Run scheme-agnostic passes to convert FHE programs that operate on scalar types to equivalent programs that operate on vectors.

This pass is intended to process FHE programs that are known to be good for SIMD, but a specific FHE scheme has not yet been chosen. It expects to handle arith ops operating on tensor types (with or without secret.generic).

The pass unrolls all loops, then applies a series of passes that convert scalar operations on tensor elements to SIMD operations on full tensors. This uses the FHE computational model common to BGV, BFV, and CKKS, in which data is packed in polynomial ciphertexts, interpreted as vectors of individual data elements, and arithmetic can be applied across entire ciphertexts, with some limited support for rotations via automorphisms of the underlying ring.

Along the way, this pipeline applies heuristic optimizations to minimize the number of rotations needed, relying on the implicit cost model that rotations are generally expensive. The specific set of passes can be found in tools/heir-opt.cpp where the pipeline is defined.


Lowers a TOSA MLIR model to func, arith, and memref.

Lowers from TOSA through linalg and affine, and converts all tensors to memrefs. Fully unrolls all loops, and forwards stores to subsequent loads whenever possible. The output is suitable as an input to heir-translate --emit-verilog. Retains affine.load and affine.store ops that cannot be removed (e.g., reading from the input and writing to the output, or loading from a memref with a variable index).

The pass pipeline assumes that the input is a valid TOSA MLIR model with stripped quantized types. The iree-import-tflite tool can lower a TFLite FlatBuffer to textual MLIR with --output-format=mlir-ir. See hello_world.tosa.mlir for an example.


Uses Yosys to booleanize and optimize MLIR functions.

This pass pipeline requires inputs to be in standard MLIR (arith, affine, func, memref). The pass imports the model to Yosys and runs passes to booleanize the circuit and then uses ABC to perform optimizations. We use standard LUT 3 cells. THe output of this pass includes arith constants and comb.truth_table ops.

The pass requires that the environment variable HEIR_ABC_BINARY contains the location of the ABC binary and that HEIR_YOSYS_SCRIPTS_DIR contains the location of the Yosys’ techlib files that are needed to execute the path.

This pass can be disabled by defining HEIR_NO_YOSYS; this will avoid Yosys library and ABC binary compilation, and avoid registration of this pass.


This is an experimental pipeline for end-to-end private inference.

Converts a TOSA MLIR model to tfhe_rust dialect defined by HEIR. It converts a tosa model to optimized boolean circuit using Yosys ABC optimizations. The resultant optimized boolean circuit in comb dialect is then converted to cggi and then to tfhe_rust exit dialect. This pipeline can be used with heir-translate –emit-tfhe-rust to generate code for tfhe-rs FHE library.

The pass requires that the environment variable HEIR_ABC_BINARY contains the location of the ABC binary and that HEIR_YOSYS_SCRIPTS_DIR contains the location of the Yosys’ techlib files that are needed to execute the path.



Code generation for the tfhe-rs FHE library. The library is based on the CGGI cryptosystem, and so this pass is most useful when paired with lowerings from the cggi dialect.

The version of tfhe-rs supported is defined in the end to end tfhe_rust tests.

Example input:

!sks = !tfhe_rust.server_key
!lut = !tfhe_rust.lookup_table
!eui3 = !tfhe_rust.eui3

func.func @test_apply_lookup_table(%sks : !sks, %lut: !lut, %input : !eui3) -> !eui3 {
  %v1 = tfhe_rust.apply_lookup_table %sks, %input, %lut : (!sks, !eui3, !lut) -> !eui3
  %v2 = tfhe_rust.add %sks, %input, %v1 : (!sks, !eui3, !eui3) -> !eui3
  %c1 = arith.constant 1 : i8
  %v3 = tfhe_rust.scalar_left_shift %sks, %v2, %c1 : (!sks, !eui3, i8) -> !eui3
  %v4 = tfhe_rust.apply_lookup_table %sks, %v3, %lut : (!sks, !eui3, !lut) -> !eui3
  return %v4 : !eui3

Example output:

use tfhe::shortint::prelude::*;

pub fn test_apply_lookup_table(
  v9: &ServerKey,
  v10: &LookupTableOwned,
  v11: &Ciphertext,
) -> Ciphertext {
  let v4 = v9.apply_lookup_table(&v11, &v10);
  let v5 = v9.unchecked_add(&v11, &v4);
  let v6 = 1;
  let v7 = v9.scalar_left_shift(&v5, v6);
  let v8 = v9.apply_lookup_table(&v7, &v10);

Note, the chosen variable names are arbitrary, and the resulting program still must be integrated with a larger Rust program.


Code generation for verilog from arith and memref. Expects a single top level func.func op as the entry point, which is converted to the output verilog module.

Example input:

module {
  func.func @main(%arg0: i8) -> (i8) {
    %c0 = arith.constant 0 : i32
    %c1 = arith.constant 1 : i32
    %c2 = arith.constant 2 : i32
    %c3 = arith.constant 3 : i32
    %0 = arith.extsi %arg0 : i8 to i32
    %1 = arith.subi %0, %c1 : i32
    %2 = arith.muli %1, %c2 : i32
    %3 = arith.addi %2, %c3 : i32
    %4 = arith.cmpi sge, %2, %c0 : i32
    %5 = arith.select %4, %c1, %c2 : i32
    %6 = arith.shrsi %3, %c1 : i32
    %7 = arith.shrui %3, %c1 : i32
    %out = arith.trunci %6 : i32 to i8
    return %out : i8


module main(
  input wire signed [7:0] arg1,
  output wire signed [7:0] _out_
  wire signed [31:0] v2;
  wire signed [31:0] v3;
  wire signed [31:0] v4;
  wire signed [31:0] v5;
  wire signed [31:0] v6;
  wire signed [31:0] v7;
  wire signed [31:0] v8;
  wire signed [31:0] v9;
  wire v10;
  wire signed [31:0] v11;
  wire signed [31:0] v12;
  wire signed [31:0] v13;
  wire signed [7:0] v14;

  assign v2 = 0;
  assign v3 = 1;
  assign v4 = 2;
  assign v5 = 3;
  assign v6 = {{24{arg1[7]}}, arg1};
  assign v7 = v6 - v3;
  assign v8 = v7 * v4;
  assign v9 = v8 + v5;
  assign v10 = v8 >= v2;
  assign v11 = v10 ? v3 : v4;
  assign v12 = v9 >>> v3;
  assign v13 = v9 >> v3;
  assign v14 = v12[7:0];
  assign _out_ = v14;


Prints a json object describing the function signatures. Used for code generation after --emit-verilog.

Example input:

module {
  func.func @main(%arg0: memref<80xi8>) -> memref<1x3x2x1xi8> {
    %alloc_0 = memref.alloc() {alignment = 64 : i64} : memref<1x3x2x1xi8>
    return %alloc_0 : memref<1x3x2x1xi8>

Example output:

  "functions": [
      "name": "main",
      "params": [
          "index": 0,
          "type": {
            "memref": {
              "element_type": {
                "integer": {
                  "is_signed": false,
                  "width": 8
              "shape": [80]
      "return_types": [{
        "memref": {
          "element_type": {
            "integer": {
              "is_signed": false,
              "width": 8
          "shape": [1, 3, 2, 1]