It looks like MCRegAliasIterator can visit the same physical register twice. When this happens in this code in LICM we end up setting the PhysRegDef and then later in the same loop visit the register again. Now we see that PhysRegDef is set from the earlier iteration so now set PhysRegClobber.
This patch splits the loop so we have one that uses the previous value of PhysRegDef to update PhysRegClobber and second loop that updates PhysRegDef.
The X86 atomic test is an improvement. I had to add sideeffect to the two shrink wrapping tests to prevent hoisting from occurring. I'm not sure about the AMDGPU tests. It looks like the branch instruction changed at end the of the loops. And in the branch-relaxation test I think there is now "and vcc, exec, -1" instruction that wasn't there before.
Differential Revision: https://reviews.llvm.org/D55102
llvm-svn: 348330
This reverts commit r348203.
Reason: this produces absolute paths in .gcno files, breaking us
internally as we rely on them being consistent with the filenames passed
in the command line.
Also reverts r348157 and r348155 to account for revert of r348154 in
clang repository.
llvm-svn: 348279
The clang frontend no longer emits the current working directory for
DIFiles containing an absolute path in the filename: and will move the
common prefix between current working directory and the file into the
directory: component.
https://reviews.llvm.org/D55085
llvm-svn: 348155
The introduction of S_{ADD|SUB}_U64_PSEUDO instructions which are decomposed
into VOP3 instruction pairs for S_ADD_U64_PSEUDO:
V_ADD_I32_e64
V_ADDC_U32_e64
and for S_SUB_U64_PSEUDO
V_SUB_I32_e64
V_SUBB_U32_e64
preclude the use of SDWA to encode a constant.
SDWA: Sub-Dword addressing is supported on VOP1 and VOP2 instructions,
but not on VOP3 instructions.
We desire to fold the bit-and operand into the instruction encoding
for the V_ADD_I32 instruction. This requires that we transform the
VOP3 into a VOP2 form of the instruction (_e32).
%19:vgpr_32 = V_AND_B32_e32 255,
killed %16:vgpr_32, implicit $exec
%47:vgpr_32, %49:sreg_64_xexec = V_ADD_I32_e64
%26.sub0:vreg_64, %19:vgpr_32, implicit $exec
%48:vgpr_32, dead %50:sreg_64_xexec = V_ADDC_U32_e64
%26.sub1:vreg_64, %54:vgpr_32, killed %49:sreg_64_xexec, implicit $exec
which then allows the SDWA encoding and becomes
%47:vgpr_32 = V_ADD_I32_sdwa
0, %26.sub0:vreg_64, 0, killed %16:vgpr_32, 0, 6, 0, 6, 0,
implicit-def $vcc, implicit $exec
%48:vgpr_32 = V_ADDC_U32_e32
0, %26.sub1:vreg_64, implicit-def $vcc, implicit $vcc, implicit $exec
Differential Revision: https://reviews.llvm.org/D54882
llvm-svn: 348132
The identity ~(x ^ y) == (~x ^ y) == (x ^ ~y) allows XNOR (XOR/NOT) to turn into NOT/XOR. Handling this case with its own split means we can make the NOT remain in the scalar unit. Previously, we split 64-bit XNOR into two 32-bit XNOR, then lowered. Now, we get three instructions (s_not, v_xor, v_xor) rather than four in the case where either of the sources is a scalar 64-bit.
Add test cases to xnor.ll to attempt XNOR Vx, Sy and XNOR Sx, Vy. Also adding test that uses the opposite identity such that (~x ^ y) on the scalar unit (or vector for gfx906) can generate XNOR. This already worked, but I didn't see a test for it.
Differential: https://reviews.llvm.org/D55071
llvm-svn: 348075
D52935 introduced the ability for SimplifyDemandedBits to call SimplifyDemandedVectorElts through BITCASTs if the demanded bit mask entirely covered the sub element.
This patch relaxes this to demanding an element if we need any bit from it.
Differential Revision: https://reviews.llvm.org/D54761
llvm-svn: 348073
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
Introduces DPP pseudo instructions and the pass that combines DPP mov with subsequent uses.
Differential revision: https://reviews.llvm.org/D53762
llvm-svn: 347993
The motivating case for this is shown in:
https://bugs.llvm.org/show_bug.cgi?id=32023
and the corresponding rot16.ll regression tests.
Because x86 scalar shift amounts are i8 values, we can end up with trunc-binop-trunc
sequences that don't get folded in IR.
As the TODO comments suggest, there will be regressions if we extend this (for x86,
we mostly seem to be missing LEA opportunities, but there are likely vector folds
missing too). I think those should be considered existing bugs because this is the
same transform that we do as an IR canonicalization in instcombine. We just need
more tests to make those visible independent of this patch.
