Fix cpu detection of sse2 on non-64 x86 (AcademySoftwareFoundation#1467)

* Fix cpu detection of sse2 on non-64 x86

Previous code accidentally elided the i386 (i.e. non-64 bit versions of
x86 compiles). This refactors and simplifies the ifdef snarl to clarify
and fix that.

Fixes AcademySoftwareFoundation#1459

Signed-off-by: Kimball Thurston <kdt3rd@gmail.com>

* Fix mismatch with windows and enable avx support check there

Signed-off-by: Kimball Thurston <kdt3rd@gmail.com>

* Disable avx detection under msvc for now

We require gcc / clang style inline asm for avx support, detecting it
properly causes problems for the dwa support which uses that. Disable
again.

Signed-off-by: Kimball Thurston <kdt3rd@gmail.com>

---------

Signed-off-by: Kimball Thurston <kdt3rd@gmail.com>

src/lib/OpenEXRCore/internal_cpuid.h

@@ -5,7 +5,13 @@
 
 #include "OpenEXRConfigInternal.h"
 
-#if defined(__x86_64__) || defined(_M_X64)
+#if defined(i386) || defined(__i386__) || defined(__i386) || defined(_M_X86) || defined(__x86_64__) || defined(_M_X64)
+#    define OPENEXR_ENABLE_X86_SIMD_CHECK 1
+#else
+#    define OPENEXR_ENABLE_X86_SIMD_CHECK 0
+#endif
+
+#if OPENEXR_ENABLE_X86_SIMD_CHECK
 #    if defined(_MSC_VER) && defined(_WIN32)
 #        include <intrin.h>
 #    else
@@ -19,30 +25,35 @@ static inline void check_for_x86_simd (int *f16c, int *avx, int *sse2)
 #ifdef __e2k__
 #    if defined(__SSE2__)
     *sse2 = 1;
+#    else
+    *sse2 = 0;
 #    endif
 #    if defined(__AVX__)
     *avx = 1;
+#    else
+    *avx = 0;
 #    endif
 #    if defined(__F16C__)
     *f16c = 1;
+#    else
+    *f16c = 0;
 #    endif
 
-#else
-#    if (defined(__x86_64__) || defined(_M_X64))
+#elif OPENEXR_ENABLE_X86_SIMD_CHECK
 
-#        if defined(__AVX__) && defined(__F16C__)
+    // shortcut if everything is turned on / compiled in
+#    if defined(__AVX__) && defined(__F16C__)
     *f16c = 1;
     *avx = 1;
     *sse2 = 1;
-#        else
-#            ifdef _WIN32
+#    elif defined(_MSC_VER) && defined(_WIN32)
     int regs[4], osxsave;
 
     __cpuid (regs, 0);
     if (regs[0] >= 1) { __cpuidex (regs, 1, 0); }
     else
         regs[2] = 0;
-#            else
+#   else
     unsigned int regs[4], osxsave;
     __get_cpuid (0, &regs[0], &regs[1], &regs[2], &regs[3]);
     if (regs[0] >= 1)
@@ -51,9 +62,16 @@ static inline void check_for_x86_simd (int *f16c, int *avx, int *sse2)
     }
     else
         regs[2] = 0;
-#            endif
-    /* AVX is indicated by bit 28, F16C by 29 of ECX (reg 2) */
+#   endif
+
+    /*
+     * linux cpuid.h for x86 has defines but not consistent cross platform
+     *
+     * see cpuid.h bit_AVX bit_F16C bit_SSE2
+     */
+
     osxsave = (regs[2] & (1 << 27)) ? 1 : 0;
+    /* AVX is indicated by bit 28, F16C by 29 of ECX (reg 2) */
     *avx    = (regs[2] & (1 << 28)) ? 1 : 0;
     *f16c   = (regs[2] & (1 << 29)) ? 1 : 0;
     /* sse2 is in EDX bit 26 */
@@ -67,40 +85,52 @@ static inline void check_for_x86_simd (int *f16c, int *avx, int *sse2)
     else
     {
         /* check extended control register */
-#            if defined(OPENEXR_IMF_HAVE_GCC_INLINE_ASM_AVX) &&                            \
-    (defined(_M_X64) || defined(__x86_64__))
-        __asm__ __volatile__("xgetbv"
-                             : /* Output  */ "=a"(regs[0]), "=d"(regs[3])
-                             : /* Input   */ "c"(0)
-                             : /* Clobber */);
+#    if defined(_M_X64) || defined(__x86_64__)
+#        if defined(_MSC_VER)
+        /* TODO: remove the following disablement once we can do inline msvc */
+#            if defined(OPENEXR_IMF_HAVE_GCC_INLINE_ASM_AVX)
+        regs[0] = _xgetbv(0);
+#            else
+        regs[0] = 0;
+#            endif
+#        else
+        __asm__ __volatile__ ("xgetbv"
+                              : /* Output  */ "=a"(regs[0]), "=d"(regs[3])
+                              : /* Input   */ "c"(0)
+                              : /* Clobber */);
+#        endif
         /* eax bit 1 - SSE managed, bit 2 - AVX managed */
         if ((regs[0] & 6) != 6)
         {
             *avx  = 0;
             *f16c = 0;
         }
-#            else
-        *avx  = 0;
-        *f16c = 0;
-#            endif
-    }
-#        endif
 #    else
+        *avx    = 0;
+        *f16c   = 0;
+#    endif
+    }
+
+#else
+    // not on x86
     *f16c = 0;
     *avx  = 0;
     *sse2 = 0;
-#    endif
 #endif
+
 }
 
 static inline int has_native_half (void)
 {
-#if defined(__x86_64__) || defined(_M_X64)
+#if OPENEXR_ENABLE_X86_SIMD_CHECK
     int sse2, avx, f16c;
     check_for_x86_simd (&f16c, &avx, &sse2);
     return avx && f16c;
+#elif defined(__aarch64__)
+    return 1;
 #else
-    // TODO: add case for neon?
     return 0;
 #endif
 }
+
+#undef OPENEXR_ENABLE_X86_SIMD_CHECK