OVERVIEW: MLIR modular optimizer driver
Available Dialects: acc, affine, amdgpu, amx, arith, arm_neon, arm_sme, arm_sve, async, bufferization, builtin, cf, complex, dlti, emitc, func, gpu, index, irdl, linalg, llvm, math, memref, ml_program, mpi, nvgpu, nvvm, omp, pdl, pdl_interp, ptr, quant, rocdl, scf, shape, shard, smt, sparse_tensor, spirv, tensor, tosa, transform, ub, vector, wasmssa, x86vector, xegpu, xevm
USAGE: mlir-opt [options] <input file>
OPTIONS:
Color Options:
--color - Use colors in output (default=autodetect)
General options:
--aarch64-neon-syntax=<value> - Choose style of NEON code to emit from AArch64 backend:
=generic - Emit generic NEON assembly
=apple - Emit Apple-style NEON assembly
--aarch64-use-aa - Enable the use of AA during codegen.
--abort-on-max-devirt-iterations-reached - Abort when the max iterations for devirtualization CGSCC repeat pass is reached
--allow-ginsert-as-artifact - Allow G_INSERT to be considered an artifact. Hack around AMDGPU test infinite loops.
--allow-unregistered-dialect - Allow operation with no registered dialects
--amdgpu-atomic-optimizer-strategy=<value> - Select DPP or Iterative strategy for scan
=DPP - Use DPP operations for scan
=Iterative - Use Iterative approach for scan
=None - Disable atomic optimizer
--amdgpu-bypass-slow-div - Skip 64-bit divide for dynamic 32-bit values
--amdgpu-disable-loop-alignment - Do not align and prefetch loops
--amdgpu-dpp-combine - Enable DPP combiner
--amdgpu-dump-hsa-metadata - Dump AMDGPU HSA Metadata
--amdgpu-enable-merge-m0 - Merge and hoist M0 initializations
--amdgpu-indirect-call-specialization-threshold=<uint> - A threshold controls whether an indirect call will be specialized
--amdgpu-kernarg-preload - Enable preload kernel arguments to SGPRs
--amdgpu-kernarg-preload-count=<uint> - How many kernel arguments to preload onto SGPRs
--amdgpu-module-splitting-max-depth=<uint> - maximum search depth. 0 forces a greedy approach. warning: the algorithm is up to O(2^N), where N is the max depth.
--amdgpu-promote-alloca-to-vector-limit=<uint> - Maximum byte size to consider promote alloca to vector
--amdgpu-promote-alloca-to-vector-max-regs=<uint> - Maximum vector size (in 32b registers) to use when promoting alloca
--amdgpu-promote-alloca-to-vector-vgpr-ratio=<uint> - Ratio of VGPRs to budget for promoting alloca to vectors
--amdgpu-sdwa-peephole - Enable SDWA peepholer
--amdgpu-use-aa-in-codegen - Enable the use of AA during codegen.
--amdgpu-verify-hsa-metadata - Verify AMDGPU HSA Metadata
--amdgpu-vgpr-index-mode - Use GPR indexing mode instead of movrel for vector indexing
--arc-contract-use-objc-claim-rv - Enable generation of calls to objc_claimAutoreleasedReturnValue
--argext-abi-check - Verify that narrow int args are properly extended per the SystemZ ABI.
--arm-add-build-attributes -
--arm-implicit-it=<value> - Allow conditional instructions outside of an IT block
=always - Accept in both ISAs, emit implicit ITs in Thumb
=never - Warn in ARM, reject in Thumb
=arm - Accept in ARM, reject in Thumb
=thumb - Warn in ARM, emit implicit ITs in Thumb
--asm-show-inst - Emit internal instruction representation to assembly file
--atomic-counter-update-promoted - Do counter update using atomic fetch add for promoted counters only
--atomic-first-counter - Use atomic fetch add for first counter in a function (usually the entry counter)
--basic-block-section-match-infer - Enable matching and inference when generating basic block sections
--bounds-checking-single-trap - Use one trap block per function
--bpf-stack-size=<int> - Specify the BPF stack size limit
--cfg-hide-cold-paths=<number> - Hide blocks with relative frequency below the given value
--cfg-hide-deoptimize-paths -
--cfg-hide-unreachable-paths -
--check-functions-filter=<regex> - Only emit checks for arguments of functions whose names match the given regular expression
--conditional-counter-update - Do conditional counter updates in single byte counters mode)
Compiler passes to run
Passes:
--acc-if-clause-lowering - Lower if clauses in ACC compute constructs
--acc-implicit-data - Generate implicit data attributes for OpenACC compute constructs
--enable-implicit-reduction-copy - Enable applying implicit copy in lieu of implicit firstprivate for reduction variables. This allows uniform treatment of reduction variables between combined constructs (e.g., 'parallel loop') and separate constructs (e.g., 'parallel' followed by 'loop'), where the OpenACC spec requires copy semantics for the former but firstprivate would normally apply for the latter.
--acc-implicit-declare - Applies implicit acc declare to globals referenced in compute and routine acc regions
--acc-implicit-routine - Generate implicit acc routine for functions in acc regions
--device-type=<value> - Target device type for implicit routine generation. Ensures that `acc routine` device_type clauses are properly considered not just default clauses.
=none - none
=host - host
=multicore - multicore
=nvidia - nvidia
=radeon - radeon
--acc-legalize-serial - Legalize OpenACC serial constructs
--acc-loop-tiling - Tile OpenACC loops with tile clauses
--default-tile-size=<int> - Default tile size to use for wildcard ('*') tile sizes
--acc-specialize-for-device - Strip OpenACC constructs inside device code
--acc-specialize-for-host - Convert OpenACC operations for host execution
--enable-host-fallback - Enable host fallback mode which converts ALL ACC operations, not just orphan operations. Use this when the `if` clause evaluates to false.
--affine-data-copy-generate - Generate explicit copying for affine memory operations
--fast-mem-capacity=<ulong> - Set fast memory space capacity in KiB (default: unlimited)
--fast-mem-space=<uint> - Fast memory space identifier for copy generation (default: 1)
--generate-dma - Generate DMA instead of point-wise copy
--min-dma-transfer=<int> - Minimum DMA transfer size supported by the target in bytes
--skip-non-unit-stride-loops - Testing purposes: avoid non-unit stride loop choice depths for copy placement
--slow-mem-space=<uint> - Slow memory space identifier for copy generation (default: 0)
--tag-mem-space=<uint> - Tag memory space identifier for copy generation (default: 0)
--affine-expand-index-ops - Lower affine operations operating on indices into more fundamental operations
--affine-expand-index-ops-as-affine - Lower affine operations operating on indices into affine.apply operations
--affine-fold-memref-alias-ops - Fold memref alias ops into affine memory ops
--affine-loop-coalescing - Coalesce nested loops with independent bounds into a single loop
--affine-loop-fusion - Fuse affine loop nests
--compute-tolerance=<number> - Fractional increase in additional computation tolerated while fusing
--fast-mem-space=<uint> - Faster memory space number to promote fusion buffers to
--local-buf-threshold=<ulong> - Threshold size (KiB) for promoting local buffers to fast memory space
--maximal - Enables maximal loop fusion
--mode=<value> - fusion mode to attempt
=greedy - Perform greedy (both producer-consumer and sibling) fusion
=producer - Perform only producer-consumer fusion
=sibling - Perform only sibling fusion
--affine-loop-invariant-code-motion - Hoist loop invariant instructions outside of affine loops
--affine-loop-normalize - Apply normalization transformations to affine loop-like ops
--promote-single-iter - Promote single iteration loops
--affine-loop-tile - Tile affine loop nests
--cache-size=<ulong> - Set size of cache to tile for in KiB (default: 512)
--separate - Separate full and partial tiles (default: false)
--tile-size=<uint> - Use this tile size for all loops
--tile-sizes=<uint> - List of tile sizes for each perfect nest (overridden by -tile-size)
--affine-loop-unroll - Unroll affine loops
--cleanup-unroll - Fully unroll the cleanup loop when possible.
--unroll-factor=<long> - Use this unroll factor for all loops being unrolled, set it to -1, and it will fully unroll.
--unroll-full-threshold=<uint> - Unroll all loops with trip count less than or equal to this
--unroll-num-reps=<uint> - Unroll innermost loops repeatedly this many times
--unroll-up-to-factor - Allow unrolling up to the factor specified
--affine-loop-unroll-jam - Unroll and jam affine loops
--unroll-jam-factor=<uint> - Use this unroll jam factor for all loops (default 4)
--affine-parallelize - Convert affine.for ops into 1-D affine.parallel
--max-nested=<uint> - Maximum number of nested parallel loops to produce. Defaults to unlimited (UINT_MAX).
--parallel-reductions - Whether to parallelize reduction loops. Defaults to false.
--affine-pipeline-data-transfer - Pipeline non-blocking data transfers between explicitly managed levels of the memory hierarchy
--affine-raise-from-memref - Turn some memref operators to affine operators where supported
--affine-scalrep - Replace affine memref accesses by scalars by forwarding stores to loads and eliminating redundant loads
--affine-simplify-min-max - Simplify affine min/max/apply
--affine-simplify-structures - Simplify affine expressions in maps/sets and normalize memrefs
--affine-super-vectorize - Vectorize to a target independent n-D vector abstraction
--test-fastest-varying=<long> - Specify a 1-D, 2-D or 3-D pattern of fastest varying memory dimensions to match. See defaultPatterns in Vectorize.cpp for a description and examples. This is used for testing purposes
--vectorize-reductions - Vectorize known reductions expressed via iter_args. Switched off by default.
--virtual-vector-size=<long> - Specify an n-D virtual vector size for vectorization. This must be greater than zero.
--amdgpu-emulate-atomics - Emulate atomic operations on chipsets that do not support them
--chipset=<string> - Chipset that these operations will run on
--amdgpu-fold-memrefs-ops - Fold memref operations into their parent operations
--amdgpu-maskedload-to-load - Lower the operations from the vector maskedload to vector load
--amdgpu-resolve-strided-metadata - Resolve memref.extract_strided_metadata on AMDGPU ops
--arith-emulate-unsupported-floats - Emulate operations on unsupported floats with extf/truncf
--source-types=<string> - MLIR types without arithmetic support on a given target
--target-type=<string> - MLIR type to convert the unsupported source types to
--arith-emulate-wide-int - Emulate 2*N-bit integer operations using N-bit operations
--widest-int-supported=<uint> - Widest integer type supported by the target
--arith-expand - Legalize Arith ops to be convertible to LLVM.
--include-bf16 - Enable the BF16 expansion patterns
--include-f4e2m1 - Enable the F4E2M1 expansion patterns
--include-f8e8m0 - Enable the F8E8M0 expansion patterns
--arith-int-range-narrowing - Reduce integer operations bitwidth based on integer range analysis
--int-bitwidths-supported=<uint> - Integer bitwidths supported
--arith-unsigned-when-equivalent - Replace signed ops with unsigned ones where they are proven equivalent
--arm-neon-2d-to-intr - Convert Arm NEON structured ops to intrinsics
--arm-sme-outer-product-fusion - Fuse 'arm_sme.outerproduct' operations into 2-way or 4-way widening variants
--arm-sme-vector-legalization - Legalize vectors for ArmSME
--arm-sve-legalize-vector-storage - Ensures stores of SVE vector types will be legal
--async-func-to-async-runtime - Lower async.func operations to the explicit async.runtime andasync.coro operations
--async-parallel-for - Convert scf.parallel operations to multiple async compute ops executed concurrently for non-overlapping iteration ranges
--async-dispatch - Dispatch async compute tasks using recursive work splitting. If `false` async compute tasks will be launched using simple for loop in the caller thread.
--min-task-size=<int> - The minimum task size for sharding parallel operation.
--num-workers=<int> - The number of available workers to execute async operations. If `-1` the value will be retrieved from the runtime.
--async-runtime-policy-based-ref-counting - Policy based reference counting for Async runtime operations
--async-runtime-ref-counting - Automatic reference counting for Async runtime operations
--async-runtime-ref-counting-opt - Optimize automatic reference counting operations for theAsync runtime by removing redundant operations
--async-to-async-runtime - Lower all high level async operations (e.g. async.execute) tothe explicit async.runtime and async.coro operations
--bubble-down-memory-space-casts - Bubbles down memory-space cast operations.
--buffer-deallocation-simplification - Optimizes `bufferization.dealloc` operation for more efficient codegen
--buffer-hoisting - Optimizes placement of allocation operations by moving them into common dominators and out of nested regions
--buffer-loop-hoisting - Optimizes placement of allocation operations by moving them out of loop nests
--buffer-results-to-out-params - Converts memref-typed function results to out-params
--add-result-attr - Add the attribute 'bufferize.result' to all output parameters.
--hoist-dynamic-allocs - Hoist dynamic allocations to call sites.
--hoist-static-allocs - Hoist static allocations to call sites.
--modify-public-functions - Modify function signatures of public functions.
--bufferization-lower-deallocations - Lowers `bufferization.dealloc` operations to `memref.dealloc`operations
--canonicalize - Canonicalize operations
--disable-patterns=<string> - Labels of patterns that should be filtered out during application
--enable-patterns=<string> - Labels of patterns that should be used during application, all other patterns are filtered out
--max-iterations=<long> - Max. iterations between applying patterns / simplifying regions
--max-num-rewrites=<long> - Max. number of pattern rewrites within an iteration
--region-simplify=<value> - Perform control flow optimizations to the region tree
=disabled - Don't run any control-flow simplification.
=normal - Perform simple control-flow simplifications (e.g. dead args elimination).
=aggressive - Perform aggressive control-flow simplification (e.g. block merging).
--test-convergence - Test only: Fail pass on non-convergence to detect cyclic pattern
--top-down - Seed the worklist in general top-down order
--composite-fixed-point-pass - Composite fixed point pass
--max-iterations=<int> - Maximum number of iterations if inner pipeline
--name=<string> - Composite pass display name
--pipeline=<string> - Composite pass inner pipeline
--control-flow-sink - Sink operations into conditional blocks
--convert-affine-for-to-gpu - Convert top-level AffineFor Ops to GPU kernels
--gpu-block-dims=<uint> - Number of GPU block dimensions for mapping
--gpu-thread-dims=<uint> - Number of GPU thread dimensions for mapping
--convert-amdgpu-to-rocdl - Convert AMDGPU dialect to ROCDL dialect
--chipset=<string> - Chipset that these operations will run on
--convert-arith-to-amdgpu - Convert Arith operations to AMDGPU-specific implementations
--allow-packed-f16-round-to-zero - Whether we should allow f32->f16 packed round-to-zero conversion
--chipset=<string> - Chipset that these operations will run on
--saturate-fp8-truncf - Use saturating truncation for 8-bit float types
--convert-arith-to-apfloat - Convert Arith ops to APFloat runtime library calls
--convert-arith-to-arm-sme - Convert Arith dialect to ArmSME dialect
--convert-arith-to-emitc - Convert Arith dialect to EmitC dialect
--convert-arith-to-llvm - Convert Arith dialect to LLVM dialect
--index-bitwidth=<uint> - Bitwidth of the index type, 0 to use size of machine word
--convert-arith-to-spirv - Convert Arith dialect to SPIR-V dialect
--emulate-lt-32-bit-scalar-types - Emulate narrower scalar types with 32-bit ones if not supported by the target
--emulate-unsupported-float-types - Emulate unsupported float types by representing them with integer types of same bit width
--convert-arm-sme-to-llvm - Lower the operations from the ArmSME dialect into the LLVM dialect
--dump-tile-live-ranges - Dump the live ranges of SME tiles (for debugging)
--convert-arm-sme-to-scf - Lower the operations from the ArmSME dialect into the SCF dialect
--convert-async-to-llvm - Convert the operations from the async dialect into the LLVM dialect
--convert-bufferization-to-memref - Convert operations from the Bufferization dialect to the MemRef dialect
--convert-cf-to-llvm - Convert ControlFlow operations to the LLVM dialect
--index-bitwidth=<uint> - Bitwidth of the index type, 0 to use size of machine word
--convert-cf-to-spirv - Convert ControlFlow dialect to SPIR-V dialect
--emulate-lt-32-bit-scalar-types - Emulate narrower scalar types with 32-bit ones if not supported by the target
--emulate-unsupported-float-types - Emulate unsupported float types by representing them with integer types of same bit width
--convert-complex-to-libm - Convert Complex dialect to libm calls
--convert-complex-to-llvm - Convert Complex dialect to LLVM dialect
--complex-range=<value> - Control the intermediate calculation of complex number division
=improved - improved
=basic - basic (default)
=none - none
--convert-complex-to-rocdl-library-calls - Convert Complex dialect to ROCDL library calls
--convert-complex-to-spirv - Convert Complex dialect to SPIRV dialect
--convert-complex-to-standard - Convert Complex dialect to standard dialect
--complex-range=<value> - Control the intermediate calculation of complex number division
=improved - improved (default)
=basic - basic
=none - none
--convert-elementwise-to-linalg - Convert ElementwiseMappable ops to linalg
--convert-func-to-emitc - Convert Func dialect to EmitC dialect
--convert-func-to-llvm - Convert from the Func dialect to the LLVM dialect
--index-bitwidth=<uint> - Bitwidth of the index type, 0 to use size of machine word
--use-bare-ptr-memref-call-conv - Replace FuncOp's MemRef arguments with bare pointers to the MemRef element types
--convert-func-to-spirv - Convert Func dialect to SPIR-V dialect
--emulate-lt-32-bit-scalar-types - Emulate narrower scalar types with 32-bit ones if not supported by the target
--emulate-unsupported-float-types - Emulate unsupported float types by representing them with integer types of same bit width
--convert-gpu-to-llvm-spv - Generate LLVM operations to be ingested by a SPIR-V backend for gpu operations
--use-64bit-index - Use 64-bit integers to convert index types
--convert-gpu-to-nvvm - Generate NVVM operations for gpu operations
--allow-pattern-rollback - Experimental performance flag to disallow pattern rollback
--allowed-dialects=<string> - Run conversion patterns of only the specified dialects
--has-redux - Target gpu supports redux
--index-bitwidth=<uint> - Bitwidth of the index type, 0 to use size of machine word
--use-bare-ptr-memref-call-conv - Replace memref arguments in GPU functions with bare pointers. All memrefs must have static shape.
