Replace bn_sub_part_words() with bn_sub().

Now that bn_sub() handles word arrays with potentially different lengths,
we no longer need bn_sub_part_words() - call bn_sub() instead. This allows
us to entirely remove the unnecessarily complex bn_sub_part_words() code.

ok tb@

4 files changed, 19 insertions, 370 deletions

diff --git a/lib/libcrypto/arch/i386/Makefile.inc b/lib/libcrypto/arch/i386/Makefile.inc
index 67c22262e61..6134dfdc155 100644
--- a/lib/libcrypto/arch/i386/Makefile.inc
+++ b/lib/libcrypto/arch/i386/Makefile.inc
@@ -1,4 +1,4 @@
-# $OpenBSD: Makefile.inc,v 1.7 2023/01/14 15:45:43 jsing Exp $
+# $OpenBSD: Makefile.inc,v 1.8 2023/02/22 05:57:19 jsing Exp $
 
 # i386-specific libcrypto build rules
 
@@ -16,7 +16,6 @@ SRCS+= bf_cbc.c
 SSLASM+= bf bf-586
 # bn
 CFLAGS+= -DOPENSSL_IA32_SSE2
-CFLAGS+= -DOPENSSL_BN_ASM_PART_WORDS
 SSLASM+= bn bn-586
 SSLASM+= bn co-586
 CFLAGS+= -DOPENSSL_BN_ASM_MONT
diff --git a/lib/libcrypto/bn/asm/bn-586.pl b/lib/libcrypto/bn/asm/bn-586.pl
index b502fe60ee2..71b775af8d6 100644
--- a/lib/libcrypto/bn/asm/bn-586.pl
+++ b/lib/libcrypto/bn/asm/bn-586.pl
@@ -17,7 +17,6 @@ for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
 &bn_div_words("bn_div_words");
 &bn_add_words("bn_add_words");
 &bn_sub_words("bn_sub_words");
-&bn_sub_part_words("bn_sub_part_words");
 
 &asm_finish();
 
