm-project/llvm/lib/Support/NativeFormatting.cpp:121:3
#2 0x1472357f in llvm::raw_ostream::operator<<(long) /code/llvm-project/llvm/lib/Support/raw_ostream.cpp:117:3
#3 0x13bb9d4 in llvm::raw_ostream::operator<<(int) /code/llvm-project/llvm/include/llvm/Support/raw_ostream.h:210:18
#4 0x3c2bc18 in void printField<unsigned int, &(amd_kernel_code_s::amd_kernel_code_version_major)>(llvm::StringRef, amd_kernel_code_s const&, llvm::raw_ostream&) /code/llvm-project/llvm/lib/Target/AMDGPU/Utils/AMDKernelCodeTUtils.cpp:78:23
#5 0x3c250ba in llvm::printAmdKernelCodeField(amd_kernel_code_s const&, int, llvm::raw_ostream&) /code/llvm-project/llvm/lib/Target/AMDGPU/Utils/AMDKernelCodeTUtils.cpp:104:5
#6 0x3c27ca3 in llvm::dumpAmdKernelCode(amd_kernel_code_s const*, llvm::raw_ostream&, char const*) /code/llvm-project/llvm/lib/Target/AMDGPU/Utils/AMDKernelCodeTUtils.cpp:113:5
#7 0x3a46e6c in llvm::AMDGPUTargetAsmStreamer::EmitAMDKernelCodeT(amd_kernel_code_s const&) /code/llvm-project/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp:161:3
#8 0xd371e4 in llvm::AMDGPUAsmPrinter::EmitFunctionBodyStart() /code/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:204:26
[...]
Uninitialized value was created by an allocation of 'KernelCode' in the stack frame of function '_ZN4llvm16AMDGPUAsmPrinter21EmitFunctionBodyStartEv'
#0 0xd36650 in llvm::AMDGPUAsmPrinter::EmitFunctionBodyStart() /code/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:192
llvm-svn: 337079
This needs to refer to arguments by their original argument
index, not the argument split index which depends on what
the type splitting decides to do.
Also avoid increment PSInputNum for each split piece.
llvm-svn: 337022
This was completely broken if there was ever a struct argument, as
this information is thrown away during the argument analysis.
The offsets as passed in to LowerFormalArguments are not useful,
as they partially depend on the legalized result register type,
and they don't consider the alignment in the first place.
Ignore the Ins array, and instead figure out from the raw IR type
what we need to do. This seems to fix the padding computation
if the DAG lowering is forced (and stops breaking arguments
following padded arguments if the arguments were only partially
lowered in the IR)
llvm-svn: 337021
Summary:
This is a follow-up to r335942.
- Merge SISubtarget into AMDGPUSubtarget and rename to GCNSubtarget
- Rename AMDGPUCommonSubtarget to AMDGPUSubtarget
- Merge R600Subtarget::Generation and GCNSubtarget::Generation into
AMDGPUSubtarget::Generation.
Reviewers: arsenm, jvesely
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D49037
llvm-svn: 336851
Summary:
We now have two sets of generated TableGen files, one for R600 and one
for GCN, so each sub-target now has its own tables of instructions,
registers, ISel patterns, etc. This should help reduce compile time
since each sub-target now only has to consider information that
is specific to itself. This will also help prevent the R600
sub-target from slowing down new features for GCN, like disassembler
support, GlobalISel, etc.
Reviewers: arsenm, nhaehnle, jvesely
Reviewed By: arsenm
Subscribers: MatzeB, kzhuravl, wdng, mgorny, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D46365
llvm-svn: 335942
We have too many mechanisms for tracking the various offsets
used for kernel arguments, so remove one. There's still a lot of
confusion with these because there are two different "implicit"
argument areas located at the beginning and end of the kernarg
segment.
Additionally, the offset was determined based on the memory
size of the split element types. This would break in a future
commit where v3i32 is decomposed into separate i32 pieces.
llvm-svn: 335830
In principle nothing should stop these from working, but
work is necessary to create an ABI for dealing with the stack
related registers.
llvm-svn: 335829
If a source of rcp instruction is a result of any conversion from
an integer convert it into rcp_iflag instruction. No FP exception
can ever happen except division by zero if a single precision rcp
argument is a representation of an integral number.
