This method uses one uint32_t * 256 table instead of 4,
which means its data cache overhead is less.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 crypto/aes.c | 41 ++++++++++++++++++++---------------------
 1 file changed, 20 insertions(+), 21 deletions(-)

diff --git a/crypto/aes.c b/crypto/aes.c
index 4438d4dcdc..914ccf38ef 100644
--- a/crypto/aes.c
+++ b/crypto/aes.c
@@ -1453,39 +1453,38 @@ aesdec_IMC_swap(AESState *r, const AESState *st, bool swap)
     bool be = HOST_BIG_ENDIAN ^ swap;
     uint32_t t;
 
-    /* Note that AES_imc is encoded for big-endian. */
-    t = (AES_imc[st->b[swap_b ^ 0x0]][0] ^
-         AES_imc[st->b[swap_b ^ 0x1]][1] ^
-         AES_imc[st->b[swap_b ^ 0x2]][2] ^
-         AES_imc[st->b[swap_b ^ 0x3]][3]);
-    if (!be) {
+    t = (      AES_imc_rot[st->b[swap_b ^ 0x0]] ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0x1]], 8) ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0x2]], 16) ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0x3]], 24));
+    if (be) {
         t = bswap32(t);
     }
     r->w[swap_w ^ 0] = t;
 
-    t = (AES_imc[st->b[swap_b ^ 0x4]][0] ^
-         AES_imc[st->b[swap_b ^ 0x5]][1] ^
-         AES_imc[st->b[swap_b ^ 0x6]][2] ^
-         AES_imc[st->b[swap_b ^ 0x7]][3]);
-    if (!be) {
+    t = (      AES_imc_rot[st->b[swap_b ^ 0x4]] ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0x5]], 8) ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0x6]], 16) ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0x7]], 24));
+    if (be) {
         t = bswap32(t);
     }
     r->w[swap_w ^ 1] = t;
 
-    t = (AES_imc[st->b[swap_b ^ 0x8]][0] ^
-         AES_imc[st->b[swap_b ^ 0x9]][1] ^
-         AES_imc[st->b[swap_b ^ 0xA]][2] ^
-         AES_imc[st->b[swap_b ^ 0xB]][3]);
-    if (!be) {
+    t = (      AES_imc_rot[st->b[swap_b ^ 0x8]] ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0x9]], 8) ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0xA]], 16) ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0xB]], 24));
+    if (be) {
         t = bswap32(t);
     }
     r->w[swap_w ^ 2] = t;
 
-    t = (AES_imc[st->b[swap_b ^ 0xC]][0] ^
-         AES_imc[st->b[swap_b ^ 0xD]][1] ^
-         AES_imc[st->b[swap_b ^ 0xE]][2] ^
-         AES_imc[st->b[swap_b ^ 0xF]][3]);
-    if (!be) {
+    t = (      AES_imc_rot[st->b[swap_b ^ 0xC]] ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0xD]], 8) ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0xE]], 16) ^
+         rol32(AES_imc_rot[st->b[swap_b ^ 0xF]], 24));
+    if (be) {
         t = bswap32(t);
     }
     r->w[swap_w ^ 3] = t;
-- 
2.34.1