cp-printf-hot — printf / fprintf in a loop

Claim (narrow)Copy Anchor Link

When fprintf runs on a truly hot path to a stream forced unbuffered (_IONBF), wall clock is several× slower than the same call sites with a large _IOFBF buffer under cc -O2 on this Linux machine. Removing the I/O entirely is another order of magnitude.

Claim we do not makeCopy Anchor Link

A cp-printf-hot lint hit alone does not imply the process is slow: dead helpers, #if 0 bodies, and dump/codegen printers fire the scanner with ~1.0× hot-path time (see STATIC-VS-O2 case3).

How measuredCopy Anchor Link

• Source: packages/cperf/examples/bench/cp003_printf_hot.c
• Build: cc -std=c11 -O2 -o cp003_printf_hot cp003_printf_hot.c
• Timing: clock_gettime(CLOCK_MONOTONIC); warm=2; median of 9 rounds; N=80000; output /dev/null
• before: setvbuf(_IONBF) + fprintf each iter
• after: setvbuf 1 MiB _IOFBF + same fprintf each iter (+ fflush)
• ref: compute-only silent loop (ceiling)
• Raw lines: cp-printf-hot.jsonl
• Companion negatives: case3_dead_printf.c / lint probe case3_if0.c (static fire, no time delta)

Runtime (this evidence file)Copy Anchor Link

Companion STATIC-VS-O2 case2 ≈ 3.8× unbuf/buf; case3 dead/#if 0 ≈ 1.0×.

Lint mappingCopy Anchor Link

cperf-lint cp-printf-hot fires on printf/fprintf/puts (and friends) inside a loop-body window, with dump/debug/print* / #if DUMP|DEBUG|… / path suppressions. Default severity is off (G3 demote) under evidence-policy: measured agreement exists for hot I/O, but correlation with “finding ⇒ cost” is partial (preprocessor-blind #if 0, dump noise). Enable per-tree with --severity cp-printf-hot=hint (or error) only when the path is known hot.

LimitsCopy Anchor Link

• Scanner is not a preprocessor; #if 0 can still produce findings.
• Buffered vs unbuffered is one fix; often the right fix is not printing on the hot path at all.
• /dev/null microbench understates real tty/file cost; still enough to show the smell matters when hot.