v0.0.290

What's Changed

• add link to fortran bindings by @vchuravy in #2967
• mlir: add enzyme.atomic_rmw op with ordering attribute by @pengmai in #2950
• mlir: integrate dataflow activity analysis by @pengmai in #2970
• mlir: llvm.select activity interface, zeroInPlace for llvm.alloca by @pengmai in #2972
• mlir: remove dependency on linalg with enzyme.fill_zero op by @pengmai in #2973
• mlir: derivative for maxnumf, pointer handling for arith.select by @pengmai in #2971
• mlir: sink constants into autodiff_region when outlining by @pengmai in #2977
• mlir: add atomic_add attribute to enzyme.autodiff and autodiff_region ops by @pengmai in #2974
• No more linalg dep for enzyme pass (now lowering to fill op) by @wsmoses in #2976
• mlir: pass to split ifs with multi-results into separate ops by @pengmai in #2975
• mlir: support multi-dim gpu allocations in ForLikeEnzymeOpsRemover by @pengmai in #2980
• Use gcp-cache for MLIR build cache by @wsmoses in #2981
• [mlir] Support multi-dim dynamic memref caches by @pengmai in #2982
• mlir: upstream interface implementations by @pengmai in #2983
• [MLIR] Split registerCoreDialectAutodiffInterfaces into its own TU; trim pass deps by @vchuravy in #2988
• Mark custream capture as inactive by @wsmoses in #2992
• mlir: support non-index IVs in loops by @pengmai in #2993
• Disable enzyme lit test ci for macos debug by @wsmoses in #2997
• Add AIR compatibility by @pw0908 in #2934
• Mark profile counter as integral by @wsmoses in #3000
• GCP cache julia action if on relevant machine by @wsmoses in #3006
• [MLIR] Add StoreLikeInterface to decouple activity analysis from stores by @vchuravy in #2978
• Update to later llvm by @wsmoses in #3007
• Fixes for mutable memory in scf.for reverse by @Pangoraw in #3009
• [MLIR][Mincut] Add test for optimality issue for mincut by @wsmoses in #3013
• mlir: mark outlined functions from enzyme.autodiff_region private by @pengmai in #3015
• [MLIR][Mincut] Prefer last equal sized element in chain to cache by @wsmoses in #3017
• mlir: mark loads of read-only pointers as movable during mincut by @pengmai in #2734
• Speed up large constant ta by @wsmoses in #3002
• [mlir] Make non-looping, non-nested RegionBranchOpInterface ops movable by min-cut by @pengmai in #3024
• Prefer OptionalPassInfoMixin by @boomanaiden154 in #2996
• Remove deprecated use of make_scope_exit by @boomanaiden154 in #3034
• Do not use deprecated LLVM Typed pointer functions by @boomanaiden154 in #3033
• mlir: loop-invariant cache for small stack allocas by @pengmai in #3030
• Avoid getTerminator assertion on incomplete derivative block by @krystophny in #3023
• Relax dynamic-loop derivative signature check by @krystophny in #3021
• Don't add non-instructions to the min-cut recompute graph by @wsmoses in #3035
• mlir: fix overly conservative check when marking readonly by @pengmai in #3039
• [TA] Don't recurse infinitely on self-referential constant globals by @wsmoses in #3038
• mlir: Fix Binomial checkpointing schedule and support for mutable memory + memory space by @wsmoses in #3042
• Make IR emission order independent of C++ argument evaluation order by @wsmoses in #3044
• Add a lint job for order-dependent IR emission by @wsmoses in #3045
• Bind IRBuilder calls that share an argument list to locals by @wsmoses in #3046
• Properly consider ptrsizehit addr by @wsmoses in #3047
• mlir: Only delegate loop-carried caches to the loop remover by @wsmoses in #3049
• mlir: Simplify enzyme ops before each remover, not only once per function by @wsmoses in #3050
• Make the min cut graph's adjacency iteration deterministic by @wsmoses in #3058
• MLIR fast math derivatives by @wsmoses in #3057
• Preserve attributes on loops built while differentiating by @wsmoses in #3059
• Add pass to hoist allocations by @Pangoraw in #3051
• Flang llvm plugin by @vchuravy in #3056
• Add more Fortran tests by @joewallwork in #3020
• Separate out Fortran tests at different optimisation levels by @joewallwork in #3054
• Add batch mode test by @joewallwork in #3053
• Fix compilation warnings by @Pangoraw in #3068
• Serialize TypeTree offsets as signed constants by @wsmoses in #3070
• Use Utils.h wrappers for version-sensitive LLVM APIs, and lint for it by @wsmoses in #3071
• Drop identity address-space casts in the Julia alloca lowering by @wsmoses in #3074
• Report mismatched activity as a warning, not just an -Rpass remark by @wsmoses in #3076
• Add exp10 intrinsic for LLVM >=18 (#2998) by @SpeedyTurtle599 in #3001
• Give math.fma a derivative by @wsmoses in #3077
• Load adjoint: a mutable type takes a shadow, not an accumulation by @wsmoses in #3078
• Differentiate an llvm.func in forward mode by @wsmoses in #3087
• Provide a fpm.toml for the Fortran interface by @vchuravy in #3080
• A result nothing differentiates has no shadow to be given by @wsmoses in #3088
• Prefer annotations for Enzyme's registration attributes, fixing templates by @wsmoses in #3085
• Join a map of sets by what the other side has by @wsmoses in #3090
• Give each differentiated function its own dataflow solver by @wsmoses in #3091
• Do not walk one map while inserting into it by @wsmoses in #3094
• Give lifetime markers a forward-mode rule by @wsmoses in #3093
• Decide whether to skip an op with the analyzer everything else uses by @wsmoses in #3095
• Say an unknown summary set the way the readers of summaries read it by @wsmoses in #3097
• Keep a branch's condition when neither successor takes an argument by @wsmoses in #3096
• Give llvm.intr.memset a forward-mode rule by @wsmoses in #3098
• Remove a pop nobody reads when its cache has no other pop by @wsmoses in #3099
• Give the derivative its own comdat rather than the primal's by @wsmoses in #3100
• MLIR dataflow activity: a summarized function still needs its own analysis by @wsmoses in #3105
• MLIR: forward-mode tangent for arith.select by @wsmoses in #3107
• MLIR: differentiating a bodyless function is an error, not a crash or a zero by @wsmoses in #3104
• Describe the _enzyme* call markers in one place by @wsmoses in #3086
• mlir: unify checkpointing implementations by @Pangoraw in #3092
• Size a cloned or zeroed alloca from its type by @wsmoses in #3108

New Contributors

• @pw0908 made their first contribution in #2934
• @boomanaiden154 made their first contribution in #2996
• @krystophny made their first contribution in #3023
• @SpeedyTurtle599 made their first contribution in #3001

Full Changelog: v0.0.289...v0.0.290