ppccap.c: engage new multipplication and squaring subroutines.

[And remove FPU mutiplication subroutine.]
Reviewed-by: NRich Salz <rsalz@openssl.org>

Configurations/00-base-templates.conf

@@ -278,7 +278,7 @@
     ppc32_asm => {
 	template	=> 1,
 	cpuid_asm_src   => "ppccpuid.s ppccap.c",
-	bn_asm_src      => "bn-ppc.s ppc-mont.s ppc64-mont.s",
+	bn_asm_src      => "bn-ppc.s ppc-mont.s",
 	aes_asm_src     => "aes_core.c aes_cbc.c aes-ppc.s vpaes-ppc.s aesp8-ppc.s",
 	sha1_asm_src    => "sha1-ppc.s sha256-ppc.s sha512-ppc.s sha256p8-ppc.s sha512p8-ppc.s",
 	modes_asm_src   => "ghashp8-ppc.s",

crypto/bn/asm/ppc-mont.pl

@@ -134,10 +134,7 @@ $code=<<___;
 .globl	.bn_mul_mont_int
 .align	5
 .bn_mul_mont_int:
-	cmpwi	$num,4
 	mr	$rp,r3		; $rp is reassigned
-	li	r3,0
-	bltlr
 ___
 $code.=<<___ if ($BNSZ==4);
 	cmpwi	$num,32		; longer key performance is not better

crypto/ppccap.c

@@ -35,38 +35,24 @@ static sigset_t all_masked;
 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
                 const BN_ULONG *np, const BN_ULONG *n0, int num)
 {
-    int bn_mul_mont_fpu64(BN_ULONG *rp, const BN_ULONG *ap,
-                          const BN_ULONG *bp, const BN_ULONG *np,
-                          const BN_ULONG *n0, int num);
     int bn_mul_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
                         const BN_ULONG *np, const BN_ULONG *n0, int num);
+    int bn_mul4x_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
+                          const BN_ULONG *np, const BN_ULONG *n0, int num);
 
-    if (sizeof(size_t) == 4) {
-# if 1 || (defined(__APPLE__) && defined(__MACH__))
-        if (num >= 8 && (num & 3) == 0 && (OPENSSL_ppccap_P & PPC_FPU64))
-            return bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
-# else
-        /*
-         * boundary of 32 was experimentally determined on Linux 2.6.22,
-         * might have to be adjusted on AIX...
-         */
-        if (num >= 32 && (num & 3) == 0 && (OPENSSL_ppccap_P & PPC_FPU64)) {
-            sigset_t oset;
-            int ret;
-
-            sigprocmask(SIG_SETMASK, &all_masked, &oset);
-            ret = bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
-            sigprocmask(SIG_SETMASK, &oset, NULL);
-
-            return ret;
-        }
-# endif
-    } else if ((OPENSSL_ppccap_P & PPC_FPU64))
-        /*
-         * this is a "must" on POWER6, but run-time detection is not
-         * implemented yet...
-         */
-        return bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
+    if (num < 4)
+        return 0;
+
+    if ((num & 3) == 0)
+        return bn_mul4x_mont_int(rp, ap, bp, np, n0, num);
+
+    /*
+     * There used to be [optional] call to bn_mul_mont_fpu64 here,
+     * but above subroutine is faster on contemporary processors.
+     * Formulation means that there might be old processors where
+     * FPU code path would be faster, POWER6 perhaps, but there was
+     * no opportunity to figure it out...
+     */
 
     return bn_mul_mont_int(rp, ap, bp, np, n0, num);
 }