The semantics of select with undefined/poison condition
are not explicitly stated in the LangRef, but this matches
comments in the code and Alive2 appears to concur:
https://alive2.llvm.org/ce/z/KXytmd
We can find this pattern after demanded elements transforms.
As noted in D101191, fuzzers are finding infinite loops because
we may not account for this pattern in other passes.
Now that we can fold some transposes into multiplies (CM: A * B^t and RM:
A^t * B), we want to move them around to create the optimal expressions:
* fold away double transposes while still using them to assert the shape
* sink transposes hoping they cancel out
* lift transposes when both operands are transposed
This also modifies the matrix remarks to include the number of exposed
transposes (i.e. transposes that we couldn't fold into a multiply).
The adjustment to the test remarks-inlining is a bit subtle: I am changing the
double transpose to a single transpose so that we don't remove it completely.
More importantly this changes some of the total instruction count, most
notable stores because we can no longer use a vector store.
Differential Revision: https://reviews.llvm.org/D102733
Nowadays LLVM does not assume that all loops are finite,
so if we want to produce a finite loop from a potentially-infinite one,
we must ensure that the original loop is known to be a finite one.
For this transform, it only matters for arithmetic right-shifts.
For them, either the function or the loop must be known to
be `mustprogress`, or the original value being shifted must be known
to be non-negative (because iff the sign bit was set,
it will never become zero, but will become `-1` in the "end").
It would be really good for alive2 to actually complain about this,
but it currently does not: https://github.com/AliveToolkit/alive2/issues/726
llvm-profgen uses profile summary based cold threshold to merge and trim cold context profile. This is to strike a good balance between profile size and performance.
We've been using 99.9% as the cutoff to save profile size without affecting performance. This change switch to use 99.9% instead of 99.9999% as default cold threshold cutoff for llvm-profgen.
Redundant switch csprof-cold-thres is also removed and tests cleaned up.
Differential Revision: https://reviews.llvm.org/D103071
This function can change regbank for registers which already have a selected
bank. Depending on the instruction where these registers were used it can
cause instruction selection to fail.
The 2nd test is based on the fuzzer example in post-commit
comments of D101191 -
https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=34661
The 1st test shows that we don't deal with this symmetrically.
We should be able to reduce both examples (possibly in
instsimplify instead of instcombine).
The parseInputFile function returns an empty unique_ptr to signal an
error, like when the input file doesn't exist, or is malformed. In this
case, the tool should exit immediately rather than segfault by
dereferencing the unique_ptr later.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D102891
We are using TOCEntry symbols like `LC..0` in TOC loads,
this is hard to read , at least requiring an additional step to figure
out the loaded symbols.
We should print out the name in comments.
Reviewed By: #powerpc, shchenz
Differential Revision: https://reviews.llvm.org/D102949
Match whats documented in the Intel AOM - the XMM variant of PSHUFB requires BOTH ports - this was being incorrectly modelled as EITHER port.
Now that we can use in-order models in llvm-mca, the atom model is a good "worst case scenario" analysis for x86.
Determined from llvm-mca analysis, AVX1 capable targets have a higher throughput for VPBLENDVB and shuffle ops, making it cheaper to perform shift+shuffle/select shift patterns.
Currently, BPF only contains three relocations:
R_BPF_NONE for no relocation
R_BPF_64_64 for LD_imm64 and normal 64-bit data relocation
R_BPF_64_32 for call insn and normal 32-bit data relocation
Also .BTF and .BTF.ext sections contain symbols in allocated
program and data sections. These two sections reserved 32bit
space to hold the offset relative to the symbol's section.
When LLVM JIT is used, the LLVM ExecutionEngine RuntimeDyld
may attempt to resolve relocations for .BTF and .BTF.ext,
which we want to prevent. So we used R_BPF_NONE for such relocations.
This all works fine until when we try to do linking of
multiple objects.
. R_BPF_64_64 handling of LD_imm64 vs. normal 64-bit data
is different, so lld target->relocate() needs more context
to do a correct job.
. The same for R_BPF_64_32. More context is needed for
lld target->relocate() to differentiate call insn vs.
normal 32-bit data relocation.
. Since relocations in .BTF and .BTF.ext are set to R_BPF_NONE,
they will not be relocated properly when multiple .BTF/.BTF.ext
sections are merged by lld.
This patch intends to address this issue by adding additional
relocation kinds:
R_BPF_64_ABS64 for normal 64-bit data relocation
R_BPF_64_ABS32 for normal 32-bit data relocation
R_BPF_64_NODYLD32 for .BTF and .BTF.ext style relocations.
The old R_BPF_64_{64,32} semantics:
R_BPF_64_64 for LD_imm64 relocation
R_BPF_64_32 for call insn relocation
The existing R_BPF_64_64/R_BPF_64_32 mapping to numeric values
is maintained. They are the most common use cases for
bpf programs and we want to maintain backward compatibility
as much as possible.
