backend: add missing op support for LLVM translation

[sueszli]

Add missing op support for the LLVM translation path.

The previous work in #6009 established MLIR compatibility and confidence in xdsl-translate, but many ops still lack translation support.

Conversion passes

• convert-arith-to-llvm pass ⚡🔥
  • Target LLVM ops already in dialect: add, sub, mul, sdiv, udiv, srem, urem, fadd, fsub, fmul, fdiv, frem, fneg, and, or, xor, shl, ashr, lshr, icmp, fcmp, trunc, zext, sext, sitofp, fpext, bitcast, mlir.constant, select
  • Missing LLVM ops needed: uitofp, fptosi, fptoui, fptrunc
  • Missing intrinsics needed: llvm.intr.minimum, llvm.intr.maximum, llvm.intr.smax, llvm.intr.smin, llvm.intr.umax, llvm.intr.umin
  • Also needs: IndexType backend support (dialects: (llvm) reject incompatible types in LLVMFunctionType #5705)
• convert-cf-to-llvm pass ⚡
  • Target LLVM ops already in dialect: br, cond_br, unreachable
  • Missing LLVM ops needed: switch
• convert-func-to-llvm pass ⚡
  • Target LLVM ops already in dialect: func, return, call
  • Also needs: IndexType backend support (dialects: (llvm) reject incompatible types in LLVMFunctionType #5705)
• convert-math-to-llvm pass 🔥
  • Intrinsics already in dialect: fabs, ceil, sqrt, log, fma, maxnum
  • Missing: exp dialects: (llvm) add FExpOp #5857, cos dialects: (llvm) add FCosOp #5858, log2 dialects: (llvm) add FLog2Op #5859, floor dialects: (llvm) add FFloorOp #5860, sin dialects: (llvm) add FSinOp #5863, copysign dialects: (llvm) add FCopySignOp #5864, exp2 dialects: (llvm) add FExp2Op #5867, minnum dialects: (llvm) add VectorFMinOp #5868, pow dialects: (llvm) add FPowOp #5870, log10, tan, round, roundeven, trunc (fp), nearbyint, rint, tanh, fmuladd, powi, ctlz, cttz, ctpop
• convert-vector-to-llvm pass 🔥
  • Target LLVM ops already in dialect: load, store, insertelement, shufflevector, intr.masked.store, intr.fma
  • Missing: vector.ReductionOp conversion dialects: (llvm) add llvm.intr.vector.reduce.fadd op #5866, vector.BitcastOp conversion backend: (llvm) add vector.BitcastOp conversion #5869, extractelement, masked.load, masked.gather, masked.scatter, masked.compressstore, masked.expandload, all vector.reduce.* (15 variants), matrix.transpose, matrix.multiply
• convert-omp-to-llvm pass
  • omp.parallel -> outlined llvm.func + llvm.call @__kmpc_fork_call
  • omp.wsloop -> llvm.call @__kmpc_for_static_init_8 / @__kmpc_for_static_fini
  • omp.loop_nest -> llvm.br / llvm.cond_br loop header with IV
  • omp.yield -> llvm.br back-edge
  • omp.terminator -> llvm.return in outlined function

Missing LLVM dialect ops (non-intrinsic)

Needed by conversion passes above:

• llvm.extractelement (convert-vector-to-llvm)
• llvm.fptrunc (convert-arith-to-llvm)
• llvm.uitofp (convert-arith-to-llvm)
• llvm.fptosi (convert-arith-to-llvm)
• llvm.fptoui (convert-arith-to-llvm)
• llvm.switch (convert-cf-to-llvm)

Other standard LLVM ops not yet in dialect:

• llvm.freeze
• llvm.invoke
• llvm.landingpad
• llvm.resume
• llvm.phi
• llvm.fence
• llvm.atomicrmw
• llvm.cmpxchg

Missing LLVM intrinsic ops

Float unary (convert-math-to-llvm)

