We would like to start pushing -mcpu=generic towards enabling the set of
features that improves performance for some CPUs, without hurting any
others. A blend of the performance options hopefully beneficial to all
CPUs. The largest part of that is enabling in-order scheduling using the
Cortex-A55 schedule model. This is similar to the Arm backend change
from eecb353d0e25ba which made -mcpu=generic perform in-order scheduling
using the cortex-a8 schedule model.
The idea is that in-order cpu's require the most help in instruction
scheduling, whereas out-of-order cpus can for the most part out-of-order
schedule around different codegen. Our benchmarking suggests that
hypothesis holds. When running on an in-order core this improved
performance by 3.8% geomean on a set of DSP workloads, 2% geomean on
some other embedded benchmark and between 1% and 1.8% on a set of
singlecore and multicore workloads, all running on a Cortex-A55 cluster.
On an out-of-order cpu the results are a lot more noisy but show flat
performance or an improvement. On the set of DSP and embedded
benchmarks, run on a Cortex-A78 there was a very noisy 1% speed
improvement. Using the most detailed results I could find, SPEC2006 runs
on a Neoverse N1 show a small increase in instruction count (+0.127%),
but a decrease in cycle counts (-0.155%, on average). The instruction
count is very low noise, the cycle count is more noisy with a 0.15%
decrease not being significant. SPEC2k17 shows a small decrease (-0.2%)
in instruction count leading to a -0.296% decrease in cycle count. These
results are within noise margins but tend to show a small improvement in
general.
When specifying an Apple target, clang will set "-target-cpu apple-a7"
on the command line, so should not be affected by this change when
running from clang. This also doesn't enable more runtime unrolling like
-mcpu=cortex-a55 does, only changing the schedule used.
A lot of existing tests have updated. This is a summary of the important
differences:
- Most changes are the same instructions in a different order.
- Sometimes this leads to very minor inefficiencies, such as requiring
an extra mov to move variables into r0/v0 for the return value of a test
function.
- misched-fusion.ll was no longer fusing the pairs of instructions it
should, as per D110561. I've changed the schedule used in the test
for now.
- neon-mla-mls.ll now uses "mul; sub" as opposed to "neg; mla" due to
the different latencies. This seems fine to me.
- Some SVE tests do not always remove movprfx where they did before due
to different register allocation giving different destructive forms.
- The tests argument-blocks-array-of-struct.ll and arm64-windows-calls.ll
produce two LDR where they previously produced an LDP due to
store-pair-suppress kicking in.
- arm64-ldp.ll and arm64-neon-copy.ll are missing pre/postinc on LPD.
- Some tests such as arm64-neon-mul-div.ll and
ragreedy-local-interval-cost.ll have more, less or just different
spilling.
- In aarch64_generated_funcs.ll.generated.expected one part of the
function is no longer outlined. Interestingly if I switch this to use
any other scheduled even less is outlined.
Some of these are expected to happen, such as differences in outlining
or register spilling. There will be places where these result in worse
codegen, places where they are better, with the SPEC instruction counts
suggesting it is not a decrease overall, on average.
Differential Revision: https://reviews.llvm.org/D110830
PR45875 notes an instance where exception handling crashes on aarch64-fuchsia
where SCS is enabled by default. The underlying issue seems to be that within libunwind,
various _Unwind_* functions, the x18 register is not updated if a function is marked
with nounwind. This removes the check for nounwind and emits the CFI instruction that updates x18.
Differential Revision: https://reviews.llvm.org/D79822
The following tests explicitly check for .loc and .file directives, which is not currently supported. Disable these tests on AIX for now.
Reviewed By: shchenz
Differential Revision: https://reviews.llvm.org/D111346
For PS shaders we can use the input SPI_PS_INPUT_ENA and SPI_PS_INPUT_ADDR
registers
Calculate the number of VGPR registers used as input VGPRs based on these
registers rather than the arguments passed in (this conservatively always
allocates the maximum).
Differential Revision: https://reviews.llvm.org/D101633
Change-Id: Idf7c060cbbd5f7e3300102c55ecee3c07f209de6
In order to not generate an unnecessary G_CTLZ, I extended the constant folder
in the CSEMIRBuilder to handle G_CTLZ. I also added some extra handing of
vector constants too. It seems we don't have any support for doing constant
folding of vector constants, so the tests show some other useless G_SUB
instructions too.
