Added full support for the SHA cpu extension

10 files changed, 454 insertions, 8 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index ee13f627..e2767686 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -303,6 +303,7 @@ set(ELFLOADER_SRC
     "${BOX64_ROOT}/src/emu/x87emu_private.c"
     "${BOX64_ROOT}/src/emu/x64primop.c"
     "${BOX64_ROOT}/src/emu/x64run_private.c"
+    "${BOX64_ROOT}/src/emu/x64shaext.c"
     "${BOX64_ROOT}/src/emu/x64syscall.c"
     "${BOX64_ROOT}/src/emu/x86syscall.c"
     "${BOX64_ROOT}/src/emu/x64tls.c"
@@ -1020,18 +1021,23 @@ add_test(x87cache ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/${BOX64}
     -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test25 -D TEST_OUTPUT=tmpfile25.txt
     -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref25.txt
     -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
-
+    
 add_test(feround ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/${BOX64}
     -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test26 -D TEST_OUTPUT=tmpfile26.txt
     -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref26.txt
     -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
-
+set_tests_properties(feround PROPERTIES ENVIRONMENT "BOX64_SYNC_ROUNDING=1")
+    
 add_test(sse4_2 ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/${BOX64}
     -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test27 -D TEST_OUTPUT=tmpfile27.txt
     -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref27.txt
     -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
-
-set_tests_properties(feround PROPERTIES ENVIRONMENT "BOX64_SYNC_ROUNDING=1")
+    
+add_test(shaext ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/${BOX64}
+    -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test28 -D TEST_OUTPUT=tmpfile28.txt
+    -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref28.txt
+    -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
+    
 
 file(GLOB extension_tests "${CMAKE_SOURCE_DIR}/tests/extensions/*.c")
 foreach(file ${extension_tests})
diff --git a/src/emu/x64run0f.c b/src/emu/x64run0f.c
index 36482ca3..739e35b1 100644
--- a/src/emu/x64run0f.c
+++ b/src/emu/x64run0f.c
@@ -22,6 +22,7 @@
 #include "my_cpuid.h"

 #include "bridge.h"

 #include "signals.h"

+#include "x64shaext.h"

 #ifdef DYNAREC

 #include "custommem.h"

 #include "../dynarec/native_lock.h"

@@ -340,6 +341,43 @@ uintptr_t Run0F(x64emu_t *emu, rex_t rex, uintptr_t addr, int *step)
                     }

                     break;

 

+                case 0xC8:  /* SHA1NEXTE Gx, Ex */

+                    nextop = F8;

+                    GETGX;

+                    GETEX(0);

+                    sha1nexte(emu, GX, EX);

+                    break;

+                case 0xC9:  /* SHA1MSG1 Gx, Ex */

+                    nextop = F8;

+                    GETGX;

+                    GETEX(0);

+                    sha1msg1(emu, GX, EX);

+                    break;

+                case 0xCA:  /* SHA1MSG2 Gx, Ex */

+                    nextop = F8;

+                    GETGX;

+                    GETEX(0);

+                    sha1msg2(emu, GX, EX);

+                    break;

+                case 0xCB:  /* SHA256RNDS2 Gx, Ex (, XMM0) */

+                    nextop = F8;

+                    GETGX;

+                    GETEX(0);

+                    sha256rnds2(emu, GX, EX);

+                    break;

+                case 0xCC:  /* SHA256MSG1 Gx, Ex */

+                    nextop = F8;

+                    GETGX;

+                    GETEX(0);

+                    sha256msg1(emu, GX, EX);

+                    break;

+                case 0xCD:  /* SHA256MSG2 Gx, Ex */

+                    nextop = F8;

+                    GETGX;

+                    GETEX(0);

+                    sha256msg2(emu, GX, EX);

+                    break;

+

                 case 0xF0: /* MOVBE Gd, Ed*/

                     nextop = F8;

                     GETGD;

@@ -389,6 +427,14 @@ uintptr_t Run0F(x64emu_t *emu, rex_t rex, uintptr_t addr, int *step)
                     }

                     break;

 