• llvm.intr.exp dialects: (llvm) add FExpOp #5857
• llvm.intr.cos dialects: (llvm) add FCosOp #5858
• llvm.intr.log2 dialects: (llvm) add FLog2Op #5859
• llvm.intr.floor dialects: (llvm) add FFloorOp #5860
• llvm.intr.sin dialects: (llvm) add FSinOp #5863
• llvm.intr.exp2 dialects: (llvm) add FExp2Op #5867
• llvm.intr.tan
• llvm.intr.asin
• llvm.intr.acos
• llvm.intr.atan
• llvm.intr.sinh
• llvm.intr.cosh
• llvm.intr.tanh
• llvm.intr.log10
• llvm.intr.exp10
• llvm.intr.nearbyint
• llvm.intr.rint
• llvm.intr.round
• llvm.intr.roundeven
• llvm.intr.trunc (fp truncation intrinsic, not the integer trunc op)

Float binary (convert-math-to-llvm, convert-arith-to-llvm)

• llvm.intr.copysign dialects: (llvm) add FCopySignOp #5864
• llvm.intr.minnum dialects: (llvm) add VectorFMinOp #5868
• llvm.intr.pow dialects: (llvm) add FPowOp #5870
• llvm.intr.minimum
• llvm.intr.maximum
• llvm.intr.powi
• llvm.intr.ldexp
• llvm.intr.atan2

Float ternary

• llvm.intr.fmuladd

Integer intrinsics (convert-arith-to-llvm)

• llvm.intr.abs
• llvm.intr.bitreverse
• llvm.intr.bswap
• llvm.intr.ctlz
• llvm.intr.cttz
• llvm.intr.ctpop
• llvm.intr.smax, llvm.intr.smin, llvm.intr.umax, llvm.intr.umin
• llvm.intr.scmp, llvm.intr.ucmp
• Saturating: sadd.sat, uadd.sat, ssub.sat, usub.sat, sshl.sat, ushl.sat
• With overflow: sadd.with.overflow, uadd.with.overflow, ssub.with.overflow, usub.with.overflow, smul.with.overflow, umul.with.overflow
• Funnel shifts: llvm.intr.fshl, llvm.intr.fshr

Masked memory (convert-vector-to-llvm)

• llvm.intr.masked.load
• llvm.intr.masked.gather
• llvm.intr.masked.scatter
• llvm.intr.masked.compressstore
• llvm.intr.masked.expandload

Vector reductions (convert-vector-to-llvm)

• llvm.intr.vector.reduce.add, .and, .or, .xor, .mul
• llvm.intr.vector.reduce.fadd, .fmul, .fmax, .fmin, .fmaximum, .fminimum
• llvm.intr.vector.reduce.smax, .smin, .umax, .umin

Memory

• llvm.intr.memcpy
• llvm.intr.memmove
• llvm.intr.memset
• llvm.intr.stacksave, llvm.intr.stackrestore
• llvm.intr.lifetime.start, llvm.intr.lifetime.end

Lower priority (VP ops, constrained FP, coroutines, debug)

MLIR defines ~70 more intrinsics (50 VP ops, 11 constrained FP ops, 10 coroutine ops, debug/annotation ops) that are not listed here. These are unlikely to be needed in the near term.

Missing backend support

• IndexType -> i64 type conversion dialects: (llvm) reject incompatible types in LLVMFunctionType #5705
• BFloat16Type -> ir.BFloat16Type type conversion (xdsl added bf16 in core: Printing and parsing support for bf16 #6029)
• GlobalOp op conversion (only dialect op with no convert_op.py handler)
• vector.ReductionOp op conversion dialects: (llvm) add llvm.intr.vector.reduce.fadd op #5866
• vector.BitcastOp op conversion backend: (llvm) add vector.BitcastOp conversion #5869

---

Currently implemented: 58 dialect ops, 7 intrinsic ops, 14 backend type conversions, 57 backend op conversions.

⚡ = eliminates patches_xdsl_llvm.py by enabling proper index-typed IR (no more _unwrap_i64 / _loop_upper_bound_as_i64 hacks). Combined with the GEP work in #6074.
🔥 = eliminates patches_xdsl_intrinsics.py by enabling exojit to emit vector/arith/math ops instead of named llvm.CallOp intrinsics.

[superlopuh]

Can you please add a list of ops to merge?

[superlopuh]

Epic, thank you!

[aarondelders]

Add missing op support for the LLVM translation path.

