linux-user/i386: Fix allocation and alignment of fp state

For modern cpus, the kernel uses xsave to store all extra
cpu state across the signal handler.  For xsave/xrstor to
work, the pointer must be 64 byte aligned.  Moreover, the
regular part of the signal frame must be 16 byte aligned.

Attempt to mirror the kernel code as much as possible.
Use enum FPStateKind instead of use_xsave() and use_fxsr().

Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1648
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

tests/tcg/x86_64/Makefile.target

@@ -13,6 +13,7 @@ X86_64_TESTS += vsyscall
 X86_64_TESTS += noexec
 X86_64_TESTS += cmpxchg
 X86_64_TESTS += adox
+X86_64_TESTS += test-1648
 TESTS=$(MULTIARCH_TESTS) $(X86_64_TESTS) test-x86_64
 else
 TESTS=$(MULTIARCH_TESTS)

tests/tcg/x86_64/test-1648.c

@@ -0,0 +1,33 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+/* See https://gitlab.com/qemu-project/qemu/-/issues/1648 */
+
+#include <signal.h>
+
+__attribute__((noinline))
+void bar(void)
+{
+    /* Success! Continue through sigreturn. */
+}
+
+/*
+ * Because of the change of ABI between foo and bar, the compiler is
+ * required to save XMM6-XMM15.  The compiler will use MOVAPS or MOVDQA,
+ * which will trap if the stack frame is not 16 byte aligned.
+ */
+__attribute__((noinline, ms_abi))
+void foo(void)
+{
+    bar();
+}
+
+void sighandler(int num)
+{
+    foo();
+}
+
+int main(void)
+{
+    signal(SIGUSR1, sighandler);
+    raise(SIGUSR1);
+    return 0;
+}