@@ -566,212 +565,3 @@ sub bn_sub_words
 
 	&function_end($name);
 	}
-
-sub bn_sub_part_words
-	{
-	local($name)=@_;
-
-	&function_begin($name,"");
-
-	&comment("");
-	$a="esi";
-	$b="edi";
-	$c="eax";
-	$r="ebx";
-	$tmp1="ecx";
-	$tmp2="edx";
-	$num="ebp";
-
-	&mov($r,&wparam(0));	# get r
-	 &mov($a,&wparam(1));	# get a
-	&mov($b,&wparam(2));	# get b
-	 &mov($num,&wparam(3));	# get num
-	&xor($c,$c);		# clear carry
-	 &and($num,0xfffffff8);	# num / 8
-
-	&jz(&label("aw_finish"));
-
-	&set_label("aw_loop",0);
-	for ($i=0; $i<8; $i++)
-		{
-		&comment("Round $i");
-
-		&mov($tmp1,&DWP($i*4,$a,"",0)); 	# *a
-		 &mov($tmp2,&DWP($i*4,$b,"",0)); 	# *b
-		&sub($tmp1,$c);
-		 &mov($c,0);
-		&adc($c,$c);
-		 &sub($tmp1,$tmp2);
-		&adc($c,0);
-		 &mov(&DWP($i*4,$r,"",0),$tmp1); 	# *r
-		}
-
-	&comment("");
-	&add($a,32);
-	 &add($b,32);
-	&add($r,32);
-	 &sub($num,8);
-	&jnz(&label("aw_loop"));
-
-	&set_label("aw_finish",0);
-	&mov($num,&wparam(3));	# get num
-	&and($num,7);
-	 &jz(&label("aw_end"));
-
-	for ($i=0; $i<7; $i++)
-		{
-		&comment("Tail Round $i");
-		&mov($tmp1,&DWP(0,$a,"",0));	# *a
-		 &mov($tmp2,&DWP(0,$b,"",0));# *b
-		&sub($tmp1,$c);
-		 &mov($c,0);
-		&adc($c,$c);
-		 &sub($tmp1,$tmp2);
-		&adc($c,0);
-		&mov(&DWP(0,$r,"",0),$tmp1);	# *r
-		&add($a, 4);
-		&add($b, 4);
-		&add($r, 4);
-		 &dec($num) if ($i != 6);
-		 &jz(&label("aw_end")) if ($i != 6);
-		}
-	&set_label("aw_end",0);
-
-	&cmp(&wparam(4),0);
-	&je(&label("pw_end"));
-
-	&mov($num,&wparam(4));	# get dl
-	&cmp($num,0);
-	&je(&label("pw_end"));
-	&jge(&label("pw_pos"));
-
-	&comment("pw_neg");
-	&mov($tmp2,0);
-	&sub($tmp2,$num);
-	&mov($num,$tmp2);
-	&and($num,0xfffffff8);	# num / 8
-	&jz(&label("pw_neg_finish"));
-
-	&set_label("pw_neg_loop",0);
-	for ($i=0; $i<8; $i++)
-	{
-	    &comment("dl<0 Round $i");
-
-	    &mov($tmp1,0);
-	    &mov($tmp2,&DWP($i*4,$b,"",0)); 	# *b
-	    &sub($tmp1,$c);
-	    &mov($c,0);
-	    &adc($c,$c);
-	    &sub($tmp1,$tmp2);
-	    &adc($c,0);
-	    &mov(&DWP($i*4,$r,"",0),$tmp1); 	# *r
-	}
-	    
-	&comment("");
-	&add($b,32);
-	&add($r,32);
-	&sub($num,8);
-	&jnz(&label("pw_neg_loop"));
-	    
-	&set_label("pw_neg_finish",0);
-	&mov($tmp2,&wparam(4));	# get dl
-	&mov($num,0);
-	&sub($num,$tmp2);
-	&and($num,7);
-	&jz(&label("pw_end"));
-	    
-	for ($i=0; $i<7; $i++)
-	{
-	    &comment("dl<0 Tail Round $i");
-	    &mov($tmp1,0);
-	    &mov($tmp2,&DWP($i*4,$b,"",0));# *b
-	    &sub($tmp1,$c);
-	    &mov($c,0);
-	    &adc($c,$c);
-	    &sub($tmp1,$tmp2);
-	    &adc($c,0);
-	    &dec($num) if ($i != 6);
