stream: csAttempt has 4 scattered bools wasting 16B and causing false sharing with mu

[gidotencate]

Summary

csAttempt (in stream.go) has four bool fields interspersed among pointer- and interface-sized fields. Each bool forces alignment padding before the next field, wasting 16 bytes total. Additionally, decompressorSet at offset 160 shares a 64-byte cache line with mu at offset 164, causing false-sharing overhead on every mu.Lock() call under concurrent load.

This is the same pattern fixed for clientStream in #9280 / #9281.

Layout (DWARF- and runtime-verified)

Current size: 232B → 240B allocator size class

Field	Offset	Padding after	Guarded by mu?
finished	120	7 B (before 16B-aligned interface)	Yes
decompressorSet	160	3 B (before sync.Mutex, align 4)	No
allowTransparentRetry	224	0 B (consecutive with drop)	No
drop	225	6 B tail	No

decompressorSet (offset 160) and mu (offset 164) are on the same 64-byte cache line (bytes 128–191).

After grouping bools at tail

Verified by constructing an equivalent mirror struct and calling unsafe.Sizeof:

Metric	Current	Proposed
Struct size	232 B	216 B
Allocator size class	240 B	224 B
Savings per csAttempt	—	16 B

Cache-line false-sharing benchmark

One goroutine spins mu.Lock()/mu.Unlock() (the hot path in finish()) while N stressor goroutines spin-write decompressorSet (the bool sharing its cache line). decompressorSet is written once per stream in recvMsg(), without holding mu. Under load, these accesses overlap across concurrent streams.

goos: linux / goarch: amd64 / cpu: AMD Ryzen 7 7800X3D
benchtime=3s, count=6, cpu=16

BenchmarkLayoutMuLockCSAttempt/stressors=0-16     3.94 ns/op   (baseline)
BenchmarkLayoutMuLockCSAttempt/stressors=1-16    18.02 ns/op   4.6× slower
BenchmarkLayoutMuLockCSAttempt/stressors=2-16    37.93 ns/op   9.6× slower
BenchmarkLayoutMuLockCSAttempt/stressors=4-16    78.34 ns/op  19.9× slower

Impact

• Allocator: 16 B saved per RPC attempt, unconditionally. Goes into the 224 B size class instead of 240 B — better span utilisation and ~6.7% less GC-scanned heap per object.
• False sharing: Eliminating decompressorSet from mu's cache line reduces the cache-miss penalty on mu.Lock() whenever another goroutine is receiving the first response message on a concurrent attempt.

Proposed fix

Group all 4 bools at the tail of csAttempt, split into // Not guarded by mu (decompressorSet, allowTransparentRetry, drop) and // Guarded by mu (finished) sections with per-field comments, following the same convention as #9281. Add a TestCSAttemptSize assertion to guard against regressions.

/cc @easwars @mbissa