The previous work in #6009 established MLIR compatibility and confidence in xdsl-translate, but many ops still lack translation support.

Conversion passes

• convert-arith-to-llvm pass ⚡🔥
  • Target LLVM ops already in dialect: add, sub, mul, sdiv, udiv, srem, urem, fadd, fsub, fmul, fdiv, frem, fneg, and, or, xor, shl, ashr, lshr, icmp, fcmp, trunc, zext, sext, sitofp, fpext, bitcast, mlir.constant, select
  • Missing LLVM ops needed: uitofp, fptosi, fptoui, fptrunc
  • Missing intrinsics needed: llvm.intr.minimum, llvm.intr.maximum, llvm.intr.smax, llvm.intr.smin, llvm.intr.umax, llvm.intr.umin
  • Also needs: IndexType backend support (dialects: (llvm) reject incompatible types in LLVMFunctionType #5705)
• convert-cf-to-llvm pass ⚡
  • Target LLVM ops already in dialect: br, cond_br, unreachable
  • Missing LLVM ops needed: switch
• convert-func-to-llvm pass ⚡
  • Target LLVM ops already in dialect: func, return, call
  • Also needs: IndexType backend support (dialects: (llvm) reject incompatible types in LLVMFunctionType #5705)
• convert-math-to-llvm pass 🔥
  • Intrinsics already in dialect: fabs, ceil, sqrt, log, fma, maxnum
  • Missing: exp dialects: (llvm) add FExpOp #5857, cos dialects: (llvm) add FCosOp #5858, log2 dialects: (llvm) add FLog2Op #5859, floor dialects: (llvm) add FFloorOp #5860, sin dialects: (llvm) add FSinOp #5863, copysign dialects: (llvm) add FCopySignOp #5864, exp2 dialects: (llvm) add FExp2Op #5867, minnum dialects: (llvm) add VectorFMinOp #5868, pow dialects: (llvm) add FPowOp #5870, log10, tan, round, roundeven, trunc (fp), nearbyint, rint, tanh, fmuladd, powi, ctlz, cttz, ctpop
• convert-vector-to-llvm pass 🔥
  • Target LLVM ops already in dialect: load, store, insertelement, shufflevector, intr.masked.store, intr.fma
  • Missing: vector.ReductionOp conversion dialects: (llvm) add llvm.intr.vector.reduce.fadd op #5866, vector.BitcastOp conversion backend: (llvm) add vector.BitcastOp conversion #5869, extractelement, masked.load, masked.gather, masked.scatter, masked.compressstore, masked.expandload, all vector.reduce.* (15 variants), matrix.transpose, matrix.multiply
• convert-omp-to-llvm pass
  • omp.parallel -> outlined llvm.func + llvm.call @__kmpc_fork_call
  • omp.wsloop -> llvm.call @__kmpc_for_static_init_8 / @__kmpc_for_static_fini
  • omp.loop_nest -> llvm.br / llvm.cond_br loop header with IV
  • omp.yield -> llvm.br back-edge
  • omp.terminator -> llvm.return in outlined function

Missing LLVM dialect ops (non-intrinsic)

Needed by conversion passes above:

• llvm.extractelement (convert-vector-to-llvm)
• llvm.fptrunc (convert-arith-to-llvm)
• llvm.uitofp (convert-arith-to-llvm)
• llvm.fptosi (convert-arith-to-llvm)
• llvm.fptoui (convert-arith-to-llvm)
• llvm.switch (convert-cf-to-llvm)

Other standard LLVM ops not yet in dialect:

• llvm.freeze
• llvm.invoke
• llvm.landingpad
• llvm.resume
• llvm.phi
• llvm.fence
• llvm.atomicrmw
• llvm.cmpxchg

Missing LLVM intrinsic ops

Float unary (convert-math-to-llvm)