Differential Revision: https://reviews.llvm.org/D111036
Make match patterns more permissive to be invariant to register
allocation choices.
Reviewed By: kaz7
Differential Revision: https://reviews.llvm.org/D111312
At this point it looks like a B extension will never exist. Instead
Zba, Zbb, Zbc, and Zbs are individual extensions being ratified
together as a package. Unknown at this time when or if the other
Zb* extensions will be ratified.
This patch removes references to the B extension. I've updated and
split tests accordingly.
This has been split from D110669 to make review a little easier.
Differential Revision: https://reviews.llvm.org/D111338
The shift libcalls have a shift amount parameter of MVT::i32, but
sometimes ExpandIntRes_Shift may be called with a node whose
second operand is a type that is larger than that. This leads to
an ABI mismatch, and for example causes a spurious zeroing of
a register in RV32 for 64-bit shifts. Note that at present regular
shift intstructions already have their shift amount operand adapted
at SelectionDAGBuilder::visitShift time, and funnelled shifts bypass that.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D110508
This change fixes a bug where the compiler generates a prologue_end
for line 0 locs. That is because line 0 is not associated with any
source location, so there should not be a prolgoue_end at a location
that doesn't correspond to a source location.
There were some LLVM tests that were explicitly checking for line 0
prologue_end's as well since I believe that to be incorrect, I had to
change those tests as well.
Patch by Shubham Rastogi!
Differential Revision: https://reviews.llvm.org/D110740
Updating the MachineDominatorTree is easy since SILowerControlFlow only
splits and removes basic blocks. This should save a bit of compile time
because previously we would recompute the dominator tree from scratch
after this pass.
Another reason for doing this is that SILowerControlFlow preserves
LiveIntervals which transitively requires MachineDominatorTree. I think
that means that SILowerControlFlow is obliged to preserve
MachineDominatorTree too as explained here:
https://lists.llvm.org/pipermail/llvm-dev/2020-November/146923.html
although it does not seem to have caused any problems in practice yet.
Differential Revision: https://reviews.llvm.org/D111313
When PHIElimination adds kills after lowering PHIs to COPYs it knows
that some instructions after the inserted COPY might use the same
SrcReg, but it was only looking at the terminator instructions at the
end of the block, not at other instructions like INLINEASM_BR that can
appear after the COPY insertion point.
Since we have already called findPHICopyInsertPoint, which knows about
INLINEASM_BR, we might as well reuse the insertion point that it
calculated when looking for instructions that might use SrcReg.
This fixes a machine verification failure if you force machine
verification to run after PHIElimination (currently it is disabled for
other reasons) when running
test/CodeGen/X86/callbr-asm-phi-placement.ll.
Differential Revision: https://reviews.llvm.org/D110834
X86InstrInfo::convertToThreeAddress would convert this:
%1:gr32 = ADD32rr killed %0:gr32(tied-def 0), %0:gr32, implicit-def dead $eflags
to this:
undef %2.sub_32bit:gr64 = COPY killed %0:gr32
undef %3.sub_32bit:gr64_nosp = COPY %0:gr32
%1:gr32 = LEA64_32r killed %2:gr64, 1, killed %3:gr64_nosp, 0, $noreg
Note that in the ADD32rr, %0 was used twice and the first use had a kill
flag, which is what MachineInstr::addRegisterKilled does.
In the converted code, each use of %0 is copied to a new reg, and the
first COPY inherits the kill flag from the ADD32rr. This causes
machine verification to fail (if you force it to run after
TwoAddressInstructionPass) because the second COPY uses %0 after it is
killed. Note that machine verification is currently disabled after
TwoAddressInstructionPass but this is a step towards being able to
enable it.
Fix this by not inserting more than one COPY from the same source
register.
Differential Revision: https://reviews.llvm.org/D110829
Print this basic block flag as inlineasm-br-indirect-target and parse
it. This allows you to write MIR test cases for INLINEASM_BR. The test
case I added is one that I wanted to precommit anyway for D110834.
Differential Revision: https://reviews.llvm.org/D111291
Currently when creating tail predicated loops, we need to validate that
all the live-outs of a loop will be equivalent with and without tail
predication, and if they are not we cannot legally create a
tail-predicated loop, leaving expensive vctp and vpst instructions in
the loop. These notably can include register-allocation instructions
like stack loads and stores, and copys lowered from COPYs to MVE_VORRs.