ExecutionEngine RuntimeDyld BPF relocations are adjusted as well.
R_BPF_64_{ABS64,ABS32} relocations will be resolved properly and
other relocations will be ignored.
Two tests are added for RuntimeDyld. Not handling R_BPF_64_NODYLD32 in
RuntimeDyldELF.cpp will result in "Relocation type not implemented yet!"
fatal error.
FK_SecRel_4 usages in BPFAsmBackend.cpp and BPFELFObjectWriter.cpp
are removed as they are not triggered in BPF backend.
BPF backend used FK_SecRel_8 for LD_imm64 instruction operands.
Differential Revision: https://reviews.llvm.org/D102712
NFC, since no instructions have their AsmMatchConverter
changed, but prepares for that to happen.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D103046
Change-Id: I6afefad899076de7b9a412374d09b95b29e012fa
rG1ad4f887bd7692a9e63fb42586f0ece366f2fe01 incorrectly assumed that vXi64 non-uniform shifts were slow like vXi32 were - but llvm-mca (+Agner) both confirm that Haswell/Broadwell are full rate.
NFC. Renames the variable in the dpp input operand generators
from DstRC to OldRC, because that is what it actually sets.
Also documents the importance of setting HasModifiers = 0 in the
dpp8 asm string.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D103047
Change-Id: Ice69ae38f644de7f228a75ca47c43e88b1f7d9e1
We can only scalarize memory accesses if we know the index is valid.
This patch adjusts canScalarizeAcceess to fall back to
computeConstantRange to check if the index is known to be valid.
Reviewed By: nlopes
Differential Revision: https://reviews.llvm.org/D102476
We could go either direction on this transform. VectorCombine already goes this
way for bitcasts (and handles more complicated cases using the cost model), so
let's try cast-first.
Deferring completely to VectorCombine is another possibility. But the backend
should be able to invert this easily when the vectors have the same shape, so
it doesn't seem like a transform that we need to avoid.
The motivating example from https://llvm.org/PR49081 has an int-to-float
sandwiched between 2 shuffles, and the backend currently does not reduce that,
so on x86, we get something like:
pshufd $249, %xmm0, %xmm0]
cvtdq2ps %xmm0, %xmm0
shufps $144, %xmm0, %xmm0
...instead of just a single conversion instruction.
Differential Revision: https://reviews.llvm.org/D103038
This adds support for the "count active bits" pattern, i.e.:
```
int countBits(unsigned val) {
int cnt = 0;
for( ; (val << cnt) != 0; ++cnt)
;
return cnt;
}
```
but a somewhat more general one:
```
int countBits(unsigned val, int start, int off) {
int cnt;
for (cnt = start; val << (cnt + off); cnt++)
;
return cnt;
}
```
alive2 is happy with all the tests there.
Note that, again, much like with the right-shift cases,
we don't require the `val != 0` guard.
This is the last pattern that was supported by
`detectShiftUntilZeroIdiom()`, which now becomes obsolete.
This adds support for the "count active bits" pattern, i.e.:
```
int countActiveBits(signed val) {
int cnt = 0;
for( ; (val >> cnt) != 0; ++cnt)
;
return cnt;
}
```
but a somewhat more general one:
```
int countActiveBits(signed val, int start, int off) {
int cnt;
for (cnt = start; val >> (cnt + off); cnt++)
;
return cnt;
}
```
This directly matches the existing 'logical right-shift until zero' idiom.
alive2 is happy with all the tests there.
Note that, again, much like with the original unsigned case,
we don't require the `val != 0` guard.
The old `detectShiftUntilZeroIdiom()` already supports this pattern,
the idea here is that the `val` must be positive (have at least one
leading zero), because otherwise the loop is non-terminating,
but since it is not `while(1)`, that would have been UB.
We really ought to support no_sanitize("coverage") in line with other
sanitizers. This came up again in discussions on the Linux-kernel
mailing lists, because we currently do workarounds using objtool to
remove coverage instrumentation. Since that support is only on x86, to
continue support coverage instrumentation on other architectures, we
must support selectively disabling coverage instrumentation via function
attributes.
Unfortunately, for SanitizeCoverage, it has not been implemented as a
sanitizer via fsanitize= and associated options in Sanitizers.def, but
rolls its own option fsanitize-coverage. This meant that we never got
"automatic" no_sanitize attribute support.
Implement no_sanitize attribute support by special-casing the string
"coverage" in the NoSanitizeAttr implementation. To keep the feature as
unintrusive to existing IR generation as possible, define a new negative
function attribute NoSanitizeCoverage to propagate the information
through to the instrumentation pass.