Differential Revision: https://reviews.llvm.org/D48569
llvm-svn: 335742
Summary:
This also removes the need for atomic pseudo instructions, since
we select the correct encoding directly in SITargetLowering::lowerImage
for dimension-aware image intrinsics.
Mesa uses dimension-aware image intrinsics since
commit a9a7993441.
Change-Id: I7473d20009476a4ed6d919cae4e6dca9ff42e77a
Reviewers: arsenm, rampitec, mareko, tpr, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D48167
llvm-svn: 335231
Summary:
Having TableGen patterns for image intrinsics is hitting limitations:
for D16 we already have to manually pre-lower the packing of data
values, and we will have to do the same for A16 eventually.
Since there is already some custom C++ code anyway, it is arguably easier
to just do everything in C++, now that we can use the beefed-up generic
tables backend of TableGen to provide all the required metadata and map
intrinsics to corresponding opcodes. With this approach, all image
intrinsic lowering happens in SITargetLowering::lowerImage. That code is
dense due to all the cases that it handles, but it should still be easier
to follow than what we had before, by virtue of it all being done in a
single location, and by virtue of not relying on the TableGen pattern
magic that very few people really understand.
This means that we will have MachineSDNodes with MIMG instructions
during DAG combining, but that seems alright: previously we had
intrinsic nodes instead, but those are similarly opaque to the generic
CodeGen infrastructure, and the final pattern matching just did a 1:1
translation to machine instructions anyway. If anything, the fact that
we now merge the address words into a vector before DAG combine should
be an advantage.
Change-Id: I417f26bd88f54ce9781c1668acc01f3f99774de6
Reviewers: arsenm, rampitec, rtaylor, tstellar
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D48017
llvm-svn: 335228
Summary:
This allows us to access rich information about MIMG opcodes from C++ code.
Simplifying the mapping between equivalent opcodes of different data size
becomes quite natural.
This also flattens the MIMG-related class and multiclass hierarchy a little,
and collapses together some of the scaffolding for sample and gather4 opcodes.
Change-Id: I1a2549fdc1e881ff100e5393d2d87e73729a0ccd
Reviewers: arsenm, rampitec
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D48016
llvm-svn: 335227
Summary:
This allows us to reduce the number of different machine instruction
opcodes, which reduces the table sizes and helps flatten the TableGen
multiclass hierarchies.
We can do this because for each hardware MIMG opcode, we have a full set
of IMAGE_xxx_Vn_Vm machine instructions for all required sizes of vdata
and vaddr registers. Instead of having separate D16 machine instructions,
a packed D16 instructions loading e.g. 4 components can simply use the
same V2 opcode variant that non-D16 instructions use.
We still require a TSFlag for D16 buffer instructions, because the
D16-ness of buffer instructions is part of the opcode. Renaming the flag
should help avoid future confusion.
The one non-obvious code change is that for gather4 instructions, the
disassembler can no longer automatically decide whether to use a V2 or
a V4 variant. The existing logic which choose the correct variant for
other MIMG instruction is extended to cover gather4 as well.
As a bonus, some of the assembler error messages are now more helpful
(e.g., complaining about a wrong data size instead of a non-existing
instruction).
While we're at it, delete a whole bunch of dead legacy TableGen code.
Change-Id: I89b02c2841c06f95e662541433e597f5d4553978
Reviewers: arsenm, rampitec, kzhuravl, artem.tamazov, dp, rtaylor
Subscribers: wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D47434
llvm-svn: 335222
This is the common case in the BE when we serialize condition and then
rematerialize it. Use either original or inverted condition.
Differential Revision: https://reviews.llvm.org/D48246
llvm-svn: 334882
Try to access pieces 4 bytes at a time. This helps
various hasOneUse extract_vector_elt combines, such
as load width reductions.
Avoids test regressions in a future commit.
llvm-svn: 334836
Summary:
The code that handles ISD:Register and ISD::CopyFromReg assumes
the target is amdgcn, so this is broken on r600. We don't
need this analysis on r600 anyway so we can safely move
it to SITargetLowering.