-	    &mov(&DWP($i*4,$r,"",0),$tmp1);	# *r
-	    &jz(&label("pw_end")) if ($i != 6);
-	}
-
-	&jmp(&label("pw_end"));
-	
-	&set_label("pw_pos",0);
-	
-	&and($num,0xfffffff8);	# num / 8
-	&jz(&label("pw_pos_finish"));
-
-	&set_label("pw_pos_loop",0);
-
-	for ($i=0; $i<8; $i++)
-	{
-	    &comment("dl>0 Round $i");
-
-	    &mov($tmp1,&DWP($i*4,$a,"",0));	# *a
-	    &sub($tmp1,$c);
-	    &mov(&DWP($i*4,$r,"",0),$tmp1);	# *r
-	    &jnc(&label("pw_nc".$i));
-	}
-	    
-	&comment("");
-	&add($a,32);
-	&add($r,32);
-	&sub($num,8);
-	&jnz(&label("pw_pos_loop"));
-	    
-	&set_label("pw_pos_finish",0);
-	&mov($num,&wparam(4));	# get dl
-	&and($num,7);
-	&jz(&label("pw_end"));
-	    
-	for ($i=0; $i<7; $i++)
-	{
-	    &comment("dl>0 Tail Round $i");
-	    &mov($tmp1,&DWP($i*4,$a,"",0));	# *a
-	    &sub($tmp1,$c);
-	    &mov(&DWP($i*4,$r,"",0),$tmp1);	# *r
-	    &jnc(&label("pw_tail_nc".$i));
-	    &dec($num) if ($i != 6);
-	    &jz(&label("pw_end")) if ($i != 6);
-	}
-	&mov($c,1);
-	&jmp(&label("pw_end"));
-
-	&set_label("pw_nc_loop",0);
-	for ($i=0; $i<8; $i++)
-	{
-	    &mov($tmp1,&DWP($i*4,$a,"",0));	# *a
-	    &mov(&DWP($i*4,$r,"",0),$tmp1);	# *r
-	    &set_label("pw_nc".$i,0);
-	}
-	    
-	&comment("");
-	&add($a,32);
-	&add($r,32);
-	&sub($num,8);
-	&jnz(&label("pw_nc_loop"));
-	    
-	&mov($num,&wparam(4));	# get dl
-	&and($num,7);
-	&jz(&label("pw_nc_end"));
-	    
-	for ($i=0; $i<7; $i++)
-	{
-	    &mov($tmp1,&DWP($i*4,$a,"",0));	# *a
-	    &mov(&DWP($i*4,$r,"",0),$tmp1);	# *r
-	    &set_label("pw_tail_nc".$i,0);
-	    &dec($num) if ($i != 6);
-	    &jz(&label("pw_nc_end")) if ($i != 6);
-	}
-
-	&set_label("pw_nc_end",0);
-	&mov($c,0);
-
-	&set_label("pw_end",0);
-
-#	&mov("eax",$c);		# $c is "eax"
-
-	&function_end($name);
-	}
-
diff --git a/lib/libcrypto/bn/bn_local.h b/lib/libcrypto/bn/bn_local.h
index 3e37238c5e8..4576c36c917 100644
--- a/lib/libcrypto/bn/bn_local.h
+++ b/lib/libcrypto/bn/bn_local.h
@@ -1,4 +1,4 @@
-/* $OpenBSD: bn_local.h,v 1.16 2023/02/22 05:46:37 jsing Exp $ */
+/* $OpenBSD: bn_local.h,v 1.17 2023/02/22 05:57:19 jsing Exp $ */
 /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
  * All rights reserved.
  *
@@ -264,8 +264,6 @@ void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
 void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b,
     int n, int tna, int tnb, BN_ULONG *t);
 void bn_sqr_recursive(BN_ULONG *r, const BN_ULONG *a, int n2, BN_ULONG *t);
-BN_ULONG bn_sub_part_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b,
-    int cl, int dl);
 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);
 
