fixes

.github/workflows/release.yml

@@ -111,6 +111,16 @@ jobs:
         run: |
           make "${{ matrix.target }}"
           file "dist/fossbench-${{ matrix.target }}.exe"
+          if [ "${{ matrix.target }}" = windows-i386 ]; then
+            pe="dist/fossbench-windows-i386.exe"
+            i686-w64-mingw32-objdump -p "$pe" > "$RUNNER_TEMP/windows-i386-pe.txt"
+            grep -Eq 'MajorSubsystemVersion[[:space:]]+5' "$RUNNER_TEMP/windows-i386-pe.txt"
+            grep -Eq 'MinorSubsystemVersion[[:space:]]+1' "$RUNNER_TEMP/windows-i386-pe.txt"
+            if grep -Eq 'GetTickCount64|InitializeCriticalSectionEx|CreateThreadpool|WaitOnAddress' "$RUNNER_TEMP/windows-i386-pe.txt"; then
+              echo 'Windows i386 release imports an API newer than Windows XP' >&2
+              exit 1
+            fi
+          fi
 
       - name: Package artifact
         run: |

Makefile

@@ -118,6 +118,12 @@ endif
 # Windows uses MinGW.
 CC_WINDOWS_AMD64 ?= x86_64-w64-mingw32-gcc
 CC_WINDOWS_I386  ?= i686-w64-mingw32-gcc
+# Keep the 32-bit executable loadable by the Windows XP loader and prevent the
+# SDK headers from exposing APIs newer than XP.  The code uses native Win32
+# threads on Windows, so this target does not acquire a libwinpthread runtime
+# dependency with a newer OS baseline.
+WINDOWS_I386_XP_CFLAGS := -D_WIN32_WINNT=0x0501 -DWINVER=0x0501 -DNTDDI_VERSION=0x05010000
+WINDOWS_I386_XP_LDFLAGS := -Wl,--major-subsystem-version,5,--minor-subsystem-version,1
 
 NATIVE_BIN := $(DIST)/fossbench-$(OSNAME)-$(HOST_ARCHNAME)
 
@@ -185,7 +191,7 @@ $(DIST)/fossbench-windows-amd64.exe: $(DRIVER) $(DRIVER_DEPS) $(ASM_AMD64) | $(D
 	@echo "built $@"
 
 $(DIST)/fossbench-windows-i386.exe: $(DRIVER) $(DRIVER_DEPS) $(ASM_I386) | $(DIST)
-	$(CC_WINDOWS_I386) -march=pentium4 $(CFLAGS) $(PTHREAD) -static -o $@ $(DRIVER) $(ASM_I386) -lm -lwinhttp -ladvapi32
+	$(CC_WINDOWS_I386) -march=pentium4 $(WINDOWS_I386_XP_CFLAGS) $(CFLAGS) -static $(WINDOWS_I386_XP_LDFLAGS) -o $@ $(DRIVER) $(ASM_I386) -lm -lwinhttp -ladvapi32
 	@echo "built $@"
 
 # Add a native rule if one was not already made above.

README.md

@@ -32,12 +32,13 @@ score weights. Version 0.1 and 0.2 scores are not comparable.
 | PowerPC 32-bit, little-endian | scalar fallback with runtime-selected extensions |
 | PowerPC 64-bit, little-endian | POWER8 with AltiVec/VSX |
 
-Linux, macOS, and Windows are supported. Send a patch if a supported target
-fails.
+Linux, macOS, and Windows are supported. The Windows i386 release has a Windows
+XP (NT 5.1) compatibility baseline. Send a patch if a supported target fails.
 
 ## Build
 
-GNU Make, a C compiler, pthreads, and the system math library are required.
+GNU Make, a C compiler, and the system math library are required. POSIX builds
+also require pthreads; Windows builds use native Win32 threads.
 Uploads use plain HTTP and require no TLS library.
 
 Build for the current machine:

src/app/benchmark.c

@@ -4,10 +4,19 @@
 #include <stdlib.h>
 #include <string.h>
 #include <stdint.h>
-#include <pthread.h>
-#include <unistd.h>
 #include <ctype.h>
 
+#if defined(_WIN32)
+#  define WIN32_LEAN_AND_MEAN
+#  include <windows.h>
+#  include <tlhelp32.h>
+#  include <winhttp.h>
+#  include <process.h>
+#else
+#  include <pthread.h>
+#  include <unistd.h>
+#endif
+
 #include "benchmark.h"
 #include "hw_detect.h"
 #if defined(__linux__)
@@ -60,10 +69,6 @@
 /* Get the current time. */
 
 #if defined(_WIN32)
-#  define WIN32_LEAN_AND_MEAN
-#  include <windows.h>
-#  include <tlhelp32.h>
-#  include <winhttp.h>
 static double now_seconds(void)
 {
 	LARGE_INTEGER f, t;
@@ -574,19 +579,28 @@ struct job {
 	uint64_t          result;
 };
 
+#if defined(_WIN32)
+static unsigned WINAPI job_entry(void *arg)
+#else
 static void *job_entry(void *arg)
+#endif
 {
 	struct job *j = arg;
 	j->result = j->run(j->n, j->ws);
-	return NULL;
+	return 0;
 }
 
 /* Run a kernel on all requested threads. */
 static uint64_t dispatch(run_fn run, uint64_t n, int threads)
 {
 	struct job *jobs = xalloc((size_t)threads * sizeof *jobs);
+#if defined(_WIN32)
+	HANDLE *tids = threads > 1
+			 ? xalloc((size_t)(threads - 1) * sizeof *tids) : NULL;
+#else
 	pthread_t  *tids = threads > 1
 			 ? xalloc((size_t)(threads - 1) * sizeof *tids) : NULL;
+#endif
 	int i, spawned = 0;
 	uint64_t agg = 0;
 
@@ -596,16 +610,30 @@ static uint64_t dispatch(run_fn run, uint64_t n, int threads)
 		jobs[i].ws  = &g_ws[i];
 	}
 	for (i = 1; i < threads; i++) {
+#if defined(_WIN32)
+		tids[spawned] = (HANDLE)_beginthreadex(NULL, 0, job_entry, &jobs[i], 0, NULL);
+		if (tids[spawned] != NULL)
+			spawned++;
+		else
+			job_entry(&jobs[i]);	/* Run it here if the thread fails. */
+#else
 		if (pthread_create(&tids[spawned], NULL, job_entry, &jobs[i]) == 0)
 			spawned++;
 		else
 			job_entry(&jobs[i]);	/* Run it here if the thread fails. */
+#endif
 	}
 
 	job_entry(&jobs[0]);			/* The main thread does the first job. */
 
-	for (i = 0; i < spawned; i++)
+	for (i = 0; i < spawned; i++) {
+#if defined(_WIN32)
+		WaitForSingleObject(tids[i], INFINITE);
+		CloseHandle(tids[i]);
+#else
 		pthread_join(tids[i], NULL);
+#endif
+	}
 	for (i = 0; i < threads; i++)
 		agg += jobs[i].result;