--convert-gpu-to-rocdl - Generate ROCDL operations for gpu operations
--allowed-dialects=<string> - Run conversion patterns of only the specified dialects
--chipset=<string> - Chipset that these operations will run on
--index-bitwidth=<uint> - Bitwidth of the index type, 0 to use size of machine word
--runtime=<value> - Runtime code will be run on (default is Unknown, can also use HIP or OpenCL)
=unknown - Unknown (default)
=HIP - HIP
=OpenCL - OpenCL
--use-bare-ptr-memref-call-conv - Replace memref arguments in GPU functions with bare pointers.All memrefs must have static shape
--convert-gpu-to-spirv - Convert GPU dialect to SPIR-V dialect
--use-64bit-index - Use 64-bit integers to convert index types
--convert-index-to-llvm - Lower the `index` dialect to the `llvm` dialect.
--index-bitwidth=<uint> - Bitwidth of the index type, 0 to use size of machine word
--convert-index-to-spirv - Lower the `index` dialect to the `spirv` dialect.
--use-64bit-index - Use 64-bit integers to convert index types
--convert-linalg-to-affine-loops - Lower the operations from the linalg dialect into affine loops
--convert-linalg-to-loops - Lower the operations from the linalg dialect into loops
--convert-linalg-to-parallel-loops - Lower the operations from the linalg dialect into parallel loops
--convert-linalg-to-std - Convert the operations from the linalg dialect into the Standard dialect
--convert-math-to-apfloat - Convert Math ops to APFloat runtime library calls
--convert-math-to-emitc - Convert some Math operations to EmitC call_opaque operations
--language-target=<value> - Select the language standard target for callees (c99 or cpp11).
=c99 - c99
=cpp11 - cpp11
--convert-math-to-funcs - Convert Math operations to calls of outlined implementations.
--convert-ctlz - Convert math.ctlz to a software implementation. Enable for targets that do not natively support ctlz.
--min-width-of-fpowi-exponent=<uint> - Convert FPowI only if the width of its exponent's integer type is greater than or equal to this value
--convert-math-to-libm - Convert Math dialect to libm calls
--convert-math-to-llvm - Convert Math dialect to LLVM dialect
--approximate-log1p - Enable approximation of Log1p.
--convert-math-to-rocdl - Convert Math dialect to ROCDL library calls
--chipset=<string> - Chipset that these operations will run on
--convert-math-to-spirv - Convert Math dialect to SPIR-V dialect
--convert-math-to-xevm - Convert (fast) math operations to native XeVM/SPIRV equivalents
--convert-arith - Convert supported Arith ops (e.g. arith.divf) as well.
--convert-memref-to-emitc - Convert MemRef dialect to EmitC dialect
--lower-to-cpp - Target C++ (true) instead of C (false)
--convert-memref-to-spirv - Convert MemRef dialect to SPIR-V dialect
--bool-num-bits=<int> - The number of bits to store a boolean value
--use-64bit-index - Use 64-bit integers to convert index types
--convert-nvgpu-to-nvvm - Convert NVGPU dialect to NVVM dialect
--convert-nvvm-to-llvm - Convert NVVM to PTX with Inline Assembly in LLVM dialect
--convert-openacc-to-scf - Convert the OpenACC ops to OpenACC with SCF dialect
--convert-openmp-to-llvm - Convert the OpenMP ops to OpenMP ops with LLVM dialect
--convert-parallel-loops-to-gpu - Convert mapped scf.parallel ops to gpu launch operations
--convert-pdl-to-pdl-interp - Convert PDL ops to PDL interpreter ops
--convert-scf-to-cf - Convert SCF dialect to ControlFlow dialect, replacing structured control flow with a CFG
--allow-pattern-rollback - Experimental performance flag to disallow pattern rollback
--convert-scf-to-emitc - Convert SCF dialect to EmitC dialect, maintaining structured control flow
--convert-scf-to-openmp - Convert SCF parallel loop to OpenMP parallel + workshare constructs.
--num-threads=<uint> - Number of threads to use
--convert-scf-to-spirv - Convert SCF dialect to SPIR-V dialect.
--convert-shape-constraints - Convert shape constraint operations to the standard dialect
--convert-shape-to-std - Convert operations from the shape dialect into the standard dialect
--convert-shard-to-mpi - Convert Shard dialect to MPI dialect.
--convert-spirv-to-llvm - Convert SPIR-V dialect to LLVM dialect
--client-api=<value> - Derive StorageClass to address space mapping from the client API
=Unknown - Unknown (default)
=Metal - Metal
=OpenCL - OpenCL
=Vulkan - Vulkan
=WebGPU - WebGPU
--convert-tensor-to-linalg - Convert some Tensor dialect ops to Linalg dialect
--convert-tensor-to-spirv - Convert Tensor dialect to SPIR-V dialect
--emulate-lt-32-bit-scalar-types - Emulate narrower scalar types with 32-bit ones if not supported by the target
--emulate-unsupported-float-types - Emulate unsupported float types by representing them with integer types of same bit width
--convert-to-emitc - Convert to EmitC dialect via dialect interfaces
--filter-dialects=<string> - Test conversion patterns of only the specified dialects
--convert-to-llvm - Convert to LLVM via dialect interfaces found in the input IR
--allow-pattern-rollback - Experimental performance flag to disallow pattern rollback
--dynamic - Use op conversion attributes to configure the conversion
--filter-dialects=<string> - Test conversion patterns of only the specified dialects
--convert-ub-to-llvm - Convert UB dialect to LLVM dialect
--index-bitwidth=<uint> - Bitwidth of the index type, 0 to use size of machine word
--convert-ub-to-spirv - Convert UB dialect to SPIR-V dialect
--convert-vector-to-amx - Lower the operations from the vector dialect into the AMX dialect
--convert-vector-to-arm-sme - Lower the operations from the vector dialect into the ArmSME dialect
--convert-vector-to-gpu - Lower the operations from the vector dialect into the GPU dialect
--use-nvgpu - convert to NvGPU ops instead of GPU dialect ops
--convert-vector-to-llvm - Lower the operations from the vector dialect into the LLVM dialect
--enable-amx - Enables the use of AMX dialect while lowering the vector dialect.
--enable-arm-bf16 - Enables the use of Arm FEAT_BF16 instructions while lowering the vector dialect.
--enable-arm-i8mm - Enables the use of Arm FEAT_I8MM instructions while lowering the vector dialect.
--enable-arm-neon - Enables the use of ArmNeon dialect while lowering the vector dialect.
--enable-arm-sve - Enables the use of ArmSVE dialect while lowering the vector dialect.
--enable-x86vector - Enables the use of X86Vector dialect while lowering the vector dialect.
--force-32bit-vector-indices - Allows compiler to assume vector indices fit in 32-bit if that yields faster code
--reassociate-fp-reductions - Allows llvm to reassociate floating-point reductions for speed
--use-vector-alignment - Use the preferred alignment of a vector type in load/store operations instead of the alignment of the element type of the memref. This flag is intended for use with hardware which requiresvector alignment, or in application contexts where it is known all vector access are naturally aligned. If operations have an alignment attribute set, the alignment attribute takes priority over this option
--vector-contract-lowering=<value> - control the lowering of `vector.contract` operations.
=dot - Progressively lower to finer grained `vector.contract` and dot-products. (default)
=llvmintr - Lower directly to `llvm.intr.matrix.multiply`.
=outerproduct - Lower to `vector.outerproduct`.
=parallelarith - Lower contract with all reduction dimensions unrolled to 1 to a vector elementwise operations.
--vector-transpose-lowering=<value> - control the lowering of `vector.transpose` operations.
=eltwise - Lower transpose into element-wise extract and inserts (default)
=llvmintr - Lower 2-D transpose directly to `llvm.intr.matrix.transpose`
=shuffle1d - Lower 2-D transpose to `vector.shuffle` on 1-D vector.
=shuffle16x16 - Lower 2-D transpose to `vector.shuffle` on 16x16 vector.
--convert-vector-to-scf - Lower the operations from the vector dialect into the SCF dialect
--full-unroll - Perform full unrolling when converting vector transfers to SCF
--lower-scalable - Add scalable vector specific lowerings (that introduce loops)
--lower-tensors - Lower transfer ops that operate on tensors
--target-rank=<uint> - Target vector rank to which transfer ops should be lowered
--convert-vector-to-spirv - Convert Vector dialect to SPIR-V dialect
--convert-vector-to-xegpu - Lower the operations from the vector dialect into the XeGPU dialect
--convert-xegpu-to-xevm - Convert XeGPU to XeVM dialect
--convert-xevm-to-llvm - Convert XeVM to LLVM dialect
--cse - Eliminate common sub-expressions
--decorate-spirv-composite-type-layout - Decorate SPIR-V composite type with layout info
--drop-equivalent-buffer-results - Remove MemRef return values that are equivalent to a bbArg
--duplicate-function-elimination - Deduplicate functions
--eliminate-empty-tensors - Try to eliminate all tensor.empty ops.
--empty-tensor-to-alloc-tensor - Replace all empty ops by alloc_tensor ops.
--enable-arm-streaming - Enable Armv9 Streaming SVE mode
--if-required-by-ops - Only apply the selected streaming/ZA modes if the function contains ops that implement the ArmSMETileOpInterface.
--if-scalable-and-supported - Only apply the selected streaming/ZA modes if the function contains supported scalable vector operations.
--streaming-mode=<value> - Select how streaming-mode is managed at the function-level.
=disabled - Streaming mode is disabled.
=streaming - Streaming mode is part of the function interface (ABI), caller manages PSTATE.SM on entry/exit.
=streaming-locally - Streaming mode is internal to the function, callee manages PSTATE.SM on entry/exit.
=streaming-compatible - Function supports both streaming and non-streaming modes.
--za-mode=<value> - Select how ZA-storage is managed at the function-level.
=disabled - ZA storage is disabled.
=new-za - The function has ZA state. The ZA state is created on entry and destroyed on exit.
=in-za - The function uses ZA state. The ZA state may be used for input.
=out-za - The function uses ZA state. The ZA state may be used for output.
=inout-za - The function uses ZA state. The ZA state may be used for input and/or output.
=preserves-za - The function shares ZA state. The ZA state may not be used for input and/or output and the function must return with ZA unchanged
--ensure-debug-info-scope-on-llvm-func - Materialize LLVM debug info subprogram attribute on every LLVMFuncOp
--emission-kind=<value> - Emission kind to generate debug info.
=None - None
=Full - Full
=LineTablesOnly - LineTablesOnly (default)
=DebugDirectivesOnly - DebugDirectivesOnly
--expand-realloc - Expand memref.realloc operations into its components
--emit-deallocs - Emit deallocation operations for the original MemRef
--expand-strided-metadata - Expand memref operations into easier to analyze constructs
--finalize-memref-to-llvm - Finalize MemRef dialect to LLVM dialect conversion
--index-bitwidth=<uint> - Bitwidth of the index type, 0 to use size of machine word
--use-aligned-alloc - Use aligned_alloc in place of malloc for heap allocations
--use-generic-functions - Use generic allocation and deallocation functions instead of the classic 'malloc', 'aligned_alloc' and 'free' functions
--flatten-memref - Flatten a multiple dimensional memref to 1-dimensional
--fold-memref-alias-ops - Fold memref alias ops into consumer load/store ops
--fold-tensor-subset-ops - Fold tensor subset ops into producer/consumer ops
--form-expressions - Form C-style expressions from C-operator ops
--generate-runtime-verification - Generate additional runtime op verification checks
--verbose-level=<uint> - Verbosity level for runtime verification messages: 0 = Minimum (only source location), 1 = Detailed (include full operation details, names, types, shapes, etc.)
--gpu-async-region - Make GPU ops async
--gpu-decompose-memrefs - Decomposes memref index computation into explicit ops.
--gpu-eliminate-barriers - Erase unnecessary barriers
--gpu-kernel-outlining - Outline gpu.launch bodies to kernel functions
--data-layout-str=<string> - String description of the data layout
--gpu-launch-sink-index-computations - Sink index computations into gpu.launch body
--gpu-map-parallel-loops - Greedily maps loops to GPU hardware dimensions.
--mapping-policy=<string> - Policy outlining how to assign loops to GPU dimensions.Supported values are `outermost-first` and `innermost-first`.
--gpu-module-to-binary - Transforms a GPU module into a GPU binary.
--format=<string> - The target representation of the compilation process.
-l <string> - Extra files to link to.
--opts=<string> - Command line options to pass to the tools.
--section=<string> - ELF section where binary is to be located.
--toolkit=<string> - Toolkit path.
--gpu-to-llvm - Convert GPU dialect to LLVM dialect with GPU runtime calls
--intersperse-sizes-for-kernels - Inserts a size_t argument following each memref argument, containing the static size in bytes of the buffer. Incompatible arguments are rejected. This is intended for use by the Vulkan runtime with the kernel bare pointer calling convention, to enable dynamic binding of buffers as arguments without static type info.
--use-bare-pointers-for-host - Use bare pointers to pass memref arguments to host functions. All memrefs must have static shape.
--use-bare-pointers-for-kernels - Use bare pointers to pass memref arguments to kernels. The kernel must use the same setting for this option.
--inline - Inline function calls
--default-pipeline=<string> - The optimizer pipeline used for callables that do not have a dedicated optimizer pipeline in opPipelineList
--inlining-threshold=<uint> - If the ratio between the number of the operations in the callee and the number of the operations in the caller exceeds this value (in percentage), then the callee is not inlined even if it is legal to inline it
--max-iterations=<uint> - Maximum number of iterations when inlining within an SCC
--op-pipelines=<pass-manager> - Callable operation specific optimizer pipelines (in the form of `dialect.op(pipeline)`)
--int-range-optimizations - Do optimizations based on integer range analysis
--lift-cf-to-scf - Lift ControlFlow dialect to SCF dialect
--linalg-block-pack-matmul - Convert linalg matmul ops to block layout and back
--allow-padding - Allow packing padding
--block-factors=<long> - Block factors (mb, nb, kb) for relayout
--lhs-transpose-inner-blocks - Transpose LHS inner block layout [mb][kb] -> [kb][mb]
--lhs-transpose-outer-blocks - Transpose LHS outer block layout [MB][KB] -> [KB][MB]
--mnk-order=<long> - Permutation of matmul (M, N, K) dimensions order
--mnk-padded-multiples=<long> - Next multiples of the packing sizes
--rhs-transpose-inner-blocks - Transpose RHS inner block layout [kb][nb] -> [nb][kb]
--rhs-transpose-outer-blocks - Transpose RHS outer block layout [KB][NB] -> [NB][KB]
--linalg-detensorize - Detensorize linalg ops
--aggressive-mode - Detensorize all ops that qualify for detensoring along with branch operands and basic-block arguments.