diff --git a/lib/libcrypto/bn/bn_mul.c b/lib/libcrypto/bn/bn_mul.c
index 1d56e57b76b..5e270b988f3 100644
--- a/lib/libcrypto/bn/bn_mul.c
+++ b/lib/libcrypto/bn/bn_mul.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: bn_mul.c,v 1.33 2023/02/15 18:10:16 jsing Exp $ */
+/* $OpenBSD: bn_mul.c,v 1.34 2023/02/22 05:57:19 jsing Exp $ */
 /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
  * All rights reserved.
  *
@@ -269,144 +269,6 @@ bn_mul_add_words(BN_ULONG *r, const BN_ULONG *a, int num, BN_ULONG w)
 }
 #endif
 
-#if defined(OPENSSL_NO_ASM) || !defined(OPENSSL_BN_ASM_PART_WORDS)
-/*
- * Here follows a specialised variant of bn_sub_words(), which has the property
- * performing operations on arrays of different sizes. The sizes of those arrays
- * is expressed through cl, which is the common length (basically,
- * min(len(a),len(b))), and dl, which is the delta between the two lengths,
- * calculated as len(a)-len(b). All lengths are the number of BN_ULONGs. For the
- * operations that require a result array as parameter, it must have the length
- * cl+abs(dl).
- */
-BN_ULONG
-bn_sub_part_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b, int cl,
-    int dl)
-{
-	BN_ULONG c, t;
-
-	assert(cl >= 0);
-	c = bn_sub_words(r, a, b, cl);
-
-	if (dl == 0)
-		return c;
-
-	r += cl;
-	a += cl;
-	b += cl;
-
-	if (dl < 0) {
-		for (;;) {
-			t = b[0];
-			r[0] = (0 - t - c) & BN_MASK2;
-			if (t != 0)
-				c = 1;
-			if (++dl >= 0)
-				break;
-
-			t = b[1];
-			r[1] = (0 - t - c) & BN_MASK2;
-			if (t != 0)
-				c = 1;
-			if (++dl >= 0)
-				break;
-
-			t = b[2];
-			r[2] = (0 - t - c) & BN_MASK2;
-			if (t != 0)
-				c = 1;
-			if (++dl >= 0)
-				break;
-
-			t = b[3];
-			r[3] = (0 - t - c) & BN_MASK2;
-			if (t != 0)
-				c = 1;
-			if (++dl >= 0)
-				break;
-
-			b += 4;
-			r += 4;
-		}
-	} else {
-		int save_dl = dl;
-		while (c) {
-			t = a[0];
-			r[0] = (t - c) & BN_MASK2;
-			if (t != 0)
-				c = 0;
-			if (--dl <= 0)
-				break;
-
-			t = a[1];
-			r[1] = (t - c) & BN_MASK2;
-			if (t != 0)
-				c = 0;
-			if (--dl <= 0)
-				break;
-
-			t = a[2];
-			r[2] = (t - c) & BN_MASK2;
-			if (t != 0)
-				c = 0;
-			if (--dl <= 0)
-				break;
-
-			t = a[3];
-			r[3] = (t - c) & BN_MASK2;
-			if (t != 0)
-				c = 0;
-			if (--dl <= 0)
-				break;
-
-			save_dl = dl;
-			a += 4;
-			r += 4;
-		}
-		if (dl > 0) {
-			if (save_dl > dl) {
-				switch (save_dl - dl) {
-				case 1:
-					r[1] = a[1];
-					if (--dl <= 0)
-						break;
-				case 2:
-					r[2] = a[2];
-					if (--dl <= 0)
-						break;
-				case 3:
-					r[3] = a[3];
-					if (--dl <= 0)
-						break;
-				}
-				a += 4;
-				r += 4;
-			}
-		}
-		if (dl > 0) {
-			for (;;) {
-				r[0] = a[0];
-				if (--dl <= 0)
-					break;
-				r[1] = a[1];
-				if (--dl <= 0)
-					break;
-				r[2] = a[2];
-				if (--dl <= 0)
-					break;
-				r[3] = a[3];
-				if (--dl <= 0)
-					break;
-
-				a += 4;
-				r += 4;
-			}
-		}
-	}
-	return c;
-}
-#endif
-
 void
 bn_mul_normal(BN_ULONG *r, BN_ULONG *a, int na, BN_ULONG *b, int nb)
 {
@@ -504,15 +366,15 @@ bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2, int dna,
 	zero = neg = 0;
 	switch (c1 * 3 + c2) {
 	case -4:
-		bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
-		bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
+		bn_sub(t, n, &a[n], tna, a, n);	/* - */
+		bn_sub(&t[n], n, b, n, &b[n], tnb);	/* - */
 		break;
 	case -3:
 		zero = 1;
 		break;
 	case -2:
-		bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
-		bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n); /* + */
+		bn_sub(t, n, &a[n], tna, a, n);	/* - */
+		bn_sub(&t[n], n, &b[n], tnb, b, n);	/* + */
 		neg = 1;
 		break;
 	case -1:
@@ -521,16 +383,16 @@ bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2, int dna,
 		zero = 1;
 		break;
 	case 2:
-		bn_sub_part_words(t, a, &(a[n]), tna, n - tna); /* + */
-		bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
+		bn_sub(t, n, a, n, &a[n], tna);	/* + */
+		bn_sub(&t[n], n, b, n, &b[n], tnb);	/* - */
 		neg = 1;
 		break;
 	case 3:
 		zero = 1;
 		break;
 	case 4:
-		bn_sub_part_words(t, a, &(a[n]), tna, n - tna);
-		bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n);
+		bn_sub(t, n, a, n, &a[n], tna);
+		bn_sub(&t[n], n, &b[n], tnb, b, n);
 		break;
 	}
 
@@ -630,14 +492,14 @@ bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n, int tna,
 	neg = 0;
 	switch (c1 * 3 + c2) {
 	case -4:
-		bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
-		bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
+		bn_sub(t, n, &a[n], tna, a, n);		/* - */
+		bn_sub(&t[n], n, b, n, &b[n], tnb);	/* - */
 		break;
 	case -3:
 		/* break; */
 	case -2:
-		bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
-		bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n); /* + */
+		bn_sub(t, n, &a[n], tna, a, n);		/* - */
+		bn_sub(&t[n], n, &b[n], tnb, b, n);	/* + */
 		neg = 1;
 		break;
 	case -1:
@@ -645,15 +507,15 @@ bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n, int tna,
 	case 1:
 		/* break; */
 	case 2:
-		bn_sub_part_words(t, a, &(a[n]), tna, n - tna); /* + */
-		bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
+		bn_sub(t, n, a, n, &a[n], tna);		/* + */
+		bn_sub(&t[n], n, b, n, &b[n], tnb);	/* - */
 		neg = 1;
 		break;
 	case 3:
 		/* break; */
 	case 4:
-		bn_sub_part_words(t, a, &(a[n]), tna, n - tna);
-		bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n);
+		bn_sub(t, n, a, n, &a[n], tna);
+		bn_sub(&t[n], n, &b[n], tnb, b, n);
 		break;
 	}
 		/* The zero case isn't yet implemented here. The speedup