Instead of trying to prove this is valid late in the pipeline, this
patch introduces a MQPRCopy pseudo instruction that COPY is lowered to.
This can then either be converted to a MVE_VORR where possible, or to a
couple of VMOVD instructions if not. This way they do not behave
differently within and outside of tail-predications regions, and we can
know by construction that they are always valid. The idea is that we can
do the same with stack load and stores, converting them to VLDR/VSTR or
VLDM/VSTM where required to prove tail predication is always valid.
This does unfortunately mean inserting multiple VMOVD instructions,
instead of a single MVE_VORR, but my experiments show it to be an
improvement in general.
Differential Revision: https://reviews.llvm.org/D111048
Fix the calculation of ReplacedAllUntiedUses when any of the tied defs
are early-clobber. The effect of this is to fix the placement of kill
flags on an instruction like this (from @f2 in
test/CodeGen/SystemZ/asm-18.ll):
INLINEASM &"stepb $1, $2" [attdialect], $0:[regdef-ec:GRH32Bit], def early-clobber %3:grh32bit, $1:[reguse tiedto:$0], killed %4:grh32bit(tied-def 3), $2:[reguse:GRH32Bit], %4:grh32bit
After TwoAddressInstruction without this patch:
%3:grh32bit = COPY killed %4:grh32bit
INLINEASM &"stepb $1, $2" [attdialect], $0:[regdef-ec:GRH32Bit], def early-clobber %3:grh32bit, $1:[reguse tiedto:$0], %3:grh32bit(tied-def 3), $2:[reguse:GRH32Bit], %4:grh32bit
Note that the COPY kills %4, even though there is a later use of %4 in
the INLINEASM. This fails machine verification if you force it to run
after TwoAddressInstruction (currently it is disabled for other
reasons).
After TwoAddressInstruction with this patch:
%3:grh32bit = COPY %4:grh32bit
INLINEASM &"stepb $1, $2" [attdialect], $0:[regdef-ec:GRH32Bit], def early-clobber %3:grh32bit, $1:[reguse tiedto:$0], %3:grh32bit(tied-def 3), $2:[reguse:GRH32Bit], %4:grh32bit
Differential Revision: https://reviews.llvm.org/D110848
This reverts c7f16ab3e3f27d944db72908c9c1b1b7366f5515 / r109694 - which
suggested this was done to improve consistency with the gdb test suite.
Possible that at the time GCC did not canonicalize integer types, and so
matching types was important for cross-compiler validity, or that it was
only a case of over-constrained test cases that printed out/tested the
exact names of integer types.
In any case neither issue seems to exist today based on my limited
testing - both gdb and lldb canonicalize integer types (in a way that
happens to match Clang's preferred naming, incidentally) and so never
print the original text name produced in the DWARF by GCC or Clang.
This canonicalization appears to be in `integer_types_same_name_p` for
GDB and in `TypeSystemClang::GetBasicTypeEnumeration` for lldb.
(I tested this with one translation unit defining 3 variables - `long`,
`long (*)()`, and `int (*)()`, and another translation unit that had
main, and a function that took `long (*)()` as a parameter - then
compiled them with mismatched compilers (either GCC+Clang, or
Clang+(Clang with this patch applied)) and no matter the combination,
despite the debug info for one CU naming the type "long int" and the
other naming it "long", both debuggers printed out the name as "long"
and were able to correctly perform overload resolution and pass the
`long int (*)()` variable to the `long (*)()` function parameter)
Did find one hiccup, identified by the lldb test suite - that CodeView
was relying on these names to map them to builtin types in that format.
So added some handling for that in LLVM. (these could be split out into
separate patches, but seems small enough to not warrant it - will do
that if there ends up needing any reverti/revisiting)
Differential Revision: https://reviews.llvm.org/D110455
Without popcnt we had a special case for using the parity flag from a single test i8 test instruction if only bits 7:0 could be non-zero. That special case is still useful when we have popcnt.
To reach this special case, we enable custom lowering of parity for i16/i32/i64 even when popcnt is enabled. The check for POPCNT being enabled is now after the special case in LowerPARITY.
Fixes PR52093
Differential Revision: https://reviews.llvm.org/D111249
Currently the max alignment representable is 1GB, see D108661.
