aesni-sha256-x86_64.pl: fix crash on AMD Jaguar.

It was also found that stich performs suboptimally on AMD Jaguar, hence
execution is limited to XOP-capable and Intel processors.
Reviewed-by: NKurt Roeckx <kurt@openssl.org>

crypto/aes/asm/aesni-sha256-x86_64.pl

@@ -140,11 +140,8 @@ $code.=<<___ if ($avx>1);
 	je	${func}_avx2
 ___
 $code.=<<___;
-	and	\$`1<<30`,%eax			# mask "Intel CPU" bit
-	and	\$`1<<28|1<<9`,%r10d		# mask AVX+SSSE3 bits
-	or	%eax,%r10d
-	cmp	\$`1<<28|1<<9|1<<30`,%r10d
-	je	${func}_avx
+	and	\$`1<<28`,%r10d			# check for AVX
+	jnz	${func}_avx
 	ud2
 ___
 						}

crypto/evp/e_aes_cbc_hmac_sha256.c

@@ -498,7 +498,18 @@ static int aesni_cbc_hmac_sha256_cipher(EVP_CIPHER_CTX *ctx,
             iv = AES_BLOCK_SIZE;
 
 #  if defined(STITCHED_CALL)
+        /*
+         * Assembly stitch handles AVX-capable processors, but its
+         * performance is not optimal on AMD Jaguar, ~40% worse, for
+         * unknown reasons. Incidentally processor in question supports
+         * AVX, but not AMD-specific XOP extension, which can be used
+         * to identify it and avoid stitch invocation. So that after we
+         * establish that current CPU supports AVX, we even see if it's
+         * either even XOP-capable Bulldozer-based or GenuineIntel one.
+         */
         if (OPENSSL_ia32cap_P[1] & (1 << (60 - 32)) && /* AVX? */
+            ((OPENSSL_ia32cap_P[1] & (1 << (43 - 32))) /* XOP? */
+             | (OPENSSL_ia32cap_P[0] & (1<<30))) &&    /* "Intel CPU"? */
             plen > (sha_off + iv) &&
             (blocks = (plen - (sha_off + iv)) / SHA256_CBLOCK)) {
             SHA256_Update(&key->md, in + iv, sha_off);