--linalg-fold-into-elementwise - Fold transpose ops into elementwise
--linalg-fold-unit-extent-dims - Remove unit-extent dimension in Linalg ops on tensors
--use-rank-reducing-slices - Generate rank-reducing slices instead of reassociative reshapes
--linalg-fuse-elementwise-ops - Fuse elementwise operations on tensors
--linalg-generalize-named-ops - Convert named ops into generic ops
--linalg-inline-scalar-operands - Inline scalar operands into linalg generic ops
--linalg-morph-ops - Convert linalg ops between forms
--category-to-generic - convert category ops e.g. `linalg.elementwise` to `linalg.generic`
--generic-to-named - convert linalg.generic to equivalent named ops
--named-to-category - convert named ops to category op e.g. `linalg.elementwise`
--named-to-generic - convert named ops e.g. `linalg.add` to `linalg.generic`
--linalg-specialize-generic-ops - Convert generic ops back to named ops
--llvm-add-comdats - Add comdats to linkonce and linkonce_odr functions
--llvm-legalize-for-export - Legalize LLVM dialect to be convertible to LLVM IR
--llvm-optimize-for-nvvm-target - Optimize NVVM IR
--llvm-request-c-wrappers - Request C wrapper emission for all functions
--llvm-target-to-data-layout - Derive data layout attributes from LLVM target attributes
--initialize-llvm-targets - Whether to pre-load all available target machines, that LLVM is configured to support, into the TargetRegistry.
--llvm-target-to-target-features - Update attached #llvm.target's features per the described target
--initialize-llvm-targets - Whether to pre-load all available target machines, that LLVM is configured to support, into the TargetRegistry.
--llvm-use-default-visibility - Update default visibility of all global values and function definitions
--visibility=<value> - Visibility to use in place of default.
=hidden - Hidden
=protected - Protected
--loop-invariant-code-motion - Hoist loop invariant instructions outside of the loop
--loop-invariant-subset-hoisting - Hoist loop invariant subset ops outside of the loop
--lower-affine - Lower Affine operations to a combination of Arith and SCF operations
--lower-host-to-llvm - Lowers the host module code and `gpu.launch_func` to LLVM
--lower-quant-ops - Lower quant.dcast and quant.qcast ops
--lower-sparse-foreach-to-scf - Decompose a complex sparse operation into multiple stages
--lower-sparse-iteration-to-scf - lower sparse_tensor.iterate/coiterate into scf loops
--lower-sparse-ops-to-foreach - Applies sparse tensor rewriting rules after sparsification
--enable-convert - Enable rewriting rules for the convert operator
--enable-runtime-library - Enable runtime library for manipulating sparse tensors
--lower-vector-mask - Lower 'vector.mask' operations
--lower-vector-multi-reduction - Lower 'vector.multi_reduction' operations
--lowering-strategy=<value> - Select the strategy to control how multi_reduction is lowered.
=inner-parallel - Lower multi_reduction into outer-reduction and inner-parallel ops.
=inner-reduction - Lower multi_reduction into outer-parallel and inner-reduction ops.
--lower-vector-to-from-elements-to-shuffle-tree - Lower `vector.to_elements` and `vector.from_elements` to a tree of `vector.shuffle` operations
--map-memref-spirv-storage-class - Map numeric MemRef memory spaces to SPIR-V storage classes
--client-api=<string> - The client API to use for populating mappings
--math-expand-ops - Expand math operations.
--ops=<string> - Operations to expand.
--math-extend-to-supported-types - Legalize floating-point math ops on low-precision floats
--extra-types=<string> - MLIR types with arithmetic support on a given target (f64 and f32 are implicitly supported)
--target-type=<string> - MLIR type to convert the unsupported source types to
--math-sincos-fusion - Fuse sin and cos operations.
--math-uplift-to-fma - Uplift arith ops to math.fma.
--mem2reg - Promotes memory slots into values.
--region-simplify - Perform control flow optimizations to the region tree
--memref-emulate-wide-int - Emulate 2*N-bit integer operations using N-bit operations
--widest-int-supported=<uint> - Widest integer type supported by the target
--memref-expand - Legalize memref operations to be convertible to LLVM.
--mlprogram-pipeline-globals - Optimize `ml_program` global operations for read and store
--normalize-memrefs - Normalize memrefs
--normalize-quant-types - Normalize generic quantized types to specific quantized types
--nvgpu-optimize-shared-memory - Optimizes accesses to shard memory memrefs in order to reduce bank conflicts.
--nvvm-attach-target - Attaches an NVVM target attribute to a GPU Module.
-O <uint> - Optimization level.
--chip=<string> - Target chip.
--fast - Enable fast math mode.
--features=<string> - Target features, default is to use the minimum PTX ISA version for the SM target.
--ftz - Enable flush to zero for denormals.
-l <string> - Extra bitcode libraries paths to link to.
--module=<string> - Regex used to identify the modules to attach the target to.
--ptxas-cmd-options=<string> - Command line options passed to downstream compiler
--triple=<string> - Target triple.
--offload-livein-value-canonicalization - Canonicalize live-in values for regions destined for outlining
--omp-offload-privatization-prepare - Prepare OpenMP maps for privatization for deferred target tasks
--one-shot-bufferize - One-Shot Bufferize
--allow-return-allocs-from-loops - Allows returning/yielding new allocations from a loop.
--allow-unknown-ops - Allows unknown (not bufferizable) ops in the input IR.
--analysis-fuzzer-seed=<uint> - Test only: Analyze ops in random order with a given seed (fuzzer)
--analysis-heuristic=<string> - Heuristic that control the IR traversal during analysis
--buffer-alignment=<ulong> - Sets the alignment of newly allocated buffers.
--bufferize-function-boundaries - Bufferize function boundaries (experimental).
--check-parallel-regions - Account for parallel regions in RaW analysis.
--copy-before-write - Skip the analysis. Make a buffer copy on every write.
--dialect-filter=<string> - Restrict bufferization to ops from these dialects.
--dump-alias-sets - Test only: Annotate tensor IR with alias sets
--function-boundary-type-conversion=<value> - Controls layout maps when bufferizing function signatures.
=infer-layout-map
=identity-layout-map
=fully-dynamic-layout-map
--must-infer-memory-space - The memory space of an memref types must always be inferred. If unset, a default memory space of 0 is used otherwise.
--no-analysis-func-filter=<string> - Skip analysis of functions with these symbol names.Set copyBeforeWrite to true when bufferizing them.
--print-conflicts - Test only: Annotate IR with RaW conflicts. Requires test-analysis-only.
--test-analysis-only - Test only: Only run inplaceability analysis and annotate IR
--unknown-type-conversion=<value> - Controls layout maps for non-inferrable memref types.
=infer-layout-map
=identity-layout-map
=fully-dynamic-layout-map
--use-encoding-for-memory-space - Use the Tensor encoding attribute for the memory space. Exclusive to the 'must-infer-memory-space' option
--openacc-legalize-data-values - Legalizes SSA values in compute regions with results from data clause operations
--apply-to-acc-data-construct - Replaces varPtr uses with accPtr for acc compute regions contained within acc.data or acc.declare region.
--host-to-device - Replace varPtr uses with accPtr if true. Replace accPtr uses with varPtr if false
--optimize-allocation-liveness - This pass optimizes the liveness of temp allocations in the input function
--outline-shape-computation - Using shape.func to preserve shape computation
--ownership-based-buffer-deallocation - Adds all required dealloc operations for all allocations in the input program
--private-function-dynamic-ownership - Allows to add additional arguments to private functions to dynamically pass ownership of memrefs to callees. This can enable earlier deallocations.
--pre-sparsification-rewrite - Applies sparse tensor rewriting rules prior to sparsification
--print-ir - Print IR on the debug stream
--label=<string> - Label
--print-op-stats - Print statistics of operations
--json - print the stats as JSON
--promote-buffers-to-stack - Promotes heap-based allocations to automatically managed stack-based allocations
--max-alloc-size-in-bytes=<uint> - Maximal size in bytes to promote allocations to stack.
--max-rank-of-allocated-memref=<uint> - Maximal memref rank to promote dynamic buffers.
--reconcile-unrealized-casts - Simplify and eliminate unrealized conversion casts
--reify-result-shapes - Reifies the results of `tensor::PadOp` and `tensor::ConcatOp`.
--remove-dead-values - Remove dead values
--canonicalize - Canonicalize region branch ops
--remove-shape-constraints - Replace all cstr_ ops with a true witness
--resolve-ranked-shaped-type-result-dims - Resolve memref.dim of result values of ranked shape type
--error-on-pattern-iteration-limit - Throw an error when pattern rewriter hits iteration limit
--resolve-shaped-type-result-dims - Resolve memref.dim of result values
--error-on-pattern-iteration-limit - Throw an error when pattern rewriter hits iteration limit
--rocdl-attach-target - Attaches a ROCDL target attribute to a GPU Module.
-O <uint> - Optimization level.
--abi=<string> - ABI version.
--chip=<string> - Target chip.
--correct-sqrt - Enable correct rounded sqrt.
--daz - Enable denormals are zero opt.
--fast - Enable fast relaxed math opt.
--features=<string> - Target features.
--finite-only - Enable finite only opt.
-l <string> - Extra bitcode libraries paths to link to.
--module=<string> - Regex used to identify the modules to attach the target to.
--triple=<string> - Target triple.
--unsafe-math - Enable unsafe math opt.
--wave64 - Use Wave64 mode.
--sccp - Sparse Conditional Constant Propagation
--scf-for-loop-canonicalization - Canonicalize operations within scf.for loop bodies
--scf-for-loop-peeling - Peel `for` loops at their upper bounds.
--peel-front - Peel the first iteration out of the loop.
--skip-partial - Do not peel loops inside of the last, partial iteration of another already peeled loop.
--scf-for-loop-range-folding - Fold add/mul ops into loop range
--scf-for-loop-specialization - Specialize `for` loops for vectorization
--scf-for-to-while - Convert SCF for loops to SCF while loops
--scf-forall-to-for - Convert SCF forall loops to SCF for loops
--scf-forall-to-parallel - Convert SCF forall loops to SCF parallel loops
--scf-parallel-for-to-nested-fors - Convert SCF parallel for loops to nested SCF for loops
--scf-parallel-loop-fusion - Fuse adjacent parallel loops
--scf-parallel-loop-specialization - Specialize parallel loops for vectorization
--scf-parallel-loop-tiling - Tile parallel loops
--no-min-max-bounds - Perform tiling with fixed upper bound with inbound check inside the internal loops
--parallel-loop-tile-sizes=<long> - Factors to tile parallel loops by
--set-llvm-module-datalayout - Attach a datalayout string as a module attribute
--data-layout=<string> - String description (LLVM format) of the data layout that is expected on the produced module
--shape-to-shape-lowering - Legalize Shape dialect to be convertible to Arith
--shard-partition - Partition a function into SPMD form.
--sharding-propagation - sharding propagation
--traversal=<value> - Traversal order to use for sharding propagation:
=forward - Forward only traversal.
=backward - backward only traversal.
=forward-backward - forward-backward traversal.
=backward-forward - backward-forward traversal.
--simplify-depthwise-conv - Simplify depthwise convolution.
--snapshot-op-locations - Generate new locations from the current IR
--filename=<string> - The filename to print the generated IR
--pretty-debuginfo - Print pretty debug info in MLIR output
--print-debuginfo - Print debug info in MLIR output
--print-local-scope - Print with local scope and inline information (eliding aliases for attributes, types, and locations
--print-op-generic - Print the generic op form
--tag=<string> - A tag to use when fusing the new locations with the original. If unset, the locations are replaced.
--sparse-assembler - Add [dis]assemble operations on external sparse tensors
--direct-out - Directly returns buffers externally
--sparse-buffer-rewrite - Rewrite sparse primitives on buffers to actual code
--enable-buffer-initialization - Enable zero-initialization of the memory buffers
--sparse-gpu-codegen - Generates GPU code during sparsification
--enable-runtime-library - Enable runtime library for manipulating sparse tensors
--num-threads=<int> - Sets the number of GPU threads
--sparse-reinterpret-map - Reinterprets sparse tensor type mappings
--loop-ordering-strategy=<value> - Set the loop ordering strategy for sparse code generation
=default - Default strategy (eagerly selects last loop in topological sort)
=dense-outer - Prefer dense, then compressed, then singleton dimensions outermost
--scope=<value> - Set the reiterpretation scope
=all - Run on every applicable operations.
=only-generic - Run only on linalg.generic operations.
=except-generic - Run on operations expect linalg.generic (e.g., foreach)
--sparse-space-collapse - sparse space collapsing pass
--sparse-storage-specifier-to-llvm - Lower sparse storage specifer to llvm structure
--sparse-tensor-codegen - Convert sparse tensors and primitives to actual code
--create-sparse-deallocs - Specify if the temporary buffers created by the sparse compiler should be deallocated. For compatibility with core bufferization passes. This option is only used when enable-runtime-library=false. See also create-deallocs for BufferizationOption.
--enable-buffer-initialization - Enable zero-initialization of the memory buffers
--sparse-tensor-conversion - Convert sparse tensors and primitives to library calls
--sparse-vectorization - Vectorizes loops after sparsification
--enable-simd-index32 - Enable i32 indexing into vectors (for efficient gather/scatter)
--enable-vla-vectorization - Enable vector length agnostic vectorization
--vl=<int> - Set the vector length (use 0 to disable vectorization)
--sparsification - Automatically generate sparse tensor code from sparse tensor types
--enable-runtime-library - Enable runtime library for manipulating sparse tensors
--parallelization-strategy=<value> - Set the parallelization strategy
=none - Turn off sparse parallelization.
=dense-outer-loop - Enable dense outer loop sparse parallelization.
=any-storage-outer-loop - Enable sparse parallelization regardless of storage for the outer loop.
=dense-any-loop - Enable dense parallelization for any loop.
=any-storage-any-loop - Enable sparse parallelization for any storage and loop.
--sparse-emit-strategy=<value> - Emit functional code or interfaces (to debug) for sparse loops
=functional - Emit functional code (with scf.for/while).
=sparse-iterator - Emit (experimental) loops (with sparse.iterate).
=debug-interface - Emit non-functional but easy-to-read interfaces to debug.
--sparsification-and-bufferization - Mini-pipeline that combines bufferization and sparsifiation
--enable-gpu-libgen - Enable GPU acceleration by means of direct library calls
--enable-simd-index32 - Enable i32 indexing into vectors (for efficient gather/scatter)
--enable-vla-vectorization - Enable vector length agnostic vectorization
--parallelization-strategy=<value> - Set the parallelization strategy
=none - Turn off sparse parallelization.
=dense-outer-loop - Enable dense outer loop sparse parallelization.
=any-storage-outer-loop - Enable sparse parallelization regardless of storage for the outer loop.
=dense-any-loop - Enable dense parallelization for any loop.
=any-storage-any-loop - Enable sparse parallelization for any storage and loop.
--sparse-emit-strategy=<value> - Emit functional code or interfaces (to debug) for sparse loops
=functional - Emit functional code (with scf.for/while).
=sparse-iterator - Emit (experimental) loops (with sparse.iterate).
=debug-interface - Emit non-functional but easy-to-read interfaces to debug.
--vl=<int> - Set the vector length (use 0 to disable vectorization)
--spirv-attach-target - Attaches an SPIR-V target attribute to a GPU Module.
