Runtime dispatch now depends on a simple cntb test rather than envvar.

Author: yapolyak

crypto/arm64cpuid.pl

@@ -120,6 +120,14 @@
 	ret
 .size	_armv8_sve2_probe,.-_armv8_sve2_probe
 
+.globl	_armv8_sve_get_vl_bytes
+.type	_armv8_sve_get_vl_bytes,%function
+_armv8_sve_get_vl_bytes:
+	AARCH64_VALID_CALL_TARGET
+	.inst	0x0420e3e0	// cntb x0
+	ret
+.size	_armv8_sve_get_vl_bytes,.-_armv8_sve_get_vl_bytes
+
 .globl	_armv8_cpuid_probe
 .type	_armv8_cpuid_probe,%function
 _armv8_cpuid_probe:

crypto/armcap.c

@@ -211,6 +211,7 @@ void _armv8_eor3_probe(void);
 void _armv8_sve_probe(void);
 void _armv8_sve2_probe(void);
 void _armv8_rng_probe(void);
+unsigned long _armv8_sve_get_vl_bytes(void);
 #  endif
 # endif /* !__APPLE__ && !OSSL_IMPLEMENT_GETAUXVAL */
 
@@ -445,9 +446,10 @@ void OPENSSL_cpuid_setup(void)
          MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_QCOMM, QCOM_CPU_PART_ORYON_X1)) &&
         (OPENSSL_armcap_P & ARMV8_SHA3))
         OPENSSL_armcap_P |= ARMV8_HAVE_SHA3_AND_WORTH_USING;
-    if ((OPENSSL_armcap_P & ARMV8_SVE2) != 0
-        && (e = getenv("OPENSSL_SVE2_POLY1305")) != NULL && atoi(e) != 0)
+    if ((OPENSSL_armcap_P & ARMV8_SVE2) && (_armv8_sve_get_vl_bytes() > 16)) {
+        // This implementation faster if vector length > 128 bits
         OPENSSL_armcap_P |= ARMV8_SVE2_POLY1305;
+    }
 # endif
 }
 #endif /* _WIN32, __ARM_MAX_ARCH__ >= 7 */