From 2f34f006bad3d80115e934371c438a83b3612fe2 Mon Sep 17 00:00:00 2001
From: Stephen Heumann <stephenheumann@gmail.com>
Date: Tue, 4 Jul 2017 14:37:17 -0500
Subject: [PATCH] Force MD5 tables to be page-aligned, which should save a
 cycle in some cases.

---
 Makefile                      | 24 ++++++++++++------------
 md5.asm                       |  2 ++
 aesalign.asm => pagealign.asm |  2 +-
 3 files changed, 15 insertions(+), 13 deletions(-)
 rename aesalign.asm => pagealign.asm (58%)

diff --git a/Makefile b/Makefile
index f72e72a..316f13b 100644
--- a/Makefile
+++ b/Makefile
@@ -6,17 +6,17 @@ PROGRAMS = aescbctest aesctrtest aestest aescrypt sha1sum sha1test \
 .PHONY: default
 default: $(PROGRAMS)
 
-aescbctest: aesalign.asm aescbctest.c aesmodes.c aes.asm aes.macros aes.h
-	occ $(CFLAGS) aesalign.asm aescbctest.c aesmodes.c aes.asm -o aescbctest
+aescbctest: pagealign.asm aescbctest.c aesmodes.c aes.asm aes.macros aes.h
+	occ $(CFLAGS) pagealign.asm aescbctest.c aesmodes.c aes.asm -o aescbctest
 
-aesctrtest: aesalign.asm aesctrtest.c aesmodes.c aes.asm aes.macros aes.h
-	occ $(CFLAGS) aesalign.asm aesctrtest.c aesmodes.c aes.asm -o aesctrtest
+aesctrtest: pagealign.asm aesctrtest.c aesmodes.c aes.asm aes.macros aes.h
+	occ $(CFLAGS) pagealign.asm aesctrtest.c aesmodes.c aes.asm -o aesctrtest
 
-aestest: aesalign.asm aestest.c aes.asm aes.macros aes.h
-	occ $(CFLAGS) aesalign.asm aestest.c aes.asm -o aestest
+aestest: pagealign.asm aestest.c aes.asm aes.macros aes.h
+	occ $(CFLAGS) pagealign.asm aestest.c aes.asm -o aestest
 
-aescrypt: aesalign.asm aescrypt.c aesmodes.c aes.asm aes.macros aes.h
-	occ $(CFLAGS) aesalign.asm aescrypt.c aesmodes.c aes.asm -o aescrypt
+aescrypt: pagealign.asm aescrypt.c aesmodes.c aes.asm aes.macros aes.h
+	occ $(CFLAGS) pagealign.asm aescrypt.c aesmodes.c aes.asm -o aescrypt
 
 sha1sum: sha1sum.c sha1.cc sha1.asm sha1.macros sha1.h
 	occ $(CFLAGS) sha1sum.c sha1.cc -o sha1sum
@@ -30,11 +30,11 @@ sha256sum: sha256sum.c sha256.cc sha256.asm sha256.macros sha256.h
 sha256test: sha256test.c sha256.cc sha256.asm sha256.macros sha256.h
 	occ $(CFLAGS) sha256test.c sha256.cc -o sha256test
 
-md5sum: md5sum.c md5.cc md5.asm md5.macros md5.h
-	occ $(CFLAGS) md5sum.c md5.cc -o md5sum
+md5sum: pagealign.asm md5sum.c md5.cc md5.asm md5.macros md5.h
+	occ $(CFLAGS) pagealign.asm md5sum.c md5.cc -o md5sum
 
-md5test: md5test.c md5.cc md5.asm md5.macros md5.h
-	occ $(CFLAGS) md5test.c md5.cc -o md5test
+md5test: pagealign.asm md5test.c md5.cc md5.asm md5.macros md5.h
+	occ $(CFLAGS) pagealign.asm md5test.c md5.cc -o md5test
 
 .PHONY: clean
 clean:
diff --git a/md5.asm b/md5.asm
index 2ef3899..6b2eb3b 100644
--- a/md5.asm
+++ b/md5.asm
@@ -39,6 +39,7 @@ m	gequ	60
 
 
 * Precomputed values of g*4 for each loop iteration, for indexing the message
+	align	256
 g_times_4 private
 	dc i4' 0,  4,  8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 60'
 	dc i4' 4, 24, 44,  0, 20, 40, 60, 16, 36, 56, 12, 32, 52,  8, 28, 48'
@@ -46,6 +47,7 @@ g_times_4 private
 	dc i4' 0, 28, 56, 20, 48, 12, 40,  4, 32, 60, 24, 52, 16, 44,  8, 36'
 	end
 
+	align	256
 k	private
 	dc i4'$d76aa478, $e8c7b756, $242070db, $c1bdceee'
 	dc i4'$f57c0faf, $4787c62a, $a8304613, $fd469501'
diff --git a/aesalign.asm b/pagealign.asm
similarity index 58%
rename from aesalign.asm
rename to pagealign.asm
index e2a2dd9..c786148 100644
--- a/aesalign.asm
+++ b/pagealign.asm
@@ -1,4 +1,4 @@
-* Dummy segment to be linked first to set the alignment.
+* Dummy segment to be linked first to set 256-byte (page) alignment.
 * This needs to be linked before the root file generated by ORCA/C.
 	align 256
 dummy	private