Implement memcpy

This can be used, for example, to copy `CALL` data (which is a slice of the caller's main memory) to the callee's `CALLDATA` segment.

evm/src/cpu/kernel/asm/basic_macros.asm

@@ -26,6 +26,24 @@
     %endrep
 %endmacro
 
+%macro pop5
+    %rep 5
+        pop
+    %endrep
+%endmacro
+
+%macro pop6
+    %rep 6
+        pop
+    %endrep
+%endmacro
+
+%macro pop7
+    %rep 7
+        pop
+    %endrep
+%endmacro
+
 %macro add_const(c)
     // stack: input, ...
     PUSH $c

evm/src/cpu/kernel/asm/memory.asm

@@ -1,4 +1,4 @@
-// Load a byte from the given segment of the current context's memory space.
+// Load a value from the given segment of the current context's memory space.
 // Note that main memory values are one byte each, but in general memory values
 // can be 256 bits. This macro deals with a single address (unlike MLOAD), so
 // if it is used with main memory, it will load a single byte.
@@ -12,7 +12,7 @@
     // stack: value
 %endmacro
 
-// Store a byte to the given segment of the current context's memory space.
+// Store a value to the given segment of the current context's memory space.
 // Note that main memory values are one byte each, but in general memory values
 // can be 256 bits. This macro deals with a single address (unlike MSTORE), so
 // if it is used with main memory, it will store a single byte.
@@ -25,3 +25,55 @@
     MSTORE_GENERAL
     // stack: (empty)
 %endmacro
+
+// Copies `count` values from
+//     SRC = (src_ctx, src_segment, src_addr)
+// to
+//     DST = (dst_ctx, dst_segment, dst_addr).
+// These tuple definitions are used for brevity in the stack comments below.
+global memcpy:
+    JUMPDEST
+    // stack: DST, SRC, count, retdest
+    DUP7
+    // stack: count, DST, SRC, count, retdest
+    ISZERO
+    // stack: count == 0, DST, SRC, count, retdest
+    %jumpi memcpy_finish
+    // stack: DST, SRC, count, retdest
+
+    // Copy the next value.
+    DUP6
+    DUP6
+    DUP6
+    // stack: SRC, DST, SRC, count, retdest
+    MLOAD_GENERAL
+    // stack: value, DST, SRC, count, retdest
+    DUP4
+    DUP4
+    DUP4
+    // stack: DST, value, DST, SRC, count, retdest
+    MSTORE_GENERAL
+    // stack: DST, SRC, count, retdest
+
+    // Increment dst_addr.
+    SWAP2
+    %add_const(1)
+    SWAP2
+    // Increment src_addr.
+    SWAP5
+    %add_const(1)
+    SWAP5
+    // Decrement count.
+    SWAP6
+    %sub_const(1)
+    SWAP6
+
+    // Recurse!
+    JUMP memcpy
+
+memcpy_finish:
+    JUMPDEST
+    // stack: DST, SRC, count, retdest
+    %pop7
+    // stack: retdest
+    JUMP