Reviewers: alex-t, arsenm, nhaehnle
Reviewed By: arsenm
Subscribers: msearles, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D46298
llvm-svn: 334607
AMDGPU inline assembler support i16, half and i128 typed variables in constraints, but they were reported as error.
Needed to fix https://github.com/RadeonOpenCompute/ROCm/issues/341,
e.g. to be able to load with global_load_dwordx4 to a 128bit integer variable
Differential Revision: https://reviews.llvm.org/D44920
llvm-svn: 334301
This has two main components. First, widen
widen short constant loads in DAG when they have
the correct alignment. This is already done a bit in
AMDGPUCodeGenPrepare, since that has access to
DivergenceAnalysis. This can't help kernarg loads
created in the DAG. Start to use DAG divergence analysis
to help this case.
The second part is to avoid kernel argument lowering
breaking the alignment of short vector elements because
calling convention lowering wants to split everything
into legal register types.
When loading a split type, load the nearest 4-byte aligned
segment and shift to get the desired bits. This extra
load of the earlier argument piece ends up merging,
and the bit extract hopefully folds out.
There are a number of improvements and regressions with
this, but I think as-is this is a better compromise between
several of the worst parts of SelectionDAG.
Particularly when i16 is legal, this produces worse code
for i8 and i16 element vector kernel arguments. This is
partially due to the very weak load merging the DAG does.
It only looks for fairly specific combines between pairs
of loads which no longer appear. In particular this
causes v4i16 loads to be split into 2 components when
previously the two halves were merged.
Worse, because of the newly introduced shifts, there
is a lot more unnecessary vector packing and unpacking code
emitted. At least some of this is due to reporting
false for isTypeDesirableForOp for i16 as a workaround for
the lack of divergence information in the DAG. The cases
where this happens it doesn't actually matter, but the
relevant code in SimplifyDemandedBits doens't have the context
to know to ignore this.
The use of the scalar cache is probably more important
than the mess of mostly scalar instructions doing this packing
and unpacking. Future work can fix this, possibly by making better
use of the new DAG divergence information for controlling promotion
decisions, or adding another version of shift + trunc + shift
combines that doesn't only know about the used types.
llvm-svn: 334180
Fixes terrible code on targets without f16 support. The
legalization creates a mess that is difficult to recover
from. Also should avoid randomly breaking these tests
multiple times in sequence in future commits.
Some regressions in cases where it happens to be better
to pull the source modifier after the conversion.
llvm-svn: 334132
Summary:
They've been deprecated in favor of UADDO/ADDCARRY or USUBO/SUBCARRY for a while.
Target that uses these opcodes are changed in order to ensure their behavior doesn't change.
Reviewers: efriedma, craig.topper, dblaikie, bkramer
Subscribers: jholewinski, arsenm, jyknight, sdardis, nemanjai, nhaehnle, kbarton, fedor.sergeev, asb, rbar, johnrusso, simoncook, jordy.potman.lists, apazos, sabuasal, niosHD, jrtc27, zzheng, edward-jones, mgrang, atanasyan, llvm-commits
Differential Revision: https://reviews.llvm.org/D47422
llvm-svn: 333748
Keep track of achieved occupancy in SIMachineFunctionInfo.
At the moment we have a lot of duplicated or even missed code to
query and maintain occupancy info. Record it in the MFI and
query in a single call. Interfaces:
- getOccupancy() - returns current recorded achieved occupancy.
- getMinAllowedOccupancy() - returns lesser of the achieved occupancy
and the lowest occupancy we are ready to tolerate. For example if
a kernel is memory bound we are ready to tolerate 4 waves.
- limitOccupancy() - record occupancy level if we have to lower it.
- increaseOccupancy() - record occupancy if scheduler managed to
increase the occupancy.
MFI takes care of integrating different checks affecting occupancy,
including LDS use and waves-per-eu attribute. Note that scheduler
starts with not yet known register pressure, so has to record either
limit or increase in occupancy after it is done. Later passes can
just query a resulting value.
New interface is used in the active scheduler and NFC wrt its work.
