Fix Int32 overflow in !range metadata for large intrinsics

[rainerrodrigues]

Fixes #3024

Summary

This PR fixes a bug where !range metadata on LLVM intrinsics was causing an Int32 overflow, which resulted in negative upper bounds and prevented LLVM from optimizing away standard bounds checks.

The Cause

LLVM !range metadata expects a half-open interval [start, stop). In CUDACore/src/device/intrinsics/indexing.jl, the exclusive upper bound is calculated as (range.stop + 1) % Int32.
For intrinsics with large maximums, like gridDim_x (2^31 - 1), adding 1 overflows the signed 32-bit integer, resulting in -2147483648. LLVM interprets this as a wrap-around range, effectively destroying bounds-check optimizations.

The Fix

Added a safeguard in _index to only attach the !range metadata if range.stop < typemax(Int32). Small-bounded intrinsics (like blockDim) keep their metadata, while large-bounded ones (like gridDim) safely omit it to prevent the overflow.

Testing

Added a unit test in test/core/device/intrinsics.jl using CUDA.code_llvm(..., dump_module=true, raw=true) to verify that:

1. !range is correctly emitted for blockDim().x.
2. !range is safely omitted for gridDim().x.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 7e44cec	Previous: d627b3e	Ratio
array/accumulate/Float32/1d	101339 ns	100853 ns	1.00
array/accumulate/Float32/dims=1	77260 ns	76362 ns	1.01
array/accumulate/Float32/dims=1L	1586550 ns	1585357 ns	1.00
array/accumulate/Float32/dims=2	144263 ns	143506 ns	1.01
array/accumulate/Float32/dims=2L	658401 ns	657416 ns	1.00
array/accumulate/Int64/1d	118883.5 ns	118459.5 ns	1.00
array/accumulate/Int64/dims=1	80473 ns	79599 ns	1.01
array/accumulate/Int64/dims=1L	1695761.5 ns	1694409 ns	1.00
array/accumulate/Int64/dims=2	156095 ns	155450.5 ns	1.00
array/accumulate/Int64/dims=2L	962295.5 ns	960958 ns	1.00
array/broadcast	20701 ns	20282 ns	1.02
array/construct	1307.7 ns	1328.2 ns	0.98
array/copy	18275 ns	17911 ns	1.02
array/copyto!/cpu_to_gpu	214231.5 ns	212984 ns	1.01
array/copyto!/gpu_to_cpu	283233 ns	280259 ns	1.01
array/copyto!/gpu_to_gpu	11021 ns	10684 ns	1.03
array/iteration/findall/bool	134462 ns	134327 ns	1.00
array/iteration/findall/int	149883 ns	149222 ns	1.00
array/iteration/findfirst/bool	81431 ns	81163 ns	1.00
array/iteration/findfirst/int	83343 ns	82984 ns	1.00
array/iteration/findmin/1d	87141.5 ns	85430.5 ns	1.02
array/iteration/findmin/2d	117107 ns	116921 ns	1.00
array/iteration/logical	199980 ns	196920 ns	1.02
array/iteration/scalar	68578 ns	67453 ns	1.02
array/permutedims/2d	52164 ns	52024.5 ns	1.00
array/permutedims/3d	52715 ns	52690 ns	1.00
array/permutedims/4d	53830.5 ns	51921.5 ns	1.04
array/random/rand/Float32	12904 ns	12625 ns	1.02
array/random/rand/Int64	24944 ns	24502 ns	1.02
array/random/rand!/Float32	8426.333333333334 ns	8842 ns	0.95
array/random/rand!/Int64	21569 ns	21520 ns	1.00
array/random/randn/Float32	37511 ns	38339 ns	0.98
array/random/randn!/Float32	30819.5 ns	30558 ns	1.01
array/reductions/mapreduce/Float32/1d	34615 ns	34493 ns	1.00
array/reductions/mapreduce/Float32/dims=1	49961 ns	39238 ns	1.27
array/reductions/mapreduce/Float32/dims=1L	51391 ns	51115 ns	1.01
array/reductions/mapreduce/Float32/dims=2	58098 ns	56362 ns	1.03
array/reductions/mapreduce/Float32/dims=2L	67940 ns	68855 ns	0.99
array/reductions/mapreduce/Int64/1d	42975.5 ns	41840 ns	1.03
array/reductions/mapreduce/Int64/dims=1	42620 ns	43041.5 ns	0.99
