This is mostly a minor convenience, but the pattern seems frequent
enough to be worthwhile (and we'll probably add more uses in the
future).
Differential Revision: https://reviews.llvm.org/D105850
Replace the experimental clang builtin and LLVM intrinsics for these
instructions with normal codegen patterns. Resolves PR50433.
Differential Revision: https://reviews.llvm.org/D105950
The data layout strings do not have any effect on llc tests and will become
misleadingly out of date as we continue to update the canonical data layout, so
remove them from the tests.
Differential Revision: https://reviews.llvm.org/D105842
MVE does not have a VMLALV instruction that can perform v16i8 -> i64
reductions, like it does for v8i16->i64 and v4i32->i64 reductions. That
means that the pattern to create them will be spilt up by type
legalization, creating a lot of instructions.
This extends the patterns for matching i64 reductions a little to handle
the v16i8->i64 case. We need to turn them into a pair of v8i16->i64
VMLALVs that each perform half of the reduction and are summed together
(so the later is a VMLALVA). The order of the lanes does not matter for
the reduction so we generate a MVEEXT for the extension, that will
either be folded into a extending load or can be optimized to a
VREV/VMOVL. Some of the resulting codegen isn't optimal, but will be
improved in a later patch.
Differential Revision: https://reviews.llvm.org/D105680
This set of folds was added recently with:
c7b658aeb526
0c400e895306
40b752d28d95
...and I noted that this wasn't likely to fire in code derived
from C/C++ source because of nsw in particular. But I didn't
notice that I had placed the code above the no-wrap block
of transforms.
This is likely the cause of regressions noted from the previous
commit because -- as shown in the test diffs -- we may have
transformed into a compare with an arbitrary constant rather
than a simpler signbit test.
This patch emits remarks for instructions that have invalid costs for
a given set of vectorization factors. Some example output:
t.c:4:19: remark: Instruction with invalid costs prevented vectorization at VF=(vscale x 1): load
dst[i] = sinf(src[i]);
^
t.c:4:14: remark: Instruction with invalid costs prevented vectorization at VF=(vscale x 1, vscale x 2, vscale x 4): call to llvm.sin.f32
dst[i] = sinf(src[i]);
^
t.c:4:12: remark: Instruction with invalid costs prevented vectorization at VF=(vscale x 1): store
dst[i] = sinf(src[i]);
^
Reviewed By: fhahn, kmclaughlin
Differential Revision: https://reviews.llvm.org/D105806
At the moment, <vscale x 1 x eltty> are not yet fully handled by the
code-generator, so to avoid vectorizing loops with that VF, we mark the
cost for these types as invalid.
The reason for not adding a new "TTI::getMinimumScalableVF" is because
the type is supposed to be a type that can be legalized. It partially is,
although the support for these types need some more work.
Reviewed By: paulwalker-arm, dmgreen
Differential Revision: https://reviews.llvm.org/D103882
The cost of the InsertSubvector shuffle kind cost is not complete and
may end up with just extracts + inserts costs in many cases. Added
a workaround to represent it as a generic PermuteSingleSrc, which is
still pessimistic but better than InsertSubvector.
Differential Revision: https://reviews.llvm.org/D105827
$ is used as PC for PowerPC inlineasm, ELF use it,
enable it for AIX XCOFF as well.
Reviewed By: #powerpc, amyk, nemanjai
Differential Revision: https://reviews.llvm.org/D105956
The CMake community Wiki has been moved to the [[ https://gitlab.kitware.com/cmake/community/wikis/home | Kitware GitLab Instance ]].
Also, the original anchor for `Information how to set up various cross compiling toolchains` section might not work as expected. The original content is now being collapsed, so browser won't navigate to the right section directly.
Hence, I think it might be better to provide the section name instead of `this section` with link to help readers find the right section by themselves.
Reviewed By: void
Differential Revision: https://reviews.llvm.org/D104996
If we try to create a new GlobalVariable on each iteration, the Module will
detect the name collision and "helpfully" rename later iterations by appending
".1" etc. But "___udivsi3.1" doesn't exist and we definitely don't want to try
to call it.
So instead check whether there's already a global with the right name in the
module and use that if so.