+                case 0xCC:  /* SHA1RNDS4 Gx, Ex, Ib */

+                    nextop = F8;

+                    GETGX;

+                    GETEX(1);

+                    tmp8u = F8;

+                    sha1rnds4(emu, GX, EX, tmp8u);

+                    break;

+

                 default:

                     return 0;

             }

diff --git a/src/emu/x64shaext.c b/src/emu/x64shaext.c
new file mode 100644
index 00000000..a0533bb3
--- /dev/null
+++ b/src/emu/x64shaext.c
@@ -0,0 +1,194 @@
+#include <stdint.h>
+
+#include "box64stack.h"
+#include "x64emu.h"
+#include "x64run_private.h"
+#include "x64emu_private.h"
+#include "x64shaext.h"
+
+static uint32_t rol(uint32_t a, int n)
+{
+    n = n&31;
+    if(!n)
+        return a;
+    return (a<<n) | (a>>(32-n));
+}
+
+static uint32_t ror(uint32_t a, int n)
+{
+    n = n&31;
+    if(!n)
+        return a;
+    return (a>>n) | (a<<(32-n));
+}
+
+static uint32_t f0(uint32_t B, uint32_t C, uint32_t D)
+{
+    return (B & C) ^ ((~B) & D);
+}
+
+static uint32_t f1(uint32_t B, uint32_t C, uint32_t D)
+{
+    return B ^ C ^ D;
+}
+
+static uint32_t f2(uint32_t B, uint32_t C, uint32_t D)
+{
+    return (B & C) ^ (B & D) ^ (C & D);
+}
+
+static uint32_t f3(uint32_t B, uint32_t C, uint32_t D)
+{
+    return B ^ C ^ D;
+}
+
+static uint32_t Ch(uint32_t E, uint32_t F, uint32_t G)
+{
+    return (E & F) ^ ((~E) & G);
+}
+
+static uint32_t Maj(uint32_t A, uint32_t B, uint32_t C)
+{
+    return (A & B) ^ (A & C) ^ (B & C);
+}
+
+static uint32_t sigma0(uint32_t A)
+{
+    return ror(A, 2) ^ ror(A, 13) ^ ror(A, 22);
+}
+static uint32_t sigma1(uint32_t E)
+{
+    return ror(E, 6) ^ ror(E, 11) ^ ror(E, 25);
+}
+static uint32_t tho0(uint32_t W)
+{
+    return ror(W, 7) ^ ror(W, 18) ^ (W>>3);
+}
+static uint32_t tho1(uint32_t W)
+{
+    return ror(W, 17) ^ ror(W, 19) ^ (W>>10);
+}
+
+static const uint32_t Ks[] = { 0x5A827999, 0x6ED9EBA1, 0X8F1BBCDC, 0xCA62C1D6 };
+
+void sha1nexte(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2)
+{
+    uint32_t tmp = rol(xmm1->ud[3], 30);
+    xmm1->ud[3] = xmm2->ud[3] + tmp;
+    xmm1->ud[2] = xmm2->ud[2];
+    xmm1->ud[1] = xmm2->ud[1];
+    xmm1->ud[0] = xmm2->ud[0];
+}
+
+void sha1msg1(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2)
+{
+    uint32_t w0 = xmm1->ud[3];
+    uint32_t w1 = xmm1->ud[2];
+    uint32_t w2 = xmm1->ud[1];
+    uint32_t w3 = xmm1->ud[0];
+    uint32_t w4 = xmm2->ud[3];
+    uint32_t w5 = xmm2->ud[2];
+    xmm1->ud[3] = w2 ^ w0;
+    xmm1->ud[2] = w3 ^ w1;
+    xmm1->ud[1] = w4 ^ w2;
+    xmm1->ud[0] = w5 ^ w3;
+}
+
+void sha1msg2(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2)
+{
+    uint32_t w13 = xmm2->ud[2];
+    uint32_t w14 = xmm2->ud[1];
+    uint32_t w15 = xmm2->ud[0];
+    uint32_t w16 = rol(xmm1->ud[3] ^ w13, 1);
+    uint32_t w17 = rol(xmm1->ud[2] ^ w14, 1);
+    uint32_t w18 = rol(xmm1->ud[1] ^ w15, 1);
+    uint32_t w19 = rol(xmm1->ud[0] ^ w16, 1);
+    xmm1->ud[3] = w16;
+    xmm1->ud[2] = w17;
+    xmm1->ud[1] = w18;
+    xmm1->ud[0] = w19;
+}
+
+void sha256msg1(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2)