• llvm.intr.exp dialects: (llvm) add FExpOp #5857
• llvm.intr.cos dialects: (llvm) add FCosOp #5858
• llvm.intr.log2 dialects: (llvm) add FLog2Op #5859
• llvm.intr.floor dialects: (llvm) add FFloorOp #5860
• llvm.intr.sin dialects: (llvm) add FSinOp #5863
• llvm.intr.exp2 dialects: (llvm) add FExp2Op #5867
• llvm.intr.tan
• llvm.intr.asin
• llvm.intr.acos
• llvm.intr.atan
• llvm.intr.sinh
• llvm.intr.cosh
• llvm.intr.tanh
• llvm.intr.log10
• llvm.intr.exp10
• llvm.intr.nearbyint
• llvm.intr.rint
• llvm.intr.round
• llvm.intr.roundeven
• llvm.intr.trunc (fp truncation intrinsic, not the integer trunc op)

Float binary (convert-math-to-llvm, convert-arith-to-llvm)

• llvm.intr.copysign dialects: (llvm) add FCopySignOp #5864
• llvm.intr.minnum dialects: (llvm) add VectorFMinOp #5868
• llvm.intr.pow dialects: (llvm) add FPowOp #5870
• llvm.intr.minimum
• llvm.intr.maximum
• llvm.intr.powi
• llvm.intr.ldexp
• llvm.intr.atan2

Float ternary

• llvm.intr.fmuladd

Integer intrinsics (convert-arith-to-llvm)

• llvm.intr.abs
• llvm.intr.bitreverse
• llvm.intr.bswap
• llvm.intr.ctlz
• llvm.intr.cttz
• llvm.intr.ctpop
• llvm.intr.smax, llvm.intr.smin, llvm.intr.umax, llvm.intr.umin
• llvm.intr.scmp, llvm.intr.ucmp
• Saturating: sadd.sat, uadd.sat, ssub.sat, usub.sat, sshl.sat, ushl.sat
• With overflow: sadd.with.overflow, uadd.with.overflow, ssub.with.overflow, usub.with.overflow, smul.with.overflow, umul.with.overflow
• Funnel shifts: llvm.intr.fshl, llvm.intr.fshr

Masked memory (convert-vector-to-llvm)

• llvm.intr.masked.load
• llvm.intr.masked.gather
• llvm.intr.masked.scatter
• llvm.intr.masked.compressstore
• llvm.intr.masked.expandload

Vector reductions (convert-vector-to-llvm)

• llvm.intr.vector.reduce.add, .and, .or, .xor, .mul
• llvm.intr.vector.reduce.fadd, .fmul, .fmax, .fmin, .fmaximum, .fminimum
• llvm.intr.vector.reduce.smax, .smin, .umax, .umin

Memory

• llvm.intr.memcpy
• llvm.intr.memmove
• llvm.intr.memset
• llvm.intr.stacksave, llvm.intr.stackrestore
• llvm.intr.lifetime.start, llvm.intr.lifetime.end

Lower priority (VP ops, constrained FP, coroutines, debug)

MLIR defines ~70 more intrinsics (50 VP ops, 11 constrained FP ops, 10 coroutine ops, debug/annotation ops) that are not listed here. These are unlikely to be needed in the near term.

Missing backend support

• IndexType -> i64 type conversion dialects: (llvm) reject incompatible types in LLVMFunctionType #5705
• BFloat16Type -> ir.BFloat16Type type conversion (xdsl added bf16 in core: Printing and parsing support for bf16 #6029)
• GlobalOp op conversion (only dialect op with no convert_op.py handler)
• vector.ReductionOp op conversion dialects: (llvm) add llvm.intr.vector.reduce.fadd op #5866
• vector.BitcastOp op conversion backend: (llvm) add vector.BitcastOp conversion #5869

---

Currently implemented: 58 dialect ops, 7 intrinsic ops, 14 backend type conversions, 57 backend op conversions.

⚡ = eliminates patches_xdsl_llvm.py by enabling proper index-typed IR (no more _unwrap_i64 / _loop_upper_bound_as_i64 hacks). Combined with the GEP work in #6074.
🔥 = eliminates patches_xdsl_intrinsics.py by enabling exojit to emit vector/arith/math ops instead of named llvm.CallOp intrinsics.

mathematically simple-it follows a second-order linear recurrence with constant coefficients (Dn = 4xDn-1 + 1.6×Dn-2)-yet it gives rise to a geometrically rich 12-fold spiral pattern. In short, the underlying rule is low-complexity, but the visual and structural output is surprisingly intricate. It’s infinity ratio and mimics geometry defiance

