Frontend clang compilation error: redefinition of 'v2_filled'

[asraa]

I got a really strange error from the openfhe backend:

Traceback (most recent call last):
  File "/usr/local/google/home/asraa/git/heir/frontend/conditional_test.py", line 11, in test_cond
    @compile(debug=True)
     ^^^^^^^^^^^^^^^^^^^
  File "/usr/local/google/home/asraa/git/heir/frontend/heir/pipeline.py", line 223, in decorator
    return run_pipeline(
           ^^^^^^^^^^^^^
  File "/usr/local/google/home/asraa/git/heir/frontend/heir/pipeline.py", line 159, in run_pipeline
    result = backend.run_backend(
             ^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/google/home/asraa/git/heir/frontend/heir/backends/openfhe/backend.py", line 267, in run_backend
    clang_backend.compile_to_shared_object(
  File "/usr/local/google/home/asraa/git/heir/frontend/heir/backends/util/clang.py", line 102, in compile_to_shared_object
    raise ValueError(message)
ValueError: Error running /usr/lib/llvm-19/bin/clang. stderr was:

/tmp/tmp7em30otb/cond.cpp:49:8: error: redefinition of 'v2_filled_n'
   49 |   auto v2_filled_n = cc->GetCryptoParameters()->GetElementParams()->GetRingDimension() / 2;
      |        ^
/tmp/tmp7em30otb/cond.cpp:16:8: note: previous definition is here
   16 |   auto v2_filled_n = cc->GetCryptoParameters()->GetElementParams()->GetRingDimension() / 2;
      |        ^
/tmp/tmp7em30otb/cond.cpp:50:8: error: redefinition of 'v2_filled'
   50 |   auto v2_filled = v2;
      |        ^
/tmp/tmp7em30otb/cond.cpp:17:8: note: previous definition is here
   17 |   auto v2_filled = v2;
      |        ^
2 errors generated.


options were:

/usr/lib/llvm-19/bin/clang /tmp/tmp7em30otb/cond.cpp -o /tmp/tmp7em30otb/cond.so -O3 -fPIC -shared -std=c++17 -I/usr/local/include/openfhe -I/usr/local/include/openfhe/binfhe -I/usr/local/include/openfhe/core -I/usr/local/include/openfhe/pke -L/usr/local/lib -lOPENFHEbinfhe -lOPENFHEcore -lOPENFHEpke

----------------------------------------------------------------------
Ran 1 test in 4.666s

The code it was trying to compile did use v2_filled twice, but within two separate function scopes: https://gist.github.com/asraa/5d85f3243e97d6732b3385b1b0c76de5

So that makes me think maybe some options or something needs to change with the call to compile?

[asraa]

Hah nevermind they are in the same function scope!! It's just two uses of that value.

I feel like this extra filling code should be done at the openfhe IR level so we can CSE it away instead of the emitter.

[asraa]

