For i64 this avoids loading a 64-bit value into register, for smaller registers this just avoids an immediate operand.
For i8+i16, limit to one use case as we save fewer bytes and these can be wasted entirely on extra register moves.
Fixes#67709
PR #66334 tried to renumber slot indexes before register allocation, but
the numbering was still affected by list entries for instructions which
had been erased. Fix this to make the register allocator's live range
length heuristics even less dependent on the history of how instructions
have been added to and removed from SlotIndexes's maps.
RegAllocGreedy uses SlotIndexes::getApproxInstrDistance to approximate
the length of a live range for its heuristics. Renumbering all slot
indexes with the default instruction distance ensures that this estimate
will be as accurate as possible, and will not depend on the history of
how instructions have been added to and removed from SlotIndexes's maps.
This also means that enabling -early-live-intervals, which runs the
SlotIndexes analysis earlier, will not cause large amounts of churn due
to different register allocator decisions.
If the upper half of an abs() is all sign bits, then we can perform the abs() using just the lower half and then zero extend.
I've limited the DAG combine to only sign_extend_inreg (and free truncate/zero_extend) to minimise any later promotion issues, but for legalization a similar fold can use ComputeNumSignBits to be more aggressive.
Alive2: https://alive2.llvm.org/ce/z/y32fS4Fixes#43370
Differential Revision: https://reviews.llvm.org/D136559
The first attempt missed changing test files for tools
(update_llc_test_checks.py).
Original commit message:
This implements the main suggested change from issue #56498.
Using the shorter (non-extending) instruction with only
-Oz ("minsize") rather than -Os ("optsize") is left as a
possible follow-up.
As noted in the bug report, the zero-extending load may have
shorter latency/better throughput across a wide range of x86
micro-arches, and it avoids a potential false dependency.
The cost is an extra instruction byte.
This could cause perf ups and downs from secondary effects,
but I don't think it is possible to account for those in
advance, and that will likely also depend on exact micro-arch.
This does bring LLVM x86 codegen more in line with existing
gcc codegen, so if problems are exposed they are more likely
to occur for both compilers.
Differential Revision: https://reviews.llvm.org/D129775
This implements the main suggested change from issue #56498.
Using the shorter (non-extending) instruction with only
-Oz ("minsize") rather than -Os ("optsize") is left as a
possible follow-up.
As noted in the bug report, the zero-extending load may have
shorter latency/better throughput across a wide range of x86
micro-arches, and it avoids a potential false dependency.
The cost is an extra instruction byte.
This could cause perf ups and downs from secondary effects,
but I don't think it is possible to account for those in
advance, and that will likely also depend on exact micro-arch.
This does bring LLVM x86 codegen more in line with existing
gcc codegen, so if problems are exposed they are more likely
to occur for both compilers.
Differential Revision: https://reviews.llvm.org/D129775
Previously we used sra+add+xor if ADDCARRY is supported. This changes
to sra+xor+sub is SUBCARRY is available.
This is consistent with the recent change to the default expansion
in LegalizeDAG.
Differential Revision: https://reviews.llvm.org/D121039
Previous we used sra (X, size(X)-1); xor (add (X, Y), Y).
By placing sub at the end, we allow RISCV to combine sign_extend_inreg
with it to form subw.
Some X86 tests for Z - abs(X) seem to have improved as well.
Other targets look to be a wash.
I had to modify ARM's abs matching code to match from sub instead of
xor. Maybe instead ISD::ABS should be made legal. I'll try that in
parallel to this patch.
This is an alternative to D119099 which was focused on RISCV only.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D119171
CMOVGE reads SF and OF. CMOVNS only reads SF. This matches with
other recent changes to use a single flag where possible. It also
matches gcc codegen.
I believe this technically changes whether the conditioanl move happens
on INT_MIN, but for INT_MIN both registers are the same so it doesn't
matter.
Differential Revision: https://reviews.llvm.org/D111826
This patch contains following enhancements to SrcRegMap and DstRegMap:
1 In findOnlyInterestingUse not only check if the Reg is two address usage,
but also check after commutation can it be two address usage.
2 If a physical register is clobbered, remove SrcRegMap entries that are
mapped to it.
3 In processTiedPairs, when create a new COPY instruction, add a SrcRegMap
entry only when the COPY instruction is coalescable. (The COPY src is
killed)
With these enhancements isProfitableToCommute can do better commute decision,
and finally more register copies are removed.
Differential Revision: https://reviews.llvm.org/D108731
This simple heuristic uses the estimated live range length combined
with the number of registers in the class to switch which heuristic to
use. This was taking the raw number of registers in the class, even
though not all of them may be available. AMDGPU heavily relies on
dynamically reserved numbers of registers based on user attributes to
satisfy occupancy constraints, so the raw number is highly misleading.
There are still a few problems here. In the original testcase that
made me notice this, the live range size is incorrect after the
scheduler rearranges instructions, since the instructions don't have
the original InstrDist offsets. Additionally, I think it would be more
appropriate to use the number of disjointly allocatable registers in
the class. For the AMDGPU register tuples, there are a large number of
registers in each tuple class, but only a small fraction can actually
be allocated at the same time since they all overlap with each
other. It seems we do not have a query that corresponds to the number
of independently allocatable registers. Relatedly, I'm still debugging
some allocation failures where overlapping tuples seem to not be
handled correctly.
The test changes are mostly noise. There are a handful of x86 tests
that look like regressions with an additional spill, and a handful
that now avoid a spill. The worst looking regression is likely
test/Thumb2/mve-vld4.ll which introduces a few additional
spills. test/CodeGen/AMDGPU/soft-clause-exceeds-register-budget.ll
shows a massive improvement by completely eliminating a large number
of spills inside a loop.
If smax() is legal, this is likely to result in smaller codegen expansion for abs(x) than the xor(add,ashr) method.
This is also what PowerPC has been doing for its abs implementation, so it lets us get rid of a load of custom lowering code there (and which was never updated when they added smax lowering).
Alive2: https://alive2.llvm.org/ce/z/xRk3cD
Differential Revision: https://reviews.llvm.org/D92095
Rather than using SELECT instructions, use SRA, UADDO/ADDCARRY and
XORs to expand ABS. This is the multi-part version of the sequence
we use in LegalizeDAG.
It's also the same as the Custom sequence uses for i64 on 32-bit
and i128 on 64-bit. So we can remove the X86 customization.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D87215
This adds the llvm.abs(), llvm.umin(), llvm.umax(), llvm.smin(),
and llvm.smax() intrinsics specified in D81829. For SelectionDAG,
the ISD opcodes and all the legalization and lowering already exist,
so this just wires them up to the intrinsic in the SDAG builder and
adds rudimentary tests. For GlobalISel only the min/max intrinsics
are wired up, as llvm.abs() will require the addition of a G_ABS op,
and corresponding legalization support.
Differential Revision: https://reviews.llvm.org/D84125