--caps=<string> - List of supported SPIR-V Capabilities
--client_api=<string> - Client API
--device_id=<uint> - Device ID
--device_type=<string> - Device Type
--exts=<string> - List of supported SPIR-V Extensions
--module=<string> - Regex used to identify the modules to attach the target to.
--vendor=<string> - Device Vendor
--ver=<string> - SPIR-V Version.
--spirv-canonicalize-gl - Canonicalize GLSL ops
--spirv-lower-abi-attrs - Decorate SPIR-V composite type with layout info
--spirv-promote-to-replicated-constants - Convert splat composite constants and spec constants to corresponding replicated constant composite ops defined by SPV_EXT_replicated_composites
--spirv-rewrite-inserts - Rewrite sequential chains of `spirv.CompositeInsert` operations into `spirv.CompositeConstruct` operations
--spirv-unify-aliased-resource - Unify access of multiple aliased resources into access of one single resource
--spirv-update-vce - Deduce and attach minimal (version, capabilities, extensions) requirements to spirv.module ops
--spirv-webgpu-prepare - Prepare SPIR-V to target WebGPU by expanding unsupported ops and replacing with supported ones
--sroa - Scalar Replacement of Aggregates
--stage-sparse-ops - Decompose a complex sparse operation into multiple stages
--strip-debuginfo - Strip debug info from all operations
--strip-func-quant-types - Strip quantized types from function headers
--symbol-dce - Eliminate dead symbols
--symbol-privatize - Mark symbols private
--exclude=<string> - Comma separated list of symbols that should not be marked private
--test-arm-sme-tile-allocation - Tests SME 'virtual tile' allocation
--dump-tile-live-ranges - Dump the live ranges of SME tiles (for debugging)
--preprocess-only - Only preprocess IR so it is ready for tile allocation (but do not allocate any tiles)
--test-scf-parallel-loop-collapsing - Test parallel loops collapsing transformation
--collapsed-indices-0=<uint> - Which loop indices to combine 0th loop index
--collapsed-indices-1=<uint> - Which loop indices to combine into the position 1 loop index
--collapsed-indices-2=<uint> - Which loop indices to combine into the position 2 loop index
--topological-sort - Sort regions without SSA dominance in topological order
--tosa-arith-const-to-tosa-const - Convert tensor arith.constant operations into tosa.const
--tosa-attach-target - Attach tosa.target_env information to the given module.
--extensions=<string> - The TOSA extension(s) that operators should conform to. TOSA profile extensions define optional operation and data type combinations.
--level=<value> - The TOSA level that operators should conform to. A TOSA level defines operator argument ranges that an implementation shall support.
=8k - Ranges are expected to be sufficient for applications with frame sizes up to 8K.
=none - Allows the full range of arguments specified by the operations according to the operation data types.
--profiles=<string> - The TOSA profile(s) that operators should conform to. TOSA profiles enable efficient implementation on different classes of device. Each profile is an independent set of operations and data type combinations.
--specification_version=<value> - The specification version that TOSA operators should conform to.
=1.0 - TOSA Specification version 1.0
=1.1.draft - TOSA Specification version 1.1.draft
--tosa-convert-integer-type-to-signless - Convert integer types to signless
--tosa-infer-shapes - Propagate shapes across TOSA operations
--tosa-layerwise-constant-fold - Fold layerwise operations on constant tensors
--aggressive-reduce-constant - Always perform the reduce constant optimizationMay add more tosa.const but would reduce runtime calculations
--tosa-make-broadcastable - TOSA rank Reshape to enable Broadcasting
--tosa-narrow-f64-to-f32 - Narrow F64 TOSA operations to F32
--aggressive-rewrite - If enabled, all TOSA operations are rewritten, regardless or whether the narrowingis safe. This option may lead to data loss if not used carefully.
--convert-function-boundaries - If enabled, the pass will convert function I/O types as well. Otherwise casts willbe inserted at the I/O boundaries.
--tosa-narrow-i64-to-i32 - Narrow I64 TOSA operations to I32
--aggressive-rewrite - If enabled, all TOSA operations are rewritten, regardless or whether the narrowingis safe. This option may lead to data loss if not used carefully.
--convert-function-boundaries - If enabled, the pass will convert function I/O types as well. Otherwise casts willbe inserted at the I/O boundaries.
--tosa-optional-decompositions - Applies Tosa operations optional decompositions
--tosa-reduce-transposes - Reduce transposes through other operators
--tosa-to-arith - Lower TOSA to the Arith dialect
--include-apply-rescale - Whether to include the lowering for tosa.apply_rescale to arith
--use-32-bit - Whether to prioritze lowering to 32-bit operations
--tosa-to-linalg - Lower TOSA to LinAlg on tensors
--aggressive-reduce-constant - Always perform the reduce constant optimization
--disable-tosa-decompositions - Disable tosa decompositions pass
--tosa-to-linalg-named - Lower TOSA to LinAlg named operations
--prefer-conv2d-kernel-layout-hwcf - Prefer generating linalg.conv_2d_nhwc_hwcf over linalg.conv_2d_nhwc_fhwc
--tosa-to-mlprogram - Lower TOSA to the MLProgram dialect
--tosa-to-scf - Lower TOSA to the SCF dialect
--tosa-to-tensor - Lower TOSA to the Tensor dialect
--tosa-validate - Validates TOSA dialect
--allow-invalid-op-datatype-combinations - Disable checks for operations that are determined to be invalid due to their operand/result datatypes not aligning with the 'Supported Data Types' sections of the specifciation
--strict-op-spec-alignment - Verify if the properties of certain operations align the spec requirement
--transform-dialect-check-uses - warn about potential use-after-free in the transform dialect
--transform-infer-effects - infer transform side effects for symbols
--transform-interpreter - transform dialect interpreter
--debug-bind-trailing-args=<string> - Binds trailing arguments of the entry point to the payload operations with specified names.
--debug-payload-root-tag=<string> - Select the operation with 'transform.target_tag' attribute having the given value as payload IR root. If empty select the pass anchor operation as the payload IR root.
--disable-expensive-checks - Disable expensive checks in the interpreter for a faster run.
--entry-point=<string> - Entry point of the pass pipeline.
--transform-preload-library - preload transform dialect library
--transform-library-paths=<string> - Optional paths to files with modules that should be merged into the transform module to provide the definitions of external named sequences.
--view-op-graph - Print Graphviz visualization of an operation
--max-label-len=<uint> - Limit attribute/type length to number of chars
--print-attrs - Print attributes of operations
--print-control-flow-edges - Print control flow edges
--print-data-flow-edges - Print data flow edges
--print-result-types - Print result types of operations
--wrap-emitc-func-in-class - Wrap functions in classes, using arguments as fields.
--xegpu-blocking - Block XeGPU ops into smaller size.
--xegpu-fold-alias-ops - Fold alias ops into XeGPU ops
--xegpu-optimize-block-loads - Optimize XeGPU block load operations
--xegpu-propagate-layout - Propagate and assign XeGPU layout information
--layout-kind=<string> - Propagate `subgroup` / `inst` / `lane` level of xegpu layouts.
--print-analysis-only - Print the result of layout propagation analysis and exit.
--xegpu-subgroup-distribute - Distribute XeGPU ops to work items
--xegpu-vector-linearize - Linearize n-D vectors to 1-D vectors
--xegpu-wg-to-sg-distribute - Transform WorkGroup level XeGPU code to SubGroup level
--xevm-attach-target - Attaches a XeVM target attribute to a GPU Module.
-O <uint> - Optimization level.
--chip=<string> - Target chip.
--cmd-options=<string> - Command line options passed to downstream compiler
-l <string> - Extra bitcode libraries paths to link to.
--module=<string> - Regex used to identify the modules to attach the target to.
--triple=<string> - Target triple.
Pass Pipelines:
--buffer-deallocation-pipeline - The default pipeline for automatically inserting deallocation operations after one-shot bufferization. Deallocation operations (except `memref.realloc`) may not be present already.
--private-function-dynamic-ownership - Allows to add additional results to private functions to return ownership of returned memrefs to callers. This can avoid spurious buffer clones in the callee.
--gpu-lower-to-nvvm-pipeline - The default pipeline lowers main dialects (arith, memref, scf, vector, gpu, and nvgpu) to NVVM. It starts by lowering GPU code to the specified compilation target (default is fatbin) then lowers the host code.
--allow-pattern-rollback - Allow pattern rollback during dialect conversion
--cubin-chip=<string> - Chip to use to serialize to cubin.
--cubin-features=<string> - Features to use to serialize to cubin.
--cubin-format=<string> - Compilation format to use to serialize to cubin.
--cubin-triple=<string> - Triple to use to serialize to cubin.
--host-bare-ptr-calling-convention - Whether to use the bareptr calling convention on the host (warning this should be false until the GPU layering is fixed)
--index-bitwidth=<long> - Bitwidth of the index type for the host (warning this should be 64 until the GPU layering is fixed)
--kernel-bare-ptr-calling-convention - Whether to use the bareptr calling convention on the kernel (warning this should be false until the GPU layering is fixed)
--opt-level=<int> - Optimization level for NVVM compilation
--ptxas-cmd-options=<string> - Command line options to pass to the downstream compiler.
--gpu-lower-to-xevm-pipeline - The default GPU to XeVM lowering pipeline. It starts by lowering GPU code to the specified compilation target (default is fatbin) then lowers the host code.
--binary-format=<string> - Final GPU binary emission format (e.g. fatbin)
--host-bare-ptr-calling-convention - Use bare pointer calling convention for host launches
--igc-cmd-options=<string> - Additional downstream compiler command line options
--kernel-bare-ptr-calling-convention - Use bare pointer calling convention for device kernels
--opt-level=<uint> - Optimization level for attached target/codegen
--use-64bit-index - Bitwidth of the index type (host & device)
--xegpu-op-level=<string> - Granularity of XeGPU operations to target: workgroup | subgroup | lane
--xevm-module-matcher=<string> - Regex to match gpu.module names for XeVM target attach
--zebin-chip=<string> - Target chip (e.g. pvc, bmg)
--zebin-triple=<string> - Target triple for XeVM codegen
--sparsifier - The standard pipeline for taking sparsity-agnostic IR using the sparse-tensor type, and lowering it to LLVM IR with concrete representations and algorithms for sparse tensors.
--create-sparse-deallocs - Specify if the temporary buffers created by the sparse compiler should be deallocated. For compatibility with core bufferization passes. This option is only used when enable-runtime-library=false.
--enable-amx - Enables the use of AMX dialect while lowering the vector dialect
--enable-arm-neon - Enables the use of ArmNeon dialect while lowering the vector dialect
--enable-arm-sve - Enables the use of ArmSVE dialect while lowering the vector dialect
--enable-buffer-initialization - Enable zero-initialization of memory buffers
--enable-gpu-libgen - Enables GPU acceleration by means of direct library calls (like cuSPARSE)
--enable-index-optimizations - Allows compiler to assume indices fit in 32-bit if that yields faster code
--enable-runtime-library - Enable runtime library for manipulating sparse tensors
--enable-x86vector - Enables the use of X86Vector dialect while lowering the vector dialect
--gpu-chip=<string> - GPU target architecture
--gpu-features=<string> - GPU target features
--gpu-format=<string> - GPU compilation format
--gpu-triple=<string> - GPU target triple
--parallelization-strategy=<value> - Set the parallelization strategy
=none - Turn off sparse parallelization.
=dense-outer-loop - Enable dense outer loop sparse parallelization.
=any-storage-outer-loop - Enable sparse parallelization regardless of storage for the outer loop.
=dense-any-loop - Enable dense parallelization for any loop.
=any-storage-any-loop - Enable sparse parallelization for any storage and loop.
--reassociate-fp-reductions - Allows llvm to reassociate floating-point reductions for speed
--sparse-emit-strategy=<value> - Emit functional code or interfaces (to debug) for sparse loops
=functional - Emit functional code (with scf.for/while).
=sparse-iterator - Emit (experimental) loops (with sparse.iterate).
=debug-interface - Emit non-functional but easy-to-read interfaces to debug.
--test-bufferization-analysis-only - Run only the inplacability analysis
--vl=<int> - Set the vector length (0 disables vectorization)
--tosa-to-linalg-pipeline - The default pipeline for converting TOSA operators to the equivalent operations using the tensor operations in LinAlg as well as LinAlg named operations.
--cost-kind=<value> - Target cost kind
=throughput - Reciprocal throughput
=latency - Instruction latency
=code-size - Code size
=size-latency - Code size and latency
=all - Print all cost kinds
--crel - Use CREL relocation format for ELF
--cs-profile-generate - Perform context sensitive PGO instrumentation
--cs-profile-path=<string> - Context sensitive profile file path
--ctx-prof-include-empty - Also write profiles with all-zero counters. Intended for testing/debugging.
--ctx-profile-force-is-specialized - Treat the given module as-if it were containing the post-thinlink module containing the root
--debugify-atoms -
--debugify-func-limit=<ulong> - Set max number of processed functions per pass.
--debugify-level=<value> - Kind of debug info to add
=locations - Locations only
=location+variables - Locations and Variables
--debugify-quiet - Suppress verbose debugify output
--devirtualize-speculatively - Enable speculative devirtualization optimization
--disable-auto-upgrade-debug-info - Disable autoupgrade of debug info
--disable-i2p-p2i-opt - Disables inttoptr/ptrtoint roundtrip optimization
--disable-promote-alloca-to-lds - Disable promote alloca to LDS
--disable-promote-alloca-to-vector - Disable promote alloca to vector
--disable-qfp-opt - Disable optimization of Qfloat operations.
--disable-qfp-opt-mul - Disable optimization of Qfloat operations for multiply.
--do-counter-promotion - Do counter register promotion
--dot-cfg-mssa=<file name for generated dot file> - file name for generated dot file
--dump-pass-pipeline - Print the pipeline that will be run
--dwarf-version=<int> - Dwarf version
--dwarf64 - Generate debugging info in the 64-bit DWARF format
--elide-all-zero-branch-weights -
--elide-resource-data-from-bytecode - Elide resources when generating bytecode
--emit-bb-hash - Emit the hash of basic block in the SHT_LLVM_BB_ADDR_MAP section.
--emit-bytecode - Emit bytecode when generating output
--emit-bytecode-version=<value> - Use specified bytecode when generating output
--emit-compact-unwind-non-canonical - Whether to try to emit Compact Unwind for non canonical entries.
--emit-dwarf-unwind=<value> - Whether to emit DWARF EH frame entries.
=always - Always emit EH frame entries
=no-compact-unwind - Only emit EH frame entries when compact unwind is not available
=default - Use target platform default
--emit-gnuas-syntax-on-zos - Emit GNU Assembly Syntax on z/OS.
--emscripten-cxx-exceptions-allowed=<string> - The list of function names in which Emscripten-style exception handling is enabled (see emscripten EMSCRIPTEN_CATCHING_ALLOWED options)
--enable-cse-in-irtranslator - Should enable CSE in irtranslator
--enable-cse-in-legalizer - Should enable CSE in Legalizer
--enable-devirtualize-speculatively - Enable speculative devirtualization optimization
--enable-emscripten-cxx-exceptions - WebAssembly Emscripten-style exception handling
--enable-emscripten-sjlj - WebAssembly Emscripten-style setjmp/longjmp handling
--enable-gvn-hoist - Enable the GVN hoisting pass (default = off)
--enable-gvn-memdep -
--enable-gvn-memoryssa -
--enable-gvn-sink - Enable the GVN sinking pass (default = off)
--enable-jump-table-to-switch - Enable JumpTableToSwitch pass (default = off)
--enable-load-in-loop-pre -
--enable-load-pre -
--enable-loop-simplifycfg-term-folding -
--enable-name-compression - Enable name/filename string compression
--enable-poison-reuse-guard - Enable poison-reuse guard
--enable-split-backedge-in-load-pre -
--enable-split-loopiv-heuristic - Enable loop iv regalloc heuristic
--enable-vtable-profile-use - If ThinLTO and WPD is enabled and this option is true, vtable profiles will be used by ICP pass for more efficient indirect call sequence. If false, type profiles won't be used.
--enable-vtable-value-profiling - If true, the virtual table address will be instrumented to know the types of a C++ pointer. The information is used in indirect call promotion to do selective vtable-based comparison.
