Merge pull request #1101 from mir-protocol/blake_fix

Blake fix
2026-01-03 06:13:07 +00:00 · 2023-06-16 07:30:47 -07:00 · 2023-06-16 07:30:47 -07:00 · 39d2237dee
commit 39d2237dee
parent 23bc390a83 ee452fc014
2 changed files with 94 additions and 47 deletions
--- a/evm/src/cpu/kernel/asm/core/precompiles/blake2_f.asm
+++ b/evm/src/cpu/kernel/asm/core/precompiles/blake2_f.asm
@ -41,84 +41,78 @@ global precompile_blake2_f:
    %sub_const(8)
    // stack: t1_addr=flag_addr-8, flag, blake2_f_contd, kexit_info

-    %stack (t1_addr) -> (@SEGMENT_CALLDATA, t1_addr, 8, t1_addr)
-    // stack: @SEGMENT_CALLDATA, t1_addr, 8, t1_addr, flag, blake2_f_contd, kexit_info
+    %stack (t1_addr) -> (@SEGMENT_CALLDATA, t1_addr, t1_addr)
+    // stack: @SEGMENT_CALLDATA, t1_addr, t1_addr, flag, blake2_f_contd, kexit_info
    GET_CONTEXT
-    // stack: ctx, @SEGMENT_CALLDATA, t1_addr, 8, t1_addr, flag, blake2_f_contd, kexit_info
-    %mload_packing
+    // stack: ctx, @SEGMENT_CALLDATA, t1_addr, t1_addr, flag, blake2_f_contd, kexit_info
+    %mload_packing_u64_LE
    // stack: t_1, t1_addr, flag, blake2_f_contd, kexit_info
    SWAP1
    // stack: t1_addr, t_1, flag, blake2_f_contd, kexit_info
    %sub_const(8)
    // stack: t0_addr=t1_addr-8, t_1, flag, blake2_f_contd, kexit_info

-    %stack (t0_addr) -> (@SEGMENT_CALLDATA, t0_addr, 8, t0_addr)
-    // stack: @SEGMENT_CALLDATA, t0_addr, 8, t0_addr, t_1, flag, blake2_f_contd, kexit_info
+    %stack (t0_addr) -> (@SEGMENT_CALLDATA, t0_addr, t0_addr)
+    // stack: @SEGMENT_CALLDATA, t0_addr, t0_addr, t_1, flag, blake2_f_contd, kexit_info
    GET_CONTEXT
-    // stack: ctx, @SEGMENT_CALLDATA, t0_addr, 8, t0_addr, t_1, flag, blake2_f_contd, kexit_info
-    %mload_packing
+    // stack: ctx, @SEGMENT_CALLDATA, t0_addr, t0_addr, t_1, flag, blake2_f_contd, kexit_info
+    %mload_packing_u64_LE
    // stack: t_0, t0_addr, t_1, flag, blake2_f_contd, kexit_info
    SWAP1
-    // stack: t0_addr, t_0, t_1, flag, blake2_f_contd, kexit_info
-    %sub_const(128) // 16 * 8
-    // stack: m0_addr=t0_addr-128, t_0, t_1, flag, blake2_f_contd, kexit_info
+    // stack: t0_addr = m0_addr + 8 * 16, t_0, t_1, flag, blake2_f_contd, kexit_info

    %rep 16
-        // stack: 68 + 8 * i, m_(i-1), ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-        PUSH 8
-        // stack: 8, 68 + 8 * i, m_(i-1), ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-        DUP2
-        // stack: 68 + 8 * i, 8, 68 + 8 * i, m_(i-1), ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+        // stack: m0_addr + 8 * (16 - i), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+        %sub_const(8)
+        // stack: m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+        DUP1
+        // stack: m0_addr + 8 * (16 - i - 1), m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
        PUSH @SEGMENT_CALLDATA
-        // stack: @SEGMENT_CALLDATA, 68 + 8 * i, 8, 68 + 8 * i, m_(i-1), ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+        // stack: @SEGMENT_CALLDATA, m0_addr + 8 * (16 - i - 1), m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
        GET_CONTEXT
-        // stack: ctx, @SEGMENT_CALLDATA, 68 + 8 * i, 8, 68 + 8 * i, m_(i-1), ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-        %mload_packing
-        // stack: m_i, 68 + 8 * i, m_(i-1), ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+        // stack: ctx, @SEGMENT_CALLDATA, m0_addr + 8 * (16 - i - 1), m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+        %mload_packing_u64_LE
+        // stack: m_i, m0_addr + 8 * (16 - i - 1), m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
        SWAP1
-        // stack: 68 + 8 * i, m_i, m_(i-1), ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-        %add_const(8)
+        // stack: m0_addr + 8 * (16 - i - 1), m_i, m_(i+1), ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
    %endrep
-    // stack: 68 + 8 * 16 = 196, m_15, ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-    %sub_const(192) // 16 * 8 (m values) + 8 * 8 (h values)
-    // stack: h0_addr, m_15, ..., m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+    // stack: m0_addr = h0_addr + 8 * 8, m_0, ..., m_15, t_0, t_1, flag, blake2_f_contd, kexit_info

