[amdgpu] Update med3 combine to skip i64
Fixes an assumption that a type which is not i32 will be i16. This asserts
when trying to sign/zero extend an i64 to i32.
Test case was cut down from an openmp application. Variations on it are hit by
other combines before reaching the problematic one, e.g. replacing the
immediate values with other function arguments changes the codegen path and
misses this combine.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D98872
Avoid revisiting nodes with the same set of defined lanes by
using a unified visited set which integrates lanes into the key.
This retains the intent of the original code by still revisiting
a subgraph if a different set of lanes is defined and hence
marking might progress differently.
Note: default size of the visited set has been confirmed to
cover >99% of invocations in large array of test shaders.
Reviewed By: piotr
Differential Revision: https://reviews.llvm.org/D98772
This adds an Mask ArrayRef to getShuffleCost, so that if an exact mask
can be provided a more accurate cost can be provided by the backend.
For example VREV costs could be returned by the ARM backend. This should
be an NFC until then, laying the groundwork for that to be added.
Differential Revision: https://reviews.llvm.org/D98206
Split out some of the instructions predicated on the dot2-insts target
feature into a new dot7-insts, in preparation for subtargets that have
some but not all of these instructions. NFCI.
Differential Revision: https://reviews.llvm.org/D98717
RA can insert something like a sub1_sub2 COPY of a wide VGPR
tuple which results in the unaligned acces with v_pk_mov_b32
after the copy is expanded. This is regression after D97316.
Differential Revision: https://reviews.llvm.org/D98549
Replace individual operands GLC, SLC, and DLC with a single cache_policy
bitmask operand. This will reduce the number of operands in MIR and I hope
the amount of code. These operands are mostly 0 anyway.
Additional advantage that parser will accept these flags in any order unlike
now.
Differential Revision: https://reviews.llvm.org/D96469
[amdgpu] Implement lower function LDS pass
Local variables are allocated at kernel launch. This pass collects global
variables that are used from non-kernel functions, moves them into a new struct
type, and allocates an instance of that type in every kernel. Uses are then
replaced with a constantexpr offset.
Prior to this pass, accesses from a function are compiled to trap. With this
pass, most such accesses are removed before reaching codegen. The trap logic
is left unchanged by this pass. It is still reachable for the cases this pass
misses, notably the extern shared construct from hip and variables marked
constant which survive the optimizer.
This is of interest to the openmp project because the deviceRTL runtime library
uses cuda shared variables from functions that cannot be inlined. Trunk llvm
therefore cannot compile some openmp kernels for amdgpu. In addition to the
unit tests attached, this patch applied to ROCm llvm with fixed-abi enabled
and the function pointer hashing scheme deleted passes the openmp suite.
This lowering will use more LDS than strictly necessary. It is intended to be
a functionally correct fallback for cases that are difficult to target from
future optimisation passes.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D94648
When tracking defined lanes through phi nodes in the live range
graph each branch of the phi must be handled independently.
Also rewrite the marking algorithm to reduce unnecessary
operations.
Previously a shared set of defined lanes was used which caused
marking to stop prematurely. This was observable in existing lit
tests, but test patterns did not cover this detail.
Reviewed By: piotr
Differential Revision: https://reviews.llvm.org/D98614
The reason for this is to avoid deep recursion in DFS() which can cause
stack overflow on large CFGs, especially on Windows.
Differential Revision: https://reviews.llvm.org/D98528
byval arguments need to be assumed writable. Only implicitly stack
passed arguments which aren't addressable in the IR can be assumed
immutable.
Mips is still broken since for some reason its doing its own thing
with the ValueHandlers (and x86 doesn't actually handle byval
arguments now, although some of the code is there).
As llvm.amdgcn.kill is lowered to a terminator it can cause
else branch annotations to end up in the wrong block.