--expand-variadics-override=<value> - Override the behaviour of expand-variadics
=unspecified - Use the implementation defaults
=disable - Disable the pass entirely
=optimize - Optimise without changing ABI
=lowering - Change variadic calling convention
--experimental-debug-variable-locations - Use experimental new value-tracking variable locations
--fatal-warnings - Treat warnings as errors
--fdpic - Use the FDPIC ABI
--force-tail-folding-style=<value> - Force the tail folding style
=none - Disable tail folding
=data - Create lane mask for data only, using active.lane.mask intrinsic
=data-without-lane-mask - Create lane mask with compare/stepvector
=data-and-control - Create lane mask using active.lane.mask intrinsic, and use it for both data and control flow
=data-and-control-without-rt-check - Similar to data-and-control, but remove the runtime check
=data-with-evl - Use predicated EVL instructions for tail folding. If EVL is unsupported, fallback to data-without-lane-mask.
--fs-profile-debug-bw-threshold=<uint> - Only show debug message if the source branch weight is greater than this value.
--fs-profile-debug-prob-diff-threshold=<uint> - Only show debug message if the branch probability is greater than this value (in percentage).
--generate-merged-base-profiles - When generating nested context-sensitive profiles, always generate extra base profile for function with all its context profiles merged into it.
--gpsize=<uint> - Global Pointer Addressing Size. The default size is 8.
--gsframe - Whether to emit .sframe unwind sections.
--hash-based-counter-split - Rename counter variable of a comdat function based on cfg hash
--hexagon-add-build-attributes -
--hexagon-rdf-limit=<uint> -
--hot-cold-split - Enable hot-cold splitting pass
--hwasan-percentile-cutoff-hot=<int> - Hot percentile cutoff.
--hwasan-random-rate=<number> - Probability value in the range [0.0, 1.0] to keep instrumentation of a function. Note: instrumentation can be skipped randomly OR because of the hot percentile cutoff, if both are supplied.
--implicit-mapsyms - Allow mapping symbol at section beginning to be implicit, lowering number of mapping symbols at the expense of some portability. Recommended for projects that can build all their object files using this option
--import-all-index - Import all external functions in index.
--incremental-linker-compatible - When used with filetype=obj, emit an object file which can be used with an incremental linker
--instcombine-code-sinking - Enable code sinking
--instcombine-guard-widening-window=<uint> - How wide an instruction window to bypass looking for another guard
--instcombine-max-num-phis=<uint> - Maximum number phis to handle in intptr/ptrint folding
--instcombine-max-sink-users=<uint> - Maximum number of undroppable users for instruction sinking
--instcombine-maxarray-size=<uint> - Maximum array size considered when doing a combine
--instcombine-negator-enabled - Should we attempt to sink negations?
--instcombine-negator-max-depth=<uint> - What is the maximal lookup depth when trying to check for viability of negation sinking.
--instrprof-atomic-counter-update-all - Make all profile counter updates atomic (for testing only)
--internalize-public-api-file=<filename> - A file containing list of symbol names to preserve
--internalize-public-api-list=<list> - A list of symbol names to preserve
--intrinsic-cost-strategy=<value> - Costing strategy for intrinsic instructions
=instruction-cost - Use TargetTransformInfo::getInstructionCost
=intrinsic-cost - Use TargetTransformInfo::getIntrinsicInstrCost
=type-based-intrinsic-cost - Calculate the intrinsic cost based only on argument types
--irdl-file=<filename> - IRDL file to register before processing the input
--iterative-counter-promotion - Allow counter promotion across the whole loop nest.
--list-passes - Print the list of registered passes and exit
--load-dialect-plugin=<string> - Load dialects from plugin library
--load-pass-plugin=<string> - Load passes from plugin library
--log-actions-to=<string> - Log action execution to a file, or stderr if '-' is passed
--log-mlir-actions-filter=<string> - Comma separated list of locations to filter actions from logging
--loongarch-use-aa - Enable the use of AA during codegen.
--lower-allow-check-percentile-cutoff-hot=<int> - Hot percentile cutoff.
--lower-allow-check-random-rate=<number> - Probability value in the range [0.0, 1.0] of unconditional pseudo-random checks.
--lto-aix-system-assembler=<path> - Path to a system assembler, picked up on AIX only
--lto-embed-bitcode=<value> - Embed LLVM bitcode in object files produced by LTO
=none - Do not embed
=optimized - Embed after all optimization passes
=post-merge-pre-opt - Embed post merge, but before optimizations
--lto-pass-remarks-filter=<regex> - Only record optimization remarks from passes whose names match the given regular expression
--lto-pass-remarks-format=<format> - The format used for serializing remarks (default: YAML)
--lto-pass-remarks-output=<filename> - Output filename for pass remarks
--matrix-default-layout=<value> - Sets the default matrix layout
=column-major - Use column-major layout
=row-major - Use row-major layout
--matrix-print-after-transpose-opt -
--max-counter-promotions=<int> - Max number of allowed counter promotions
--max-counter-promotions-per-loop=<uint> - Max number counter promotions per loop to avoid increasing register pressure too much
--mc-relax-all - When used with filetype=obj, relax all fixups in the emitted object file
--mcabac - tbd
--merror-missing-parenthesis - Error for missing parenthesis around predicate registers
--merror-noncontigious-register - Error for register names that aren't contigious
--mhvx - Enable Hexagon Vector eXtensions
--mhvx=<value> - Enable Hexagon Vector eXtensions
=v60 - Build for HVX v60
=v62 - Build for HVX v62
=v65 - Build for HVX v65
=v66 - Build for HVX v66
=v67 - Build for HVX v67
=v68 - Build for HVX v68
=v69 - Build for HVX v69
=v71 - Build for HVX v71
=v73 - Build for HVX v73
=v75 - Build for HVX v75
=v79 - Build for HVX v79
=v81 - Build for HVX v81
--mips-compact-branches=<value> - MIPS Specific: Compact branch policy.
=never - Do not use compact branches if possible.
=optimal - Use compact branches where appropriate (default).
=always - Always use compact branches if possible.
--mips16-constant-islands - Enable mips16 constant islands.
--mips16-hard-float - Enable mips16 hard float.
--mir-strip-debugify-only - Should mir-strip-debug only strip debug info from debugified modules by default
--misexpect-tolerance=<uint> - Prevents emitting diagnostics when profile counts are within N% of the threshold..
--mlir-debug-counter=<string> - Comma separated list of debug counter skip and count arguments
--mlir-diagnostic-verbosity-level=<value> - Choose level of diagnostic information
=errors - Errors only
=warnings - Errors and warnings
=remarks - Errors, warnings and remarks
--mlir-disable-diagnostic-notes - Disable diagnostic notes.
--mlir-disable-threading - Disable multi-threading within MLIR, overrides any further call to MLIRContext::enableMultiThreading()
--mlir-elide-elementsattrs-if-larger=<uint> - Elide ElementsAttrs with "..." that have more elements than the given upper limit
--mlir-elide-resource-strings-if-larger=<uint> - Elide printing value of resources if string is too long in chars.
--mlir-enable-debugger-hook - Enable Debugger hook for debugging MLIR Actions
--mlir-generate-reproducer=<filename> - Generate an mlir reproducer at the provided filename (no crash required)
--mlir-output-format=<value> - Output format for timing data
=text - display the results in text format
=json - display the results in JSON format
--mlir-pass-pipeline-crash-reproducer=<string> - Generate a .mlir reproducer file at the given output path if the pass manager crashes or fails
--mlir-pass-pipeline-local-reproducer - When generating a crash reproducer, attempt to generated a reproducer with the smallest pipeline.
--mlir-pass-statistics - Display the statistics of each pass
--mlir-pass-statistics-display=<value> - Display method for pass statistics
=list - display the results in a merged list sorted by pass name
=pipeline - display the results with a nested pipeline view
--mlir-pretty-debuginfo - Print pretty debug info in MLIR output
--mlir-print-debug-counter - Print out debug counter information after all counters have been accumulated
--mlir-print-debuginfo - Print debug info in MLIR output
--mlir-print-elementsattrs-with-hex-if-larger=<long> - Print DenseElementsAttrs with a hex string that have more elements than the given upper limit (use -1 to disable)
--mlir-print-ir-after=<pass-arg> - Print IR after specified passes
--mlir-print-ir-after-all - Print IR after each pass
--mlir-print-ir-after-change - When printing the IR after a pass, only print if the IR changed
--mlir-print-ir-after-failure - When printing the IR after a pass, only print if the pass failed
--mlir-print-ir-before=<pass-arg> - Print IR before specified passes
--mlir-print-ir-before-all - Print IR before each pass
--mlir-print-ir-module-scope - When printing IR for print-ir-[before|after]{-all} always print the top-level operation
--mlir-print-ir-tree-dir=<string> - When printing the IR before/after a pass, print file tree rooted at this directory. Use in conjunction with mlir-print-ir-* flags
--mlir-print-local-scope - Print with local scope and inline information (eliding aliases for attributes, types, and locations)
--mlir-print-op-on-diagnostic - When a diagnostic is emitted on an operation, also print the operation as an attached note
--mlir-print-skip-regions - Skip regions when printing ops.
--mlir-print-stacktrace-on-diagnostic - When a diagnostic is emitted, also print the stack trace as an attached note
--mlir-print-unique-ssa-ids - Print unique SSA ID numbers for values, block arguments and naming conflicts across all regions
--mlir-print-value-users - Print users of operation results and block arguments as a comment
--mlir-timing - Display execution times
--mlir-timing-display=<value> - Display method for timing data
=list - display the results in a list sorted by total time
=tree - display the results ina with a nested tree view
--mlir-use-nameloc-as-prefix - Print SSA IDs using NameLocs as prefixes
--mlir-very-unsafe-disable-verifier-on-parsing - Disable the verifier on parsing (very unsafe)
--mno-compound - Disable looking for compound instructions for Hexagon
--mno-fixup - Disable fixing up resolved relocations for Hexagon
--mno-ldc1-sdc1 - Expand double precision loads and stores to their single precision counterparts
--mno-pairing - Disable looking for duplex instructions for Hexagon
--ms-secure-hotpatch-functions-file=<filename> - A file containing list of mangled function names to mark for Windows Secure Hot-Patching
--ms-secure-hotpatch-functions-list=<list> - A list of mangled function names to mark for Windows Secure Hot-Patching
--mwarn-missing-parenthesis - Warn for missing parenthesis around predicate registers
--mwarn-noncontigious-register - Warn for register names that aren't contigious
--mwarn-sign-mismatch - Warn for mismatching a signed and unsigned value
--no-deprecated-warn - Suppress all deprecated warnings
--no-discriminators - Disable generation of discriminator information.
--no-implicit-module - Disable implicit addition of a top-level module op during parsing
--no-type-check - Suppress type errors (Wasm)
--no-warn - Suppress all warnings
--nvptx-approx-log2f32 - NVPTX Specific: whether to use lg2.approx for log2
--nvptx-sched4reg - NVPTX Specific: schedule for register pressue
-o <filename> - Output filename
--object-size-offset-visitor-max-visit-instructions=<uint> - Maximum number of instructions for ObjectSizeOffsetVisitor to look at
--output-split-marker=<string> - Split marker to use for merging the ouput
--pass-pipeline=<string> - Textual description of the pass pipeline to run
--pgo-block-coverage - Use this option to enable basic block coverage instrumentation
--pgo-temporal-instrumentation - Use this option to enable temporal instrumentation
--pgo-view-block-coverage-graph - Create a dot file of CFGs with block coverage inference information
--print-pipeline-passes - Print a '-passes' compatible string describing the pipeline (best-effort only).
--profcheck-annotate-select - Also inject (if missing) and verify MD_prof for `select` instructions
--profcheck-default-function-entry-count=<long> -
--profcheck-default-select-false-weight=<uint> - When annotating `select` instructions, this value will be used for the second ('false') case.
--profcheck-default-select-true-weight=<uint> - When annotating `select` instructions, this value will be used for the first ('true') case.
--profcheck-weights-for-test - Generate weights with small values for tests.
--profile-actions-to=<string> - Profile action execution to a file, or stderr if '-' is passed
--profile-correlate=<value> - Use debug info or binary file to correlate profiles.
=<empty> - No profile correlation
=debug-info - Use debug info to correlate
=binary - Use binary to correlate
--promote-alloca-vector-loop-user-weight=<uint> - The bonus weight of users of allocas within loop when sorting profitable allocas
--propeller-infer-threshold=<number> - Threshold for infer stale profile
--r600-ir-structurize - Use StructurizeCFG IR pass
--riscv-add-build-attributes -
--riscv-use-aa - Enable the use of AA during codegen.