+{
+    uint32_t w4 = xmm2->ud[0];
+    uint32_t w3 = xmm1->ud[3];
+    uint32_t w2 = xmm1->ud[2];
+    uint32_t w1 = xmm1->ud[1];
+    uint32_t w0 = xmm1->ud[0];
+    xmm1->ud[3] = w3 + tho0(w4);
+    xmm1->ud[2] = w2 + tho0(w3);
+    xmm1->ud[1] = w1 + tho0(w2);
+    xmm1->ud[0] = w0 + tho0(w1);
+}
+
+void sha256msg2(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2)
+{
+    uint32_t w14 = xmm2->ud[2];
+    uint32_t w15 = xmm2->ud[3];
+    uint32_t w16 = xmm1->ud[0] + tho1(w14);
+    uint32_t w17 = xmm1->ud[1] + tho1(w15);
+    uint32_t w18 = xmm1->ud[2] + tho1(w16);
+    uint32_t w19 = xmm1->ud[3] + tho1(w17);
+    xmm1->ud[3] = w19;
+    xmm1->ud[2] = w18;
+    xmm1->ud[1] = w17;
+    xmm1->ud[0] = w16;
+}
+
+void sha1rnds4(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2, uint8_t ib)
+{
+    uint32_t K = Ks[ib&3];
+    uint32_t(*f)(uint32_t , uint32_t , uint32_t) = NULL;
+    switch (ib&3) {
+        case 0: f = f0; break;
+        case 1: f = f1; break;
+        case 2: f = f2; break;
+        case 3: f = f3; break;
+    }
+    uint32_t A = xmm1->ud[3];
+    uint32_t B = xmm1->ud[2];
+    uint32_t C = xmm1->ud[1];
+    uint32_t D = xmm1->ud[0];
+    uint32_t E = 0;
+    for(int i=0; i<4; ++i) {
+        uint32_t new_A = f(B, C, D) + rol(A, 5) + xmm2->ud[3-i] + E + K;
+        E = D;
+        D = C;
+        C = rol(B, 30);
+        B = A;
+        A = new_A;
+    }
+    xmm1->ud[3] = A;
+    xmm1->ud[2] = B;
+    xmm1->ud[1] = C;
+    xmm1->ud[0] = D;
+}
+
+void sha256rnds2(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2)
+{
+    uint32_t A = xmm2->ud[3];
+    uint32_t B = xmm2->ud[2];
+    uint32_t C = xmm1->ud[3];
+    uint32_t D = xmm1->ud[2];
+    uint32_t E = xmm2->ud[1];
+    uint32_t F = xmm2->ud[0];
+    uint32_t G = xmm1->ud[1];
+    uint32_t H = xmm1->ud[0];
+    for(int i=0; i<2; ++i) {
+        uint32_t new_A = Ch(E, F, G) + sigma1(E) + emu->xmm[0].ud[i] + H + Maj(A, B, C) + sigma0(A);
+        uint32_t new_E = Ch(E, F, G) + sigma1(E) + emu->xmm[0].ud[i] + H + D;
+        H = G;
+        G = F;
+        F = E;
+        E = new_E;
+        D = C;
+        C = B;
+        B = A;
+        A = new_A;
+    }
+    xmm1->ud[3] = A;
+    xmm1->ud[2] = B;
+    xmm1->ud[1] = E;
+    xmm1->ud[0] = F;
+}
\ No newline at end of file
diff --git a/src/emu/x64shaext.h b/src/emu/x64shaext.h
new file mode 100644
index 00000000..a86102dc
--- /dev/null
+++ b/src/emu/x64shaext.h
@@ -0,0 +1,18 @@
+#ifndef __X64_SHAEXT_H__
+#define __X64_SHAEXT_H__
+
+#include <stdint.h>
+
+#include "regs.h"
+
+typedef struct x64emu_s x64emu_t;
+
+void sha1nexte(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2);
+void sha1msg1(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2);
+void sha1msg2(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2);
+void sha256msg1(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2);
+void sha256msg2(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2);
+void sha1rnds4(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2, uint8_t ib);
+void sha256rnds2(x64emu_t* emu, sse_regs_t* xmm1, sse_regs_t* xmm2);