Do not annotate conditionals as else branches where there is
a kill to avoid this.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D97427
This instruction is only valid on 2D MSAA and 2D MSAA Array
surfaces. Remove intrinsic support for other dimension types,
and block assembly for unsupported dimensions.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D98397
We have amdgpu_gfx functions that have high register pressure. If
we do not reserve VGPR for SGPR spill, we will fall into the path
to spill the SGPR to memory, which does not only have correctness issue,
but also have really bad performance.
I don't know why there is the check for hasStackObjects(), in our case,
we don't have stack objects at the time of finalizeLowering(). So just
remove the check that we always reserve a VGPR for possible SGPR spill
in non-entry functions.
Reviewed by: arsenm
Differential Revision: https://reviews.llvm.org/D98345
I met some code generation behavior change when I tried to remove
the hasStackObject() check when reserving VGPR for SGPR spill.
For example, the function `callee_no_stack_no_fp_elim_all` in the lit
test file `callee-frame-setup.ll`.
The generated code changed from:
```
s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
s_mov_b32 s4, s33
s_mov_b32 s33, s32
s_mov_b32 s33, s4
s_setpc_b64 s[30:31]
```
into something like:
```
s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
v_writelane_b32 v63, s33, 0
s_mov_b32 s33, s32
v_readlane_b32 s33, v63, 0
s_setpc_b64 s[30:31]
```
I think we still prefer the old version where only scalar instructions are needed.
The idea here is free the reserved VGPR if no SGPR spills. So we will very likely
to use a free SGPR for fp/sp spill.
Reviewed by: arsenm
Differential Revision: https://reviews.llvm.org/D98344
These patterns are obviously dead, they are using format
operand which is not selected and we have no corresponding
SelectMUBUF() function.
Differential Revision: https://reviews.llvm.org/D98451
As we may overwrite inactive lanes of a caller-save-vgpr, we should
always save/restore the reserved vgpr for sgpr spill.
Reviewed by: arsenm
Differential Revision: https://reviews.llvm.org/D98319
D57708 changed SIInstrInfo::isReallyTriviallyReMaterializable to reject
V_MOVs with extra implicit operands, but it accidentally rejected all
V_MOVs because of their implicit use of exec. Fix it but avoid adding a
moderately expensive call to MI.getDesc().getNumImplicitUses().
In real graphics shaders this changes quite a few vgpr copies into move-
immediates, which is good for avoiding stalls on GFX10.
Differential Revision: https://reviews.llvm.org/D98347
Refactor and add comments to explain where the magic numbers come from
in terms of the instruction cache line size. NFC.
Differential Revision: https://reviews.llvm.org/D98266
AMDGPU target tries to handle the SGPR and VGPR spills in a
custom pass before the actual frame lowering pass. Once they
are handled and the respective frames are eliminated in the
custom pass, certain uses of them still remain. For instance,
the DBG_VALUE instructions inserted by the allocator alongside
the spill instruction will use the corresponding frame index.
They become dead later during PEI and causes a crash while trying to
replace the frame indices. We should possibly avoid this custom pass.
For now, replacing such dead references with null register value.
Reviewed By: arsenm, scott.linder
Differential Revision: https://reviews.llvm.org/D98038
This is already deprecated, so remove code working on this.
Also update the tests by using S_CBRANCH_EXECZ instead of SI_MASK_BRANCH.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D97545
gfx1030 added a new way to implement readcyclecounter using the
SHADER_CYCLES hardware register, but the s_memtime instruction still
exists, so the MC layer should still accept it and the
llvm.amdgcn.s.memtime intrinsic should still work.
Differential Revision: https://reviews.llvm.org/D97928
Same as other memory instructions, ds instructions add latency even if
exec is zero. Jumping over them if exec=0 is cheaper than executing
them.
With this change, the branch instruction that skips over a basic block
if exec=0 is not removed when the block contains a ds instruction.
Differential Revision: https://reviews.llvm.org/D97922
Recommit bf5a582650. Depends on
4c8fb7ddd6 which was reverted.
