Jump to content
Setting up an AI agent? ReverseEngineering.wiki skill (works from the command line, no configuration): https://reverseengineering.wiki/skill/rewiki/SKILL.md · setup for every client, including MCP: https://reverseengineering.wiki/llms.txt · REWiki:Connect

Technique:Recognizing compiler-generated code/Switch jump tables

From ReverseEngineering.wiki

How compilers lower a dense C switch to an unsigned bounds check followed by an indirect jump through a table, and how the table entries are encoded on x86, x64 (image-relative 32-bit) and AArch64 (byte, halfword or word offsets from a label).

A dense switch becomes: subtract the lowest case value, unsigned-compare with the case count, branch to the default case when above, then an indirect jump through a table of code addresses. The table is data in .rdata / .rodata that disassemblers often fail to follow.

Shapes

  • x86 (absolute): lea eax, [ecx-1]; cmp eax, 6; ja default; jmp dword ptr [table + eax*4]; table entries are absolute addresses (they have base relocations in a PE).
  • x64 MSVC (relative): cmp eax, 6; ja default; lea rdx, __ImageBase; mov ecx, [rdx+rax*4+table_rva]; add rcx, rdx; jmp rcx: entries are 32-bit RVAs from the image base. Some MSVC tables are split by an additional byte-index table (movzx eax, byte ptr [index+rax]) when several cases share targets (two-level: byte table selects an entry of the dword table). Unverified: from memory of MSVC output.
  • x64 GCC/Clang, PIC: lea rdx, table(%rip); movslq (%rdx,%rax,4), %rax; add %rdx, %rax; jmp *%rax: entries are signed 32-bit offsets relative to the table start (.long .L5-.L4). Non-PIC code can use absolute jmp *table(,%rax,8). Unverified: from memory of GCC output.
  • AArch64 (LLVM): the JumpTableDest8/16/32 pseudo lowers to adr x, <label>; ldrb|ldrh|ldrsw w/x, [table, index, ...]; add x, x, w, sxtw #2; br x. Entries are 1, 2 or 4 bytes; for the 1- and 2-byte forms they count instructions (4-byte steps) from the adr base label, so the offset is label + entry*4 (ldrb with no scaling, ldrh scaled by 2, ldrsw scaled by 4). The 4-byte form is a signed word. (Read in AArch64AsmPrinter::LowerJumpTableDest: the final add uses shift 0 for 4-byte entries and shift 2 for the compressed forms.)

Recovery

  1. Find the bounds check: the ja/b.hi target is the default case; the immediate is the number of entries minus one.
  2. Read the table at the address in the instruction, N entries of the stride the code uses, decode them as above, and label the targets as cases.
  3. Case values are index plus the constant subtracted before the compare. Sparse switches use a binary search of compares instead, or a hash/lookup table; cases sharing a handler repeat the same target.
  4. Ghidra and IDA recover most tables; when they miss (indirect jump shown as jmp rax), define the table manually and set the switch override.

Evidence & references

  • llvm-project llvm/lib/Target/AArch64/AArch64AsmPrinter.cpp (LowerJumpTableDest).
  • x86/x64 shapes: Unverified, not read in a source for this page.