    %rep 8
-        // stack: 4 + 8 * i, h_(i-1), ..., h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-        PUSH 8
-        // stack: 8, 4 + 8 * i, h_(i-1), ..., h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-        DUP2
-        // stack: 4 + 8 * i, 8, 4 + 8 * i, h_(i-1), ..., h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+        // stack: h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+        %sub_const(8)
+        // stack: h0_addr + 8 * (8 - i - 1), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+        DUP1
+        // stack: h0_addr + 8 * (8 - i), h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
        PUSH @SEGMENT_CALLDATA
-        // stack: @SEGMENT_CALLDATA, 4 + 8 * i, 8, 4 + 8 * i, h_(i-1), ..., h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+        // stack: @SEGMENT_CALLDATA, h0_addr + 8 * (8 - i), h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
        GET_CONTEXT
-        // stack: ctx, @SEGMENT_CALLDATA, 4 + 8 * i, 8, 4 + 8 * i, h_(i-1), ..., h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-        %mload_packing
-        // stack: h_i, 4 + 8 * i, h_(i-1), ..., h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+        // stack: ctx, @SEGMENT_CALLDATA, h0_addr + 8 * (8 - i), h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
+        %mload_packing_u64_LE
+        // stack: h_i, h0_addr + 8 * (8 - i), h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
        SWAP1
-        // stack: 4 + 8 * i, h_i, h_(i-1), ..., h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
-        %add_const(8)
+        // stack: h0_addr + 8 * (8 - i), h_i, h_(i+1), ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
    %endrep
-    // stack: 4 + 8 * 8 = 68, h_7, ..., h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+    // stack: h0_addr + 8 * 8 = 68, h_0, ..., h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
    POP

    %stack () -> (@SEGMENT_CALLDATA, 0, 4)
    GET_CONTEXT
-    // stack: ctx, @SEGMENT_CALLDATA, 0, 4, h_7..h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+    // stack: ctx, @SEGMENT_CALLDATA, 0, 4, h_0..h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
    %mload_packing
-    // stack: rounds, h_7..h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+    // stack: rounds, h_0..h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
    
    DUP1
-    // stack: rounds, rounds, h_7..h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+    // stack: rounds, rounds, h_0..h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
    %charge_gas
    
-    // stack: rounds, h_7..h_0, m_15..m_0, t_0, t_1, flag, blake2_f_contd, kexit_info
+    // stack: rounds, h_0..h_7, m_0..m_15, t_0, t_1, flag, blake2_f_contd, kexit_info
    %jump(blake2_f)
 blake2_f_contd:
    // stack: h_0', h_1', h_2', h_3', h_4', h_5', h_6', h_7', kexit_info
-    // Store the result hash to the parent's return data using `mstore_unpacking`.
+    // Store the result hash to the parent's return data using `mstore_unpacking_u64_LE`.

-    %mstore_parent_context_metadata(@CTX_METADATA_RETURNDATA_SIZE, 32)
+    %mstore_parent_context_metadata(@CTX_METADATA_RETURNDATA_SIZE, 64)
    PUSH 0
    // stack: addr_0=0, h_0', h_1', h_2', h_3', h_4', h_5', h_6', h_7', kexit_info
    %mload_context_metadata(@CTX_METADATA_PARENT_CONTEXT)
@ -126,11 +120,10 @@ blake2_f_contd:

