fix(compat): pin SinCosLUT to ~3.2 to dodge Julia 1.10 SIMD codegen crash

[zsoerenm]

Problem

The Julia 1.10 - ubuntu-latest CI job intermittently (~85% of runs) crashes during test codegen:

error: couldn't allocate output register for constraint 'x'

This is an LLVM 15 register-allocation failure on 16-register AVX2 targets while compiling the Int16 / bit-wise downconvert kernels — a compiler crash, not a test failure. It affects the whole branch (reproduced on unmodified base) and blocks #194.

Root cause

downconvert_and_correlate_int16.jl (and the one-/two-bit kernels) call SinCosLUT.generate_carrier_signs!. SinCosLUT v3.3.0 (released 2026-07-08) rewrote that function with new AVX2/AVX-512 sign-pack SIMD (vpmovmskb/vpmovb2m, per its changelog: "branch-free flip loop and table-free sign packs"). That inline-asm codegen trips the LLVM 15 register allocator on Julia 1.10.

CI last passed on 2026-07-04, when the SinCosLUT = "3" range resolved to 3.2.0 (the pre-rewrite generate_carrier_signs!). Pkg.test re-resolves dependencies per run, so runs after 2026-07-08 pull 3.3.0 and crash.

Fix

Pin SinCosLUT = "~3.2" ([3.2.0, 3.3.0)) to restore the last-known-good SIMD path on Julia 1.10. Every SinCosLUT symbol this package uses exists in 3.2.x (base built and passed with 3.2.0), so nothing else changes.

This is a workaround. The proper fix is in SinCosLUT's 3.3.0 codegen (make the new sign-pack SIMD survive LLVM 15 / Julia 1.10 register allocation), or dropping Julia 1.10 from support. Lift the pin once SinCosLUT ships a 1.10-safe release.

Verification

Since the crash is AVX2-specific and this box only offers AVX-512 (32 registers, which hides it) or a non-native target (which trips an unrelated SinCosLUT fallback bug), it can't be reproduced locally — verified via CI: this PR's Julia 1.10 job should pass reliably across re-runs, vs ~85% failure on the unpinned base.

🤖 Generated with Claude Code

[zsoerenm]

Verified: pin makes Julia 1.10 reliably green

Ran the Julia 1.10 - ubuntu-latest job 3 times in a row on this pinned build — all passed. Against the unpinned base's ~85% failure rate, three consecutive passes is ~0.3% by luck, so the SinCosLUT = "~3.2" pin reliably clears the LLVM 15 register-allocation crash.

[github-actions]

Benchmark Results (minimum time) — macos-14

Reporting the minimum over all samples (robust to shared-runner contention), not the median.

Alternative backends vs Float32 (track!, PR head)

Legend — backends: F32 Float32 (default) · I16 Int16 · 1b OneBit · 2b TwoBit (2-bit measurement + 2-bit carrier). Time columns are the minimum track! time; ×B = F32 / B (so >1 ⇒ backend B is faster than Float32), ✅ ≥ 5 % faster, ⚠️ ≥ 5 % slower. 1b/2b are BPSK-only, so their cells are blank for CBOC (Galileo E1B) scenarios.

Scenario	F32	I16	1b	2b	×I16	×1b	×2b
4-antenna @ 5 MHz	11.9 μs	7.76 μs	4.3 μs	8.22 μs	1.53 ✅	2.76 ✅	1.44 ✅
GPS L1CA, 8 sats @ 40 MHz	371.0 μs	144.0 μs	63.6 μs	97.5 μs	2.58 ✅	5.83 ✅	3.8 ✅
GPS L1CA, 8 sats @ 5 MHz	53.1 μs	24.1 μs	15.7 μs	20.8 μs	2.2 ✅	3.39 ✅	2.56 ✅
Galileo E1B, 4 sats @ 25 MHz	141.0 μs	69.3 μs			2.04 ✅
dynamic taps @ 5 MHz (kernel)	6.68 μs	2.56 μs	1.53 μs		2.62 ✅	4.36 ✅
multi-signal N=3 @ 5 MHz	10.7 μs	6.14 μs	3.93 μs	8.08 μs	1.74 ✅	2.71 ✅	1.32 ✅

Time benchmarks (base vs PR head)

Ratio = 81e5299… / e661bc1…: >1 means the PR is faster. ✅ ≥ 5 % faster, ⚠️ ≥ 5 % slower. A blank cell means the benchmark exists on only one revision (🆕 = new on the PR, 🗑 = removed).