This has been a work-in-progress for a long time...we finally have all of
the pieces in place to handle vectorization of compare code as shown in:
https://llvm.org/PR41312
To do this (see PhaseOrdering tests), we converted SimplifyCFG and
InstCombine to the poison-safe (select) forms of the logic ops, so now we
need to have SLP recognize those patterns and insert a freeze op to make
a safe reduction:
https://alive2.llvm.org/ce/z/NH54Ah
We get the minimal patterns with this patch, but the PhaseOrdering tests
show that we still need adjustments to get the ideal IR in some or all of
the motivating cases.
Differential Revision: https://reviews.llvm.org/D105730
This new MIR pass removes redundant DBG_VALUEs.
After the register allocator is done, more precisely, after
the Virtual Register Rewriter, we end up having duplicated
DBG_VALUEs, since some virtual registers are being rewritten
into the same physical register as some of existing DBG_VALUEs.
Each DBG_VALUE should indicate (at least before the LiveDebugValues)
variables assignment, but it is being clobbered for function
parameters during the SelectionDAG since it generates new DBG_VALUEs
after COPY instructions, even though the parameter has no assignment.
For example, if we had a DBG_VALUE $regX as an entry debug value
representing the parameter, and a COPY and after the COPY,
DBG_VALUE $virt_reg, and after the virtregrewrite the $virt_reg gets
rewritten into $regX, we'd end up having redundant DBG_VALUE.
This breaks the definition of the DBG_VALUE since some analysis passes
might be built on top of that premise..., and this patch tries to fix
the MIR with the respect to that.
This first patch performs bacward scan, by trying to detect a sequence of
consecutive DBG_VALUEs, and to remove all DBG_VALUEs describing one
variable but the last one:
For example:
(1) DBG_VALUE $edi, !"var1", ...
(2) DBG_VALUE $esi, !"var2", ...
(3) DBG_VALUE $edi, !"var1", ...
...
in this case, we can remove (1).
By combining the forward scan that will be introduced in the next patch
(from this stack), by inspecting the statistics, the RemoveRedundantDebugValues
removes 15032 instructions by using gdb-7.11 as a testbed.
Differential Revision: https://reviews.llvm.org/D105279
As discussed on PR50183, we already fold to prefer 'select-of-idx' vs 'select-of-gep':
define <4 x i32>* @select0a(<4 x i32>* %a0, i64 %a1, i1 %a2, i64 %a3) {
%gep0 = getelementptr inbounds <4 x i32>, <4 x i32>* %a0, i64 %a1
%gep1 = getelementptr inbounds <4 x i32>, <4 x i32>* %a0, i64 %a3
%sel = select i1 %a2, <4 x i32>* %gep0, <4 x i32>* %gep1
ret <4 x i32>* %sel
}
-->
define <4 x i32>* @select1a(<4 x i32>* %a0, i64 %a1, i1 %a2, i64 %a3) {
%sel = select i1 %a2, i64 %a1, i64 %a3
%gep = getelementptr inbounds <4 x i32>, <4 x i32>* %a0, i64 %sel
ret <4 x i32>* %gep
}
This patch adds basic handling for the 'fallthrough' cases where the gep idx == 0 has been folded away to the base address:
define <4 x i32>* @select0(<4 x i32>* %a0, i64 %a1, i1 %a2) {
%gep = getelementptr inbounds <4 x i32>, <4 x i32>* %a0, i64 %a1
%sel = select i1 %a2, <4 x i32>* %a0, <4 x i32>* %gep
ret <4 x i32>* %sel
}
-->
define <4 x i32>* @select1(<4 x i32>* %a0, i64 %a1, i1 %a2) {
%sel = select i1 %a2, i64 0, i64 %a1
%gep = getelementptr inbounds <4 x i32>, <4 x i32>* %a0, i64 %sel
ret <4 x i32>* %gep
}
Reapplied with a fix for the bpf "-bpf-disable-avoid-speculation" tests
Differential Revision: https://reviews.llvm.org/D105901
We know that "CVTTPS2SI" returns 0x80000000 for out of range inputs (and for FP_TO_UINT, negative float values are undefined). We can use this to make unsigned conversions from vXf32 to vXi32 more efficient, particularly on targets without blend using the following logic:
small := CVTTPS2SI(x);
fp_to_ui(x) := small | (CVTTPS2SI(x - 2^31) & ARITHMETIC_RIGHT_SHIFT(small, 31))
Even on targets where "PBLENDVPS"/"PBLENDVB" exists, it is often a latency 2, low throughput instruction so this logic is applied there too (in particular for AVX2 also). It furthermore gets rid of one high latency floating point comparison in the previous lowering.
