This was [addressed for AArch64
here](https://github.com/llvm/llvm-project/pull/79004), but the same
applies to ARM.
Move the enablement of neon+fp64 to `-mcpu=cortex-r52`, which optionally
supports these features.
Vectorized ADDRSPACECASTs were not supported by the type legalizer.
This patch adds the support for:
- splitting the vector result: <2 x ptr> => 2 x <1 x ptr>
- scalarization: <1 x ptr> => ptr
- widening: <3 x ptr> => <4 x ptr>
This is all exercised by the added NVPTX tests.
This patch limits the icmp_i16(x,c) -> icmp_i32(ext(x),ext(c)) fold to CPUs that aren't known to have fast handling for length-changing prefixes for imm16 operands.
We are always assuming that 66/67h length-changing prefixes cause severe stalls and we should always extend imm16 operands and use a i32 icmp instead, the only exception being Intel Bonnell CPUs.
Agner makes this clear (see microarchitecture.pdf) that there are no stalls for any of the Intel Atom family (at least as far as Tremont - not sure about Gracemont or later). This is also true for AMD Bobcat/Jaguar and Ryzen families.
Recent performance Intel CPUs are trickier - Core2/Nehalem and earlier could have a 6-11cy stall, while SandyBridge onwards this is reduced to 3cy or less. I'm not sure if we should accept this as fast or not, we only use this flag for the icmp_i16 case, so that might be acceptable? If so, we should add this to x86-64-v3/v4 tuning as well.
Part of #90355 + #62952
This removes the GISel versions of isREVMask, isTRNMask, isUZPMask and
isZipMask. They are combined with the existing versions from SDAG into
AArch64PerfectShuffle.h.
Currently RISCVDeadRegisterDefinitions runs after vsetvli insertion, but
in #70549 vsetvli insertion runs after vector regalloc and as a result
we no longer convert some vsetvli a0, a0s to vsetvli x0, a0. This patch
moves it to after vector regalloc, but before scalar regalloc so we
still get the benefits of reducing register pressure.
If we don't demand the same element from both single source shuffles (permutes), then attempt to blend the sources together first and then perform a merged permute.
For vXi16 blends we have to be careful as these are much more likely to involve byte/word vector shuffles that will result in the creation of additional shuffle instructions.
This fold might be worth it for VSELECT with constant masks on AVX512 targets, but I haven't investigated this yet, but I've tried to write combineBlendOfPermutes so to be prepared for this.
The PR34592 -O0 regression is an unfortunate failure to cleanup with a later pass that calls SimplifyDemandedElts like the -O3 does - I'm not sure how worried we should be tbh.
Change definition of expandBitCastI128ToF128 and expandBitCastF128ToI128
to allow for simplified use in atomic load/store.
Update logic to split 128-bit loads and stores in DAGCombine to also
handle the f128 case where appropriate. This fixes the regressions
introduced by recent atomic load/store patches.
Referring to RISC-V, adding an MI level pass to optimize *W instructions
for LoongArch.
First it removes unneeded sext(addi.w rd, rs, 0) instructions. Either
because the sign extended bits aren't consumed or because the input was
already sign extended by an earlier instruction.
Then:
1. Unless explicit disabled or the target prefers instructions with W
suffix, it removes the -w suffix from opw instructions whenever all
users are dependent only on the lower word of the result of the
instruction. The cases handled are:
* addi.w because it helps reduce test differences between LA32 and LA64
w/o being a pessimization.
2. Or if explicit enabled or the target prefers instructions with W
suffix, it adds the W suffix to the instruction whenever all users are
dependent only on the lower word of the result of the instruction. The
cases handled are:
* add.d/addi.d/sub.d/mul.d.
* slli.d with imm < 32.
* ld.d/ld.wu.
This is really a workaround to allow control flow lowering in the
presence of convergence control tokens. Control-flow intrinsics in LLVM
IR are convergent because they indirectly represent the wave CFG, i.e.,
sets of threads that are "converged" or "execute in lock-step". But they
exist during a small window in the lowering process, inserted after the
structurizer and then translated to equivalent MIR pseudos. So rather
than create convergence tokens for these builtins, we simply mark them
as not convergent.
The corresponding MIR pseudos are marked as having side effects, which
is sufficient to prevent optimizations without having to mark them as
convergent.
This reverts commit a415b4dfcc02e3e82b8c8a7836f7c04b9d65dc9b.
Modify the instruction in place to transform it into a REG_SEQUENCE,
which is what other implementations of foldImmediate do. Also start
erasing the def instruction if there are no other uses.
Fixes#91110.
Revert "Revert 4 last AMDGPU commits to unbreak Windows bots"
This reverts commit 0d493ed2c6e664849a979b357a606dcd8273b03f.
MSVC does not like constexpr on the definition after an extern
declaration of a global.
In DAGCombiner, the `performCONDCombine` function attempts to remove AND
instructions in front of SUBS (cmp) instructions for which the AND is
transparent. The rules for that are correct, but it fails to take into
account the case where the SUBS instruction has multiple users with
different condition codes for comparison and simply removes the AND for
all of them. This causes a miscompilation in the attached test case.
Relax the constraint on the fold to allow any value for X and Y iff all upper 32-bits are known zero
Minor code-size saving by moving to i32 bit instructions
Revert "AMDGPU: Try to fix build error with old gcc"
This reverts commit c7ad12d0d7606b0b9fb531b0b273bdc5f1490ddb.
Revert "AMDGPU: Use umin in set.rounding expansion"
This reverts commit a56f0b51dd988ad2b533de759c98457c1ed42456.
Revert "AMDGPU: Optimize set_rounding if input is known to fit in 2 bits (#88588)"
This reverts commit b4e751e2ab0ff152ed18dea59ebf9691e963e1dd.
Revert "AMDGPU: Implement llvm.set.rounding (#88587)"
This reverts commit 9731b77e80261c627d79980f8c275700bdaf6591.
If the shuffle mask contains no undef elements, then we can move the freeze through a shuffle node.
This requires special case handling to create a new ShuffleVectorSDNode.
Includes VECTOR_SHUFFLE support for isGuaranteedNotToBeUndefOrPoison / canCreateUndefOrPoison.
We need it to test isel related passes. Currently
`verifyMachineFunction` is incomplete (no LiveIntervals support), but is
enough for testing isel pass, will migrate to complete
`MachineVerifierPass` in future.
These check lines break as of 91446e2aa687e due to changes in how LLVM
handles debug information. Since debug informaiton isn't important to
what this test is verifying we can remove the check lines.
This PR will fix the si-mode-register pass which is inserting an extra
setreg instruction in case of constrained FP operations. This pass will
be ignored for strictfp functions.