+
+#endif //__X64_SHAEXT_H__
\ No newline at end of file
diff --git a/src/include/debug.h b/src/include/debug.h
index ee4d7700..d5e4bb91 100644
--- a/src/include/debug.h
+++ b/src/include/debug.h
@@ -37,6 +37,8 @@ extern int arm64_aes;
 extern int arm64_pmull;
 extern int arm64_crc32;
 extern int arm64_atomics;
+extern int arm64_sha1;
+extern int arm64_sha2;
 extern int arm64_flagm;
 extern int arm64_flagm2;
 #elif defined(RV64)
diff --git a/src/main.c b/src/main.c
index 022b92df..4fae8d7f 100644
--- a/src/main.c
+++ b/src/main.c
@@ -77,6 +77,8 @@ int arm64_aes = 0;
 int arm64_pmull = 0;
 int arm64_crc32 = 0;
 int arm64_atomics = 0;
+int arm64_sha1 = 0;
+int arm64_sha2 = 0;
 int arm64_uscat = 0;
 int arm64_flagm = 0;
 int arm64_flagm2 = 0;
@@ -377,6 +379,14 @@ HWCAP2_ECV
         arm64_aes = 1;
     if(hwcap&HWCAP_ATOMICS)
         arm64_atomics = 1;
+    #ifdef HWCAP_SHA1
+    if(hwcap&HWCAP_SHA1)
+        arm64_sha1 = 1;
+    #endif
+    #ifdef HWCAP_SHA2
+    if(hwcap&HWCAP_SHA2)
+        arm64_sha2 = 1;
+    #endif
     #ifdef HWCAP_USCAT
     if(hwcap&HWCAP_USCAT)
         arm64_uscat = 1;
@@ -407,6 +417,10 @@ HWCAP2_ECV
         printf_log(LOG_INFO, " PMULL");
     if(arm64_atomics)
         printf_log(LOG_INFO, " ATOMICS");
+    if(arm64_sha1)
+        printf_log(LOG_INFO, " SHA1");
+    if(arm64_sha2)
+        printf_log(LOG_INFO, " SHA2");
     if(arm64_uscat)
         printf_log(LOG_INFO, " USCAT");
     if(arm64_flagm)
diff --git a/src/tools/my_cpuid.c b/src/tools/my_cpuid.c
index 81399633..b4bcbcbb 100644
--- a/src/tools/my_cpuid.c
+++ b/src/tools/my_cpuid.c
@@ -270,10 +270,11 @@ void my_cpuid(x64emu_t* emu, uint32_t tmp32u)
             R_EDX = 0;
             break;
         case 0x7:   // extended bits...
-            if(R_ECX==1) {
-                R_EAX = 0; // Bit 5 is avx512_bf16
-            } else 
-                R_EAX = R_ECX = R_EBX = R_EDX = 0; // TODO
+            if(R_ECX==0) {
+                R_EAX = 0;
+                R_EBX = 0 |
+                        1<<29;  // SHA extension
+            } else {R_EAX = R_ECX = R_EBX = R_EDX = 0;}
             break;
         case 0xB:   // Extended Topology Enumeration Leaf
             //TODO!
diff --git a/tests/ref28.txt b/tests/ref28.txt
new file mode 100644
index 00000000..033d0f49
--- /dev/null
+++ b/tests/ref28.txt
@@ -0,0 +1,31 @@
+test SHA Ext
+sha1rnds4(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001, 0x0) => f40757f7-b4b82290-eab46b51-56a09e67
+sha1rnds4(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001, 0x1) => 9477347e-209fe171-5285d814-5bb67ae9
+sha1rnds4(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001, 0x2) => e27962a9-e186daec-1ca4d63c-23c6ef38
+sha1rnds4(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001, 0x3) => f32a4da8-6c2529f4-85aebf4c-b298b076
+sha1rnds4(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000, 0x0) => 986d7873-b5db5395-2ab46b29-16a09e66
+sha1rnds4(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000, 0x1) => 78dd5f0b-21c31271-9285d7eb-1bb67ae8
+sha1rnds4(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000, 0x2) => c6df831c-e2aa0bf1-5ca4d613-e3c6ef36
