Add rules for BigFloat integer constructors

[kshyatt]

BigFloat(x::Clong) and BigFloat(x::Culong) lower to mpfr_set_si / mpfr_set_ui, for which there is no derivative, so any active use of an integer-constructed BigFloat failed with EnzymeNoDerivativeError.

The argument is an integer, so the result never carries a derivative. This cannot be expressed as inactive, though: the result is a fresh mutable BigFloat that subsequent MPFR calls may write into, so it still needs a shadow to exist -- the same reason the existing no-argument BigFloat() rule allocates one. The shadow is built through the same constructor with a zero value, so it lands at the requested precision as a properly initialised MPFR value rather than raw zeroed memory.

Having a rule here also keeps Enzyme from compiling the body of these constructors at all, which matters beyond the missing derivative: eps(::Type{BigFloat}) is nextfloat(BigFloat(1)) - BigFloat(1), and compiling that construction inside an Enzyme-generated function miscompiles it into a BigFloat with an invalid d::Memory, segfaulting on first use. This repro segfaults before the change and passes after:

f(out) = (out[1] = eps(BigFloat) > 0 ? 1.0 : 0.0; nothing)
Enzyme.autodiff(Forward, Const(f), Const(zeros(3)))

Developed with 🤖

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic main) to apply these changes.

Click here to view the suggested changes.

diff --git a/src/internal_rules/bigfloat.jl b/src/internal_rules/bigfloat.jl
index 1c2c93e0..c8454596 100644
--- a/src/internal_rules/bigfloat.jl
+++ b/src/internal_rules/bigfloat.jl
@@ -71,12 +71,12 @@ end
 # for which there is no derivative, so any active use of an integer-constructed
 # BigFloat otherwise fails with `EnzymeNoDerivativeError`.
 function EnzymeRules.forward(
-    config::EnzymeRules.FwdConfig,
-    Ty::Const{Type{BigFloat}},
-    RT::Type{<:Union{DuplicatedNoNeed,Duplicated,BatchDuplicated,BatchDuplicatedNoNeed}},
-    x::Const{Ti},
-    rs::Const{Base.MPFR.MPFRRoundingMode}...;
-    kwargs...,
+        config::EnzymeRules.FwdConfig,
+        Ty::Const{Type{BigFloat}},
+        RT::Type{<:Union{DuplicatedNoNeed, Duplicated, BatchDuplicated, BatchDuplicatedNoNeed}},
+        x::Const{Ti},
+        rs::Const{Base.MPFR.MPFRRoundingMode}...;
+        kwargs...,
     ) where {Ti <: Union{Clong, Culong}}
     rvals = map(r -> r.val, rs)
     if EnzymeRules.needs_primal(config) && EnzymeRules.needs_shadow(config)
@@ -109,12 +109,12 @@ function EnzymeRules.forward(
 end
 
 function EnzymeRules.augmented_primal(
-    config::EnzymeRules.RevConfig,
-    Ty::Const{Type{BigFloat}},
-    RT::Type{<:Union{DuplicatedNoNeed,Duplicated,BatchDuplicated,BatchDuplicatedNoNeed}},
-    x::Const{Ti},
-    rs::Const{Base.MPFR.MPFRRoundingMode}...;
-    kwargs...,
+        config::EnzymeRules.RevConfig,
+        Ty::Const{Type{BigFloat}},
+        RT::Type{<:Union{DuplicatedNoNeed, Duplicated, BatchDuplicated, BatchDuplicatedNoNeed}},
+        x::Const{Ti},
+        rs::Const{Base.MPFR.MPFRRoundingMode}...;
+        kwargs...,
     ) where {Ti <: Union{Clong, Culong}}
     rvals = map(r -> r.val, rs)
     primal = if EnzymeRules.needs_primal(config)
@@ -136,14 +136,14 @@ function EnzymeRules.augmented_primal(
 end
 
 function EnzymeRules.reverse(
-    config::EnzymeRules.RevConfig,
-    Ty::Const{Type{BigFloat}},
-    RT::Type{<:Union{DuplicatedNoNeed,Duplicated,BatchDuplicated,BatchDuplicatedNoNeed}},
-    tape,
-    x::Const{<:Union{Culong, Clong}},
-    rs::Const{Base.MPFR.MPFRRoundingMode}...;
-    kwargs...,
-)
+        config::EnzymeRules.RevConfig,
+        Ty::Const{Type{BigFloat}},
+        RT::Type{<:Union{DuplicatedNoNeed, Duplicated, BatchDuplicated, BatchDuplicatedNoNeed}},
+        tape,
+        x::Const{<:Union{Culong, Clong}},
+        rs::Const{Base.MPFR.MPFRRoundingMode}...;
+        kwargs...,
+    )
     # the integer argument carries no derivative
     return ntuple(Returns(nothing), Val(1 + length(rs)))
 end

[github-actions]

Benchmark Results

	main	2223e29...	main / 2223e29...
basics/make_zero/namedtuple	0.0529 ± 0.0024 μs	0.0546 ± 0.0021 μs	0.969 ± 0.058
basics/make_zero/struct	0.259 ± 0.0067 μs	0.268 ± 0.0069 μs	0.963 ± 0.035
basics/overhead	5.26 ± 0.01 ns	4.33 ± 0.01 ns	1.21 ± 0.0036
basics/remake_zero!/namedtuple	0.224 ± 0.0052 μs	0.225 ± 0.0085 μs	0.992 ± 0.044
basics/remake_zero!/struct	0.225 ± 0.0062 μs	0.225 ± 0.0067 μs	1 ± 0.041
fold_broadcast/multidim_sum_bcast/1D	10.3 ± 1.8 μs	10.3 ± 0.46 μs	0.994 ± 0.18
fold_broadcast/multidim_sum_bcast/2D	10.2 ± 0.16 μs	10.2 ± 0.2 μs	1 ± 0.025
time_to_load	1.52 ± 0.015 s	1.52 ± 0.018 s	1 ± 0.015

Benchmark Plots

A plot of the benchmark results has been uploaded as an artifact at https://github.com/EnzymeAD/Enzyme.jl/actions/runs/36103894614/artifacts/10850443528.

[codecov]

Codecov Report

❌ Patch coverage is 24.24242% with 25 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.44%. Comparing base (481fb6f) to head (c2ed260).

Files with missing lines	Patch %	Lines
src/internal_rules/bigfloat.jl	24.24%	25 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3461      +/-   ##
==========================================
+ Coverage   77.43%   78.44%   +1.01%     
==========================================
  Files          68       68              
  Lines       23278    22966     -312     
==========================================
- Hits        18025    18016       -9     
+ Misses       5253     4950     -303     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[wsmoses]

closures are bad

[vchuravy]

closures are bad

Citation needed?