[BUG] descr() causes hard R crash (exit code -1073741819) when passed an empty data frame

[hatschibratschi]

Calling descr() (collapse 2.1.6) on a zero-row data frame — produced by subsetting — causes R to crash entirely with a fatal exit rather than returning an error or an empty result gracefully.

Exit code -1073741819 (0xC0000005) is STATUS_ACCESS_VIOLATION on Windows, indicating an out-of-bounds memory access (equivalent to a segfault on Linux/macOS).

library(collapse)

d = data.frame(a = 1:3)
descr(d[d$a < 1, ])

# R 4.5.3 exited unexpectedly: exit code -1073741819

descr() should either:

• return an empty / zero-row statistics table, or
• emit an informative R warning/error such as "No observations to describe."

[SebKrantz]

@claude please debug and propose a minimal fix.

[claude]

Claude finished @SebKrantz's task in 3m 42s —— View job • claude/issue-831-20260513-1600 • Create PR ➔

---

Debugging descr() crash on empty data frame

• Read R/descr.R to understand the R layer
• Find corresponding C code and identify crash location
• Implement minimal fix in src/fbstats.cpp
• Commit and push

Root Cause

The crash originates in fbstatstemp() (src/fbstats.cpp). When x has zero elements (l = 0) and there are no groups (ng == 0), the function computes j = l - 1 = -1, then immediately tries to access x[j] = x[-1] — one element before the start of the array:

int j = l - 1;  // j = -1 when l == 0
while(std::isnan(x[j]) && j != 0) --j;  // x[-1] → STATUS_ACCESS_VIOLATION

The existing guard j != 0 only stops the loop when j reaches exactly 0; with j = -1 the condition -1 != 0 is true, so the out-of-bounds read happens immediately.

The same pattern is present in both the ext = false (basic stats) and ext = true (higher-order stats, skewness/kurtosis) branches.

This is not descr()-specific — qsu() on any empty numeric vector would crash the same way.

Fix

Added an early-return guard in fbstatstemp (src/fbstats.cpp) right after the weights flag is set:

if(l == 0 && ng == 0) {
    int d = ((ext) ? 7 : 5) + weights;
    NumericVector result(d, NA_REAL);
    result[0] = 0;   // N = 0, all other stats NA
    if(setn) {
        Rf_namesgets(result, get_stats_names(d));
        Rf_classgets(result, CharacterVector::create("qsu","table"));
    }
    return result;
}

This single branch covers both ext modes, weighted and unweighted, and produces a result consistent with the all-NA path already used when every observation is missing. fndistinctC and fquantileC already guard against empty input, so no other changes are needed.

After the fix, descr(d[d$a < 1, ]) returns a valid descr object with N = 0 and all statistics NA instead of crashing.

[SebKrantz]

Thanks. Will be fixed in 2.1.7 which will be submitted today.