[Rushaway]

Summary

On css-ze, Hook_UserMessage (the SayText2 hook) uses 152 s of CPU over 7 days in our production profiler: 6.76M calls, 22.6 µs per call. Over the same period Event_PlayerSay ran 135k times, so the hook runs about 50 times per chat message.

That is how the game sends chat: Host_Say sends a separate SayText2 to each recipient, then fires player_say. Each of these copies runs the full formatting in our hook again: tag and color key lookups, the replace list, %t translation, and CFormatColor() on the final string. Each copy builds the same g_msgFinal, because the result only depends on the message fields and the author's settings, not on the recipient. Event_PlayerSay then sends that one result to everyone.

Changes:

• Hook_UserMessage stores what it read (author, chat flag, name, sender, trimmed text) and the formatted name/sender/text. A later call in the same tick with the same fields restores these values and returns Plugin_Handled right away. g_msgFinal already holds the right result.
• Event_PlayerSay (the end of a chat message) clears this cache. The next message is always formatted again, even if the player sends the same text again in the same tick.
• The early returns (invalid author, empty name/sender, name change, empty text) run before the cache check and are unchanged.
• Bump version: 8.0.22 -> 8.0.23

Expected: about 49 out of 50 calls become a few string compares and copies, so most of the hook's cost goes away.

Behaviour

Unchanged. A cache hit leaves g_msgName, g_msgSender, g_msgText and g_msgFinal exactly as a full run would, and returns the same value (Plugin_Handled). The formatting reads nothing that can change between the copies of one message: they are sent in a loop with no commands or events in between. The cache only applies within the same tick, so it is also safe for SayText2 messages sent by other plugins without USERMSG_BLOCKHOOKS, which are not followed by player_say.

Test plan

• Compiles with spcomp 1.12.0.7259 using the same dependencies as CI (no errors, no warnings)
• CI build passes
• On a test server: check that chat from a player with tag + name/chat colors, say_team, /me, >greentext, replace-list words and dead chat still shows the same for every recipient. Check that a SelfMute ignore still hides the message. Then compare Hook_UserMessage in the profiler.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

[WareZ57]

Checked what the formatting depends on: the author's tag and colors, IsClientEnabled() (author flags), the replace list, SetGlobalTransTarget(g_msgAuthor) and CFormatColor(..., g_msgAuthor). All of them depend on the author, never on the recipient. g_msgIsTeammate is set in the say/say_team command listener (Command_Say) before any SayText2, and isn't touched here. So within one Host_Say loop every copy builds the same g_msgFinal, and a hit leaves the globals exactly as a full run would.

Edge cases look fine:

• Cache in-fields are compared after TrimString, i.e. trimmed text against trimmed text.
• If formatting throws (e.g. a missing phrase), g_bMsgCacheValid stays false, so the next copy runs the full path again, as before.
• Two messages from the same player in the same tick are separated by player_say, and messages from different players differ in g_msgAuthor.
• SayText2 sent by other plugins without USERMSG_BLOCKHOOKS with different text per recipient (per-client %t) is a miss, and identical text is a correct hit.

LGTM.