Here was the backend MLIR

!Z1073750017_i64 = !mod_arith.int<1073750017 : i64>
!Z65537_i64 = !mod_arith.int<65537 : i64>
!Z67239937_i64 = !mod_arith.int<67239937 : i64>
!cc = !openfhe.crypto_context
!params = !openfhe.cc_params
!pk = !openfhe.public_key
!sk = !openfhe.private_key
#full_crt_packing_encoding = #lwe.full_crt_packing_encoding<scaling_factor = 0>
#key = #lwe.key<>
#modulus_chain_L1_C0 = #lwe.modulus_chain<elements = <67239937 : i64, 1073750017 : i64>, current = 0>
#modulus_chain_L1_C1 = #lwe.modulus_chain<elements = <67239937 : i64, 1073750017 : i64>, current = 1>
!rns_L0 = !rns.rns<!Z67239937_i64>
!rns_L1 = !rns.rns<!Z67239937_i64, !Z1073750017_i64>
#ring_Z65537_i64_1_x1024 = #polynomial.ring<coefficientType = !Z65537_i64, polynomialModulus = <1 + x**1024>>
#ring_rns_L0_1_x1024 = #polynomial.ring<coefficientType = !rns_L0, polynomialModulus = <1 + x**1024>>
#ring_rns_L1_1_x1024 = #polynomial.ring<coefficientType = !rns_L1, polynomialModulus = <1 + x**1024>>
!pt = !lwe.new_lwe_plaintext<application_data = <message_type = i1>, plaintext_space = <ring = #ring_Z65537_i64_1_x1024, encoding = #full_crt_packing_encoding>>
!pt1 = !lwe.new_lwe_plaintext<application_data = <message_type = i64>, plaintext_space = <ring = #ring_Z65537_i64_1_x1024, encoding = #full_crt_packing_encoding>>
#ciphertext_space_L0 = #lwe.ciphertext_space<ring = #ring_rns_L0_1_x1024, encryption_type = lsb>
#ciphertext_space_L1 = #lwe.ciphertext_space<ring = #ring_rns_L1_1_x1024, encryption_type = lsb>
!ct_L0 = !lwe.new_lwe_ciphertext<application_data = <message_type = i64>, plaintext_space = <ring = #ring_Z65537_i64_1_x1024, encoding = #full_crt_packing_encoding>, ciphertext_space = #ciphertext_space_L0, key = #key, modulus_chain = #modulus_chain_L1_C0>
!ct_L1 = !lwe.new_lwe_ciphertext<application_data = <message_type = i1>, plaintext_space = <ring = #ring_Z65537_i64_1_x1024, encoding = #full_crt_packing_encoding>, ciphertext_space = #ciphertext_space_L1, key = #key, modulus_chain = #modulus_chain_L1_C1>
!ct_L1_1 = !lwe.new_lwe_ciphertext<application_data = <message_type = i64>, plaintext_space = <ring = #ring_Z65537_i64_1_x1024, encoding = #full_crt_packing_encoding>, ciphertext_space = #ciphertext_space_L1, key = #key, modulus_chain = #modulus_chain_L1_C1>
module attributes {scheme.bgv} {
  func.func @cond(%cc: !cc, %arg0: i64, %arg1: i64, %ct: !ct_L1) -> !ct_L0 {
    %cst = arith.constant dense<1> : tensor<1024xi64>
    %pt = openfhe.make_packed_plaintext %cc, %cst : (!cc, tensor<1024xi64>) -> !pt
    %ct_0 = openfhe.negate %cc, %ct : (!cc, !ct_L1) -> !ct_L1
    %ct_1 = openfhe.add_plain %cc, %ct_0, %pt : (!cc, !ct_L1, !pt) -> !ct_L1
    %ct_2 = lwe.reinterpret_application_data %ct : !ct_L1 to !ct_L1_1
    %splat = tensor.splat %arg0 : tensor<1024xi64>
    %pt_3 = openfhe.make_packed_plaintext %cc, %splat : (!cc, tensor<1024xi64>) -> !pt1
    %ct_4 = openfhe.mul_plain %cc, %ct_2, %pt_3 : (!cc, !ct_L1_1, !pt1) -> !ct_L1_1
    %ct_5 = lwe.reinterpret_application_data %ct_1 : !ct_L1 to !ct_L1_1
    %splat_6 = tensor.splat %arg1 : tensor<1024xi64>
    %pt_7 = openfhe.make_packed_plaintext %cc, %splat_6 : (!cc, tensor<1024xi64>) -> !pt1
    %ct_8 = openfhe.mul_plain %cc, %ct_5, %pt_7 : (!cc, !ct_L1_1, !pt1) -> !ct_L1_1
    %ct_9 = openfhe.add %cc, %ct_4, %ct_8 : (!cc, !ct_L1_1, !ct_L1_1) -> !ct_L1_1
    %pt_10 = openfhe.make_packed_plaintext %cc, %cst : (!cc, tensor<1024xi64>) -> !pt1
    %ct_11 = openfhe.add_plain %cc, %ct_9, %pt_10 : (!cc, !ct_L1_1, !pt1) -> !ct_L1_1
    %ct_12 = openfhe.mod_reduce %cc, %ct_11 : (!cc, !ct_L1_1) -> !ct_L0
    return %ct_12 : !ct_L0
  }
  func.func @cond__encrypt__arg2(%cc: !cc, %arg0: i1, %pk: !pk) -> !ct_L1 {
    %splat = tensor.splat %arg0 : tensor<1024xi1>
    %0 = arith.extui %splat : tensor<1024xi1> to tensor<1024xi64>
    %pt = openfhe.make_packed_plaintext %cc, %0 : (!cc, tensor<1024xi64>) -> !pt
    %ct = openfhe.encrypt %cc, %pt, %pk : (!cc, !pt, !pk) -> !ct_L1
    return %ct : !ct_L1
  }
  func.func @cond__decrypt__result0(%cc: !cc, %ct: !ct_L0, %sk: !sk) -> i64 {
    %pt = openfhe.decrypt %cc, %ct, %sk : (!cc, !ct_L0, !sk) -> !pt1
    %0 = lwe.rlwe_decode %pt {encoding = #full_crt_packing_encoding, ring = #ring_Z65537_i64_1_x1024} : !pt1 -> i64
    return %0 : i64
  }
  func.func @cond__generate_crypto_context() -> !cc {
    %params = openfhe.gen_params  {encryptionTechniqueExtended = false, evalAddCount = 2 : i64, insecure = false, keySwitchCount = 0 : i64, mulDepth = 1 : i64, plainMod = 65537 : i64} : () -> !params
    %cc = openfhe.gen_context %params {supportFHE = false} : (!params) -> !cc
    return %cc : !cc
  }
  func.func @cond__configure_crypto_context(%cc: !cc, %sk: !sk) -> !cc {
    return %cc : !cc
  }
}

So it seems that the plaintext was generated with two different levels.

[AlexanderViand-Intel]

I feel like this extra filling code should be done at the openfhe IR level so we can CSE it away instead of the emitter.

+1
Beyond CSE/avoiding this bug, I'm interested in hoisting any application data -> plaintext conversion code outside the function (i.e.,rewriting the function to take in pt2 and pt3 instead, and adding make_pt2 and make_pt3 client side functions, and having this stuff happen in the IR rather than in the emitter seems like it'd make that easier :)

[ZenithalHourlyRate]

v2_filled exists purely out of the need of implicit cyclic repetition assumption. With the layout packing pass making it obvious, such variable should vanish one day.

[ZenithalHourlyRate]

heir/lib/Target/Lattigo/LattigoEmitter.cpp

Lines 407 to 408 in 56d6fbf

	auto packedName =
	getName(op.getValue()) + "_" + getName(op.getPlaintext()) + "_packed";

A workaround is specifying both the pt and target ct in the name to make it unique.