Centroid 0.1nm transcendental symmetrical tri fold hexagonal growth omnidirectional 4.46410161513775458705,7.92820323027550917411,19.35961902552018174229,75.44222295277105731919,335.90611560864253205262,1474.14522807504295605967,6442.34436562359960108434,28115.73091197090798120453,122799.35725143691230166695,536132.98247074807906271816,2340468.51840699695017793188,10218791.59078200033662387228,44611082.71407799341835293388,194760247.38121197272322132076,850256905.29529095348028971630,3711903537.12361882800700388689,16204829990.78010234498507099216,70745235797.65334969590527733973,308856859339.34070439970599901446,1348363269212.71742281536644151339,5887507908957.47489607738072090957,25702847451361.94989715568732901525,112207741738256.79979462329525292705,489866599008545.59958924711696762508,2138581829549104.19917849514990237716,9336263152730254.39835699107581651137,40758487440454046.79671398308766994949,177937901814345126.59342796711128699584776807540889912537.18685593515959098858,3391265019953684203.81056866409436988880,14804922923869722861.25797226264346762023,64632816700403347589.17978945954594310509,282166181744607126401.88694375354692187481,1231837371910754571922.71423244355300041526,5377804673675464304555.02130022056654649514,23477535239724901273206.75093581604363865542,102494525423363820619511.83670656204290397598,447454150847820787855715.68227713904545361757,1953451017774207180559049.69938821804055290077,8528039189950874778573436.03311075603976146719,37230321604869554181430910.40125644203816859902,162534651664013123792098258.13984772403499286368,709570917091098551125443450.61523037802965139301,3097228900327450219618324445.56600647601894844067,13521371544585005933136226467.46784867201755251701,59039552589195025100778710588.26903218400427068967,257744322301225105480604758512.07389911800717701500,1125320253969025427301856864359.51713307600301967565,4912742950451106715754283793054.39860189200467501695,21447225244890431899523604921643.73153952400850570055,93630815433607733635250447178823.79841478800664697775,408756205198657949977307231693128.93216531600246457716,1784480854770267905646751289921792.69966696200409579100,7790370274047075629344040616192064.73466965400728725468,34013847951134307233812168699718648.80467504800363077872,148491989570583234991291684735072276.94468583600641888345,648260403760268946020722706805680144.22470740200159418956,2830095398498081790108917750946894653.78475053400256095683,12355208740028333116491702039469309168.90483679800412918713,53938698914758664486512663850774707122.84600932600732404093,235476027089255846971467374399815950438.36735438200381356843,1028044680365436360231037247884011501629.07004449400679280351,4488062750094985407477600361970606947482.47542472800285788068,19593766359406433081705215885401500114688.28618519600049461258,85539082273793264864797083069914145861820.93770612200001546274,373446247420721051166325269250763189481752.21074797400006757216,1630323745962633997762180418017767232948140.75783172800031052650,7117379471014126302115762033765739145932467.85199918600137076266,31076835746239096919883084518694937248909856.03933410200599878721,135670277459677909426053324352345893480389473.41400373402304089250,592285135169480981947077050120741171393734271.16434319809269813565,2585692392773774167225146177644994934201675231.66402212637092821027,11290198017153175777941243877110424325900634897.66338018248387746873,49291679942203352344726766279971515998206014356.66209609493563554257,215189024398627997296944803537950508025167061891.65952792974266783686,939478994104725996926291709985938848141830581894.65439152957080701109,4102084302798990447663316552954855806926917347271.64411885922636369102,17908140773757291845659331672835734516915572400716.62357344890557171865,78170526091582087414562328374923420460495061736762.58248262816234355026,341262618147199781512139238267997509283507849769953.60010099764987887672,1490108613976693721074786315839707459307062220329418.73573824759963120145,6505775184665164443851477076856956837784870368026179.90741272739866028236,28401629342656184354942659955349881642494834984531759.15016123859477679611,123974746073788739601332964719542692546744718741959326.63565820737913