sct: crypto_hash sha3-* ref and bmi1

formosa-crypto · Jun 26, 2024 · 43fd681 · 43fd681
1 parent b639b73
commit 43fd681
Show file tree

Hide file tree

Showing 11 changed files with 35 additions and 11 deletions.
diff --git a/src/common/keccak/keccak1600/amd64/bmi1/keccak1600.jinc b/src/common/keccak/keccak1600/amd64/bmi1/keccak1600.jinc
@@ -88,7 +88,7 @@ inline fn __absorb_bmi1(
   reg   u64 rate // rate already in bytes -- it is returned bc of spills
 ) -> reg ptr u64[25], reg u64
 {
-  stack u64 s_in s_inlen s_rate;
+  #mmx reg u64 s_in s_inlen s_rate;
   reg u8 trail_byte;
 
   // intermediate blocks
@@ -174,13 +174,13 @@ inline fn __xtr_bytes_bmi1(
 
 inline fn __squeeze_bmi1(
   reg mut ptr u64[25] state,
-  stack   u64 s_out,
-  reg     u64 outlen,
-  reg     u64 rate
+  #mmx reg u64 s_out,
+  reg      u64 outlen,
+  reg      u64 rate
 )
 {
   reg u64 out;
-  stack u64 s_outlen s_rate;
+  #mmx reg u64 s_outlen s_rate;
 
   // intermediate blocks
   while ( outlen > rate )
@@ -213,7 +213,7 @@ inline fn __keccak1600_bmi1(reg u64 out outlen in inlen, reg u8 trail_byte, reg
 {
   stack u64[25] _state;
   reg ptr u64[25] state;
-  stack u64 s_out s_outlen;
+  #mmx reg u64 s_out s_outlen;
   stack u8 s_trail_byte;
 
   s_out = out;

diff --git a/src/common/keccak/keccak1600/amd64/bmi1/keccakf1600.jinc b/src/common/keccak/keccak1600/amd64/bmi1/keccakf1600.jinc
@@ -129,7 +129,7 @@ inline fn __round_bmi1(reg ptr u64[25] e a, reg u64 rc) -> reg ptr u64[25]
 inline fn __keccakf1600_bmi1(reg ptr u64[25] a) -> reg ptr u64[25]
 {
   reg ptr u64[24] RC;
-  stack ptr u64[24] s_RC;
+  #mmx reg ptr u64[24] s_RC;
   stack u64[25] s_e;
   reg ptr u64[25] e;
   reg u64 c rc;

diff --git a/src/common/keccak/keccak1600/amd64/ref/keccak1600.jinc b/src/common/keccak/keccak1600/amd64/ref/keccak1600.jinc
@@ -87,7 +87,7 @@ inline fn __absorb_ref(
   reg   u64 rate // rate already in bytes -- it is returned bc of spills
 ) -> stack u64[25], reg u64
 {
-  stack u64 s_in s_inlen s_rate;
+  #mmx reg u64 s_in s_inlen s_rate;
   reg u8 trail_byte;
 
   // intermediate blocks
@@ -171,13 +171,13 @@ inline fn __xtr_bytes_ref(
 
 inline fn __squeeze_ref(
   stack u64[25] state,
-  stack u64 s_out,
+  #mmx reg u64 s_out,
   reg   u64 outlen,
   reg   u64 rate
 )
 {
   reg u64 out;
-  stack u64 s_outlen s_rate;
+  #mmx reg u64 s_outlen s_rate;
 