@TomHender checked the correctness of this for all possible floats between -1 and 2^32 (both ends excluded).
Original Patch by @TomHender (Tom Hender)
Differential Revision: https://reviews.llvm.org/D89697
ACC regsiters are a combination of 4 consecutive vector regsiters and therefore
somtimes require special treatment for register allocation. This patch only
adds a test.
This patch fixes code that incorrectly handled dbg.values with duplicate
location operands, i.e. !DIArgList(i32 %a, i32 %a). The errors in
question were caused by either applying an update to dbg.value multiple
times when the update is only valid once, or by updating the
DIExpression for only the first instance of a value that appears
multiple times.
Differential Revision: https://reviews.llvm.org/D105831
Using positive zero as the neutral element in 'fadd' reductions, while
it generates better code, is incorrect. The correct neutral element is
negative zero: 0.0 + -0.0 = 0.0, whereas -0.0 + -0.0 = -0.0.
There are perhaps more optimal lowerings of negative zero avoiding
constant-pool loads which could be left as future work.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D105902
If no scratch or flat instructions are used, we do not need to
initialize the flat scratch hardware register.
Differential Revision: https://reviews.llvm.org/D105920
SME introduces the ZA array, a new piece of architectural register state
consisting of a matrix of [SVLb x SVLb] bytes, where SVL is the
implementation defined Streaming SVE vector length and SVLb is the
number of 8-bit elements in a vector of SVL bits.
SME instructions consist of three types of matrix operands:
* Tiles: a ZA tile is a square, two-dimensional sub-array of elements
within the ZA array. These tiles make up the larger accumulator array
and the granularity varies based on the element size, i.e.
- ZAQ0..ZAQ15 (smallest tile granule)
- ZAD0..ZAD7
- ZAS0..ZAS3
- ZAH0..ZAH1
or ZAB0 (largest tile granule, single tile)
* Tile vectors: similar to regular tiles, but have an extra 'h' or 'v'
to tell how the vector at [reg+offset] is layed out in the tile,
horizontally or vertically. E.g. za1h.h or za15v.q, which corresponds
to vectors in registers ZAH1 and ZAQ15, respectively.
* Accumulator matrix: this is the entire accumulator array ZA.
This patch adds the register classes and related operands and parsing
for SME instructions operating on the accumulator array.
The ADDHA and ADDVA instructions which operate on tiles are also added
in this patch to make some use of the code added, later patches will
make use of the other operands introduced here.
The reference can be found here:
https://developer.arm.com/documentation/ddi0602/2021-06
Co-authored by: Sander de Smalen (@sdesmalen)
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D105570
This patch aims to revert the changes introduced by D70781 D71192 D76364
D70781 was introduced to fix hardware hang where we do not insert exp-
null-done for a kill inside infinit loop. At that time we have not added
exp-null-done for kill early termination, but I believe as for now, we will
always add the exp-null-done for early termination case in LaterBranchLowering.
D71192 was introduced to handle the only_kill case, which is also been
handled by the kill early termination work.
D76364 was used to fix a regression by D71192, where we cleared the done
bit of the export in the existing program and not let the normal return
block branching to the new unified return block.
With this change, we just trust frontends have setup exp-done correctly
which is true for all existing frontends. The backend only inserts
exp-null-done for the kill cases which is handled in SILateBranchLowering.cpp.
Reviewed by: critson
Differential Revision: https://reviews.llvm.org/D105610
Currently we are resolving lane/subregister conflict by visiting
instructions sequentially in current block to see whether there is any
use of the tainted lanes. To save compile time, we are not doing further
check in successor blocks. This sounds reasonable without subgregister liveness.
But since we have added subregister liveness tracking capability to
register coalescer, we can easily determine whether we have subregister
liveness conflict by checking subranges. This would help coalescing more
COPYs for target that enables subregister liveness tracking.
