vfs: read-only Windows streams on their own handle by default

Read-only streams on Windows NT take their own handle, position and
buffer over the C runtime's FILE by default; VFS_WIN32_NO_NATIVE keeps
the C runtime. Measured with vfs_native_bench on Windows: 1-byte
sequential reads 16.0 -> 6.8 ns, 16-byte 16.6 -> 7.5 ns, 256-byte 40.0
-> 30.4 ns; 4 KiB and up, random reads, reads at offsets, threads and
small-file opens the same within run-to-run spread.

The cross-libc Windows job's second VFS pass is now the C runtime one
(make VFS_NATIVE=0); vfs_native_bench_native is the default build and
vfs_native_bench the C runtime one.

.github/workflows/Cross-libc-vfs-samples.yml

@@ -235,11 +235,11 @@ jobs:
           ./vfs_positional_nommap_test.exe
           ./vfs_positional_ntverify_test.exe
           ./vfs_positional_pathreopen_test.exe
-          # The same with read-only streams on a handle of their own
-          # (VFS_WIN32_NATIVE, opt-in): the stream contract, the
-          # path an open takes, large files, offsets from threads.
+          # The same with read-only streams on the C runtime
+          # (VFS_WIN32_NO_NATIVE): the stream contract, the path an
+          # open takes, large files, offsets from threads.
           make clean MMAP=0
-          make MMAP=0 VFS_NATIVE=1
+          make MMAP=0 VFS_NATIVE=0
           ./vfs_bulk_read_test.exe
           ./vfs_seek_contract_test.exe
           ./vfs_large_file_test.exe

.github/workflows/Linux-libretro-common-samples.yml

@@ -547,8 +547,8 @@ jobs:
           # self-contained regression tests came to be compiled every
           # run and never executed.
           declare -a NORUN_TARGETS=(
-            # Benchmarks: C runtime streams against VFS_WIN32_NATIVE,
-            # for timing on Windows.
+            # Benchmarks: C runtime streams (VFS_WIN32_NO_NATIVE)
+            # against the stream's own handle, for timing on Windows.
             vfs_native_bench
             vfs_native_bench_native
             # Fixture-driven checks run in Linux-samples-record.yml.

libretro-common/include/vfs/vfs.h

@@ -199,7 +199,7 @@ struct libretro_vfs_implementation_file
    HANDLE fh;
    /* CreateFileMapping handle backing `mapped`; NULL when unmapped */
    HANDLE map_handle;
-   /* A read-only stream on a handle of its own (VFS_WIN32_NATIVE): its
+   /* A read-only stream on a handle of its own (not VFS_WIN32_NO_NATIVE): its
     * position, the part of the file `buf` holds, and a read error */
    uint64_t npos;
    uint64_t nbuf_at;

libretro-common/samples/file/vfs/Makefile

@@ -136,11 +136,12 @@ ifeq ($(DESCRIPTOR_IO),1)
    CFLAGS += -DVFS_HAVE_DESCRIPTOR_IO=1
 endif
 
-# 'make VFS_NATIVE=1' builds the suite with read-only Windows streams on
-# a handle of their own (VFS_WIN32_NATIVE); elsewhere it changes nothing.
+# 'make VFS_NATIVE=0' builds the suite with read-only Windows streams on
+# the C runtime (VFS_WIN32_NO_NATIVE) rather than a handle of their own;
+# elsewhere it changes nothing.
 VFS_NATIVE ?=
-ifeq ($(VFS_NATIVE),1)
-   CFLAGS += -DVFS_WIN32_NATIVE
+ifeq ($(VFS_NATIVE),0)
+   CFLAGS += -DVFS_WIN32_NO_NATIVE
 endif
 
 # FAKE_CFR=1 stands in a copy_file_range that copies to EOF whatever
@@ -258,22 +259,23 @@ $(TARGET_TEST18): vfs_positional_required.o $(NTVERIFY_OBJS)
 	$(CC) -o $@ $^ $(LDFLAGS) -lpthread
 
