perf: small native tls optimisation (#2512)

This folds a 3 instruction tls load on aarch64 to a single mrs.
On x86-64 gnu ld *should* have relaxed it to a single fs anyway, but
this guarantees it.

---------

Signed-off-by: Marc <m@pyc.ac>

Author: henderkes

build-static.sh

@@ -195,6 +195,9 @@ if [ -n "${COMPRESS}" ] && [ -z "${DEBUG_SYMBOLS}" ] && [ "${os}" = "linux" ]; t
 	SPC_OPT_INSTALL_ARGS="${SPC_OPT_INSTALL_ARGS} upx"
 fi
 
+MTLS_CFLAGS="$(sh "${CURRENT_DIR}/mtls-cflags.sh")"
+export CGO_CFLAGS="${CGO_CFLAGS} ${MTLS_CFLAGS}"
+
 export SPC_DEFAULT_C_FLAGS="-fPIC -O2"
 if [ -n "${DEBUG_SYMBOLS}" ]; then
 	SPC_CMD_VAR_PHP_MAKE_EXTRA_CFLAGS="${SPC_CMD_VAR_PHP_MAKE_EXTRA_CFLAGS} -fPIE -g"

docs/performance.md

@@ -18,7 +18,7 @@ We strongly recommend changing these values. For best system stability, it is re
 To find the right values, it's best to run load tests simulating real traffic.
 [k6](https://k6.io) and [Gatling](https://gatling.io) are good tools for this.
 
-To configure the number of threads, use the `num_threads` option of the `php_server` and `php` directives.
+To configure the number of threads, use the `num_threads` option of the global `frankenphp` directive.
 To change the number of workers, use the `num` option of the `worker` section of the `frankenphp` directive.
 
 ### `max_threads`

frankenphp.c

@@ -98,21 +98,27 @@ bool original_user_abort_setting = 0;
 frankenphp_interned_strings_t frankenphp_strings = {0};
 HashTable *main_thread_env = NULL;
 
-__thread uintptr_t thread_index;
-__thread bool is_worker_thread = false;
-__thread HashTable *sandboxed_env = NULL;
+#if defined(__ELF__) && defined(__GNUC__)
+#define THREAD_LOCAL __thread __attribute__((tls_model("local-exec")))
+#else
+#define THREAD_LOCAL __thread
+#endif
+
+static THREAD_LOCAL uintptr_t thread_index;
+static THREAD_LOCAL bool is_worker_thread = false;
+static THREAD_LOCAL HashTable *sandboxed_env = NULL;
 /* prepared_env holds entries from php(_server)'s `env KEY VAL`, exposed to
  * getenv() and merged into $_ENV when 'E' is in variables_order. Separate from
  * putenv() so those don't leak into $_ENV. */
-__thread HashTable *prepared_env = NULL;
+static THREAD_LOCAL HashTable *prepared_env = NULL;
 
 /* Published via SG(server_context) so ext-parallel children, which inherit
  * the parent's SG(server_context), can route SAPI callbacks back to the
  * parent's thread_index instead of their zero-initialized TLS. */
 typedef struct {
   uintptr_t thread_index;
 } frankenphp_server_ctx;
-static __thread frankenphp_server_ctx frankenphp_local_server_ctx;
+static THREAD_LOCAL frankenphp_server_ctx frankenphp_local_server_ctx;
 
 static inline uintptr_t frankenphp_thread_index(void) {
   frankenphp_server_ctx *ctx = (frankenphp_server_ctx *)SG(server_context);

go.sh

@@ -2,6 +2,6 @@
 # Runs the go command with the proper Go and cgo flags.
 
 GOFLAGS="$GOFLAGS -tags=nobadger,nomysql,nopgx" \
-	CGO_CFLAGS="$CGO_CFLAGS $(php-config --includes)" \
+	CGO_CFLAGS="$CGO_CFLAGS $(php-config --includes) $(sh "$(dirname "$0")/mtls-cflags.sh")" \
 	CGO_LDFLAGS="$CGO_LDFLAGS $(php-config --ldflags) $(php-config --libs)" \
 	go "$@"

mtls-cflags.sh

@@ -0,0 +1,6 @@
+#!/bin/sh
+# Prints -mtls-size=12 if the toolchain compiles and links local-exec TLS with it (excludes non-AArch64 and ld.lld).
+d="$(mktemp -d)" || exit 0
+printf '__thread __attribute__((tls_model("local-exec"))) int v;\nint main(void){v=1;return v;}\n' >"$d/p.c"
+"${CC:-cc}" -O2 -mtls-size=12 "$d/p.c" -o "$d/p" >/dev/null 2>&1 && printf -- -mtls-size=12
+rm -rf "$d"