	81e5299…	e661bc1…	81e5299… / e661bc1…
downconvert and correlate/CPU/Float32	2.5 μs	2.51 μs	0.996
downconvert and correlate/CPU/Float32 4ant	4.94 μs	5.07 μs	0.974
downconvert and correlate/CPU/Float64	2.71 μs	2.69 μs	1.01
downconvert and correlate/CPU/Int16	2.72 μs	2.62 μs	1.04
downconvert and correlate/CPU/Int16 4ant	4.88 μs	5.04 μs	0.969
downconvert and correlate/CPU/Int32	2.64 μs	2.62 μs	1.01
fused kernel/1-ant dynamic taps	2.56 μs	2.57 μs	0.996
fused kernel/1-ant static taps	2.13 μs	2.13 μs	1.0
fused kernel/4-ant dynamic taps	7.78 μs	7.81 μs	0.996
fused kernel/4-ant static taps	4.43 μs	4.44 μs	0.999
fused tuple kernel/1-ant N=2	3.28 μs	3.27 μs	1.0
fused tuple kernel/1-ant N=3	3.41 μs	3.41 μs	1.0
fused tuple kernel/2-ant N=2	6.32 μs	6.33 μs	0.999
fused tuple kernel/2-ant N=3	5.62 μs	5.79 μs	0.97
fused tuple kernel/4-ant N=2	11.6 μs	12.0 μs	0.969
fused tuple kernel/4-ant N=3	10.8 μs	10.5 μs	1.03
track/1. Float32/2K – track	2.75 μs	2.66 μs	1.03
track/1. Float32/2K – track!	2.92 μs	2.78 μs	1.05
track/2. L1 8sat/5K – track	50.5 μs	49.9 μs	1.01
track/2. L1 8sat/5K – track!	49.7 μs	49.4 μs	1.01
track/2. L1 8sat/5K – track!-threaded	51.9 μs	50.0 μs	1.04
track/2. L1 8sat/5K – track-threaded	50.3 μs	51.8 μs	0.97
track/3. E1B 4sat/25K – track	138.0 μs	136.0 μs	1.02
track/3. E1B 4sat/25K – track!	135.0 μs	136.0 μs	0.998
track/3. E1B 4sat/25K – track!-threaded	136.0 μs	136.0 μs	0.999
track/3. E1B 4sat/25K – track-threaded	136.0 μs	136.0 μs	1.0
track/4. 8L1+8E1B/25K – track	491.0 μs	490.0 μs	1.0
track/4. 8L1+8E1B/25K – track!	491.0 μs	490.0 μs	1.0
track/4. 8L1+8E1B/25K – track!-threaded	493.0 μs	491.0 μs	1.01
track/4. 8L1+8E1B/25K – track-threaded	492.0 μs	493.0 μs	0.996
track/5. multi-signal N=1/5K – track	6.56 μs	6.58 μs	0.997
track/5. multi-signal N=1/5K – track!	6.56 μs	6.72 μs	0.975
track/6. multi-signal N=2/5K – track	11.0 μs	10.6 μs	1.04
track/6. multi-signal N=2/5K – track!	10.9 μs	11.0 μs	0.996
track/7. multi-signal N=3/5K – track	11.9 μs	11.8 μs	1.01
track/7. multi-signal N=3/5K – track!	12.2 μs	12.3 μs	0.997
time_to_load	175.0 μs	304.0 μs	0.576 ⚠️

Memory benchmarks (base vs PR head)