Reviewed by: arsenm, qcolombet
Differential Revision: https://reviews.llvm.org/D104509
In the device runtime there are many function calls to `__kmpc_is_spmd_exec_mode`
to query the execution mode of current kernels. In many cases, user programs
only contain target region executing in one mode. As a consequence, those runtime
function calls will only return one value. If we can get rid of these function
calls during compliation, it can potentially improve performance.
In this patch, we use `AAKernelInfo` to analyze kernel execution. Basically, for
each kernel (device) function `F`, we collect all kernel entries `K` that can
reach `F`. A new AA, `AAFoldRuntimeCall`, is created for each call site. In each
iteration, it will check all reaching kernel entries, and update the folded value
accordingly.
In the future we will support more function.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D105787
This is split from D105216, but the code is hoisted much earlier into
the path where we can actually get a zero stride flowing through. Some
fairly simple proofs handle the cases which show up in practice. The
only test changes are the cases where we really do need a non-zero
divider to produce the right result.
Recommitting with isLoopInvariant() check.
Differential Revision: https://reviews.llvm.org/D105921
Previously we reliedy on pseudo probe descriptors to look up precomputed GUID during probe emission for inlined probes. Since we are moving to always using unique linkage names, GUID for functions can be computed in place from dwarf names. This eliminates the need of importing pseudo probe descs in thinlto, since those descs should be emitted by the original modules.
This significantly reduces thinlto memory footprint in some extreme case where the number of imported modules for a single module is massive.
Test Plan:
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D105248
The linker or post-link optimizer can create an ELF image with multiple executable segments each of which will be loaded separately at run time. This breaks the assumption of llvm-profgen that currently only supports one base load address. What it ends up with is that the subsequent mmap events will be treated as an overwrite of the first mmap event which will in turn screw up address mapping. While it is non-trivial to support multiple separate load addresses and given that on x64 those segments will always be loaded at consecutive addresses (though via separate mmap
sys calls), I'm adding an error checking logic to bail out if that's violated and keep using a single load address which is the address of the first executable segment.
Also changing the disassembly output from printing section offset to printing the virtual address instead, which matches the behavior of objdump.
Differential Revision: https://reviews.llvm.org/D103178
Allow
```
%x:_<2 x p0> = G_INTTOPTR %y:_<2 x s64>
```
This shows up when building clang for AArch64 with GlobalISel.
Also show that we can select it.
This should match SDAG's behaviour: https://godbolt.org/z/33oqYoaYv
Differential Revision: https://reviews.llvm.org/D105944
To help with debugging non-trivial unswitching issues.
Don't care about the legacy pass, nobody is using it.
If a pass's string params are empty (e.g. "simple-loop-unswitch"), don't
default to the empty constructor for the pass params. We should still
let the parser take care of it in case the parser has its own defaults.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D105933
AMDGPU normally spills SGPRs to VGPRs. Previously, since all register
classes are handled at the same time, this was problematic. We don't
know ahead of time how many registers will be needed to be reserved to
handle the spilling. If no VGPRs were left for spilling, we would have
to try to spill to memory. If the spilled SGPRs were required for exec
mask manipulation, it is highly problematic because the lanes active
at the point of spill are not necessarily the same as at the restore
point.
Avoid this problem by fully allocating SGPRs in a separate regalloc
run from VGPRs. This way we know the exact number of VGPRs needed, and
can reserve them for a second run. This fixes the most serious
issues, but it is still possible using inline asm to make all VGPRs
unavailable. Start erroring in the case where we ever would require
memory for an SGPR spill.
This is implemented by giving each regalloc pass a callback which
reports if a register class should be handled or not. A few passes
need some small changes to deal with leftover virtual registers.
In the AMDGPU implementation, a new pass is introduced to take the
place of PrologEpilogInserter for SGPR spills emitted during the first
run.
One disadvantage of this is currently StackSlotColoring is no longer
used for SGPR spills. It would need to be run again, which will
require more work.
Error if the standard -regalloc option is used. Introduce new separate
-sgpr-regalloc and -vgpr-regalloc flags, so the two runs can be
controlled individually. PBQB is not currently supported, so this also
prevents using the unhandled allocator.