Changes are also made to experimental schedulers to use it and record
an occupancy after they are done. Before the change waves-per-eu was
ignored by experimental schedulers and tolerance window for memory
bound kernels was not used.
Differential Revision: https://reviews.llvm.org/D47509
llvm-svn: 333629
This was just emitting loads with the ABI alignment
for the raw type. The true alignment is often better,
especially when an illegal vector type was scalarized.
The better alignment allows using a scalar load
more often.
llvm-svn: 333558
This usually results in better code. Fixes using
inline asm with short2, and also fixes having a different
ABI for function parameters between VI and gfx9.
Partially cleans up the mess used for lowering of the d16
operations. Making v4f16 legal will help clean this up more,
but this requires additional work.
llvm-svn: 332953
Summary:
MCTargetDesc/AMDGPUMCTargetDesc.h contains enums for all the instuction
and register defintions, which are huge so we only want to include
them where needed.
This will also make it easier if we want to split the R600 and GCN
definitions into separate tablegenerated files.
I was unable to remove AMDGPUMCTargetDesc.h from SIMachineFunctionInfo.h
because it uses some enums from the header to initialize default values
for the SIMachineFunction class, so I ended up having to remove includes of
SIMachineFunctionInfo.h from headers too.
Reviewers: arsenm, nhaehnle
Reviewed By: nhaehnle
Subscribers: MatzeB, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D46272
llvm-svn: 332930
It is legal for the type passed to isLegalAddressingMode to be
unsized or, more specifically, VoidTy. In this case, we must
check the legality of load / stores for all legal types. Directly
trying to call getTypeStoreSize is incorrect, and leads to breakage
in e.g. Loop Strength Reduction. This change guards against that
behaviour.
Differential Revision: https://reviews.llvm.org/D40405
llvm-svn: 332409
Summary: Adding support for Fast flags in the SDNode to leverage fast math sub flag usage.
Reviewers: spatel, arsenm, jbhateja, hfinkel, escha, qcolombet, echristo, wristow, javed.absar
Reviewed By: spatel
Subscribers: llvm-commits, rampitec, nhaehnle, tstellar, FarhanaAleen, nemanjai, javed.absar, jbhateja, hfinkel, wdng
Differential Revision: https://reviews.llvm.org/D45710
llvm-svn: 331547
Summary: performAddCombine should run after DAG is legalized; Otherwise generic optimization
in the DAGCombiner can optimize an addcarry+trunc into an addcarry instruction with
illegal types.
Author: FarhanaAleen
Reviewed By: rampitec
Subscribers: llvm-commits, AMDGPU
Differential Revision: https://reviews.llvm.org/D46337
llvm-svn: 331368
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\\brief'); do perl -pi -e 's/\\brief //g' $i & done
Differential Revision: https://reviews.llvm.org/D46290
llvm-svn: 331272
Author: Samuel Pitoiset
ds_read_b128 and ds_write_b128 have been recently enabled
under the amdgpu-ds128 option because the performance benefit
is unclear.
Though, using 128-bit loads/stores for the local address space
appears to introduce regressions in tessellation shaders. Not
sure what is broken, but as ds_read_b128/ds_write_b128 are not
enabled by default, just introduce a global option and enable
128-bit only if requested (until it's fixed/used correctly).
v2: - fix regressions in merge-stores.ll and multiple_tails.ll
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105464
llvm-svn: 329764
Author: Samuel Pitoiset
ds_read_b128 and ds_write_b128 have been recently enabled
under the amdgpu-ds128 option because the performance benefit
is unclear.
Though, using 128-bit loads/stores for the local address space
appears to introduce regressions in tessellation shaders. Not
sure what is broken, but as ds_read_b128/ds_write_b128 are not
enabled by default, just introduce a global option and enable
128-bit only if requested (until it's fixed/used correctly).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105464
llvm-svn: 329591
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
Summary: There are no packed instructions for min3 or max3. So, performMinMaxCombine should not optimize vectors of f16 to min3/max3.
