P-code
P-code is the language everything above the processor layer speaks. Every instruction of every processor is translated (“lifted”) into a short sequence of p-code operations, and analysis, emulation and the decompiler work on those alone. That is how one decompiler serves 39 processor families.
Basilisk’s p-code is Ghidra’s: the same operations with the same meaning, down to the result of every arithmetic corner case.
Varnodes
Section titled “Varnodes”P-code has one kind of operand, the varnode: a run of bytes somewhere.
(register,0x0,4) 4 bytes at offset 0 of the register space(ram,0x401000,8) 8 bytes of memory at address 0x401000(const,0x1,4) the constant 1, 4 bytes wide(unique,0x58300,4) a temporary inside one instructionA varnode is an address space, an offset and a size. It has no type and no name. That
(register,0x0,4) is called EAX is something the language adds for display.
| Space | Holds |
|---|---|
ram (the name varies by processor) |
The program’s memory |
register |
The processor’s registers, laid out by the language |
const |
Constants: the offset is the value |
unique |
Temporaries that live within one instruction |
Some processors have more: separate code and data memories, for example.
Operations
Section titled “Operations”An operation has an opcode, input varnodes and at most one output. An example, x86’s ADD EAX,0x1, in
part:
(register,0x200,1,CF) = INT_CARRY (register,0x0,4,EAX), (const,0x1,4)(register,0x0,4,EAX) = INT_ADD (register,0x0,4,EAX), (const,0x1,4)(register,0x206,1,ZF) = INT_EQUAL (register,0x0,4,EAX), (const,0x0,4)Nothing is implicit. Flags are ordinary one-byte varnodes, computed by ordinary operations.
| Family | Operations |
|---|---|
| Moving data | COPY, LOAD, STORE |
| Integer arithmetic | INT_ADD, INT_SUB, INT_MULT, INT_DIV, INT_SDIV, INT_REM, INT_SREM, INT_NEGATE, INT_2COMP |
| Bits and shifts | INT_AND, INT_OR, INT_XOR, INT_LEFT, INT_RIGHT, INT_SRIGHT, POPCOUNT, LZCOUNT |
| Comparison | INT_EQUAL, INT_NOTEQUAL, INT_LESS, INT_SLESS, INT_LESSEQUAL, INT_SLESSEQUAL, INT_CARRY, INT_SCARRY, INT_SBORROW |
| Changing size | INT_ZEXT, INT_SEXT, SUBPIECE, PIECE |
| Booleans | BOOL_AND, BOOL_OR, BOOL_XOR, BOOL_NEGATE |
| Floating point | FLOAT_ADD, FLOAT_SUB, FLOAT_MULT, FLOAT_DIV, FLOAT_NEG, FLOAT_ABS, FLOAT_SQRT, comparisons, conversions, rounding |
| Control flow | BRANCH, CBRANCH, BRANCHIND, CALL, CALLIND, RETURN |
| Escape | CALLOTHER: an operation the specification names but does not define, such as a system call or a vector instruction |
The decompiler adds a few of its own while it works (MULTIEQUAL and INDIRECT for SSA form, PTRADD,
PTRSUB and CAST for types); instructions never lift to those.
Two properties worth knowing
Section titled “Two properties worth knowing”Sizes are explicit. Every varnode has a size, and every operation has rules for the sizes of its inputs and output. A 32-bit add of 64-bit registers does not exist: the specification must extract, add and extend, and each step is visible.
Addresses of p-code. One instruction lifts to several operations, so an operation is identified by the
instruction’s address and its index within the instruction. A branch whose target is a constant is a jump
within the instruction’s own p-code; that is how a specification writes loops and conditions, as in
x86’s REP prefixes.
In Basilisk
Section titled “In Basilisk”| Crate | Part |
|---|---|
bk-core |
Address spaces, addresses, ranges and varnode data |
bk-pcode |
The opcodes, the arithmetic of every operation, and the floating-point formats |
The arithmetic matters for parity. When the decompiler folds constants, or an emulator runs, the result of
INT_SDIV on the most negative number, or of a float-to-integer conversion that overflows, must be what
Ghidra computes. bk-pcode is tested against tables generated from Ghidra’s own implementation for every
operation and size.
See p-code for your own bytes with basilisk sleigh decode.