Fixes: https://bugs.llvm.org/show_bug.cgi?id=49035
Reviewed By: vitalybuka, morehouse
Differential Revision: https://reviews.llvm.org/D102772
Determined from llvm-mca analysis, AVX2+ capable targets have a higher throughput for VPBLENDVB and VPMOVZX ops, making it cheaper to perform shift+select patterns for vXi8 shifts or extend/shift/truncate for vXi16 shifts. Similarly AVX512BW can perform vXi8 as extend/shift/truncate patterns.
Currently the vector load + extract gets lowered to a single scalar
store, not accounting for the fact that the index could be
out-of-bounds, which is poison, not UB.
See PR50382.
The change is currently NFC, but exploited by the depending D102954.
Code to handle constants is borrowed from the general implementation
of Value::doRAUW().
Differential Revision: https://reviews.llvm.org/D103051
I really needed this, like, factually, yesterday,
when verifying dependency breaking idioms for AMD Zen 3 scheduler model.
Consider the following example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=duplicate
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-4a7e50.o
---
mode: inverse_throughput
key:
instructions:
- 'VPXORYrr YMM0 YMM0 YMM0'
config: ''
register_initial_values: []
cpu_name: znver3
llvm_triple: x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
- { key: inverse_throughput, value: 0.31025, per_snippet_value: 0.31025 }
error: ''
info: ''
assembled_snippet: C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C3
...
```
What does it tell us?
So wait, it can only execute ~3 x86 AVX YMM PXOR zero-idioms per cycle?
That doesn't seem right. That's even less than there are pipes supporting this type of op.
Now, second example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=loop
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-2418b5.o
---
mode: inverse_throughput
key:
instructions:
- 'VPXORYrr YMM0 YMM0 YMM0'
config: ''
register_initial_values: []
cpu_name: znver3
llvm_triple: x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
- { key: inverse_throughput, value: 1.00011, per_snippet_value: 1.00011 }
error: ''
info: ''
assembled_snippet: 49B80800000000000000C5FDEFC0C5FDEFC04983C0FF75F2C3
...
```
Now that's just worse. Due to the looping, the throughput completely plummeted,
and now we can only do a single instruction/cycle!?
That's not great.
And final example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=loop --loop-body-size=1000
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-c402e2.o
---
mode: inverse_throughput
key:
instructions:
- 'VPXORYrr YMM0 YMM0 YMM0'
config: ''
register_initial_values: []
cpu_name: znver3
llvm_triple: x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
- { key: inverse_throughput, value: 0.167087, per_snippet_value: 0.167087 }
error: ''
info: ''
assembled_snippet: 49B80800000000000000C5FDEFC0C5FDEFC04983C0FF75F2C3
...
```
So if we merge the previous two approaches, do duplicate this single-instruction snippet 1000x
(loop-body-size/instruction count in snippet), and run a loop with 1000 iterations
over that duplicated/unrolled snippet, the measured throughput goes through the roof,
up to 5.9 instructions/cycle, which finally tells us that this idiom is zero-cycle!
Reviewed By: courbet
Differential Revision: https://reviews.llvm.org/D102522
I cannot find documentation on this CPU, and it
is not supported by the Arm Compiler 5 product either.
It was likely a mistake or a different name for the
"ep9312", which is an Arm based Cirrus Logic chip.
Reviewed By: peter.smith
Differential Revision: https://reviews.llvm.org/D103024
The D82085 "allow TRE for non-capturing calls" caused failure during bootstrap.
This patch does the same as D82085 plus fixes bootstrap error.
The problem with D82085 is that it does not create copies for byval
operands, while replacing function call with a branch.
Consider following example:
```
int zoo ( S p1 );
int foo ( int count, S p1 ) {
if ( count > 10 )
return zoo(p1);
// temporarily variable created for passing byvalue parameter
// p1 could be used when zoo(p1) is called(after TRE is done).
// lifetime.start p1.byvalue.temp
return foo(count+1, p1);
// lifetime.end p1.byvalue.temp
}
```
After recursive call to foo is replaced with a jump into
start of the function, its parameters could be passed to
zoo function. i.e. temporarily variable created for byvalue
parameter "p1" could be passed to zoo. Finally zoo receives
broken operand:
```
int foo ( int count, S p1 ) {
:tailrecurse
p1_tr = phi p1, p1.byvalue.temp
if ( count > 10 )
return zoo(p1_tr);
// temporarily variable created for passing byvalue parameter
// p1 could be used when zoo(p1) is called(after TRE is done).
lifetime.start p1.byvalue.temp
memcpy (p1.byvalue.temp, p1_tr)
count = count + 1
lifetime.end p1.byvalue.temp
br tailrecurse
}
```
To prevent using p1.byvalue.temp after its scope finished by
lifetime.end marker this patch copies value from p1.byvalue.temp
into another temporarily variable and then copies this variable
into the input parameter for next iteration.
This patch passes bootstrap build and bootstrap build with AddressSanitizer.
Differential Revision: https://reviews.llvm.org/D85614