Author: FarhanaAleen
Reviewed By: arsenm
Subscribers: llvm-commits, AMDGPU
Differential Revision: https://reviews.llvm.org/D45219
llvm-svn: 329131
Summary:
Avoids having to list all intrinsics manually.
This is in preparation for the new dimension-aware image intrinsics,
which I'd rather not have to list here by hand.
Change-Id: If7ced04998397ef68c4cb8f7de66b5050fb767e5
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, mgorny, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44937
llvm-svn: 328938
Currently EVT is in the IR layer only because of Function.cpp needing a very small piece of the functionality of EVT::getEVTString(). The rest of EVT is used in codegen making CodeGen a better place for it.
The previous code converted a Type* to EVT and then called getEVTString. This was only expected to handle the primitive types from Type*. Since there only a few primitive types, we can just print them as strings directly.
Differential Revision: https://reviews.llvm.org/D45017
llvm-svn: 328806
The combine on a select of a load only triggers for
addrspace 0, and discards the MachinePointerInfo. The
conservative default needs to be used for this.
llvm-svn: 328652
In a function, s5 is used as the frame base SGPR. If a function
is calling another function, during the call sequence
it is copied to a preserved SGPR and restored.
Before it was possible for the scheduler to move stack operations
before the restore of s5, since there's nothing to associate
a frame index access with the restore.
Add an implicit use of s5 to the adjcallstack pseudo which ends
the call sequence to preven this from happening. I'm not 100%
satisfied with this solution, but I'm not sure what else would be
better.
llvm-svn: 328650
This is used by llvm tblgen as well as by LLVM Targets, so the only
common place is Support for now. (maybe we need another target for these
sorts of things - but for now I'm at least making them correct & we can
make them better if/when people have strong feelings)
llvm-svn: 328395
Summary: Starting from GCN 2nd generation, ISA supports ds_read_b128 on top of ds_read_b64.
This patch supports ds_read_b128 instruction pattern and generation of this instruction.
In the vectorizer, this patch also widen the vector length so that vectorizer generates
128 bit loads for local address-space which gets translated to ds_read_b128.
Since the performance benefit is not clear; compiler generates ds_read_b128 under -amdgpu-ds128.
Author: FarhanaAleen
Reviewed By: rampitec, arsenm
Subscribers: llvm-commits, AMDGPU
Differential Revision: https://reviews.llvm.org/D44210
llvm-svn: 327153
Summary: GCN ISA supports instructions that can read 16 consecutive dwords from memory through the scalar data cache;
loadstoreVectorizer should take advantage of the wider vector length and pack 16/8 elements of dwords/quadwords.
Author: FarhanaAleen
Reviewed By: rampitec
Subscribers: llvm-commits, AMDGPU
Differential Revision: https://reviews.llvm.org/D44179
llvm-svn: 326910
The code checks Level == AfterLegalizeDAG which is the fourth and last of the possible DAG combine stages that we have.
There is a Level called AfterLegalVectorOps, but that's the third DAG combine and it doesn't always run.
A function called isAfterLegalVectorOps should imply it returns true in either of the DAG combines that runs after the legalize vector ops stage, but that's not what this function does.
llvm-svn: 326832
i16 capable ASICs do not support i16 operands for this instruction.
Add tablegen pattern to merge chained i16 additions.
Differential Revision: https://reviews.llvm.org/D43985
llvm-svn: 326535
Summary:
In the current implementation of GPR Indexing Mode when the index is of non-uniform, the s_set_gpr_idx_off instruction
is incorrectly inserted after the loop. This will lead the instructions with vgpr operands (v_readfirstlane for example) to read incorrect
vgpr.
In this patch, we fix the issue by inserting s_set_gpr_idx_on/off immediately around the interested instruction.
Reviewers:
rampitec
Differential Revision:
https://reviews.llvm.org/D43297
llvm-svn: 325355
Kernel arguments likely read by all workitems and should not bypass
cache. Fixes performance hit in sub-dword argument loads.
Differential Revision: https://reviews.llvm.org/D43249
llvm-svn: 325146
Note: This is a candidate for LLVM 6.0, because it was planned to be
in that release but was delayed due to a long review period.
Merge conflict in release_60 - resolution:
Add "-p6:32:32" into the second (non-amdgiz) string.