+sha1rnds4(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000, 0x3) => 57906d5d-6d485af4-c5aebf23-7298b075
+sha1rnds4(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000, 0x0) => d8ae9505-35dd5bf7-eab46f39-96a09e86
+sha1rnds4(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000, 0x1) => f91d79e9-61c513cc-4285dbed-1bb67b08
+sha1rnds4(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000, 0x2) => 4721274f-62ac1813-9ca4da2b-a3c6ef57
+sha1rnds4(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000, 0x3) => 57d20b8d-ad4a5c6d-b5aec325-7298b095
+sha1msg1(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001) => 00000000-00000000-00000004-00000003
+sha1msg1(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000) => 00000000-00000000-ffffffff-00000000
+sha1msg1(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000) => 00000006-00000002-fffffffd-00000001
+sha1msg2(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001) => 00000006-00000004-00000002-0000000c
+sha1msg2(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000) => 00000000-02468ace-00000001-00000000
+sha1msg2(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000) => 00000008-02468ac8-00000005-00000012
+sha1nexte(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001) => 00000004-00000003-00000002-00000001
+sha1nexte(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000) => ffffffff-00000000-01234567-80000000
+sha1nexte(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000) => 00000000-00000000-01234567-80000000
+sha256msg1(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001) => 02004000-00000000-00000000-00000000
+sha256msg1(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000) => 11002000-00000000-00000000-00000000
+sha256msg1(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000) => 11002004-08010003-0600c002-04008001
+sha256msg2(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001) => 100000a1-cc000078-00028000-0001e000
+sha256msg2(00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000) => 00006fe7-00000000-003fffff-00000000
+sha256msg2(00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000) => 0000b02c-0000a003-00400001-00000001
+sha256rnds2(00000000-00000000-00000000-00000000 ,00000004-00000003-00000002-00000001 ,ffffffff-00000000-01234567-80000000) => 6da40f0a-88601101-0353cda8-88400100
+sha256rnds2(00000004-00000003-00000002-00000001 ,00000000-00000000-00000000-00000000 ,ffffffff-00000000-01234567-80000000) => fc0f56a9-80000003-1bf348ad-80000006
+sha256rnds2(ffffffff-00000000-01234567-80000000 ,00000004-00000003-00000002-00000001 ,00000000-00000000-00000000-00000000) => c8b008c8-8983566d-ec5ad91b-89634665
diff --git a/tests/test28 b/tests/test28
new file mode 100755
index 00000000..43aee286
--- /dev/null
+++ b/tests/test28
Binary files differdiff --git a/tests/test28.c b/tests/test28.c
new file mode 100644
index 00000000..86df4a90
--- /dev/null
+++ b/tests/test28.c
@@ -0,0 +1,134 @@
+// build with  gcc -O0 -g -msha -msse4.2 test28.c -o test28
+// and -m32 for 32bits version