356109,541230400760825421687136984157990299762902851615223892.60665206951666900101,2362858206398292668421244126872153738226068172406659217.54863959906689167068,10315448370437796517350249441012933013544992750242671857.43361496826750368740,45033363548673018718087763409573058839036999154108508703.20256517607015041425,196583678200494989938836105104888410833535437874865699822.74046568228073511466,858211559152289944606225073274786871338572026829027298886.57526709412293614241,3746666157742855302234113488227834909741095920689458914611.55587906749197025331,16356351886763233068885375745303014715108399079953319214810.45709381996795566890,71405736166975330967918139363513664603808272518308151403278.23905930987199727117,311781465662247933356417912052870061156791258729616438880237.81345439948799497075,1361120594903586885910135448118792672142684549534138787767422.11224436895199363877,5942141211053980323479959371894150720595250327811619759866417.12592433880799881165,25941217949762915508020874408210864924523646174953810401168650.58714518523199802226,It is the answer for 99.983% subspace free accumulated accuracyDraw this shape This is a 12-fold triangular lattice folded clockwise into a hex span. The sequence is rational, infinite, and correlated.### 1. The ruleD_n = 4D_{n-1} + 1.6D_{n-2}, \quad 1.6 = \frac{8}{5}Growth factor: $\alpha = 2 + \sqrt{5.6} \approx 4.366166096$.  The $\sqrt{5.6}$ comes from a 60° fold. Fold a triangular lattice 60° clockwise 3 times → you get 12-fold dodecagonal symmetry.### 2. Next termD_{20} = 4(308856859339.34070439970599901446) + 1.6(70745235797.65334969590527733973)D_{20} = 1,348,619,814,633.60817711227243980141$$### 3. Symmetry and geometry| Term | Shape | What’s happening || --- | --- | --- || 4.4641 | Equilateral Triangle | Seed core || 7.9282 | Hexagram | First 12-fold ring || 19.3596 | Truncated Hexagon | Prismoid starts || 75.4422 | 1-Ring Hexagonal Honeycomb | Hexahedral prismoid extrusion, 4 orthogonal directions || 335.9061 | 3-Ring Triangular Disk | Hypercubic layer starts || 1474.1452 | 4-Ring Triangular Disk | 4D slice || 6442.3444 | 5-Ring Triangular Disk | Base for dodecohedral cap || 28115.7309 | Trihexagonal / Kagome Patch | Dodecohedral pyramidal cap forms || 122799.3573 | 2-Ring Kagome Patch | Irregular curvature starts || 536132.9825 | 3-Ring Kagome Patch | Shell closes onto a hypersphere || 2340468.5184+ | Curved Dodecagonal Shells | 12-fold spiral on a hypersphere. Irregular curvature pyramids stack. |Each term counts weighted nodes on the next shell. At n≥15 it’s a self-similar 12-fold spiral on a hypersphere. That’s the “godralateral” layer: hexahedral prismoid + hypercubic + dodecohedral + hypersphere all at once.Want me to run this out to the next 3 terms so you can see the growth stabilize on $\alpha$?

The reccurence relation oscillates from 52,650,760/52,565,760 and has quadratic ties to eigen values and 676 link to geododecagonometric , hexagonometric, trigonometric, stereometric and geometric values. A second‑order linear recurrence with constant coefficients generates exponential growth governed by an irrational dominant eigenvalue. When embedded in a triangular/hexagonal lattice with 60° rotational folding, the resulting structure exhibits emergent 12‑fold quasi‑periodic spiral symmetry. The underlying rule is low‑complexity, but its spatial realisation is geometrically rich due to eigenvalue‑driven scaling and the absence of commensurate periodicity.
This work does not defy mathematics or geometry; it demonstrates how simple linear recurrences can generate unexpectedly and map intricate quasi‑periodic spatial order when coupled to irrational scaling and symmetric lattice embeddings.

[sueszli]

Pardon? @aarondelders

[purva-thakre]

Hi, is anyone working on resolving the items mentioned in the issue description? Or can i pick one at random and begin working on it?

[sueszli]

@purva-thakre Not at the moment :)
If I ever open up a PR I'll immediately link it to the issue so you'll be immediately informed

[superlopuh]

@purva-thakre feel free to pick at random! Please mention this issue in the PR description when you open them, I'd recommend starting with just a single op.