stan/jdk-24
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

8295698: AArch64: test/jdk/sun/security/ec/ed/EdDSATest.java failed with -XX:+UseSHA3Intrinsics

Browse Source 
Reviewed-by: haosun, aph
This commit is contained in:
Dong Bo 2022-11-17 09:05:43 +00:00 committed by Tobias Hartmann
parent b9db16ab09
commit 2f728d0cbb
4 changed files with 55 additions and 50 deletions

69
src/hotspot/cpu/aarch64/stubGenerator_aarch64.cpp

@ -3821,8 +3821,8 @@ class StubGenerator: public StubCodeGenerator {
//
// Inputs:
// c_rarg0 - byte[] source+offset
// c_rarg1 - byte[] SHA.state
// c_rarg2 - int digest_length
// c_rarg1 - byte[] SHA.state
// c_rarg2 - int block_size
// c_rarg3 - int offset
// c_rarg4 - int limit
//
@ -3844,12 +3844,12 @@ class StubGenerator: public StubCodeGenerator {
Register buf = c_rarg0;
Register state = c_rarg1;
Register digest_length = c_rarg2;
Register block_size = c_rarg2;
Register ofs = c_rarg3;
Register limit = c_rarg4;
Label sha3_loop, rounds24_loop;
Label sha3_512, sha3_384_or_224, sha3_256;
Label sha3_512_or_sha3_384, shake128;
__ stpd(v8, v9, __ pre(sp, -64));
__ stpd(v10, v11, Address(sp, 16));
@ -3885,46 +3885,54 @@ class StubGenerator: public StubCodeGenerator {
__ eor(v5, __ T8B, v5, v30);
__ eor(v6, __ T8B, v6, v31);
// digest_length == 64, SHA3-512
__ tbnz(digest_length, 6, sha3_512);
// block_size == 72, SHA3-512; block_size == 104, SHA3-384
__ tbz(block_size, 7, sha3_512_or_sha3_384);
__ ld1(v25, v26, v27, v28, __ T8B, __ post(buf, 32));
__ ld1(v29, v30, __ T8B, __ post(buf, 16));
__ ld1(v29, v30, v31, __ T8B, __ post(buf, 24));
__ eor(v7, __ T8B, v7, v25);
__ eor(v8, __ T8B, v8, v26);
__ eor(v9, __ T8B, v9, v27);
__ eor(v10, __ T8B, v10, v28);
__ eor(v11, __ T8B, v11, v29);
__ eor(v12, __ T8B, v12, v30);
__ eor(v13, __ T8B, v13, v31);
// digest_length == 28, SHA3-224; digest_length == 48, SHA3-384
__ tbnz(digest_length, 4, sha3_384_or_224);
__ ld1(v25, v26, v27, __ T8B, __ post(buf, 24));
__ eor(v14, __ T8B, v14, v25);
__ eor(v15, __ T8B, v15, v26);
__ eor(v16, __ T8B, v16, v27);
// SHA3-256
__ ld1(v25, v26, v27, v28, __ T8B, __ post(buf, 32));
__ eor(v13, __ T8B, v13, v25);
__ eor(v14, __ T8B, v14, v26);
__ eor(v15, __ T8B, v15, v27);
__ eor(v16, __ T8B, v16, v28);
// block_size == 136, bit4 == 0 and bit5 == 0, SHA3-256 or SHAKE256
__ andw(c_rarg5, block_size, 48);
__ cbzw(c_rarg5, rounds24_loop);
__ tbnz(block_size, 5, shake128);
// block_size == 144, bit5 == 0, SHA3-244
__ ldrd(v28, __ post(buf, 8));
__ eor(v17, __ T8B, v17, v28);
__ b(rounds24_loop);
__ BIND(sha3_384_or_224);
__ tbz(digest_length, 2, rounds24_loop); // bit 2 cleared? SHA-384
__ BIND(shake128);
__ ld1(v28, v29, v30, v31, __ T8B, __ post(buf, 32));
__ eor(v17, __ T8B, v17, v28);
__ eor(v18, __ T8B, v18, v29);
__ eor(v19, __ T8B, v19, v30);
__ eor(v20, __ T8B, v20, v31);
__ b(rounds24_loop); // block_size == 168, SHAKE128
// SHA3-224
__ ld1(v25, v26, v27, v28, __ T8B, __ post(buf, 32));
__ ld1(v29, __ T8B, __ post(buf, 8));
__ eor(v13, __ T8B, v13, v25);
__ eor(v14, __ T8B, v14, v26);
__ eor(v15, __ T8B, v15, v27);
__ eor(v16, __ T8B, v16, v28);
__ eor(v17, __ T8B, v17, v29);
__ b(rounds24_loop);
__ BIND(sha3_512);
__ BIND(sha3_512_or_sha3_384);
__ ld1(v25, v26, __ T8B, __ post(buf, 16));
__ eor(v7, __ T8B, v7, v25);
__ eor(v8, __ T8B, v8, v26);
__ tbz(block_size, 5, rounds24_loop); // SHA3-512
// SHA3-384
__ ld1(v27, v28, v29, v30, __ T8B, __ post(buf, 32));
__ eor(v9, __ T8B, v9, v27);
__ eor(v10, __ T8B, v10, v28);
__ eor(v11, __ T8B, v11, v29);
__ eor(v12, __ T8B, v12, v30);
__ BIND(rounds24_loop);
__ subw(rscratch2, rscratch2, 1);
@ -4009,10 +4017,7 @@ class StubGenerator: public StubCodeGenerator {
__ cbnzw(rscratch2, rounds24_loop);
if (multi_block) {
// block_size = 200 - 2 * digest_length, ofs += block_size
__ add(ofs, ofs, 200);
__ sub(ofs, ofs, digest_length, Assembler::LSL, 1);
__ add(ofs, ofs, block_size);
__ cmp(ofs, limit);
__ br(Assembler::LE, sha3_loop);
__ mov(c_rarg0, ofs); // return ofs