Differential Revision: https://reviews.llvm.org/D54640
llvm-svn: 347917
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
This patch adds support for S_ANDN2, S_ORN2 32-bit and 64-bit instructions and adds splits to move them to the vector unit (for which there is no equivalent instruction). It modifies the way that the more complex scalar instructions are lowered to vector instructions by first breaking them down to sequences of simpler scalar instructions which are then lowered through the existing code paths. The pattern for S_XNOR has also been updated to apply inversion to one input rather than the output of the XOR as the result is equivalent and may allow leaving the NOT instruction on the scalar unit.
A new tests for NAND, NOR, ANDN2 and ORN2 have been added, and existing tests now hit the new instructions (and have been modified accordingly).
Differential: https://reviews.llvm.org/D54714
llvm-svn: 347877
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
Summary:
MachineLoopInfo cannot be relied on for correctness, because it cannot
properly recognize loops in irreducible control flow which can be
introduced by late machine basic block optimization passes. See the new
test case for the reduced form of an example that occurred in practice.
Use a simple fixpoint iteration instead.
In order to facilitate this change, refactor WaitcntBrackets so that it
only tracks pending events and registers, rather than also maintaining
state that is relevant for the high-level algorithm. Various accessor
methods can be removed or made private as a consequence.
Affects (in radv):
- dEQP-VK.glsl.loops.special.{for,while}_uniform_iterations.select_iteration_count_{fragment,vertex}
Fixes: r345719 ("AMDGPU: Rewrite SILowerI1Copies to always stay on SALU")
Reviewers: msearles, rampitec, scott.linder, kanarayan
Subscribers: arsenm, kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits, hakzsam
Differential Revision: https://reviews.llvm.org/D54231
llvm-svn: 347853
Summary:
Reduce the statefulness of the algorithm in two ways:
1. More clearly split generateWaitcntInstBefore into two phases: the
first one which determines the required wait, if any, without changing
the ScoreBrackets, and the second one which actually inserts the wait
and updates the brackets.
2. Communicate pre-existing s_waitcnt instructions using an argument to
generateWaitcntInstBefore instead of through the ScoreBrackets.
To simplify these changes, a Waitcnt structure is introduced which carries
the counts of an s_waitcnt instruction in decoded form.
There are some functional changes:
1. The FIXME for the VCCZ bug workaround was implemented: we only wait for
SMEM instructions as required instead of waiting on all counters.
2. We now properly track pre-existing waitcnt's in all cases, which leads
to less conservative waitcnts being emitted in some cases.
s_load_dword ...
s_waitcnt lgkmcnt(0) <-- pre-existing wait count
ds_read_b32 v0, ...
ds_read_b32 v1, ...
s_waitcnt lgkmcnt(0) <-- this is too conservative
use(v0)
more code
use(v1)
This increases code size a bit, but the reduced latency should still be a
win in basically all cases. The worst code size regressions in my shader-db
are:
WORST REGRESSIONS - Code Size
Before After Delta Percentage
1724 1736 12 0.70 % shaders/private/f1-2015/1334.shader_test [0]
2276 2284 8 0.35 % shaders/private/f1-2015/1306.shader_test [0]
4632 4640 8 0.17 % shaders/private/ue4_elemental/62.shader_test [0]
2376 2384 8 0.34 % shaders/private/f1-2015/1308.shader_test [0]
3284 3292 8 0.24 % shaders/private/talos_principle/1955.shader_test [0]
Reviewers: msearles, rampitec, scott.linder, kanarayan
Subscribers: arsenm, kzhuravl, jvesely, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits, hakzsam
Differential Revision: https://reviews.llvm.org/D54226
llvm-svn: 347848
This works if DAG combiner is enabled, but without combining
we cannot select scalar_to_vector of <2 x half> and <2 x i16>.
Differential Revision: https://reviews.llvm.org/D54718
llvm-svn: 347259
This allows to avoid scratch use or indirect VGPR addressing for
small vectors.
Differential Revision: https://reviews.llvm.org/D54606
llvm-svn: 347231
Sadly, this duplicates (twice) the logic from InstSimplify. There
might be some way to at least share the DAG versions of the code,
but copying the folds seems to be the standard method to ensure
that we don't miss these folds.
Unlike in IR, we don't run DAGCombiner to fixpoint, so there's no
way to ensure that we do these kinds of simplifications unless the
code is repeated at node creation time and during combines.
There were other tests that would become worthless with this
improvement that I changed as pre-commits:
rL347161
rL347164
rL347165
rL347166
rL347167
I'm not sure how to salvage the remaining tests (diffs in this patch).
So the x86 tests verify that the new code is working as intended.