	81e5299…	e661bc1…
downconvert and correlate/CPU/Float32	2 allocs: 576 B	2 allocs: 576 B
downconvert and correlate/CPU/Float32 4ant	2 allocs: 848 B	2 allocs: 848 B
downconvert and correlate/CPU/Float64	2 allocs: 576 B	2 allocs: 576 B
downconvert and correlate/CPU/Int16	2 allocs: 576 B	2 allocs: 576 B
downconvert and correlate/CPU/Int16 4ant	2 allocs: 848 B	2 allocs: 848 B
downconvert and correlate/CPU/Int32	2 allocs: 576 B	2 allocs: 576 B
fused kernel/1-ant dynamic taps	0 allocs: 0 B	0 allocs: 0 B
fused kernel/1-ant static taps	0 allocs: 0 B	0 allocs: 0 B
fused kernel/4-ant dynamic taps	0 allocs: 0 B	0 allocs: 0 B
fused kernel/4-ant static taps	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/1-ant N=2	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/1-ant N=3	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/2-ant N=2	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/2-ant N=3	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/4-ant N=2	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/4-ant N=3	0 allocs: 0 B	0 allocs: 0 B
track/1. Float32/2K – track	9 allocs: 944 B	9 allocs: 944 B
track/1. Float32/2K – track!	1 allocs: 26 B	1 allocs: 26 B
track/2. L1 8sat/5K – track	10 allocs: 4656 B	10 allocs: 4656 B
track/2. L1 8sat/5K – track!	0 allocs: 0 B	0 allocs: 0 B
track/2. L1 8sat/5K – track!-threaded	0 allocs: 0 B	0 allocs: 0 B
track/2. L1 8sat/5K – track-threaded	10 allocs: 4656 B	10 allocs: 4656 B
track/3. E1B 4sat/25K – track	10 allocs: 3056 B	10 allocs: 3056 B
track/3. E1B 4sat/25K – track!	0 allocs: 0 B	0 allocs: 0 B
track/3. E1B 4sat/25K – track!-threaded	0 allocs: 0 B	0 allocs: 0 B
track/3. E1B 4sat/25K – track-threaded	10 allocs: 3056 B	10 allocs: 3056 B
track/4. 8L1+8E1B/25K – track	26 allocs: 10464 B	26 allocs: 10464 B
track/4. 8L1+8E1B/25K – track!	0 allocs: 0 B	0 allocs: 0 B
track/4. 8L1+8E1B/25K – track!-threaded	0 allocs: 0 B	0 allocs: 0 B
track/4. 8L1+8E1B/25K – track-threaded	26 allocs: 10464 B	26 allocs: 10464 B
track/5. multi-signal N=1/5K – track	12 allocs: 1024 B	12 allocs: 1024 B
track/5. multi-signal N=1/5K – track!	3 allocs: 80 B	3 allocs: 80 B
track/6. multi-signal N=2/5K – track	15 allocs: 1568 B	15 allocs: 1568 B
track/6. multi-signal N=2/5K – track!	6 allocs: 160 B	6 allocs: 160 B
track/7. multi-signal N=3/5K – track	18 allocs: 2000 B	18 allocs: 2000 B
track/7. multi-signal N=3/5K – track!	9 allocs: 240 B	9 allocs: 240 B
time_to_load	196 allocs: 13984 B	196 allocs: 13984 B

[github-actions]

Benchmark Results (minimum time) — ubuntu-latest

Reporting the minimum over all samples (robust to shared-runner contention), not the median.

Alternative backends vs Float32 (track!, PR head)

Legend — backends: F32 Float32 (default) · I16 Int16 · 1b OneBit · 2b TwoBit (2-bit measurement + 2-bit carrier). Time columns are the minimum track! time; ×B = F32 / B (so >1 ⇒ backend B is faster than Float32), ✅ ≥ 5 % faster, ⚠️ ≥ 5 % slower. 1b/2b are BPSK-only, so their cells are blank for CBOC (Galileo E1B) scenarios.

Scenario	F32	I16	1b	2b	×I16	×1b	×2b
4-antenna @ 5 MHz	12.2 μs	6.3 μs	5.26 μs	9.72 μs	1.94 ✅	2.32 ✅	1.26 ✅
GPS L1CA, 8 sats @ 40 MHz	331.0 μs	129.0 μs	76.0 μs	111.0 μs	2.56 ✅	4.36 ✅	2.97 ✅
GPS L1CA, 8 sats @ 5 MHz	49.1 μs	27.6 μs	20.6 μs	27.8 μs	1.78 ✅	2.39 ✅	1.77 ✅
Galileo E1B, 4 sats @ 25 MHz	123.0 μs	55.0 μs			2.23 ✅
dynamic taps @ 5 MHz (kernel)	5.77 μs	3.34 μs	2.04 μs		1.73 ✅	2.83 ✅
multi-signal N=3 @ 5 MHz	9.55 μs	5.78 μs	5.21 μs	9.3 μs	1.65 ✅	1.83 ✅	1.03

Time benchmarks (base vs PR head)

Ratio = 81e5299… / e661bc1…: >1 means the PR is faster. ✅ ≥ 5 % faster, ⚠️ ≥ 5 % slower. A blank cell means the benchmark exists on only one revision (🆕 = new on the PR, 🗑 = removed).