Only scalar loads support 32-bit pointers. An address in a VGPR will
fail to compile. That's OK because the results of loads will only be used
in places where VGPRs are forbidden.
Updated AMDGPUAliasAnalysis and used SReg_64_XEXEC.
The tests cover all uses cases we need for Mesa.
Reviewers: arsenm, nhaehnle
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D41651
llvm-svn: 324487
- using qualified pointer addrspace in intrinsics class to avoid .f32 mangling
- changed too common atomic mangling to ds
- added missing intrinsics to AMDGPUTTIImpl::getTgtMemIntrinsic
Reviewed by: b-sumner
Differential Revision: https://reviews.llvm.org/D42383
llvm-svn: 323516
Summary:
This patch implements d16 support for image load, image store and image sample intrinsics.
Reviewers:
Matt, Brian.
Differential Revision:
https://reviews.llvm.org/D3991
llvm-svn: 322903
Currently all images are lowered to have a single
image PseudoSourceValue. Image stores happen to have
overly strict mayLoad/mayStore/hasSideEffects flags
set on them, so this happens to work. When these
are fixed to be correct, the scheduler breaks
this because the identical PSVs are assumed to
be the same address. These need to be unique
to the image resource value.
llvm-svn: 321555
Rather than adding more bits to express every
MMO flag you could want, just directly use the
MMO flags. Also fixes using a bunch of bool arguments to
getMemIntrinsicNode.
On AMDGPU, buffer and image intrinsics should always
have MODereferencable set, but currently there is no
way to do that directly during the initial intrinsic
lowering.
llvm-svn: 320746
Stores failed to decode at all since they didn't have a
DecoderNamespace set. Loads worked, but did not change
the register width displayed to match the numbmer of
enabled channels.
The number of printed registers for vaddr is still wrong,
but I don't think that's encoded in the instruction so
there's not much we can do about that.
Image atomics are still broken. MIMG is the same
encoding for SI/VI, but the image atomic classes
are split up into encoding specific versions unlike
every other MIMG instruction. They have isAsmParserOnly
set on them for some reason. dmask is also special for
these, so we probably should not have it as an explicit
operand as it is now.
llvm-svn: 320614
Move the entire optimization to one place. Before it was possible
to adjust dmask without changing the register class of the output
instruction, since they were done in separate places. Fix all
lane sizes and move all of the optimization into the DAG folding.
llvm-svn: 319705
The object can't straddle the address space
wrap around, so I think it's OK to assume any
offsets added to the base object pointer can't
overflow. Similar logic already appears to be
applied in SelectionDAGBuilder when lowering
aggregate returns.
llvm-svn: 319272
As part of the unification of the debug format and the MIR format,
always print registers as lowercase.
* Only debug printing is affected. It now follows MIR.
Differential Revision: https://reviews.llvm.org/D40417
llvm-svn: 319187
SITargetLowering::LowerCall uses dummy pointer info for byval argument, which causes
flat load instead of buffer load.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D40040
llvm-svn: 318844
All these headers already depend on CodeGen headers so moving them into
CodeGen fixes the layering (since CodeGen depends on Target, not the
other way around).
llvm-svn: 318490
Use VOP3 add/addc like usual.
This has some tradeoffs. Inline immediates fold
a little better, but other constants are worse off.
SIShrinkInstructions could be made smarter to handle
these cases.
This allows us to avoid selecting scalar adds where we
need to track the carry in scc and replace its users.
This makes it easier to use the carryless VALU adds.
llvm-svn: 318340
This was using a custom function that didn't handle the
addressing modes properly for private. Use
isLegalAddressingMode to avoid duplicating this.
Additionally, skip the combine if there is only one use
since the standard combine will handle it.
llvm-svn: 318013
The backend assumes pointer in default addr space is 32 bit, which is not
true for the new addr space mapping and causes assertion for unresolved
functions.
This patch fixes that.
Differential Revision: https://reviews.llvm.org/D39643
llvm-svn: 317476
Summary:
Kill the thread if operand 0 == false.
llvm.amdgcn.wqm.vote can be applied to the operand.
