SSE2 accelerated bn_mul_add_words. Code is currently disabled till proper
authorAndy Polyakov <appro@openssl.org>
Thu, 6 May 2004 10:36:49 +0000 (10:36 +0000)
committerAndy Polyakov <appro@openssl.org>
Thu, 6 May 2004 10:36:49 +0000 (10:36 +0000)
config and run-time support is added.
PR: 788
Submitted by: <dean@arctic.org>
Reviewed by: <appro>

Obtained from: http://arctic.org/~dean/crypto/rsa.html

crypto/bn/asm/bn-586.pl

index 33f6125..9a7c64c 100644 (file)
@@ -5,6 +5,8 @@ require "x86asm.pl";
 
 &asm_init($ARGV[0],$0);
 
+$sse2=0;       # SSE2 is temporarily disabled...
+
 &bn_mul_add_words("bn_mul_add_words");
 &bn_mul_words("bn_mul_words");
 &bn_sqr_words("bn_sqr_words");
@@ -42,6 +44,83 @@ sub bn_mul_add_words
 
        &jz(&label("maw_finish"));
 
+       if ($sse2) {
+               &picmeup("eax","OPENSSL_ia32cap");
+               &bt(&DWP(0,"eax"),26);
+               &jnc(&label("maw_loop"));
+
+               &movd("mm0",$w);                # mm0 = w
+               &pxor("mm1","mm1");             # mm1 = carry_in
+
+               &set_label("maw_sse2_loop",0);
+               &movd("mm3",&DWP(0,$r,"",0));   # mm3 = r[0]
+               &paddq("mm1","mm3");            # mm1 = carry_in + r[0]
+               &movd("mm2",&DWP(0,$a,"",0));   # mm2 = a[0]
+               &pmuludq("mm2","mm0");          # mm2 = w*a[0]
+               &movd("mm4",&DWP(4,$a,"",0));   # mm4 = a[1]
+               &pmuludq("mm4","mm0");          # mm4 = w*a[1]
+               &movd("mm6",&DWP(8,$a,"",0));   # mm6 = a[2]
+               &pmuludq("mm6","mm0");          # mm6 = w*a[2]
+               &movd("mm7",&DWP(12,$a,"",0));  # mm7 = a[3]
+               &pmuludq("mm7","mm0");          # mm7 = w*a[3]
+               &paddq("mm1","mm2");            # mm1 = carry_in + r[0] + w*a[0]
+               &movd("mm3",&DWP(4,$r,"",0));   # mm3 = r[1]
+               &paddq("mm3","mm4");            # mm3 = r[1] + w*a[1]
+               &movd("mm5",&DWP(8,$r,"",0));   # mm5 = r[2]
+               &paddq("mm5","mm6");            # mm5 = r[2] + w*a[2]
+               &movd("mm4",&DWP(12,$r,"",0));  # mm4 = r[3]
+               &paddq("mm7","mm4");            # mm7 = r[3] + w*a[3]
+               &movd(&DWP(0,$r,"",0),"mm1");
+               &movd("mm2",&DWP(16,$a,"",0));  # mm2 = a[4]
+               &pmuludq("mm2","mm0");          # mm2 = w*a[4]
+               &psrlq("mm1",32);               # mm1 = carry0
+               &movd("mm4",&DWP(20,$a,"",0));  # mm4 = a[5]
+               &pmuludq("mm4","mm0");          # mm4 = w*a[5]
+               &paddq("mm1","mm3");            # mm1 = carry0 + r[1] + w*a[1]
+               &movd("mm6",&DWP(24,$a,"",0));  # mm6 = a[6]
+               &pmuludq("mm6","mm0");          # mm6 = w*a[6]
+               &movd(&DWP(4,$r,"",0),"mm1");
+               &psrlq("mm1",32);               # mm1 = carry1
+               &movd("mm3",&DWP(28,$a,"",0));  # mm3 = a[7]
+               &add($a,32);
+               &pmuludq("mm3","mm0");          # mm3 = w*a[7]
+               &paddq("mm1","mm5");            # mm1 = carry1 + r[2] + w*a[2]
+               &movd("mm5",&DWP(16,$r,"",0));  # mm5 = r[4]
+               &paddq("mm2","mm5");            # mm2 = r[4] + w*a[4]
+               &movd(&DWP(8,$r,"",0),"mm1");
+               &psrlq("mm1",32);               # mm1 = carry2
+               &paddq("mm1","mm7");            # mm1 = carry2 + r[3] + w*a[3]
+               &movd("mm5",&DWP(20,$r,"",0));  # mm5 = r[5]
+               &paddq("mm4","mm5");            # mm4 = r[5] + w*a[5]
+               &movd(&DWP(12,$r,"",0),"mm1");
+               &psrlq("mm1",32);               # mm1 = carry3
+               &paddq("mm1","mm2");            # mm1 = carry3 + r[4] + w*a[4]
+               &movd("mm5",&DWP(24,$r,"",0));  # mm5 = r[6]
+               &paddq("mm6","mm5");            # mm6 = r[6] + w*a[6]
+               &movd(&DWP(16,$r,"",0),"mm1");
+               &psrlq("mm1",32);               # mm1 = carry4
+               &paddq("mm1","mm4");            # mm1 = carry4 + r[5] + w*a[5]
+               &movd("mm5",&DWP(28,$r,"",0));  # mm5 = r[7]
+               &paddq("mm3","mm5");            # mm3 = r[7] + w*a[7]
+               &movd(&DWP(20,$r,"",0),"mm1");
+               &psrlq("mm1",32);               # mm1 = carry5
+               &paddq("mm1","mm6");            # mm1 = carry5 + r[6] + w*a[6]
+               &movd(&DWP(24,$r,"",0),"mm1");
+               &psrlq("mm1",32);               # mm1 = carry6
+               &paddq("mm1","mm3");            # mm1 = carry6 + r[7] + w*a[7]
+               &movd(&DWP(28,$r,"",0),"mm1");
+               &add($r,32);
+               &psrlq("mm1",32);               # mm1 = carry_out
+
+               &sub("ecx",8);
+               &jnz(&label("maw_sse2_loop"));
+
+               &movd($c,"mm1");                # c = carry_out
+               &emms();
+
+               &jmp(&label("maw_finish"));
+       }
+
        &set_label("maw_loop",0);
 
        &mov(&swtmp(0),"ecx");  #