 # vfs_native_bench / vfs_native_bench_native: read-only stream timings
-# through the C runtime, and through VFS_WIN32_NATIVE - run both on
-# Windows and compare. Elsewhere the two are the same build.
-NATIVE_VFS_OBJ := vfs_implementation_native.o
-BENCH_OBJS     := $(filter-out %/vfs_implementation.o,$(COMMON_OBJS)) \
+# through the C runtime (VFS_WIN32_NO_NATIVE), and through the stream's
+# own handle - run both on Windows and compare. Elsewhere the two are
+# the same build.
+CRT_VFS_OBJ := vfs_implementation_crt.o
+BENCH_OBJS  := $(filter-out %/vfs_implementation.o,$(COMMON_OBJS)) \
 	rthreads_ra.o retro_eventcount_ra.o
 
-$(NATIVE_VFS_OBJ): $(LIBRETRO_COMM_DIR)/vfs/vfs_implementation.c
-	$(CC) -c -o $@ $< $(CFLAGS) -DVFS_WIN32_NATIVE
+$(CRT_VFS_OBJ): $(LIBRETRO_COMM_DIR)/vfs/vfs_implementation.c
+	$(CC) -c -o $@ $< $(CFLAGS) -DVFS_WIN32_NO_NATIVE
 
-vfs_native_bench_native.o: vfs_native_bench.c
-	$(CC) -c -o $@ $< $(CFLAGS) -DVFS_WIN32_NATIVE
+vfs_native_bench_crt.o: vfs_native_bench.c
+	$(CC) -c -o $@ $< $(CFLAGS) -DVFS_WIN32_NO_NATIVE
 
-$(BENCH_CRT): vfs_native_bench.o $(COMMON_OBJS) rthreads_ra.o retro_eventcount_ra.o
+$(BENCH_CRT): vfs_native_bench_crt.o $(BENCH_OBJS) $(CRT_VFS_OBJ)
 	$(CC) -o $@ $^ $(LDFLAGS) $(RA_LIBS) -lpthread
 
-$(BENCH_NATIVE): vfs_native_bench_native.o $(BENCH_OBJS) $(NATIVE_VFS_OBJ)
+$(BENCH_NATIVE): vfs_native_bench.o $(COMMON_OBJS) rthreads_ra.o retro_eventcount_ra.o
 	$(CC) -o $@ $^ $(LDFLAGS) $(RA_LIBS) -lpthread
 
 $(TARGET_TEST10): vfs_view_test.o $(VIEW_OBJS)
@@ -355,7 +357,7 @@ clean:
 		$(TARGET_TEST4) $(TARGET_TEST5) $(TARGET_TEST6) \
 		$(TARGET_TEST7) $(TARGET_TEST8) $(TARGET_TEST9) $(TARGET_TEST10) \
 		$(TARGET_TEST11) $(TARGET_TEST12) $(TARGET_TEST13) $(TARGET_TEST14) $(TARGET_TEST15) $(TARGET_TEST16) $(TARGET_TEST17) $(TARGET_TEST18) $(BENCH_CRT) $(BENCH_NATIVE) \
-	vfs_native_bench.o vfs_native_bench_native.o $(NATIVE_VFS_OBJ) $(CLOSE_VFS_OBJ) \
+	vfs_native_bench.o vfs_native_bench_crt.o $(CRT_VFS_OBJ) $(CLOSE_VFS_OBJ) \
 		$(OBJS) $(FAULT_FS_OBJ) $(RENAME_VFS_OBJ) \
 		$(RA_VFS_OBJ) $(HOOK_VFS_OBJ) rthreads_ra.o retro_eventcount_ra.o \
 		$(NOMMAP_VFS_OBJ) vfs_positional_required.o $(OFF32_VFS_OBJ) \