	81e5299…	e661bc1…	81e5299… / e661bc1…
downconvert and correlate/CPU/Float32	2.34 μs	2.34 μs	1.0
downconvert and correlate/CPU/Float32 4ant	4.77 μs	4.73 μs	1.01
downconvert and correlate/CPU/Float64	2.76 μs	2.76 μs	1.0
downconvert and correlate/CPU/Int16	2.43 μs	2.41 μs	1.01
downconvert and correlate/CPU/Int16 4ant	5.19 μs	5.2 μs	0.998
downconvert and correlate/CPU/Int32	2.4 μs	2.41 μs	0.997
fused kernel/1-ant dynamic taps	2.15 μs	2.16 μs	0.999
fused kernel/1-ant static taps	1.88 μs	1.88 μs	1.0
fused kernel/4-ant dynamic taps	5.27 μs	5.28 μs	0.998
fused kernel/4-ant static taps	4.15 μs	4.14 μs	1.0
fused tuple kernel/1-ant N=2	2.35 μs	2.35 μs	1.0
fused tuple kernel/1-ant N=3	2.82 μs	2.85 μs	0.989
fused tuple kernel/2-ant N=2	3.62 μs	3.61 μs	1.0
fused tuple kernel/2-ant N=3	4.67 μs	4.54 μs	1.03
fused tuple kernel/4-ant N=2	6.01 μs	6.07 μs	0.99
fused tuple kernel/4-ant N=3	9.72 μs	9.74 μs	0.998
track/1. Float32/2K – track	2.57 μs	2.55 μs	1.01
track/1. Float32/2K – track!	2.65 μs	2.65 μs	0.998
track/2. L1 8sat/5K – track	48.5 μs	48.0 μs	1.01
track/2. L1 8sat/5K – track!	47.0 μs	47.2 μs	0.996
track/2. L1 8sat/5K – track!-threaded	47.2 μs	47.3 μs	0.999
track/2. L1 8sat/5K – track-threaded	47.8 μs	48.1 μs	0.994
track/3. E1B 4sat/25K – track	117.0 μs	120.0 μs	0.982
track/3. E1B 4sat/25K – track!	117.0 μs	119.0 μs	0.982
track/3. E1B 4sat/25K – track!-threaded	117.0 μs	119.0 μs	0.985
track/3. E1B 4sat/25K – track-threaded	117.0 μs	120.0 μs	0.983
track/4. 8L1+8E1B/25K – track	433.0 μs	438.0 μs	0.99
track/4. 8L1+8E1B/25K – track!	432.0 μs	436.0 μs	0.992
track/4. 8L1+8E1B/25K – track!-threaded	432.0 μs	436.0 μs	0.99
track/4. 8L1+8E1B/25K – track-threaded	434.0 μs	438.0 μs	0.992
track/5. multi-signal N=1/5K – track	6.23 μs	6.22 μs	1.0
track/5. multi-signal N=1/5K – track!	6.2 μs	6.23 μs	0.996
track/6. multi-signal N=2/5K – track	8.58 μs	8.56 μs	1.0
track/6. multi-signal N=2/5K – track!	8.64 μs	8.76 μs	0.986
track/7. multi-signal N=3/5K – track	10.9 μs	11.0 μs	0.993
track/7. multi-signal N=3/5K – track!	11.1 μs	11.3 μs	0.988
time_to_load	144.0 μs	101.0 μs	1.42 ✅

Memory benchmarks (base vs PR head)