[truncated]
Options (600)
| Option | Description | Source |
|---|---|---|
--color | - Use colors in output (default=autodetect) | help |
--aarch64-neon-syntax value | - Choose style of NEON code to emit from AArch64 backend: =generic - Emit generic NEON assembly =apple - Emit Apple-style NEON assembly | help |
--aarch64-use-aa | - Enable the use of AA during codegen. | help |
--abort-on-max-devirt-iterations-reached | - Abort when the max iterations for devirtualization CGSCC repeat pass is reached | help |
--allow-ginsert-as-artifact | - Allow G_INSERT to be considered an artifact. | help |
--allow-unregistered-dialect | - Allow operation with no registered dialects | help |
--amdgpu-atomic-optimizer-strategy value | - Select DPP or Iterative strategy for scan =DPP - Use DPP operations for scan =Iterative - Use Iterative approach for scan =None - Disable atomic optimizer | help |
--amdgpu-bypass-slow-div | - Skip 64-bit divide for dynamic 32-bit values | help |
--amdgpu-disable-loop-alignment | - Do not align and prefetch loops | help |
--amdgpu-dpp-combine | - Enable DPP combiner | help |
--amdgpu-dump-hsa-metadata | - Dump AMDGPU HSA Metadata | help |
--amdgpu-enable-merge-m0 | - Merge and hoist M0 initializations | help |
--amdgpu-indirect-call-specialization-threshold uint | - A threshold controls whether an indirect call will be specialized | help |
--amdgpu-kernarg-preload | - Enable preload kernel arguments to SGPRs | help |
--amdgpu-kernarg-preload-count uint | - How many kernel arguments to preload onto SGPRs | help |
--amdgpu-module-splitting-max-depth uint | - maximum search depth. | help |
--amdgpu-promote-alloca-to-vector-limit uint | - Maximum byte size to consider promote alloca to vector | help |
--amdgpu-promote-alloca-to-vector-max-regs uint | - Maximum vector size (in 32b registers) to use when promoting alloca | help |
--amdgpu-promote-alloca-to-vector-vgpr-ratio uint | - Ratio of VGPRs to budget for promoting alloca to vectors | help |
--amdgpu-sdwa-peephole | - Enable SDWA peepholer | help |
--amdgpu-use-aa-in-codegen | - Enable the use of AA during codegen. | help |
--amdgpu-verify-hsa-metadata | - Verify AMDGPU HSA Metadata | help |
--amdgpu-vgpr-index-mode | - Use GPR indexing mode instead of movrel for vector indexing | help |
--arc-contract-use-objc-claim-rv | - Enable generation of calls to objc_claimAutoreleasedReturnValue | help |
--argext-abi-check | - Verify that narrow int args are properly extended per the SystemZ ABI. | help |
--arm-add-build-attributes | - | help |
--arm-implicit-it value | - Allow conditional instructions outside of an IT block =always - Accept in both ISAs, emit implicit ITs in Thumb =never - Warn in ARM, reject in Thumb =arm - Accept in ARM, reject in Thumb =thumb -… | help |
--asm-show-inst | - Emit internal instruction representation to assembly file | help |
--atomic-counter-update-promoted | - Do counter update using atomic fetch add for promoted counters only | help |
--atomic-first-counter | - Use atomic fetch add for first counter in a function (usually the entry counter) | help |
--basic-block-section-match-infer | - Enable matching and inference when generating basic block sections | help |
--bounds-checking-single-trap | - Use one trap block per function | help |
--bpf-stack-size int | - Specify the BPF stack size limit | help |
--cfg-hide-cold-paths number | - Hide blocks with relative frequency below the given value | help |
--cfg-hide-deoptimize-paths | - | help |
--cfg-hide-unreachable-paths | - | help |
--check-functions-filter regex | - Only emit checks for arguments of functions whose names match the given regular expression | help |
--conditional-counter-update | - Do conditional counter updates in single byte counters mode) | help |
--acc-if-clause-lowering | - Lower if clauses in ACC compute constructs | help |
--acc-implicit-data | - Generate implicit data attributes for OpenACC compute constructs | help |
--enable-implicit-reduction-copy | - Enable applying implicit copy in lieu of implicit firstprivate for reduction variables. | help |
--acc-implicit-declare | - Applies implicit acc declare to globals referenced in compute and routine acc regions | help |
--acc-implicit-routine | - Generate implicit acc routine for functions in acc regions | help |
--device-type value | - Target device type for implicit routine generation. | help |
--acc-legalize-serial | - Legalize OpenACC serial constructs | help |
--acc-loop-tiling | - Tile OpenACC loops with tile clauses | help |
--default-tile-size int | - Default tile size to use for wildcard ('*') tile sizes | help |
--acc-specialize-for-device | - Strip OpenACC constructs inside device code | help |
--acc-specialize-for-host | - Convert OpenACC operations for host execution | help |
--enable-host-fallback | - Enable host fallback mode which converts ALL ACC operations, not just orphan operations. | help |
--affine-data-copy-generate | - Generate explicit copying for affine memory operations | help |
--fast-mem-capacity ulong | - Set fast memory space capacity in KiB (default: unlimited) | help |
--fast-mem-space uint | - Fast memory space identifier for copy generation (default: 1) | help |
--generate-dma | - Generate DMA instead of point-wise copy | help |
--min-dma-transfer int | - Minimum DMA transfer size supported by the target in bytes | help |
--skip-non-unit-stride-loops | - Testing purposes: avoid non-unit stride loop choice depths for copy placement | help |
--slow-mem-space uint | - Slow memory space identifier for copy generation (default: 0) | help |
--tag-mem-space uint | - Tag memory space identifier for copy generation (default: 0) | help |
--affine-expand-index-ops | - Lower affine operations operating on indices into more fundamental operations | help |
--affine-expand-index-ops-as-affine | - Lower affine operations operating on indices into affine.apply operations | help |
--affine-fold-memref-alias-ops | - Fold memref alias ops into affine memory ops | help |
--affine-loop-coalescing | - Coalesce nested loops with independent bounds into a single loop | help |
--affine-loop-fusion | - Fuse affine loop nests | help |
--compute-tolerance number | - Fractional increase in additional computation tolerated while fusing | help |
--local-buf-threshold ulong | - Threshold size (KiB) for promoting local buffers to fast memory space | help |
--maximal | - Enables maximal loop fusion | help |
--mode value | - fusion mode to attempt | help |
--affine-loop-invariant-code-motion | - Hoist loop invariant instructions outside of affine loops | help |
--affine-loop-normalize | - Apply normalization transformations to affine loop-like ops | help |
--promote-single-iter | - Promote single iteration loops | help |
--affine-loop-tile | - Tile affine loop nests | help |
--cache-size ulong | - Set size of cache to tile for in KiB (default: 512) | help |
--separate | - Separate full and partial tiles (default: false) | help |
--tile-size uint | - Use this tile size for all loops | help |
--tile-sizes uint | - List of tile sizes for each perfect nest (overridden by -tile-size) | help |
--affine-loop-unroll | - Unroll affine loops | help |
--cleanup-unroll | - Fully unroll the cleanup loop when possible. | help |
--unroll-factor long | - Use this unroll factor for all loops being unrolled, set it to -1, and it will fully unroll. | help |
--unroll-full-threshold uint | - Unroll all loops with trip count less than or equal to this | help |
--unroll-num-reps uint | - Unroll innermost loops repeatedly this many times | help |
--unroll-up-to-factor | - Allow unrolling up to the factor specified | help |
--affine-loop-unroll-jam | - Unroll and jam affine loops | help |
--unroll-jam-factor uint | - Use this unroll jam factor for all loops (default 4) | help |
--affine-parallelize | - Convert affine.for ops into 1-D affine.parallel | help |
--max-nested uint | - Maximum number of nested parallel loops to produce. | help |
--parallel-reductions | - Whether to parallelize reduction loops. | help |
--affine-pipeline-data-transfer | - Pipeline non-blocking data transfers between explicitly managed levels of the memory hierarchy | help |
--affine-raise-from-memref | - Turn some memref operators to affine operators where supported | help |
--affine-scalrep | - Replace affine memref accesses by scalars by forwarding stores to loads and eliminating redundant loads | help |
--affine-simplify-min-max | - Simplify affine min/max/apply | help |
--affine-simplify-structures | - Simplify affine expressions in maps/sets and normalize memrefs | help |
--affine-super-vectorize | - Vectorize to a target independent n-D vector abstraction | help |
--test-fastest-varying long | - Specify a 1-D, 2-D or 3-D pattern of fastest varying memory dimensions to match. | help |
--vectorize-reductions | - Vectorize known reductions expressed via iter_args. | help |
--virtual-vector-size long | - Specify an n-D virtual vector size for vectorization. | help |
--amdgpu-emulate-atomics | - Emulate atomic operations on chipsets that do not support them | help |
--chipset string | - Chipset that these operations will run on | help |
--amdgpu-fold-memrefs-ops | - Fold memref operations into their parent operations | help |
--amdgpu-maskedload-to-load | - Lower the operations from the vector maskedload to vector load | help |
--amdgpu-resolve-strided-metadata | - Resolve memref.extract_strided_metadata on AMDGPU ops | help |
--arith-emulate-unsupported-floats | - Emulate operations on unsupported floats with extf/truncf | help |
--source-types string | - MLIR types without arithmetic support on a given target | help |
--target-type string | - MLIR type to convert the unsupported source types to | help |
--arith-emulate-wide-int | - Emulate 2*N-bit integer operations using N-bit operations | help |
--widest-int-supported uint | - Widest integer type supported by the target | help |
--arith-expand | - Legalize Arith ops to be convertible to LLVM. | help |
--include-bf16 | - Enable the BF16 expansion patterns | help |
--include-f4e2m1 | - Enable the F4E2M1 expansion patterns | help |
--include-f8e8m0 | - Enable the F8E8M0 expansion patterns | help |
--arith-int-range-narrowing | - Reduce integer operations bitwidth based on integer range analysis | help |
--int-bitwidths-supported uint | - Integer bitwidths supported | help |
--arith-unsigned-when-equivalent | - Replace signed ops with unsigned ones where they are proven equivalent | help |
--arm-neon-2d-to-intr | - Convert Arm NEON structured ops to intrinsics | help |
--arm-sme-outer-product-fusion | - Fuse 'arm_sme.outerproduct' operations into 2-way or 4-way widening variants | help |
--arm-sme-vector-legalization | - Legalize vectors for ArmSME | help |
--arm-sve-legalize-vector-storage | - Ensures stores of SVE vector types will be legal | help |
--async-func-to-async-runtime | - Lower async.func operations to the explicit async.runtime andasync.coro operations | help |
--async-parallel-for | - Convert scf.parallel operations to multiple async compute ops executed concurrently for non-overlapping iteration ranges | help |
--async-dispatch | - Dispatch async compute tasks using recursive work splitting. | help |
--min-task-size int | - The minimum task size for sharding parallel operation. | help |
--num-workers int | - The number of available workers to execute async operations. | help |
--async-runtime-policy-based-ref-counting | - Policy based reference counting for Async runtime operations | help |
--async-runtime-ref-counting | - Automatic reference counting for Async runtime operations | help |
--async-runtime-ref-counting-opt | - Optimize automatic reference counting operations for theAsync runtime by removing redundant operations | help |
--async-to-async-runtime | - Lower all high level async operations (e.g. async.execute) tothe explicit async.runtime and async.coro operations | help |
--bubble-down-memory-space-casts | - Bubbles down memory-space cast operations. | help |
--buffer-deallocation-simplification | - Optimizes `bufferization.dealloc` operation for more efficient codegen | help |
--buffer-hoisting | - Optimizes placement of allocation operations by moving them into common dominators and out of nested regions | help |
--buffer-loop-hoisting | - Optimizes placement of allocation operations by moving them out of loop nests | help |
--buffer-results-to-out-params | - Converts memref-typed function results to out-params | help |
--add-result-attr | - Add the attribute 'bufferize.result' to all output parameters. | help |
--hoist-dynamic-allocs | - Hoist dynamic allocations to call sites. | help |
--hoist-static-allocs | - Hoist static allocations to call sites. | help |
--modify-public-functions | - Modify function signatures of public functions. | help |
--bufferization-lower-deallocations | - Lowers `bufferization.dealloc` operations to `memref.dealloc`operations | help |
--canonicalize | - Canonicalize operations | help |
--disable-patterns string | - Labels of patterns that should be filtered out during application | help |
--enable-patterns string | - Labels of patterns that should be used during application, all other patterns are filtered out | help |
--max-iterations long | - Max. | help |
--max-num-rewrites long | - Max. | help |
--region-simplify value | - Perform control flow optimizations to the region tree | help |
--test-convergence | - Test only: Fail pass on non-convergence to detect cyclic pattern | help |
--top-down | - Seed the worklist in general top-down order | help |
--composite-fixed-point-pass | - Composite fixed point pass | help |
--name string | - Composite pass display name | help |
--pipeline string | - Composite pass inner pipeline | help |
--control-flow-sink | - Sink operations into conditional blocks | help |
--convert-affine-for-to-gpu | - Convert top-level AffineFor Ops to GPU kernels | help |
--gpu-block-dims uint | - Number of GPU block dimensions for mapping | help |
--gpu-thread-dims uint | - Number of GPU thread dimensions for mapping | help |
--convert-amdgpu-to-rocdl | - Convert AMDGPU dialect to ROCDL dialect | help |
--convert-arith-to-amdgpu | - Convert Arith operations to AMDGPU-specific implementations | help |
--allow-packed-f16-round-to-zero | - Whether we should allow f32->f16 packed round-to-zero conversion | help |
--saturate-fp8-truncf | - Use saturating truncation for 8-bit float types | help |
--convert-arith-to-apfloat | - Convert Arith ops to APFloat runtime library calls | help |
--convert-arith-to-arm-sme | - Convert Arith dialect to ArmSME dialect | help |
--convert-arith-to-emitc | - Convert Arith dialect to EmitC dialect | help |
--convert-arith-to-llvm | - Convert Arith dialect to LLVM dialect | help |
--index-bitwidth uint | - Bitwidth of the index type, 0 to use size of machine word | help |
--convert-arith-to-spirv | - Convert Arith dialect to SPIR-V dialect | help |
--emulate-lt-32-bit-scalar-types | - Emulate narrower scalar types with 32-bit ones if not supported by the target | help |
--emulate-unsupported-float-types | - Emulate unsupported float types by representing them with integer types of same bit width | help |
--convert-arm-sme-to-llvm | - Lower the operations from the ArmSME dialect into the LLVM dialect | help |
--dump-tile-live-ranges | - Dump the live ranges of SME tiles (for debugging) | help |
--convert-arm-sme-to-scf | - Lower the operations from the ArmSME dialect into the SCF dialect | help |
--convert-async-to-llvm | - Convert the operations from the async dialect into the LLVM dialect | help |
--convert-bufferization-to-memref | - Convert operations from the Bufferization dialect to the MemRef dialect | help |
--convert-cf-to-llvm | - Convert ControlFlow operations to the LLVM dialect | help |
--convert-cf-to-spirv | - Convert ControlFlow dialect to SPIR-V dialect | help |
--convert-complex-to-libm | - Convert Complex dialect to libm calls | help |
--convert-complex-to-llvm | - Convert Complex dialect to LLVM dialect | help |
--complex-range value | - Control the intermediate calculation of complex number division | help |
--convert-complex-to-rocdl-library-calls | - Convert Complex dialect to ROCDL library calls | help |
--convert-complex-to-spirv | - Convert Complex dialect to SPIRV dialect | help |
--convert-complex-to-standard | - Convert Complex dialect to standard dialect | help |
--convert-elementwise-to-linalg | - Convert ElementwiseMappable ops to linalg | help |
--convert-func-to-emitc | - Convert Func dialect to EmitC dialect | help |
--convert-func-to-llvm | - Convert from the Func dialect to the LLVM dialect | help |
--use-bare-ptr-memref-call-conv | - Replace FuncOp's MemRef arguments with bare pointers to the MemRef element types | help |
--convert-func-to-spirv | - Convert Func dialect to SPIR-V dialect | help |
--convert-gpu-to-llvm-spv | - Generate LLVM operations to be ingested by a SPIR-V backend for gpu operations | help |
--use-64bit-index | - Use 64-bit integers to convert index types | help |
--convert-gpu-to-nvvm | - Generate NVVM operations for gpu operations | help |
--allow-pattern-rollback | - Experimental performance flag to disallow pattern rollback | help |
--allowed-dialects string | - Run conversion patterns of only the specified dialects | help |
--has-redux | - Target gpu supports redux | help |
--convert-gpu-to-rocdl | - Generate ROCDL operations for gpu operations | help |
--runtime value | - Runtime code will be run on (default is Unknown, can also use HIP or OpenCL) | help |
--convert-gpu-to-spirv | - Convert GPU dialect to SPIR-V dialect | help |
--convert-index-to-llvm | - Lower the `index` dialect to the `llvm` dialect. | help |
--convert-index-to-spirv | - Lower the `index` dialect to the `spirv` dialect. | help |
--convert-linalg-to-affine-loops | - Lower the operations from the linalg dialect into affine loops | help |
--convert-linalg-to-loops | - Lower the operations from the linalg dialect into loops | help |
--convert-linalg-to-parallel-loops | - Lower the operations from the linalg dialect into parallel loops | help |
--convert-linalg-to-std | - Convert the operations from the linalg dialect into the Standard dialect | help |