Also allow kill in all shader stages.
Reviewers: arsenm, nhaehnle
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D38544
llvm-svn: 316427
Summary:
The interpolation mode workaround ensures that at least one
interpolation mode is enabled in PSInputAddr. It does not also check
PSInputEna on the basis that the user might enable bits in that
depending on run-time state.
However, for amdpal os type, the user does not enable some bits after
compilation based on run-time states; the register values being
generated here are the final ones set in the hardware. Therefore, apply
the workaround to PSInputAddr and PSInputEnable together. (The case
where a bit is set in PSInputAddr but not in PSInputEnable is where the
frontend set up an input arg for a particular interpolation mode, but
nothing uses that input arg. Really we should have an earlier pass that
removes such an arg.)
Reviewers: arsenm, nhaehnle, dstuttard
Subscribers: kzhuravl, wdng, yaxunl, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D37758
llvm-svn: 315591
The hardware will only forward EXEC_LO; the high 32 bits will be zero.
Additionally, inline constants do not work. At least,
v_addc_u32_e64 v0, vcc, v0, v1, -1
which could conceivably be used to combine (v0 + v1 + 1) into a single
instruction, acts as if all carry-in bits are zero.
The llvm.amdgcn.ps.live test is adjusted; it would be nice to combine
s_mov_b64 s[0:1], exec
v_cndmask_b32_e64 v0, v1, v2, s[0:1]
into
v_mov_b32 v0, v3
but it's not particularly high priority.
Fixes dEQP-GLES31.functional.shaders.helper_invocation.value.*
llvm-svn: 314522
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
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
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
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
Changing mask argument type from const SmallVectorImpl<int>& to
ArrayRef<int>.
This came up in D35700 where a mask is received as an ArrayRef<int> and
we want to pass it to TargetLowering::isShuffleMaskLegal().
Also saves a few lines of code.
llvm-svn: 309085
This patch makes LSR generate better code for SystemZ in the cases of memory
intrinsics, Load->Store pairs or comparison of immediate with memory.
In order to achieve this, the following common code changes were made:
* New TTI hook: LSRWithInstrQueries(), which defaults to false. Controls if
LSR should do instruction-based addressing evaluations by calling
isLegalAddressingMode() with the Instruction pointers.
* In LoopStrengthReduce: handle address operands of memset, memmove and memcpy
as address uses, and call isFoldableMemAccessOffset() for any LSRUse::Address,
not just loads or stores.
SystemZ changes:
* isLSRCostLess() implemented with Insns first, and without ImmCost.
* New function supportedAddressingMode() that is a helper for TTI methods
looking at Instructions passed via pointers.
Review: Ulrich Weigand, Quentin Colombet
https://reviews.llvm.org/D35262https://reviews.llvm.org/D35049
llvm-svn: 308729
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
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
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
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
Going through the Constant methods requires redetermining that the Constant is a ConstantInt and then calling isZero/isOne/isMinusOne.
llvm-svn: 307292
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
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
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
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
I did this a long time ago with a janky python script, but now
clang-format has built-in support for this. I fed clang-format every
line with a #include and let it re-sort things according to the precise
LLVM rules for include ordering baked into clang-format these days.
I've reverted a number of files where the results of sorting includes
isn't healthy. Either places where we have legacy code relying on
particular include ordering (where possible, I'll fix these separately)
or where we have particular formatting around #include lines that
I didn't want to disturb in this patch.
This patch is *entirely* mechanical. If you get merge conflicts or
anything, just ignore the changes in this patch and run clang-format
over your #include lines in the files.
Sorry for any noise here, but it is important to keep these things
stable. I was seeing an increasing number of patches with irrelevant
re-ordering of #include lines because clang-format was used. This patch
at least isolates that churn, makes it easy to skip when resolving
conflicts, and gets us to a clean baseline (again).
llvm-svn: 304787
Various address spaces on the SI and R600 subtargets have stricter
limits on memory access size that other address spaces. Use
canMergeStoresTo predicate to prevent the DAGCombiner from creating
these stores as they will be split up during legalization.
llvm-svn: 303767
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
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
Changpeng Fang