Xtensa in the ESP Chips

Introduction

Xtensa is configured per chip. The table says which options each ESP chip's core has. The ESP32, ESP32-S2 and ESP32-S3 columns are what the Espressif toolchain's objdump for that chip decodes (the toolchain selects each chip's configuration with a wrapper); the ESP32 and ESP32-S3 were also read from Tensilica's own configuration files (XCHAL_HAVE_*), and the ESP8266 only from its configuration file. The ESP32-C2, C3, C6, H2 and P4 use RISC-V and are on the RISC-V pages.

Options by chip

What an ELF file says about the chip

Nothing directly. The ESP toolchains write e_machine 94 and e_flags 0x300 (EF_XTENSA_XT_INSN and EF_XTENSA_XT_LIT, both "the linker may relax instructions and literals"); the machine field of e_flags (EF_XTENSA_MACH, the low four bits) is 0 for every ESP chip. A .xtensa.info note holds the ABI (0 windowed, 1 call0) and USE_ABSOLUTE_LITERALS. So the explorer's Mode choice is how you say which chip, and "From the file" means the ESP32.

How the decoder was compared

Each chip's objdump was run over a raw binary with every word placed at a fixed 8-byte stride (padding that decodes in whole instructions, so a rejected word cannot throw the next one out of step) and the lines matched by address. The words: every 24-bit word of op0 0 to 7 (all 2²⁰ words of each), every 16-bit word of op0 8 to 13, and 4096 words of each of op0 14 and 15 (on the ESP32-S3, 40,000 random 32-bit words of each). Texts are compared after collapsing white space; lengths are compared too.

The explained differences are all the same two behaviours of objdump:

• For an encoding that is not an instruction, the ESP32 and ESP32-S3 builds print the first opcode of their table, lsi, with the fields they find; the ESP32-S2 build prints a one-byte .byte. The explorer keeps the length the opcode gives (3 or 2 bytes) so the instructions that follow line up as the hardware reads them.
• Objdump decodes only the bits that tell its opcodes apart, so it reads ret, retw, ret.n and retw.n whatever their s field holds (60 words in each run). The explorer calls a nonzero s reserved.

On the ESP32-S3, 190 of the 217 EE.* mnemonics are decoded. A random sample of 80,000 32-bit words showed no difference from objdump in the decoded ones; the 27 missing mnemonics are rare 32-bit forms whose encodings did not turn up. Not every word of the 32-bit space was compared (it has 2²⁹ words).

Not described

Known, not yet covered

• Reference and viewer: the generic Xtensa configuration's cache, prefetch and MMU-register instructions, and the FLIX bundles of its objdump; no ESP chip has them.
• Viewer: the ESP8266 has no independent decoder to compare with here beyond the configuration file.

Unknown

• Reference: whether the ESP32-S2 lacks S32C1I and the loop option is established by objdump's behaviour; the S2 core file was not on this machine.

Links