+#include <inttypes.h>
+#include <string.h>
+#include <stdio.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <math.h>
+#include <pmmintrin.h>
+#include <immintrin.h>
+
+typedef unsigned char u8x16 __attribute__ ((vector_size (16)));
+typedef unsigned short u16x8 __attribute__ ((vector_size (16)));
+typedef unsigned int  u32x4 __attribute__ ((vector_size (16)));
+typedef unsigned long int  u64x2 __attribute__ ((vector_size (16)));
+typedef float  f32x4 __attribute__ ((vector_size (16)));
+typedef double d64x2 __attribute__ ((vector_size (16)));
+
+typedef union {
+        __m128i mm;
+        __m128  mf;
+        __m128d md;
+        u8x16   u8;
+        u16x8   u16;
+        u32x4   u32;
+        u64x2   u64;
+        f32x4   f32;
+        d64x2   d64;
+} v128;
+
+static const uint32_t A[] = {0, 0, 0, 0};
+static const uint32_t B[] = {1, 2, 3, 4};
+static const uint32_t C[] = {0x80000000, 0x1234567, 0, 0xffffffff};
+
+static void print_u32(v128 a) {
+  printf("%08x-%08x-%08x-%08x", a.u32[3], a.u32[2], a.u32[1], a.u32[0]);
+}
+
+static void print_u32_res(v128 a, v128 b, const char* op, v128 res) {
+  printf("%s(", op);
+  print_u32(a);
+  printf(" ,");
+  print_u32(b);
+  printf(") => ");
+  print_u32(res);
+  printf("\n");
+}
+static void print_u32_u8_res(v128 a, v128 b, const char* op, uint8_t ib, v128 res) {
+  printf("%s(", op);
+  print_u32(a);
+  printf(" ,");
+  print_u32(b);
+  printf(", 0x%x) => ", ib);
+  print_u32(res);
+  printf("\n");
+}
+
+static void print_3u32_res(v128 a, v128 b, v128 c, const char* op, v128 res) {
+  printf("%s(", op);
+  print_u32(a);
+  printf(" ,");
+  print_u32(b);
+  printf(" ,");
+  print_u32(c);
+  printf(") => ");
+  print_u32(res);
+  printf("\n");
+}
+
+int main(int argc, const char** argv)
+{
+  printf("test SHA Ext\n");
+
+  v128 a, b, c, d;
+  int ret;
+  #define LOAD(a, A) a.u32[0] = A[0]; a.u32[1] = A[1]; a.u32[2] = A[2]; a.u32[3] = A[3]
+
+  #define GO2I_(A, B, C, I)  \
+    LOAD(a, A); \
+    LOAD(b, B); \
+    c.mm = _mm_##C##_epu32(a.mm, b.mm, I); \
+    print_u32_u8_res(a, b, #C, I, c)
+
+  #define GO2(A, B, C)  \
+    LOAD(a, A); \
+    LOAD(b, B); \
+    c.mm = _mm_##C##_epu32(a.mm, b.mm); \
+    print_u32_res(a, b, #C, c)
+
+  #define GO3(A, B, C, D)  \
+    LOAD(a, A); \
+    LOAD(b, B); \
+    LOAD(c, C); \
+    d.mm = _mm_##D##_epu32(a.mm, b.mm, c.mm); \
+    print_3u32_res(a, b, c, #D, d)
+
+
+  #define GO2I(A, B, C)   \
+  GO2I_(A, B, C, 0x00);   \
+  GO2I_(A, B, C, 0x01);   \
+  GO2I_(A, B, C, 0x02);   \
+  GO2I_(A, B, C, 0x03)
+
+  GO2I(A, B, sha1rnds4);
+  GO2I(A, C, sha1rnds4);
+  GO2I(B, C, sha1rnds4);
+
+  GO2(A, B, sha1msg1);
+  GO2(A, C, sha1msg1);
+  GO2(B, C, sha1msg1);
+
+  GO2(A, B, sha1msg2);
+  GO2(A, C, sha1msg2);
+  GO2(B, C, sha1msg2);
+
+  GO2(A, B, sha1nexte);
+  GO2(A, C, sha1nexte);
+  GO2(B, C, sha1nexte);
+
+  GO2(A, B, sha256msg1);
+  GO2(A, C, sha256msg1);
+  GO2(B, C, sha256msg1);
+
+  GO2(A, B, sha256msg2);
+  GO2(A, C, sha256msg2);
+  GO2(B, C, sha256msg2);
+
+  GO3(A, B, C, sha256rnds2);
+  GO3(B, A, C, sha256rnds2);
+  GO3(C, B, A, sha256rnds2);
+
+  return 0;
+}