Setting the align of an object to 4GB is desirable in some cases to make sure the lower 32 bits are clear which can be used for some optimizations, e.g. https://crbug.com/1016945.
This uses an extra bit in instructions that carry an alignment. We can store 15 bits of "free" information, and with this change some instructions (e.g. AtomicCmpXchgInst) use 14 bits.
We can increase the max alignment representable above 4GB (up to 2^62) since we're only using 33 of the 64 values, but I've just limited it to 4GB for now.
The one place we have to update the bitcode format is for the alloca instruction. It stores its alignment into 5 bits of a 32 bit bitfield. I've added another field which is 8 bits and should be future proof for a while. For backward compatibility, we check if the old field has a value and use that, otherwise use the new field.
Updating clang's max allowed alignment will come in a future patch.
Reviewed By: hans
Differential Revision: https://reviews.llvm.org/D110451
Lowering of byval parameters with sizes that are not represented by a single
store require multiple stores to properly address the correct size of the
parameter.
Sizes that cannot be done with a single store are 3 bytes, 5 bytes, 6 bytes,
7 bytes. It is not correct to simply perform an 8 byte store and for these
elements because then the store would be larger than the element and alias
analysis would assume that this is undefined behaivour and return NoAlias
for them.
This patch adds the correct stores so that the size of the store is not larger
than the size of the element.
Reviewed By: nemanjai, #powerpc
Differential Revision: https://reviews.llvm.org/D108795
Currently the max alignment representable is 1GB, see D108661.
Setting the align of an object to 4GB is desirable in some cases to make sure the lower 32 bits are clear which can be used for some optimizations, e.g. https://crbug.com/1016945.
This uses an extra bit in instructions that carry an alignment. We can store 15 bits of "free" information, and with this change some instructions (e.g. AtomicCmpXchgInst) use 14 bits.
We can increase the max alignment representable above 4GB (up to 2^62) since we're only using 33 of the 64 values, but I've just limited it to 4GB for now.
The one place we have to update the bitcode format is for the alloca instruction. It stores its alignment into 5 bits of a 32 bit bitfield. I've added another field which is 8 bits and should be future proof for a while. For backward compatibility, we check if the old field has a value and use that, otherwise use the new field.
Updating clang's max allowed alignment will come in a future patch.
Reviewed By: hans
Differential Revision: https://reviews.llvm.org/D110451
D100244 missed a check on the ResNo of the extract's operand 0 when finding a
pair of extracts to combine into a VMOVRRD (extract(x, n); extract(x, n+1) ->
VMOVRRD(extract x, n/2)). As a result, it can incorrectly pair an extract(x, n)
with another extract(x:3, n+1) for example. This patch fixes the bug by adding
the proper check on ResNo.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D111188
Currently the max alignment representable is 1GB, see D108661.
Setting the align of an object to 4GB is desirable in some cases to make sure the lower 32 bits are clear which can be used for some optimizations, e.g. https://crbug.com/1016945.
This uses an extra bit in instructions that carry an alignment. We can store 15 bits of "free" information, and with this change some instructions (e.g. AtomicCmpXchgInst) use 14 bits.
We can increase the max alignment representable above 4GB (up to 2^62) since we're only using 33 of the 64 values, but I've just limited it to 4GB for now.
The one place we have to update the bitcode format is for the alloca instruction. It stores its alignment into 5 bits of a 32 bit bitfield. I've added another field which is 8 bits and should be future proof for a while. For backward compatibility, we check if the old field has a value and use that, otherwise use the new field.
Updating clang's max allowed alignment will come in a future patch.
Reviewed By: hans
Differential Revision: https://reviews.llvm.org/D110451
This updates a few more check lines, in some mte tests that were close
to auto generated already and some CodeGenPrepare/consthoist tests where
being able to see the entire code sequence is useful for determining
whether code differences are improvements or not.
This reverts commit d95cd81141a4e398e0d3337cb2e6617281d06278.
Re-land the original patch now that the bug this exposed in selection has been
fixed by 6bc64e24c38a
Seem to cause test failures in compiler-rt.
Revert "[SystemZ] Implement memcmp of variable length with CLC."
This reverts commit 7a4e9a0c73667cb80e4572d41535a9e48f1ed9ef.
Revert "[SystemZ] Implement memcpy of variable length with MVC."
This reverts commit c6c13c58eebda605a9a05f1f13cac1e46407afc7.