   // intermediate blocks
   while ( outlen > rate )
@@ -205,7 +205,7 @@ inline fn __squeeze_ref(
 inline fn __keccak1600_ref(reg u64 out outlen in inlen, reg u8 trail_byte, reg u64 rate)
 {
   stack u64[25] state;
-  stack u64 s_out s_outlen;
+  #mmx reg u64 s_out s_outlen;
   stack u8 s_trail_byte;
 
   s_out = out;

diff --git a/src/crypto_hash/sha3-224/amd64/bmi1/hash.jazz b/src/crypto_hash/sha3-224/amd64/bmi1/hash.jazz
@@ -3,6 +3,9 @@ require "sha3-224.jinc"
 export fn jade_hash_sha3_224_amd64_bmi1(reg u64 hash input input_length) -> reg u64
 {
   reg u64 r;
+
+  _ = #init_msf();
+
   __sha3_224_bmi1(hash, input, input_length);
   ?{}, r = #set0();
   return r;

diff --git a/src/crypto_hash/sha3-224/amd64/ref/hash.jazz b/src/crypto_hash/sha3-224/amd64/ref/hash.jazz
@@ -3,6 +3,9 @@ require "sha3-224.jinc"
 export fn jade_hash_sha3_224_amd64_ref(reg u64 hash input input_length) -> reg u64
 {
   reg u64 r;
+
+  _ = #init_msf();
+
   __sha3_224_ref(hash, input, input_length);
   ?{}, r = #set0();
   return r;

diff --git a/src/crypto_hash/sha3-256/amd64/bmi1/hash.jazz b/src/crypto_hash/sha3-256/amd64/bmi1/hash.jazz
@@ -3,6 +3,9 @@ require "sha3-256.jinc"
 export fn jade_hash_sha3_256_amd64_bmi1(reg u64 hash input input_length) -> reg u64
 {
   reg u64 r;
+
+  _ = #init_msf();
+
   __sha3_256_bmi1(hash, input, input_length);
   ?{}, r = #set0();
   return r;

diff --git a/src/crypto_hash/sha3-256/amd64/ref/hash.jazz b/src/crypto_hash/sha3-256/amd64/ref/hash.jazz
@@ -3,6 +3,9 @@ require "sha3-256.jinc"
 export fn jade_hash_sha3_256_amd64_ref(reg u64 hash input input_length) -> reg u64
 {
   reg u64 r;
+
+  _ = #init_msf();
+
   __sha3_256_ref(hash, input, input_length);
   ?{}, r = #set0();
   return r;

diff --git a/src/crypto_hash/sha3-384/amd64/bmi1/hash.jazz b/src/crypto_hash/sha3-384/amd64/bmi1/hash.jazz
@@ -3,6 +3,9 @@ require "sha3-384.jinc"
 export fn jade_hash_sha3_384_amd64_bmi1(reg u64 hash input input_length) -> reg u64
 {
   reg u64 r;
+
+  _ = #init_msf();
+
   __sha3_384_bmi1(hash, input, input_length);
   ?{}, r = #set0();
   return r;

diff --git a/src/crypto_hash/sha3-384/amd64/ref/hash.jazz b/src/crypto_hash/sha3-384/amd64/ref/hash.jazz
@@ -3,6 +3,9 @@ require "sha3-384.jinc"
 export fn jade_hash_sha3_384_amd64_ref(reg u64 hash input input_length) -> reg u64
 {
   reg u64 r;
+
+  _ = #init_msf();
+
   __sha3_384_ref(hash, input, input_length);
   ?{}, r = #set0();
   return r;

diff --git a/src/crypto_hash/sha3-512/amd64/bmi1/hash.jazz b/src/crypto_hash/sha3-512/amd64/bmi1/hash.jazz
@@ -3,6 +3,9 @@ require "sha3-512.jinc"
 export fn jade_hash_sha3_512_amd64_bmi1(reg u64 hash input input_length) -> reg u64
 {
   reg u64 r;
+
+  _ = #init_msf();
+
   __sha3_512_bmi1(hash, input, input_length);
   ?{}, r = #set0();
   return r;

diff --git a/src/crypto_hash/sha3-512/amd64/ref/hash.jazz b/src/crypto_hash/sha3-512/amd64/ref/hash.jazz
@@ -3,6 +3,9 @@ require "sha3-512.jinc"
 export fn jade_hash_sha3_512_amd64_ref(reg u64 hash input input_length) -> reg u64
 {
   reg u64 r;
+
+  _ = #init_msf();
+
   __sha3_512_ref(hash, input, input_length);
   ?{}, r = #set0();
   return r;