main/streams/streams: optimized code path for unfiltered regular files

If we know the exact file size already, we can allocate the right
buffer size and issue only one read() system call.  This eliminates
unnecessary buffer allocations (file size plus 8 kB), eliminates
buffer reallocations via zend_string_truncate(), and eliminates the
last read() system call.

Benchmark:

 echo Hello World >/tmp/hello.txt
 <?php for ($i = 1; $i <= 1000000; $i++) file_get_contents('/tmp/hello.txt');

Previously:

 openat(AT_FDCWD, "/tmp/hello.txt", O_RDONLY) = 3
 newfstatat(3, "", {st_mode=S_IFREG|0600, st_size=12, ...}, AT_EMPTY_PATH) = 0
 lseek(3, 0, SEEK_CUR)                   = 0
 newfstatat(3, "", {st_mode=S_IFREG|0600, st_size=12, ...}, AT_EMPTY_PATH) = 0
 read(3, "Hello World\n", 8204)          = 12
 read(3, "", 8192)                       = 0
 close(3)                                = 0

          4,036.55 msec task-clock:u              #    0.995 CPUs utilized
     4,683,289,286      cycles:u                  #    1.160 GHz
     7,954,269,313      instructions:u            #    1.70  insn per cycle
     1,165,461,174      branches:u                #  288.727 M/sec
         1,431,759      branch-misses:u           #    0.12% of all branches

Optimized:

 openat(AT_FDCWD, "/tmp/hello.txt", O_RDONLY) = 3
 newfstatat(3, "", {st_mode=S_IFREG|0600, st_size=12, ...}, AT_EMPTY_PATH) = 0
 lseek(3, 0, SEEK_CUR)                   = 0
 newfstatat(3, "", {st_mode=S_IFREG|0600, st_size=12, ...}, AT_EMPTY_PATH) = 0
 read(3, "Hello World\n", 12)            = 12
 close(3)                                = 0

          3,473.51 msec task-clock:u              #    0.997 CPUs utilized
     4,132,407,666      cycles:u                  #    1.190 GHz
     6,841,273,246      instructions:u            #    1.66  insn per cycle
     1,024,464,752      branches:u                #  294.936 M/sec
         1,478,403      branch-misses:u           #    0.14% of all branches

For this micro-benchmark, the improvement is 14%.

This patch needs to adjust the unit test "bug54946.phpt" because the
EBADF warning is no longer logged.

Author: MaxKellermann

ext/standard/tests/streams/bug54946.phpt

@@ -31,11 +31,6 @@ fclose($stream);
 unlink($filename);
 ?>
 --EXPECTF--
-Notice: stream_get_contents(): Read of 8192 bytes failed with errno=9 Bad file descriptor in %s on line %d
 string(0) ""
-
-Notice: stream_get_contents(): Read of 8192 bytes failed with errno=9 Bad file descriptor in %s on line %d
 string(0) ""
-
-Notice: stream_get_contents(): Read of 8192 bytes failed with errno=9 Bad file descriptor in %s on line %d
 string(0) ""

main/streams/streams.c

@@ -1474,6 +1474,26 @@ PHPAPI zend_string *_php_stream_copy_to_mem(php_stream *src, size_t maxlen, int
 		return ZSTR_EMPTY_ALLOC();
 	}
 
+	if (php_stream_is(src, PHP_STREAM_IS_STDIO) &&
+	    src->readfilters.head == NULL &&
+	    php_stream_stat(src, &ssbuf) == 0 &&
+#ifdef S_ISREG
+	    S_ISREG(ssbuf.sb.st_mode) &&
+#endif
+	    ssbuf.sb.st_size >= 0) {
+		/* optimized code path for unfiltered regular files:
+		 * if we know the exact size, we can allocate the
+		 * right buffer size and issue only one read() system
+		 * call */
+
+		if (ssbuf.sb.st_size <= src->position)
+			return ZSTR_EMPTY_ALLOC();
+
+		const size_t remaining = ssbuf.sb.st_size - src->position;
+		if (remaining < maxlen)
+			maxlen = remaining;
+	}
+
 	if (maxlen == PHP_STREAM_COPY_ALL) {
 		maxlen = 0;
 	}