30
src/hotspot/share/opto/library_call.cpp

@ -7001,7 +7001,7 @@ bool LibraryCallKit::inline_digestBase_implCompress(vmIntrinsics::ID id) {
src = must_be_not_null(src, true);
Node* src_start = array_element_address(src, ofs, src_elem);
Node* state = NULL;
Node* digest_length = NULL;
Node* block_size = NULL;
address stubAddr;
const char *stubName;
@ -7035,8 +7035,8 @@ bool LibraryCallKit::inline_digestBase_implCompress(vmIntrinsics::ID id) {
state = get_state_from_digest_object(digestBase_obj, T_BYTE);
stubAddr = StubRoutines::sha3_implCompress();
stubName = "sha3_implCompress";
digest_length = get_digest_length_from_digest_object(digestBase_obj);
if (digest_length == NULL) return false;
block_size = get_block_size_from_digest_object(digestBase_obj);
if (block_size == NULL) return false;
break;
default:
fatal_unexpected_iid(id);
@ -7049,14 +7049,14 @@ bool LibraryCallKit::inline_digestBase_implCompress(vmIntrinsics::ID id) {
// Call the stub.
Node* call;
if (digest_length == NULL) {
if (block_size == NULL) {
call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::digestBase_implCompress_Type(false),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, state);
} else {
call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::digestBase_implCompress_Type(true),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, state, digest_length);
src_start, state, block_size);
}
return true;
@ -7168,15 +7168,15 @@ bool LibraryCallKit::inline_digestBase_implCompressMB(Node* digestBase_obj, ciIn
Node* state = get_state_from_digest_object(digest_obj, elem_type);
if (state == NULL) return false;
Node* digest_length = NULL;
Node* block_size = NULL;
if (strcmp("sha3_implCompressMB", stubName) == 0) {
digest_length = get_digest_length_from_digest_object(digest_obj);
if (digest_length == NULL) return false;
block_size = get_block_size_from_digest_object(digest_obj);
if (block_size == NULL) return false;
}
// Call the stub.
Node* call;
if (digest_length == NULL) {
if (block_size == NULL) {
call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::digestBase_implCompressMB_Type(false),
stubAddr, stubName, TypePtr::BOTTOM,
@ -7185,7 +7185,7 @@ bool LibraryCallKit::inline_digestBase_implCompressMB(Node* digestBase_obj, ciIn
call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::digestBase_implCompressMB_Type(true),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, state, digest_length, ofs, limit);
src_start, state, block_size, ofs, limit);
}
// return ofs (int)
@ -7342,11 +7342,11 @@ Node * LibraryCallKit::get_state_from_digest_object(Node *digest_object, BasicTy
return state;
}
//------------------------------get_digest_length_from_sha3_object----------------------------------
Node * LibraryCallKit::get_digest_length_from_digest_object(Node *digest_object) {
Node* digest_length = load_field_from_object(digest_object, "digestLength", "I");
assert (digest_length != NULL, "sanity");
return digest_length;
//------------------------------get_block_size_from_sha3_object----------------------------------
Node * LibraryCallKit::get_block_size_from_digest_object(Node *digest_object) {
Node* block_size = load_field_from_object(digest_object, "blockSize", "I");
assert (block_size != NULL, "sanity");
return block_size;
}
//----------------------------inline_digestBase_implCompressMB_predicate----------------------------

2
src/hotspot/share/opto/library_call.hpp

@ -299,7 +299,7 @@ class LibraryCallKit : public GraphKit {
BasicType elem_type, address stubAddr, const char *stubName,
Node* src_start, Node* ofs, Node* limit);
Node* get_state_from_digest_object(Node *digestBase_object, BasicType elem_type);
Node* get_digest_length_from_digest_object(Node *digestBase_object);
Node* get_block_size_from_digest_object(Node *digestBase_object);
Node* inline_digestBase_implCompressMB_predicate(int predicate);
bool inline_encodeISOArray(bool ascii);
bool inline_updateCRC32();

4
src/hotspot/share/opto/runtime.cpp

@ -1018,7 +1018,7 @@ const TypeFunc* OptoRuntime::digestBase_implCompress_Type(bool is_sha3) {
int argp = TypeFunc::Parms;
fields[argp++] = TypePtr::NOTNULL; // buf
fields[argp++] = TypePtr::NOTNULL; // state
if (is_sha3) fields[argp++] = TypeInt::INT; // digest_length
if (is_sha3) fields[argp++] = TypeInt::INT; // block_size
assert(argp == TypeFunc::Parms+argcnt, "correct decoding");
const TypeTuple* domain = TypeTuple::make(TypeFunc::Parms+argcnt, fields);
@ -1040,7 +1040,7 @@ const TypeFunc* OptoRuntime::digestBase_implCompressMB_Type(bool is_sha3) {
int argp = TypeFunc::Parms;
fields[argp++] = TypePtr::NOTNULL; // buf
fields[argp++] = TypePtr::NOTNULL; // state
if (is_sha3) fields[argp++] = TypeInt::INT; // digest_length
if (is_sha3) fields[argp++] = TypeInt::INT; // block_size
fields[argp++] = TypeInt::INT; // ofs
fields[argp++] = TypeInt::INT; // limit
assert(argp == TypeFunc::Parms+argcnt, "correct decoding");