The AMDGPU test is actually similar to my motivating case: we have
some undef value that has survived to machine IR in an x86 test, and
then it gets folded in some weird way, or we crash if we don't transfer
the undef flag. But we would have been better off never getting to that
point by doing these simplifications.
This will lead back to PR32023 someday...
https://bugs.llvm.org/show_bug.cgi?id=32023
llvm-svn: 347170
If a block had one of the _term instructions used for gluing
exec modifying instructions to the end of the block,
analyzeBranch would fail, preventing the verifier from catching
a broken successor list.
llvm-svn: 347027
Add a pass to fixup various vector ISel issues.
Currently we handle converting GLOBAL_{LOAD|STORE}_*
and GLOBAL_Atomic_* instructions into their _SADDR variants.
This involves feeding the sreg into the saddr field of the new instruction.
llvm-svn: 347008
The machine scheduler currently biases register copies to/from
physical registers to be closer to their point of use / def to
minimize their live ranges. This change extends this to also physical
register assignments from immediate values.
This causes a reduction in reduction in overall register pressure and
minor reduction in spills and indirectly fixes an out-of-registers
assertion (PR39391).
Most test changes are from minor instruction reorderings and register
name selection changes and direct consequences of that.
Reviewers: MatzeB, qcolombet, myatsina, pcc
Subscribers: nemanjai, jvesely, nhaehnle, eraman, hiraditya,
javed.absar, arphaman, jfb, jsji, llvm-commits
Differential Revision: https://reviews.llvm.org/D54218
llvm-svn: 346894
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
Legalizer used to request an ext load from i8 to i1 when promoting
vector element type to i8. Fixed.
Differential Revision: https://reviews.llvm.org/D54440
llvm-svn: 346795
Sometimes after basic block placement we end up with a code like:
sreg = s_mov_b64 -1
vcc = s_and_b64 exec, sreg
s_cbranch_vccz
This happens as a join of a block assigning -1 to a saved mask and
another block which consumes that saved mask with s_and_b64 and a
branch.
This is essentially a single s_cbranch_execz instruction when moved
into a single new basic block.
Differential Revision: https://reviews.llvm.org/D54164
llvm-svn: 346690
The function only checks that instruction reads a super-register
containing requested physical register. In case if a sub-register
if being read that is also a use of a super-reg, so added the check.
In particular MI->readsRegister() is broken because of the missing
check. The resulting check is essentially regsOverlap().
Differential Revision: https://reviews.llvm.org/D54128
llvm-svn: 346686
Promote alloca can vectorize a small array by bitcasting it to a
vector type. Extend vectorization for the case when alloca is
already a vector type. We still want to replace GEPs with an
insert/extract element instructions in this case.
Differential Revision: https://reviews.llvm.org/D54219
llvm-svn: 346376
This allows testing AMDGPU alias analysis like any
other alias analysis pass. This fixes the existing
test pointlessly running opt -O3 when it really
just wants to run the one analysis.
Before there was no way to test this using -aa-eval
with opt, since the default constructed pass
is run. The wrapper subclass allows the
default constructor to pass the necessary callback.
llvm-svn: 346353
Set `LiveReg::PhysReg` to zero when freeing a register instead of
removing it from the entry from `LiveRegMap`. This way no iterators get
invalidated and we can avoid passing around and updating iterators all
over the place.
This does not change any allocator decisions. It is not completely NFC
because the arbitrary iteration order through `LiveRegMap` in
`spillAll()` changes so we may get a different order in those spill
sequences (the amount of spills does not change).
This is in preparation of https://reviews.llvm.org/D52010.
llvm-svn: 346298
Add this option for debugging and providing workaround.
By default it is off so no behavior change in backend.
Differential Revision: https://reviews.llvm.org/D54158
llvm-svn: 346267
This patch makes LICM use `ICFLoopSafetyInfo` that is a smarter version
of LoopSafetyInfo that leverages power of Implicit Control Flow Tracking
to keep track of throwing instructions and give less pessimistic answers
to queries related to throws.
The ICFLoopSafetyInfo itself has been introduced in rL344601. This patch
enables it in LICM only.
Differential Revision: https://reviews.llvm.org/D50377
Reviewed By: apilipenko
llvm-svn: 346201
The new atomic optimizer I previously added in D51969 did not work
correctly when a pixel shader was using derivatives, and had helper
lanes active.
To fix this we add an llvm.amdgcn.ps.live call that guards a branch
around the entire atomic operation - ensuring that all helper lanes are
inactive within the wavefront when we compute our atomic results.
I've added a test case that can cause derivatives, and exposes the
problem.
Differential Revision: https://reviews.llvm.org/D53930
llvm-svn: 346128