    %rep 8
        // stack: parent_ctx, addr_i, h_i', ..., h_7', kexit_info
-        %stack (ctx, addr, h_i) -> (ctx, @SEGMENT_RETURNDATA, addr, h_i, 4, addr, ctx)
-        // stack: parent_ctx, @SEGMENT_RETURNDATA, addr_i, h_i', 4, addr_i, parent_ctx, h_(i+1)', ..., h_7', kexit_info
-        %mstore_unpacking
+        %stack (ctx, addr, h_i) -> (ctx, @SEGMENT_RETURNDATA, addr, h_i, addr, ctx)
+        %mstore_unpacking_u64_LE
        // stack: addr_i, parent_ctx, h_(i+1)', ..., h_7', kexit_info
-        %add_const(4)
+        %add_const(8)
        // stack: addr_(i+1), parent_ctx, h_(i+1)', ..., h_7', kexit_info
        SWAP1
        // stack: parent_ctx, addr_(i+1), h_(i+1)', ..., h_7', kexit_info
--- a/evm/src/cpu/kernel/asm/memory/packing.asm
+++ b/evm/src/cpu/kernel/asm/memory/packing.asm
@ -49,6 +49,25 @@ mload_packing_return:
 %%after:
 %endmacro

+global mload_packing_u64_LE:
+    // stack: context, segment, offset, retdest
+    DUP3                DUP3 DUP3 MLOAD_GENERAL
+    DUP4 %add_const(1)  DUP4 DUP4 MLOAD_GENERAL %shl_const( 8) ADD
+    DUP4 %add_const(2)  DUP4 DUP4 MLOAD_GENERAL %shl_const(16) ADD
+    DUP4 %add_const(3)  DUP4 DUP4 MLOAD_GENERAL %shl_const(24) ADD
+    DUP4 %add_const(4)  DUP4 DUP4 MLOAD_GENERAL %shl_const(32) ADD
+    DUP4 %add_const(5)  DUP4 DUP4 MLOAD_GENERAL %shl_const(40) ADD
+    DUP4 %add_const(6)  DUP4 DUP4 MLOAD_GENERAL %shl_const(48) ADD
+    DUP4 %add_const(7)  DUP4 DUP4 MLOAD_GENERAL %shl_const(56) ADD
+    %stack (value, context, segment, offset, retdest) -> (retdest, value)
+    JUMP
+
+%macro mload_packing_u64_LE
+    %stack (addr: 3) -> (addr, %%after)
+    %jump(mload_packing_u64_LE)
+%%after:
+%endmacro
+
 // Pre stack: context, segment, offset, value, len, retdest
 // Post stack: offset'
 global mstore_unpacking:
@ -94,3 +113,38 @@ mstore_unpacking_finish:
    %jump(mstore_unpacking)
 %%after:
 %endmacro
+
+// Pre stack: context, segment, offset, value, retdest
+// Post stack: offset'
+global mstore_unpacking_u64_LE:
+    %stack (context, segment, offset, value) -> (0xff, value, context, segment, offset, value)
+    AND
+    DUP4 DUP4 DUP4 MSTORE_GENERAL // First byte
+    %stack (context, segment, offset, value) -> (0xff00, value, context, segment, offset, value)
+    AND %shr_const(8)
+    DUP4 %add_const(1) DUP4 DUP4 MSTORE_GENERAL // Second byte
+    %stack (context, segment, offset, value) -> (0xff0000, value, context, segment, offset, value)
+    AND %shr_const(16)
+    DUP4 %add_const(2) DUP4 DUP4 MSTORE_GENERAL // Third byte
+    %stack (context, segment, offset, value) -> (0xff000000, value, context, segment, offset, value)
+    AND %shr_const(24)
+    DUP4 %add_const(3) DUP4 DUP4 MSTORE_GENERAL // Fourth byte
+    %stack (context, segment, offset, value) -> (0xff00000000, value, context, segment, offset, value)
+    AND %shr_const(32)
+    DUP4 %add_const(4) DUP4 DUP4 MSTORE_GENERAL // Fifth byte
+    %stack (context, segment, offset, value) -> (0xff0000000000, value, context, segment, offset, value)
+    AND %shr_const(40)
+    DUP4 %add_const(5) DUP4 DUP4 MSTORE_GENERAL // Sixth byte
+    %stack (context, segment, offset, value) -> (0xff000000000000, value, context, segment, offset, value)
+    AND %shr_const(48)
+    DUP4 %add_const(6) DUP4 DUP4 MSTORE_GENERAL // Seventh byte
+    %stack (context, segment, offset, value) -> (0xff00000000000000, value, context, segment, offset, value)
+    AND %shr_const(56)
+    DUP4 %add_const(7) DUP4 DUP4 MSTORE_GENERAL // Eighth byte
+    %pop4 JUMP
+
+%macro mstore_unpacking_u64_LE
+    %stack (addr: 3, value) -> (addr, value, %%after)
+    %jump(mstore_unpacking_u64_LE)
+%%after:
+%endmacro