--convert-math-to-apfloat | - Convert Math ops to APFloat runtime library calls | help |
--convert-math-to-emitc | - Convert some Math operations to EmitC call_opaque operations | help |
--language-target value | - Select the language standard target for callees (c99 or cpp11). | help |
--convert-math-to-funcs | - Convert Math operations to calls of outlined implementations. | help |
--convert-ctlz | - Convert math.ctlz to a software implementation. | help |
--min-width-of-fpowi-exponent uint | - Convert FPowI only if the width of its exponent's integer type is greater than or equal to this value | help |
--convert-math-to-libm | - Convert Math dialect to libm calls | help |
--convert-math-to-llvm | - Convert Math dialect to LLVM dialect | help |
--approximate-log1p | - Enable approximation of Log1p. | help |
--convert-math-to-rocdl | - Convert Math dialect to ROCDL library calls | help |
--convert-math-to-spirv | - Convert Math dialect to SPIR-V dialect | help |
--convert-math-to-xevm | - Convert (fast) math operations to native XeVM/SPIRV equivalents | help |
--convert-arith | - Convert supported Arith ops (e.g. arith.divf) as well. | help |
--convert-memref-to-emitc | - Convert MemRef dialect to EmitC dialect | help |
--lower-to-cpp | - Target C++ (true) instead of C (false) | help |
--convert-memref-to-spirv | - Convert MemRef dialect to SPIR-V dialect | help |
--bool-num-bits int | - The number of bits to store a boolean value | help |
--convert-nvgpu-to-nvvm | - Convert NVGPU dialect to NVVM dialect | help |
--convert-nvvm-to-llvm | - Convert NVVM to PTX with Inline Assembly in LLVM dialect | help |
--convert-openacc-to-scf | - Convert the OpenACC ops to OpenACC with SCF dialect | help |
--convert-openmp-to-llvm | - Convert the OpenMP ops to OpenMP ops with LLVM dialect | help |
--convert-parallel-loops-to-gpu | - Convert mapped scf.parallel ops to gpu launch operations | help |
--convert-pdl-to-pdl-interp | - Convert PDL ops to PDL interpreter ops | help |
--convert-scf-to-cf | - Convert SCF dialect to ControlFlow dialect, replacing structured control flow with a CFG | help |
--convert-scf-to-emitc | - Convert SCF dialect to EmitC dialect, maintaining structured control flow | help |
--convert-scf-to-openmp | - Convert SCF parallel loop to OpenMP parallel + workshare constructs. | help |
--num-threads uint | - Number of threads to use | help |
--convert-scf-to-spirv | - Convert SCF dialect to SPIR-V dialect. | help |
--convert-shape-constraints | - Convert shape constraint operations to the standard dialect | help |
--convert-shape-to-std | - Convert operations from the shape dialect into the standard dialect | help |
--convert-shard-to-mpi | - Convert Shard dialect to MPI dialect. | help |
--convert-spirv-to-llvm | - Convert SPIR-V dialect to LLVM dialect | help |
--client-api value | - Derive StorageClass to address space mapping from the client API | help |
--convert-tensor-to-linalg | - Convert some Tensor dialect ops to Linalg dialect | help |
--convert-tensor-to-spirv | - Convert Tensor dialect to SPIR-V dialect | help |
--convert-to-emitc | - Convert to EmitC dialect via dialect interfaces | help |
--filter-dialects string | - Test conversion patterns of only the specified dialects | help |
--convert-to-llvm | - Convert to LLVM via dialect interfaces found in the input IR | help |
--dynamic | - Use op conversion attributes to configure the conversion | help |
--convert-ub-to-llvm | - Convert UB dialect to LLVM dialect | help |
--convert-ub-to-spirv | - Convert UB dialect to SPIR-V dialect | help |
--convert-vector-to-amx | - Lower the operations from the vector dialect into the AMX dialect | help |
--convert-vector-to-arm-sme | - Lower the operations from the vector dialect into the ArmSME dialect | help |
--convert-vector-to-gpu | - Lower the operations from the vector dialect into the GPU dialect | help |
--use-nvgpu | - convert to NvGPU ops instead of GPU dialect ops | help |
--convert-vector-to-llvm | - Lower the operations from the vector dialect into the LLVM dialect | help |
--enable-amx | - Enables the use of AMX dialect while lowering the vector dialect. | help |
--enable-arm-bf16 | - Enables the use of Arm FEAT_BF16 instructions while lowering the vector dialect. | help |
--enable-arm-i8mm | - Enables the use of Arm FEAT_I8MM instructions while lowering the vector dialect. | help |
--enable-arm-neon | - Enables the use of ArmNeon dialect while lowering the vector dialect. | help |
--enable-arm-sve | - Enables the use of ArmSVE dialect while lowering the vector dialect. | help |
--enable-x86vector | - Enables the use of X86Vector dialect while lowering the vector dialect. | help |
--force-32bit-vector-indices | - Allows compiler to assume vector indices fit in 32-bit if that yields faster code | help |
--reassociate-fp-reductions | - Allows llvm to reassociate floating-point reductions for speed | help |
--use-vector-alignment | - Use the preferred alignment of a vector type in load/store operations instead of the alignment of the element type of the memref. | help |
--vector-contract-lowering value | - control the lowering of `vector.contract` operations. | help |
--vector-transpose-lowering value | - control the lowering of `vector.transpose` operations. | help |
--convert-vector-to-scf | - Lower the operations from the vector dialect into the SCF dialect | help |
--full-unroll | - Perform full unrolling when converting vector transfers to SCF | help |
--lower-scalable | - Add scalable vector specific lowerings (that introduce loops) | help |
--lower-tensors | - Lower transfer ops that operate on tensors | help |
--target-rank uint | - Target vector rank to which transfer ops should be lowered | help |
--convert-vector-to-spirv | - Convert Vector dialect to SPIR-V dialect | help |
--convert-vector-to-xegpu | - Lower the operations from the vector dialect into the XeGPU dialect | help |
--convert-xegpu-to-xevm | - Convert XeGPU to XeVM dialect | help |
--convert-xevm-to-llvm | - Convert XeVM to LLVM dialect | help |
--cse | - Eliminate common sub-expressions | help |
--decorate-spirv-composite-type-layout | - Decorate SPIR-V composite type with layout info | help |
--drop-equivalent-buffer-results | - Remove MemRef return values that are equivalent to a bbArg | help |
--duplicate-function-elimination | - Deduplicate functions | help |
--eliminate-empty-tensors | - Try to eliminate all tensor.empty ops. | help |
--empty-tensor-to-alloc-tensor | - Replace all empty ops by alloc_tensor ops. | help |
--enable-arm-streaming | - Enable Armv9 Streaming SVE mode | help |
--if-required-by-ops | - Only apply the selected streaming/ZA modes if the function contains ops that implement the ArmSMETileOpInterface. | help |
--if-scalable-and-supported | - Only apply the selected streaming/ZA modes if the function contains supported scalable vector operations. | help |
--streaming-mode value | - Select how streaming-mode is managed at the function-level. | help |
--za-mode value | - Select how ZA-storage is managed at the function-level. | help |
--ensure-debug-info-scope-on-llvm-func | - Materialize LLVM debug info subprogram attribute on every LLVMFuncOp | help |
--emission-kind value | - Emission kind to generate debug info. | help |
--expand-realloc | - Expand memref.realloc operations into its components | help |
--emit-deallocs | - Emit deallocation operations for the original MemRef | help |
--expand-strided-metadata | - Expand memref operations into easier to analyze constructs | help |
--finalize-memref-to-llvm | - Finalize MemRef dialect to LLVM dialect conversion | help |
--use-aligned-alloc | - Use aligned_alloc in place of malloc for heap allocations | help |
--use-generic-functions | - Use generic allocation and deallocation functions instead of the classic 'malloc', 'aligned_alloc' and 'free' functions | help |
--flatten-memref | - Flatten a multiple dimensional memref to 1-dimensional | help |
--fold-memref-alias-ops | - Fold memref alias ops into consumer load/store ops | help |
--fold-tensor-subset-ops | - Fold tensor subset ops into producer/consumer ops | help |
--form-expressions | - Form C-style expressions from C-operator ops | help |
--generate-runtime-verification | - Generate additional runtime op verification checks | help |
--verbose-level uint | - Verbosity level for runtime verification messages: 0 = Minimum (only source location), 1 = Detailed (include full operation details, names, types, shapes, etc.) | help |
--gpu-async-region | - Make GPU ops async | help |
--gpu-decompose-memrefs | - Decomposes memref index computation into explicit ops. | help |
--gpu-eliminate-barriers | - Erase unnecessary barriers | help |
--gpu-kernel-outlining | - Outline gpu.launch bodies to kernel functions | help |
--data-layout-str string | - String description of the data layout | help |
--gpu-launch-sink-index-computations | - Sink index computations into gpu.launch body | help |
--gpu-map-parallel-loops | - Greedily maps loops to GPU hardware dimensions. | help |
--mapping-policy string | - Policy outlining how to assign loops to GPU dimensions.Supported values are `outermost-first` and `innermost-first`. | help |
--gpu-module-to-binary | - Transforms a GPU module into a GPU binary. | help |
--format string | - The target representation of the compilation process. | help |
-l string | - Extra files to link to. | help |
--opts string | - Command line options to pass to the tools. | help |
--section string | - ELF section where binary is to be located. | help |
--toolkit string | - Toolkit path. | help |
--gpu-to-llvm | - Convert GPU dialect to LLVM dialect with GPU runtime calls | help |
--intersperse-sizes-for-kernels | - Inserts a size_t argument following each memref argument, containing the static size in bytes of the buffer. | help |
--use-bare-pointers-for-host | - Use bare pointers to pass memref arguments to host functions. | help |
--use-bare-pointers-for-kernels | - Use bare pointers to pass memref arguments to kernels. | help |
--inline | - Inline function calls | help |
--default-pipeline string | - The optimizer pipeline used for callables that do not have a dedicated optimizer pipeline in opPipelineList | help |
--inlining-threshold uint | - If the ratio between the number of the operations in the callee and the number of the operations in the caller exceeds this value (in percentage), then the callee is not inlined even if it is legal… | help |
--op-pipelines pass-manager | - Callable operation specific optimizer pipelines (in the form of `dialect.op(pipeline)`) | help |
--int-range-optimizations | - Do optimizations based on integer range analysis | help |
--lift-cf-to-scf | - Lift ControlFlow dialect to SCF dialect | help |
--linalg-block-pack-matmul | - Convert linalg matmul ops to block layout and back | help |
--allow-padding | - Allow packing padding | help |
--block-factors long | - Block factors (mb, nb, kb) for relayout | help |
--lhs-transpose-inner-blocks | - Transpose LHS inner block layout [mb][kb] -> [kb][mb] | help |
--lhs-transpose-outer-blocks | - Transpose LHS outer block layout [MB][KB] -> [KB][MB] | help |
--mnk-order long | - Permutation of matmul (M, N, K) dimensions order | help |
--mnk-padded-multiples long | - Next multiples of the packing sizes | help |
--rhs-transpose-inner-blocks | - Transpose RHS inner block layout [kb][nb] -> [nb][kb] | help |
--rhs-transpose-outer-blocks | - Transpose RHS outer block layout [KB][NB] -> [NB][KB] | help |
--linalg-detensorize | - Detensorize linalg ops | help |
--aggressive-mode | - Detensorize all ops that qualify for detensoring along with branch operands and basic-block arguments. | help |
--linalg-fold-into-elementwise | - Fold transpose ops into elementwise | help |
--linalg-fold-unit-extent-dims | - Remove unit-extent dimension in Linalg ops on tensors | help |
--use-rank-reducing-slices | - Generate rank-reducing slices instead of reassociative reshapes | help |
--linalg-fuse-elementwise-ops | - Fuse elementwise operations on tensors | help |
--linalg-generalize-named-ops | - Convert named ops into generic ops | help |
--linalg-inline-scalar-operands | - Inline scalar operands into linalg generic ops | help |
--linalg-morph-ops | - Convert linalg ops between forms | help |
--category-to-generic | - convert category ops e.g. `linalg.elementwise` to `linalg.generic` | help |
--generic-to-named | - convert linalg.generic to equivalent named ops | help |
--named-to-category | - convert named ops to category op e.g. `linalg.elementwise` | help |
--named-to-generic | - convert named ops e.g. `linalg.add` to `linalg.generic` | help |
--linalg-specialize-generic-ops | - Convert generic ops back to named ops | help |
--llvm-add-comdats | - Add comdats to linkonce and linkonce_odr functions | help |
--llvm-legalize-for-export | - Legalize LLVM dialect to be convertible to LLVM IR | help |
--llvm-optimize-for-nvvm-target | - Optimize NVVM IR | help |
--llvm-request-c-wrappers | - Request C wrapper emission for all functions | help |
--llvm-target-to-data-layout | - Derive data layout attributes from LLVM target attributes | help |
--initialize-llvm-targets | - Whether to pre-load all available target machines, that LLVM is configured to support, into the TargetRegistry. | help |
--llvm-target-to-target-features | - Update attached #llvm.target's features per the described target | help |
--llvm-use-default-visibility | - Update default visibility of all global values and function definitions | help |
--visibility value | - Visibility to use in place of default. | help |
--loop-invariant-code-motion | - Hoist loop invariant instructions outside of the loop | help |
--loop-invariant-subset-hoisting | - Hoist loop invariant subset ops outside of the loop | help |
--lower-affine | - Lower Affine operations to a combination of Arith and SCF operations | help |
--lower-host-to-llvm | - Lowers the host module code and `gpu.launch_func` to LLVM | help |
--lower-quant-ops | - Lower quant.dcast and quant.qcast ops | help |
--lower-sparse-foreach-to-scf | - Decompose a complex sparse operation into multiple stages | help |
--lower-sparse-iteration-to-scf | - lower sparse_tensor.iterate/coiterate into scf loops | help |
--lower-sparse-ops-to-foreach | - Applies sparse tensor rewriting rules after sparsification | help |
--enable-convert | - Enable rewriting rules for the convert operator | help |
--enable-runtime-library | - Enable runtime library for manipulating sparse tensors | help |
--lower-vector-mask | - Lower 'vector.mask' operations | help |
--lower-vector-multi-reduction | - Lower 'vector.multi_reduction' operations | help |
--lowering-strategy value | - Select the strategy to control how multi_reduction is lowered. | help |
--lower-vector-to-from-elements-to-shuffle-tree | - Lower `vector.to_elements` and `vector.from_elements` to a tree of `vector.shuffle` operations | help |
--map-memref-spirv-storage-class | - Map numeric MemRef memory spaces to SPIR-V storage classes | help |
--math-expand-ops | - Expand math operations. | help |
--ops string | - Operations to expand. | help |
--math-extend-to-supported-types | - Legalize floating-point math ops on low-precision floats | help |
--extra-types string | - MLIR types with arithmetic support on a given target (f64 and f32 are implicitly supported) | help |
--math-sincos-fusion | - Fuse sin and cos operations. | help |
--math-uplift-to-fma | - Uplift arith ops to math.fma. | help |
--mem2reg | - Promotes memory slots into values. | help |
--memref-emulate-wide-int | - Emulate 2*N-bit integer operations using N-bit operations | help |
--memref-expand | - Legalize memref operations to be convertible to LLVM. | help |
--mlprogram-pipeline-globals | - Optimize `ml_program` global operations for read and store | help |
--normalize-memrefs | - Normalize memrefs | help |
--normalize-quant-types | - Normalize generic quantized types to specific quantized types | help |
--nvgpu-optimize-shared-memory | - Optimizes accesses to shard memory memrefs in order to reduce bank conflicts. | help |
--nvvm-attach-target | - Attaches an NVVM target attribute to a GPU Module. | help |
-O uint | - Optimization level. | help |
--chip string | - Target chip. | help |
--fast | - Enable fast math mode. | help |
--features string | - Target features, default is to use the minimum PTX ISA version for the SM target. | help |
--ftz | - Enable flush to zero for denormals. | help |
--module string | - Regex used to identify the modules to attach the target to. | help |
--ptxas-cmd-options string | - Command line options passed to downstream compiler | help |
--triple string | - Target triple. | help |
--offload-livein-value-canonicalization | - Canonicalize live-in values for regions destined for outlining | help |
--omp-offload-privatization-prepare | - Prepare OpenMP maps for privatization for deferred target tasks | help |
--one-shot-bufferize | - One-Shot Bufferize | help |
--allow-return-allocs-from-loops | - Allows returning/yielding new allocations from a loop. | help |
--allow-unknown-ops | - Allows unknown (not bufferizable) ops in the input IR. | help |
--analysis-fuzzer-seed uint | - Test only: Analyze ops in random order with a given seed (fuzzer) | help |
--analysis-heuristic string | - Heuristic that control the IR traversal during analysis | help |
--buffer-alignment ulong | - Sets the alignment of newly allocated buffers. | help |
--bufferize-function-boundaries | - Bufferize function boundaries (experimental). | help |
--check-parallel-regions | - Account for parallel regions in RaW analysis. | help |
--copy-before-write | - Skip the analysis. | help |
--dialect-filter string | - Restrict bufferization to ops from these dialects. | help |
--dump-alias-sets | - Test only: Annotate tensor IR with alias sets | help |
--function-boundary-type-conversion value | - Controls layout maps when bufferizing function signatures. | help |
--must-infer-memory-space | - The memory space of an memref types must always be inferred. | help |