libretro-common/samples/file/vfs/vfs_bulk_read_test.c

@@ -272,13 +272,13 @@ static void test_path_selection(void)
    }
 
    /* No hint: unchanged, buffered - on a handle of its own where
-    * VFS_WIN32_NATIVE takes read-only streams off the C runtime. */
+    * Windows takes read-only streams off the C runtime. */
    s = retro_vfs_file_open_impl(path, RETRO_VFS_FILE_ACCESS_READ,
          RETRO_VFS_FILE_ACCESS_HINT_NONE);
    CHECK(s != NULL, "plain open failed");
    if (s)
    {
-#if defined(VFS_WIN32_NATIVE) && defined(_WIN32) && !defined(_XBOX)
+#if !defined(VFS_WIN32_NO_NATIVE) && defined(_WIN32) && !defined(_XBOX)
       CHECK(s->fp == NULL && s->fh != NULL,
             "plain read-only open is not on its own handle");
 #else

libretro-common/samples/file/vfs/vfs_native_bench.c

@@ -21,9 +21,9 @@
  */
 
 /* Read-only stream throughput through filestream: the C runtime path
- * (vfs_native_bench) against VFS_WIN32_NATIVE (vfs_native_bench_native),
- * the same source built twice. On anything but Windows the two are the
- * same build.
+ * (vfs_native_bench, VFS_WIN32_NO_NATIVE) against the stream's own
+ * handle (vfs_native_bench_native), the same source built twice. On
+ * anything but Windows the two are the same build.
  *
  *   sequential reads of 1, 16, 256, 4096 and 65536 bytes, and 1 MiB
  *   4 KiB at random offsets by seek and read, and by filestream_read_at
@@ -273,10 +273,10 @@ int main(int argc, char **argv)
    if (size < 8192)
       return 1;
 
-#ifdef VFS_WIN32_NATIVE
-   printf("VFS_WIN32_NATIVE, %lld bytes\n", (long long)size);
-#else
+#ifdef VFS_WIN32_NO_NATIVE
    printf("C runtime streams, %lld bytes\n", (long long)size);
+#else
+   printf("native streams, %lld bytes\n", (long long)size);
 #endif
    sequential(path, 1);
    sequential(path, 16);

libretro-common/vfs/vfs_implementation.c

@@ -339,15 +339,18 @@ static void vfs_unix_to_filetime(int64_t unix_s, FILETIME *ft)
 
 #define RFILE_HINT_UNBUFFERED (1 << 8)
 
-/* VFS_WIN32_NATIVE: a read-only stream on Windows NT opens a handle of
- * its own rather than a C runtime FILE, keeps its own position and a
- * 64 KiB buffer, and reads at offsets - with no CRT lock or buffer on
- * the way, and its reads at offsets on the same handle. Opt-in until
- * measured; see samples/file/vfs/vfs_native_bench.c. */
-#if defined(VFS_WIN32_NATIVE) && defined(_WIN32) && !defined(_XBOX) \
+/* A read-only stream on Windows NT opens a handle of its own rather
+ * than a C runtime FILE, keeps its own position and a 64 KiB buffer,
+ * and reads at offsets - with no CRT lock or buffer on the way, and its
+ * reads at offsets on the same handle. VFS_WIN32_NO_NATIVE keeps the C
+ * runtime; samples/file/vfs/vfs_native_bench.c times the two. */
+#if !defined(VFS_WIN32_NO_NATIVE) && defined(_WIN32) && !defined(_XBOX) \
    && !defined(__WINRT__) && defined(VFS_HAVE_AT_HANDLE)
 #define VFS_NATIVE 1
 #define RFILE_HINT_WIN32_NATIVE (1 << 9)
+#ifndef INVALID_FILE_SIZE
+#define INVALID_FILE_SIZE ((DWORD)0xFFFFFFFF)
+#endif
 #define VFS_NATIVE_BUF          0x10000
 static bool    vfs_native_open(libretro_vfs_implementation_file *stream,
       const char *path);