Initialize guest floating-point state explicitly

CMakeLists.txt

@@ -495,6 +495,7 @@ if (WIBO_ENABLE_FIXTURE_TESTS)
         wibo_add_fixture_bin(NAME test_tls SOURCES test/test_tls.c)
         wibo_add_fixture_bin(NAME test_tls_reloc SOURCES test/test_tls_reloc.c)
         wibo_add_fixture_bin(NAME test_fls SOURCES test/test_fls.c)
+        wibo_add_fixture_bin(NAME test_fpu_state SOURCES test/test_fpu_state.c)
         wibo_add_fixture_bin(NAME test_handleapi SOURCES test/test_handleapi.c)
         wibo_add_fixture_bin(NAME test_findfile SOURCES test/test_findfile.c)
         wibo_add_fixture_bin(NAME test_getfileattributesex SOURCES test/test_getfileattributesex.c)

src/macros.S

@@ -31,6 +31,12 @@
 
 .intel_syntax noprefix
 
+#ifdef __x86_64__
+#define SP rsp
+#else
+#define SP esp
+#endif
+
 #ifdef __x86_64__
 
 .macro LJMP32 teb_reg

src/main.cpp

@@ -86,7 +86,11 @@ bool wibo::installTibForCurrentThread(TEB *tibPtr) {
 		return false;
 	}
 	currentThreadTeb = tibPtr;
-	return tebThreadSetup(tibPtr);
+	if (!tebThreadSetup(tibPtr)) {
+		return false;
+	}
+	initFpState();
+	return true;
 }
 
 void wibo::uninstallTebForCurrentThread() {

src/setup.S

@@ -6,6 +6,19 @@
 .no_dead_strip _wibo_reserve
 #endif
 
+# void initFpState()
+# Configures 53-bit x87 precision and default SSE control state
+ASM_GLOBAL(initFpState, @function)
+	sub SP, 8
+	mov word ptr [SP], 0x027f
+	mov dword ptr [SP+4], 0x1f80
+	fninit
+	fldcw word ptr [SP]
+	ldmxcsr dword ptr [SP+4]
+	add SP, 8
+	ret
+ASM_END(initFpState)
+
 #if defined(__x86_64__) && defined(__linux__)
 
 # void installSelectors(TEB *teb)

src/setup.h

@@ -10,6 +10,7 @@ extern "C" {
 
 bool tebThreadSetup(TEB *teb);
 bool tebThreadTeardown(TEB *teb);
+void initFpState();
 
 #ifdef __cplusplus
 }

test/test_fpu_state.c

@@ -0,0 +1,43 @@
+#include <stdint.h>
+#include <stdlib.h>
+#include <windows.h>
+
+#include "test_assert.h"
+
+static uint16_t g_x87ControlWord;
+static uint32_t g_mxcsr;
+static uint16_t g_threadX87ControlWord;
+static uint32_t g_threadMxcsr;
+
+static void NTAPI capture_fpu_state(PVOID module, DWORD reason, PVOID reserved) {
+	(void)module;
+	(void)reserved;
+	if (reason != DLL_PROCESS_ATTACH) {
+		return;
+	}
+	__asm__ volatile("fnstcw %0" : "=m"(g_x87ControlWord));
+	__asm__ volatile("stmxcsr %0" : "=m"(g_mxcsr));
+}
+
+PIMAGE_TLS_CALLBACK g_fpuStateCallback __attribute__((section(".CRT$XLB"), used)) = capture_fpu_state;
+
+static DWORD WINAPI capture_thread_fpu_state(LPVOID param) {
+	(void)param;
+	__asm__ volatile("fnstcw %0" : "=m"(g_threadX87ControlWord));
+	__asm__ volatile("stmxcsr %0" : "=m"(g_threadMxcsr));
+	return 0;
+}
+
+int main(void) {
+	TEST_CHECK_EQ(0x027f, g_x87ControlWord);
+	// Exception status flags may be raised before the process-attach callback runs.
+	TEST_CHECK_EQ(0x1f80, g_mxcsr & ~0x3f);
+
+	HANDLE thread = CreateThread(NULL, 0, capture_thread_fpu_state, NULL, 0, NULL);
+	TEST_CHECK(thread != NULL);
+	TEST_CHECK_EQ(WAIT_OBJECT_0, WaitForSingleObject(thread, INFINITE));
+	TEST_CHECK(CloseHandle(thread));
+	TEST_CHECK_EQ(0x027f, g_threadX87ControlWord);
+	TEST_CHECK_EQ(0x1f80, g_threadMxcsr & ~0x3f);
+	return EXIT_SUCCESS;
+}