RegBankSelect creates zext and trunc when it selects banks for uniform i1.
Add zext_trunc_fold from generic combiner to post RegBankSelect combiner.
Differential Revision: https://reviews.llvm.org/D95432
The hazard where a VMEM reads an SGPR written by a VALU counts as a data
dependency hazard, so no nops are required on GFX10. Tested with Vulkan
CTS on GFX10.1 and GFX10.3.
Differential Revision: https://reviews.llvm.org/D97926
RegBankSelect creates zext and trunc when it selects banks for uniform i1.
Add zext_trunc_fold from generic combiner to post RegBankSelect combiner.
Differential Revision: https://reviews.llvm.org/D95432
Honor always_inline attribute when processing -amdgpu-inline-max-bb.
It was lost during the ports of the heuristic. There is no reason
to honor inline hint, but not always inline.
Differential Revision: https://reviews.llvm.org/D97790
Refactor insertion of the asserting ops. This enables using them for
AMDGPU.
This code should essentially be the same for every target. Mips, X86
and ARM all have different code there now, but this seems to be an
accident. The assignment functions are called with different types
than they would be in the DAG, so this is all likely an assortment of
hacks to get around that.
* Add amdgcn_strict_wqm intrinsic.
* Add a corresponding STRICT_WQM machine instruction.
* The semantic is similar to amdgcn_strict_wwm with a notable difference that not all threads will be forcibly enabled during the computations of the intrinsic's argument, but only all threads in quads that have at least one thread active.
* The difference between amdgc_wqm and amdgcn_strict_wqm, is that in the strict mode an inactive lane will always be enabled irrespective of control flow decisions.
Reviewed By: critson
Differential Revision: https://reviews.llvm.org/D96258
* Introduce the new intrinsic amdgcn_strict_wwm
* Deprecate the old intrinsic amdgcn_wwm
The change is done for consistency as the "strict"
prefix will become an important, distinguishing factor
between amdgcn_wqm and amdgcn_strictwqm in the future.
The "strict" prefix indicates that inactive lanes do not
take part in control flow, specifically an inactive lane
enabled by a strict mode will always be enabled irrespective
of control flow decisions.
The amdgcn_wwm will be removed, but doing so in two steps
gives users time to switch to the new name at their own pace.
Reviewed By: critson
Differential Revision: https://reviews.llvm.org/D96257
While the underlying instruction is called image_msaa_load,
the resource must be x component only.
Rename the intrinsic for clarity.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D97829
This merges more AMDGPU ABI lowering code into the generic call
lowering. Start cleaning up by factoring away more of the pack/unpack
logic into the buildCopy{To|From}Parts functions. These could use more
improvement, and the SelectionDAG versions are significantly more
complex, and we'll eventually have to emulate all of those cases too.
This is mostly NFC, but does result in some minor instruction
reordering. It also removes some of the limitations with mismatched
sizes the old code had. However, similarly to the merge on the input,
this is forcing gfx6/gfx7 to use the gfx8+ ABI (which is what we
actually want, but SelectionDAG is stuck using the weird emergent
ABI).
This also changes the load/store size for stack passed EVTs for
AArch64, which makes it consistent with the DAG behavior.
To do this while supporting the existing functionality in SelectionDAG of using
PGO info, we add the ProfileSummaryInfo and LazyBlockFrequencyInfo analysis
dependencies to the instruction selector pass.
Then, use the predicate to generate constant pool loads for f32 materialization,
if we're targeting optsize/minsize.
Differential Revision: https://reviews.llvm.org/D97732
Make OMod explicit instead of implied by HasModifiers in the
operand list. Requires explicitly setting HasOMod=1 for
irregular OMod usage in instruction V_CVT_{U,I}*
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D97587
Change-Id: I230e1476f529e816eec60e242531f23a99e3839f
Stanislav Mekhanoshin