array/reductions/mapreduce/Int64/dims=1L	87221 ns	87062 ns	1.00
array/reductions/mapreduce/Int64/dims=2	60865 ns	59400.5 ns	1.02
array/reductions/mapreduce/Int64/dims=2L	84250 ns	84650 ns	1.00
array/reductions/reduce/Float32/1d	35253.5 ns	34748 ns	1.01
array/reductions/reduce/Float32/dims=1	48912 ns	42567 ns	1.15
array/reductions/reduce/Float32/dims=1L	51580 ns	51269 ns	1.01
array/reductions/reduce/Float32/dims=2	58356 ns	56421.5 ns	1.03
array/reductions/reduce/Float32/dims=2L	68854 ns	69423 ns	0.99
array/reductions/reduce/Int64/1d	42724 ns	41786 ns	1.02
array/reductions/reduce/Int64/dims=1	42554 ns	41701 ns	1.02
array/reductions/reduce/Int64/dims=1L	87190 ns	87004 ns	1.00
array/reductions/reduce/Int64/dims=2	60933 ns	59173 ns	1.03
array/reductions/reduce/Int64/dims=2L	84180 ns	84474 ns	1.00
array/reverse/1d	17974 ns	17689 ns	1.02
array/reverse/1dL	68610 ns	68239 ns	1.01
array/reverse/1dL_inplace	65750 ns	65528 ns	1.00
array/reverse/1d_inplace	10291.666666666666 ns	8294.666666666666 ns	1.24
array/reverse/2d	20772 ns	20296 ns	1.02
array/reverse/2dL	72889 ns	72502 ns	1.01
array/reverse/2dL_inplace	65846 ns	65556 ns	1.00
array/reverse/2d_inplace	10049 ns	9743 ns	1.03
array/sorting/1d	2735535 ns	2734808 ns	1.00
array/sorting/2d	1069343 ns	1071711 ns	1.00
array/sorting/by	3304052 ns	3304331 ns	1.00
cuda/synchronization/context/auto	1171.7 ns	1162.7 ns	1.01
cuda/synchronization/context/blocking	923.5 ns	928.3666666666667 ns	0.99
cuda/synchronization/context/nonblocking	7085.6 ns	6978.1 ns	1.02
cuda/synchronization/stream/auto	998.9642857142858 ns	994.2105263157895 ns	1.00
cuda/synchronization/stream/blocking	804.0309278350516 ns	840.6543209876543 ns	0.96
cuda/synchronization/stream/nonblocking	6989.2 ns	7241.299999999999 ns	0.97
integration/byval/reference	143828 ns	143552.5 ns	1.00
integration/byval/slices=1	145614 ns	145385 ns	1.00
integration/byval/slices=2	284608 ns	284202.5 ns	1.00
integration/byval/slices=3	423154 ns	422620 ns	1.00
integration/cudadevrt	102393 ns	102213 ns	1.00
integration/volumerhs	23422519 ns	23474593 ns	1.00
kernel/indexing	13226 ns	12944 ns	1.02
kernel/indexing_checked	13872 ns	13691.5 ns	1.01
kernel/launch	2106.8888888888887 ns	2106.5555555555557 ns	1.00
kernel/occupancy	664.8993710691824 ns	665.54375 ns	1.00
kernel/rand	14214 ns	13998 ns	1.02
latency/import	3821530614.5 ns	3829156614.5 ns	1.00
latency/precompile	4585635970.5 ns	4603348569 ns	1.00
latency/ttfp	4391840056 ns	4403053778.5 ns	1.00

This comment was automatically generated by workflow using github-action-benchmark.

[maleadt]

CI failures look related.

[codecov]

Codecov Report

❌ Patch coverage is 0% with 23 lines in your changes missing coverage. Please review.
✅ Project coverage is 16.43%. Comparing base (d627b3e) to head (7e44cec).

Files with missing lines	Patch %	Lines
lib/cusparse/src/array.jl	0.00%	23 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3119      +/-   ##
==========================================
+ Coverage   16.41%   16.43%   +0.01%     
==========================================
  Files         123      123              
  Lines        9678     9670       -8     
==========================================
  Hits         1589     1589              
+ Misses       8089     8081       -8     

☔ View full report in Codecov by Sentry.
📢 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.