it *should* work now

Author: picnixz

Makefile.pre.in

@@ -1938,6 +1938,9 @@ Python/sysmodule.o: $(srcdir)/Python/sysmodule.c Makefile $(srcdir)/Include/pydt
 		$(MULTIARCH_CPPFLAGS) \
 		-o $@ $(srcdir)/Python/sysmodule.c
 
+Python/cpuinfo.o: $(srcdir)/Python/cpuinfo.c Makefile
+	$(CC) -c $(PY_CORE_CFLAGS) @CORE_CPUINFO_CFLAGS@ -o $@ $(srcdir)/Python/cpuinfo.c
+
 $(IO_OBJS): $(IO_H)
 
 .PHONY: regen-pegen-metaparser

Python/cpuinfo.c

@@ -25,8 +25,6 @@
 #  undef HAS_XGETBV_SUPPORT
 #endif
 
-#undef HAS_XGETBV_SUPPORT
-
 // Below, we declare macros for guarding the detection of SSE, AVX/AVX2
 // and AVX-512 instructions. If the compiler does not even recognize the
 // corresponding flags or if we are not on an 64-bit platform we do not
@@ -162,6 +160,18 @@ get_xgetbv(uint32_t index)
 {
     assert(index == 0); // only XCR0 is supported for now
 #  if defined(HAS_CPUID_SUPPORT) && defined(__x86_64__) && defined(__GNUC__)
+#    if defined(__clang__)
+#       if _Py__has_builtin(__builtin_ia32_xgetbv)
+    return (uint64_t)_xgetbv(index);
+#       else
+    /*
+     * Without -mxsave support, directly using xgetbv() with raw opcode
+     * may still fail on some platforms (e.g., AMD64 + FreeBSD + clang).
+     * To be on the safe side, we assume that XGETBV is not supported.
+     */
+    return 0;
+#       endif
+#    else /* gcc & icc */
     uint32_t eax = 0, edx = 0;
     __asm__ volatile(
         /* raw opcode for xgetbv for compatibility with older toolchains */
@@ -170,6 +180,7 @@ get_xgetbv(uint32_t index)
         : "c" (index)
     );
     return ((uint64_t)edx << 32) | eax;
+#    endif
 #  elif defined(HAS_CPUID_SUPPORT) && defined(_M_X64)
     return (uint64_t)_xgetbv(index);
 #  else

configure

@@ -8003,7 +8003,7 @@ AC_DEFUN([PY_SIMD_DETECT], [
 # we do not necessarily know which instruction sets will be used,
 # we disable SIMD support on some older Android platforms.
 #
-# See py_cpuid_features in pycore_cpuinfo.h for how to order fields
+# See _Py_cpuid_features in pycore_cpuinfo.h for how to order fields
 # and where to put blank lines to separate processor generations for
 # AVX-512 instructions.
 #
@@ -8058,6 +8058,14 @@ then
   PY_SIMD_DETECT([AVX512_VP2INTERSECT], [-mavx512vp2intersect])
 fi
 
+dnl Check that -mxsave can be used for cpuinfo.c as the latter
+dnl requires to be compiled with this option for xgetbv() support.
+AX_CHECK_COMPILE_FLAG([-mxsave],
+    [AS_VAR_SET([CORE_CPUINFO_CFLAGS], [-mxsave])],
+    [AS_VAR_SET([CORE_CPUINFO_CFLAGS], [])],
+    [-Werror])
+AC_SUBST([CORE_CPUINFO_CFLAGS])
+
 ###############################################################################
 # HACL* compilation and linking configuration (contact: @picnixz)
 #