	81e5299…	e661bc1…
downconvert and correlate/CPU/Float32	2 allocs: 576 B	2 allocs: 576 B
downconvert and correlate/CPU/Float32 4ant	2 allocs: 848 B	2 allocs: 848 B
downconvert and correlate/CPU/Float64	2 allocs: 576 B	2 allocs: 576 B
downconvert and correlate/CPU/Int16	2 allocs: 576 B	2 allocs: 576 B
downconvert and correlate/CPU/Int16 4ant	2 allocs: 848 B	2 allocs: 848 B
downconvert and correlate/CPU/Int32	2 allocs: 576 B	2 allocs: 576 B
fused kernel/1-ant dynamic taps	0 allocs: 0 B	0 allocs: 0 B
fused kernel/1-ant static taps	0 allocs: 0 B	0 allocs: 0 B
fused kernel/4-ant dynamic taps	0 allocs: 0 B	0 allocs: 0 B
fused kernel/4-ant static taps	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/1-ant N=2	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/1-ant N=3	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/2-ant N=2	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/2-ant N=3	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/4-ant N=2	0 allocs: 0 B	0 allocs: 0 B
fused tuple kernel/4-ant N=3	0 allocs: 0 B	0 allocs: 0 B
track/1. Float32/2K – track	9 allocs: 944 B	9 allocs: 944 B
track/1. Float32/2K – track!	1 allocs: 26 B	1 allocs: 26 B
track/2. L1 8sat/5K – track	10 allocs: 4536 B	10 allocs: 4536 B
track/2. L1 8sat/5K – track!	0 allocs: 0 B	0 allocs: 0 B
track/2. L1 8sat/5K – track!-threaded	0 allocs: 0 B	0 allocs: 0 B
track/2. L1 8sat/5K – track-threaded	10 allocs: 4536 B	10 allocs: 4536 B
track/3. E1B 4sat/25K – track	10 allocs: 2808 B	10 allocs: 2808 B
track/3. E1B 4sat/25K – track!	0 allocs: 0 B	0 allocs: 0 B
track/3. E1B 4sat/25K – track!-threaded	0 allocs: 0 B	0 allocs: 0 B
track/3. E1B 4sat/25K – track-threaded	10 allocs: 2808 B	10 allocs: 2808 B
track/4. 8L1+8E1B/25K – track	26 allocs: 10352 B	26 allocs: 10352 B
track/4. 8L1+8E1B/25K – track!	0 allocs: 0 B	0 allocs: 0 B
track/4. 8L1+8E1B/25K – track!-threaded	0 allocs: 0 B	0 allocs: 0 B
track/4. 8L1+8E1B/25K – track-threaded	26 allocs: 10352 B	26 allocs: 10352 B
track/5. multi-signal N=1/5K – track	12 allocs: 1024 B	12 allocs: 1024 B
track/5. multi-signal N=1/5K – track!	3 allocs: 80 B	3 allocs: 80 B
track/6. multi-signal N=2/5K – track	15 allocs: 1568 B	15 allocs: 1568 B
track/6. multi-signal N=2/5K – track!	6 allocs: 160 B	6 allocs: 160 B
track/7. multi-signal N=3/5K – track	18 allocs: 2000 B	18 allocs: 2000 B
track/7. multi-signal N=3/5K – track!	9 allocs: 240 B	9 allocs: 240 B
time_to_load	145 allocs: 11216 B	145 allocs: 11216 B

[zsoerenm]

Upstream issue for the root cause (SinCosLUT v3.3.0 generate_carrier_signs! AVX2 codegen crash on Julia 1.10), with a minimal reproducer: JuliaGNSS/SinCosLUT.jl#25. Lift this pin once a 1.10-safe SinCosLUT release ships.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.78%. Comparing base (81e5299) to head (e661bc1).
⚠️ Report is 2 commits behind head on ss/two-bit-downconvert-and-correlate.

Additional details and impacted files

@@                           Coverage Diff                            @@
##           ss/two-bit-downconvert-and-correlate     #196      +/-   ##
========================================================================
- Coverage                                 98.05%   97.78%   -0.28%     
========================================================================
  Files                                        29       29              
  Lines                                      2933     2933              
========================================================================
- Hits                                       2876     2868       -8     
- Misses                                       57       65       +8     

☔ 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.

[zsoerenm]

No longer needed: SinCosLUT v3.3.1 fixes the root cause upstream ("avx2: narrow flip-loop accumulator to Vec{32} to avoid LLVM-15 register exhaustion", closes #25). Tracking's existing SinCosLUT = "3" now resolves the fixed 3.3.1, so no pin is required. Closing; verifying #194 goes green with 3.3.1.