--no-analysis-func-filter string | - Skip analysis of functions with these symbol names.Set copyBeforeWrite to true when bufferizing them. | help |
--print-conflicts | - Test only: Annotate IR with RaW conflicts. | help |
--test-analysis-only | - Test only: Only run inplaceability analysis and annotate IR | help |
--unknown-type-conversion value | - Controls layout maps for non-inferrable memref types. | help |
--use-encoding-for-memory-space | - Use the Tensor encoding attribute for the memory space. | help |
--openacc-legalize-data-values | - Legalizes SSA values in compute regions with results from data clause operations | help |
--apply-to-acc-data-construct | - Replaces varPtr uses with accPtr for acc compute regions contained within acc.data or acc.declare region. | help |
--host-to-device | - Replace varPtr uses with accPtr if true. | help |
--optimize-allocation-liveness | - This pass optimizes the liveness of temp allocations in the input function | help |
--outline-shape-computation | - Using shape.func to preserve shape computation | help |
--ownership-based-buffer-deallocation | - Adds all required dealloc operations for all allocations in the input program | help |
--private-function-dynamic-ownership | - Allows to add additional arguments to private functions to dynamically pass ownership of memrefs to callees. | help |
--pre-sparsification-rewrite | - Applies sparse tensor rewriting rules prior to sparsification | help |
--print-ir | - Print IR on the debug stream | help |
--label string | - Label | help |
--print-op-stats | - Print statistics of operations | help |
--json | - print the stats as JSON | help |
--promote-buffers-to-stack | - Promotes heap-based allocations to automatically managed stack-based allocations | help |
--max-alloc-size-in-bytes uint | - Maximal size in bytes to promote allocations to stack. | help |
--max-rank-of-allocated-memref uint | - Maximal memref rank to promote dynamic buffers. | help |
--reconcile-unrealized-casts | - Simplify and eliminate unrealized conversion casts | help |
--reify-result-shapes | - Reifies the results of `tensor::PadOp` and `tensor::ConcatOp`. | help |
--remove-dead-values | - Remove dead values | help |
--remove-shape-constraints | - Replace all cstr_ ops with a true witness | help |
--resolve-ranked-shaped-type-result-dims | - Resolve memref.dim of result values of ranked shape type | help |
--error-on-pattern-iteration-limit | - Throw an error when pattern rewriter hits iteration limit | help |
--resolve-shaped-type-result-dims | - Resolve memref.dim of result values | help |
--rocdl-attach-target | - Attaches a ROCDL target attribute to a GPU Module. | help |
--abi string | - ABI version. | help |
--correct-sqrt | - Enable correct rounded sqrt. | help |
--daz | - Enable denormals are zero opt. | help |
--finite-only | - Enable finite only opt. | help |
--unsafe-math | - Enable unsafe math opt. | help |
--wave64 | - Use Wave64 mode. | help |
--sccp | - Sparse Conditional Constant Propagation | help |
--scf-for-loop-canonicalization | - Canonicalize operations within scf.for loop bodies | help |
--scf-for-loop-peeling | - Peel `for` loops at their upper bounds. | help |
--peel-front | - Peel the first iteration out of the loop. | help |
--skip-partial | - Do not peel loops inside of the last, partial iteration of another already peeled loop. | help |
--scf-for-loop-range-folding | - Fold add/mul ops into loop range | help |
--scf-for-loop-specialization | - Specialize `for` loops for vectorization | help |
--scf-for-to-while | - Convert SCF for loops to SCF while loops | help |
--scf-forall-to-for | - Convert SCF forall loops to SCF for loops | help |
--scf-forall-to-parallel | - Convert SCF forall loops to SCF parallel loops | help |
--scf-parallel-for-to-nested-fors | - Convert SCF parallel for loops to nested SCF for loops | help |
--scf-parallel-loop-fusion | - Fuse adjacent parallel loops | help |
--scf-parallel-loop-specialization | - Specialize parallel loops for vectorization | help |
--scf-parallel-loop-tiling | - Tile parallel loops | help |
--no-min-max-bounds | - Perform tiling with fixed upper bound with inbound check inside the internal loops | help |
--parallel-loop-tile-sizes long | - Factors to tile parallel loops by | help |
--set-llvm-module-datalayout | - Attach a datalayout string as a module attribute | help |
--data-layout string | - String description (LLVM format) of the data layout that is expected on the produced module | help |
--shape-to-shape-lowering | - Legalize Shape dialect to be convertible to Arith | help |
--shard-partition | - Partition a function into SPMD form. | help |
--sharding-propagation | - sharding propagation | help |
--traversal value | - Traversal order to use for sharding propagation: | help |
--simplify-depthwise-conv | - Simplify depthwise convolution. | help |
--snapshot-op-locations | - Generate new locations from the current IR | help |
--filename string | - The filename to print the generated IR | help |
--pretty-debuginfo | - Print pretty debug info in MLIR output | help |
--print-debuginfo | - Print debug info in MLIR output | help |
--print-local-scope | - Print with local scope and inline information (eliding aliases for attributes, types, and locations | help |
--print-op-generic | - Print the generic op form | help |
--tag string | - A tag to use when fusing the new locations with the original. | help |
--sparse-assembler | - Add [dis]assemble operations on external sparse tensors | help |
--direct-out | - Directly returns buffers externally | help |
--sparse-buffer-rewrite | - Rewrite sparse primitives on buffers to actual code | help |
--enable-buffer-initialization | - Enable zero-initialization of the memory buffers | help |
--sparse-gpu-codegen | - Generates GPU code during sparsification | help |
--sparse-reinterpret-map | - Reinterprets sparse tensor type mappings | help |
--loop-ordering-strategy value | - Set the loop ordering strategy for sparse code generation | help |
--scope value | - Set the reiterpretation scope | help |
--sparse-space-collapse | - sparse space collapsing pass | help |
--sparse-storage-specifier-to-llvm | - Lower sparse storage specifer to llvm structure | help |
--sparse-tensor-codegen | - Convert sparse tensors and primitives to actual code | help |
--create-sparse-deallocs | - Specify if the temporary buffers created by the sparse compiler should be deallocated. | help |
--sparse-tensor-conversion | - Convert sparse tensors and primitives to library calls | help |
--sparse-vectorization | - Vectorizes loops after sparsification | help |
--enable-simd-index32 | - Enable i32 indexing into vectors (for efficient gather/scatter) | help |
--enable-vla-vectorization | - Enable vector length agnostic vectorization | help |
--vl int | - Set the vector length (use 0 to disable vectorization) | help |
--sparsification | - Automatically generate sparse tensor code from sparse tensor types | help |
--parallelization-strategy value | - Set the parallelization strategy | help |
--sparse-emit-strategy value | - Emit functional code or interfaces (to debug) for sparse loops | help |
--sparsification-and-bufferization | - Mini-pipeline that combines bufferization and sparsifiation | help |
--enable-gpu-libgen | - Enable GPU acceleration by means of direct library calls | help |
--spirv-attach-target | - Attaches an SPIR-V target attribute to a GPU Module. | help |
--caps string | - List of supported SPIR-V Capabilities | help |
--client_api string | - Client API | help |
--device_id uint | - Device ID | help |
--device_type string | - Device Type | help |
--exts string | - List of supported SPIR-V Extensions | help |
--vendor string | - Device Vendor | help |
--ver string | - SPIR-V Version. | help |
--spirv-canonicalize-gl | - Canonicalize GLSL ops | help |
--spirv-lower-abi-attrs | - Decorate SPIR-V composite type with layout info | help |
--spirv-promote-to-replicated-constants | - Convert splat composite constants and spec constants to corresponding replicated constant composite ops defined by SPV_EXT_replicated_composites | help |
--spirv-rewrite-inserts | - Rewrite sequential chains of `spirv.CompositeInsert` operations into `spirv.CompositeConstruct` operations | help |
--spirv-unify-aliased-resource | - Unify access of multiple aliased resources into access of one single resource | help |
--spirv-update-vce | - Deduce and attach minimal (version, capabilities, extensions) requirements to spirv.module ops | help |
--spirv-webgpu-prepare | - Prepare SPIR-V to target WebGPU by expanding unsupported ops and replacing with supported ones | help |
--sroa | - Scalar Replacement of Aggregates | help |
--stage-sparse-ops | - Decompose a complex sparse operation into multiple stages | help |
--strip-debuginfo | - Strip debug info from all operations | help |
--strip-func-quant-types | - Strip quantized types from function headers | help |
--symbol-dce | - Eliminate dead symbols | help |
--symbol-privatize | - Mark symbols private | help |
--exclude string | - Comma separated list of symbols that should not be marked private | help |
--test-arm-sme-tile-allocation | - Tests SME 'virtual tile' allocation | help |
--preprocess-only | - Only preprocess IR so it is ready for tile allocation (but do not allocate any tiles) | help |
--test-scf-parallel-loop-collapsing | - Test parallel loops collapsing transformation | help |
--collapsed-indices-0 uint | - Which loop indices to combine 0th loop index | help |
--collapsed-indices-1 uint | - Which loop indices to combine into the position 1 loop index | help |
--collapsed-indices-2 uint | - Which loop indices to combine into the position 2 loop index | help |
--topological-sort | - Sort regions without SSA dominance in topological order | help |
--tosa-arith-const-to-tosa-const | - Convert tensor arith.constant operations into tosa.const | help |
--tosa-attach-target | - Attach tosa.target_env information to the given module. | help |
--extensions string | - The TOSA extension(s) that operators should conform to. | help |
--level value | - The TOSA level that operators should conform to. | help |
--profiles string | - The TOSA profile(s) that operators should conform to. | help |
--specification_version value | - The specification version that TOSA operators should conform to. | help |
--tosa-convert-integer-type-to-signless | - Convert integer types to signless | help |
--tosa-infer-shapes | - Propagate shapes across TOSA operations | help |
--tosa-layerwise-constant-fold | - Fold layerwise operations on constant tensors | help |
--aggressive-reduce-constant | - Always perform the reduce constant optimizationMay add more tosa.const but would reduce runtime calculations | help |
--tosa-make-broadcastable | - TOSA rank Reshape to enable Broadcasting | help |
--tosa-narrow-f64-to-f32 | - Narrow F64 TOSA operations to F32 | help |
--aggressive-rewrite | - If enabled, all TOSA operations are rewritten, regardless or whether the narrowingis safe. | help |
--convert-function-boundaries | - If enabled, the pass will convert function I/O types as well. | help |
--tosa-narrow-i64-to-i32 | - Narrow I64 TOSA operations to I32 | help |
--tosa-optional-decompositions | - Applies Tosa operations optional decompositions | help |
--tosa-reduce-transposes | - Reduce transposes through other operators | help |
--tosa-to-arith | - Lower TOSA to the Arith dialect | help |
--include-apply-rescale | - Whether to include the lowering for tosa.apply_rescale to arith | help |
--use-32-bit | - Whether to prioritze lowering to 32-bit operations | help |
--tosa-to-linalg | - Lower TOSA to LinAlg on tensors | help |
--disable-tosa-decompositions | - Disable tosa decompositions pass | help |
--tosa-to-linalg-named | - Lower TOSA to LinAlg named operations | help |
--prefer-conv2d-kernel-layout-hwcf | - Prefer generating linalg.conv_2d_nhwc_hwcf over linalg.conv_2d_nhwc_fhwc | help |
--tosa-to-mlprogram | - Lower TOSA to the MLProgram dialect | help |
--tosa-to-scf | - Lower TOSA to the SCF dialect | help |
--tosa-to-tensor | - Lower TOSA to the Tensor dialect | help |
--tosa-validate | - Validates TOSA dialect | help |
--allow-invalid-op-datatype-combinations | - Disable checks for operations that are determined to be invalid due to their operand/result datatypes not aligning with the 'Supported Data Types' sections of the specifciation | help |
--strict-op-spec-alignment | - Verify if the properties of certain operations align the spec requirement | help |
--transform-dialect-check-uses | - warn about potential use-after-free in the transform dialect | help |
--transform-infer-effects | - infer transform side effects for symbols | help |
--transform-interpreter | - transform dialect interpreter | help |
--debug-bind-trailing-args string | - Binds trailing arguments of the entry point to the payload operations with specified names. | help |
--debug-payload-root-tag string | - Select the operation with 'transform.target_tag' attribute having the given value as payload IR root. | help |
--disable-expensive-checks | - Disable expensive checks in the interpreter for a faster run. | help |
--entry-point string | - Entry point of the pass pipeline. | help |
--transform-preload-library | - preload transform dialect library | help |
--transform-library-paths string | - Optional paths to files with modules that should be merged into the transform module to provide the definitions of external named sequences. | help |
--view-op-graph | - Print Graphviz visualization of an operation | help |
--max-label-len uint | - Limit attribute/type length to number of chars | help |
--print-attrs | - Print attributes of operations | help |
--print-control-flow-edges | - Print control flow edges | help |
--print-data-flow-edges | - Print data flow edges | help |
--print-result-types | - Print result types of operations | help |
--wrap-emitc-func-in-class | - Wrap functions in classes, using arguments as fields. | help |
--xegpu-blocking | - Block XeGPU ops into smaller size. | help |
--xegpu-fold-alias-ops | - Fold alias ops into XeGPU ops | help |
--xegpu-optimize-block-loads | - Optimize XeGPU block load operations | help |
--xegpu-propagate-layout | - Propagate and assign XeGPU layout information | help |
--layout-kind string | - Propagate `subgroup` / `inst` / `lane` level of xegpu layouts. | help |
--print-analysis-only | - Print the result of layout propagation analysis and exit. | help |
--xegpu-subgroup-distribute | - Distribute XeGPU ops to work items | help |
--xegpu-vector-linearize | - Linearize n-D vectors to 1-D vectors | help |
--xegpu-wg-to-sg-distribute | - Transform WorkGroup level XeGPU code to SubGroup level | help |
--xevm-attach-target | - Attaches a XeVM target attribute to a GPU Module. | help |
--cmd-options string | - Command line options passed to downstream compiler | help |
--buffer-deallocation-pipeline | - The default pipeline for automatically inserting deallocation operations after one-shot bufferization. | help |
--gpu-lower-to-nvvm-pipeline | - The default pipeline lowers main dialects (arith, memref, scf, vector, gpu, and nvgpu) to NVVM. | help |
--cubin-chip string | - Chip to use to serialize to cubin. | help |
--cubin-features string | - Features to use to serialize to cubin. | help |
--cubin-format string | - Compilation format to use to serialize to cubin. | help |
--cubin-triple string | - Triple to use to serialize to cubin. | help |
--host-bare-ptr-calling-convention | - Whether to use the bareptr calling convention on the host (warning this should be false until the GPU layering is fixed) | help |
--kernel-bare-ptr-calling-convention | - Whether to use the bareptr calling convention on the kernel (warning this should be false until the GPU layering is fixed) | help |
--opt-level int | - Optimization level for NVVM compilation | help |
--gpu-lower-to-xevm-pipeline | - The default GPU to XeVM lowering pipeline. | help |
--binary-format string | - Final GPU binary emission format (e.g. fatbin) | help |
--igc-cmd-options string | - Additional downstream compiler command line options | help |
--xegpu-op-level string | - Granularity of XeGPU operations to target: workgroup | subgroup | lane | help |
--xevm-module-matcher string | - Regex to match gpu.module names for XeVM target attach | help |
--zebin-chip string | - Target chip (e.g. pvc, bmg) | help |
--zebin-triple string | - Target triple for XeVM codegen | help |
--sparsifier | - The standard pipeline for taking sparsity-agnostic IR using the sparse-tensor type, and lowering it to LLVM IR with concrete representations and algorithms for sparse tensors. | help |
--enable-index-optimizations | - Allows compiler to assume indices fit in 32-bit if that yields faster code | help |
--gpu-chip string | - GPU target architecture | help |
--gpu-features string | - GPU target features | help |
--gpu-format string | - GPU compilation format | help |
--gpu-triple string | - GPU target triple | help |
--test-bufferization-analysis-only | - Run only the inplacability analysis | help |
--tosa-to-linalg-pipeline | - The default pipeline for converting TOSA operators to the equivalent operations using the tensor operations in LinAlg as well as LinAlg named operations. | help |
--cost-kind value | - Target cost kind =throughput - Reciprocal throughput =latency - Instruction latency =code-size - Code size =size-latency - Code size and latency =all - Print all cost kinds | help |
--crel | - Use CREL relocation format for ELF | help |
--cs-profile-generate | - Perform context sensitive PGO instrumentation | help |
--cs-profile-path string | - Context sensitive profile file path | help |
--ctx-prof-include-empty | - Also write profiles with all-zero counters. | help |
--ctx-profile-force-is-specialized | - Treat the given module as-if it were containing the post-thinlink module containing the root | help |
--debugify-atoms | - | help |
--debugify-func-limit ulong | - Set max number of processed functions per pass. | help |
--debugify-level value | - Kind of debug info to add =locations - Locations only =location+variables - Locations and Variables | help |
--debugify-quiet | - Suppress verbose debugify output | help |