7.10 CFI directives

7.10.1 .cfi_sections section_list

.cfi_sections may be used to specify whether CFI directives should emit .eh_frame section and/or .debug_frame section. If section_list is .eh_frame, .eh_frame is emitted, if section_list is .debug_frame, .debug_frame is emitted. To emit both use .eh_frame, .debug_frame. The default if this directive is not used is .cfi_sections .eh_frame.

On targets that support compact unwinding tables these can be generated by specifying .eh_frame_entry instead of .eh_frame.

Some targets may support an additional name, such as .c6xabi.exidx which is used by the target.

The .cfi_sections directive can be repeated, with the same or different arguments, provided that CFI generation has not yet started. Once CFI generation has started however the section list is fixed and any attempts to redefine it will result in an error.

7.10.2 .cfi_startproc [simple]

.cfi_startproc is used at the beginning of each function that should have an entry in .eh_frame. It initializes some internal data structures. Don’t forget to close the function by .cfi_endproc.

Unless .cfi_startproc is used along with parameter simple it also emits some architecture dependent initial CFI instructions.

7.10.3 .cfi_endproc

.cfi_endproc is used at the end of a function where it closes its unwind entry previously opened by .cfi_startproc, and emits it to .eh_frame.

7.10.4 .cfi_personality encoding [, exp]

.cfi_personality defines personality routine and its encoding. encoding must be a constant determining how the personality should be encoded. If it is 255 (DW_EH_PE_omit), second argument is not present, otherwise second argument should be a constant or a symbol name. When using indirect encodings, the symbol provided should be the location where personality can be loaded from, not the personality routine itself. The default after .cfi_startproc is .cfi_personality 0xff, no personality routine.

7.10.5 .cfi_personality_id id

cfi_personality_id defines a personality routine by its index as defined in a compact unwinding format. Only valid when generating compact EH frames (i.e. with .cfi_sections eh_frame_entry.

7.10.6 .cfi_fde_data [opcode1 [, …]]

cfi_fde_data is used to describe the compact unwind opcodes to be used for the current function. These are emitted inline in the .eh_frame_entry section if small enough and there is no LSDA, or in the .gnu.extab section otherwise. Only valid when generating compact EH frames (i.e. with .cfi_sections eh_frame_entry.

7.10.7 .cfi_lsda encoding [, exp]

.cfi_lsda defines LSDA and its encoding. encoding must be a constant determining how the LSDA should be encoded. If it is 255 (DW_EH_PE_omit), the second argument is not present, otherwise the second argument should be a constant or a symbol name. The default after .cfi_startproc is .cfi_lsda 0xff, meaning that no LSDA is present.

7.10.8 .cfi_inline_lsda [align]

.cfi_inline_lsda marks the start of a LSDA data section and switches to the corresponding .gnu.extab section. Must be preceded by a CFI block containing a .cfi_lsda directive. Only valid when generating compact EH frames (i.e. with .cfi_sections eh_frame_entry.

The table header and unwinding opcodes will be generated at this point, so that they are immediately followed by the LSDA data. The symbol referenced by the .cfi_lsda directive should still be defined in case a fallback FDE based encoding is used. The LSDA data is terminated by a section directive.

The optional align argument specifies the alignment required. The alignment is specified as a power of two, as with the .p2align directive.

7.10.9 .cfi_def_cfa register, offset

.cfi_def_cfa defines a rule for computing CFA as: take address from register and add offset to it.

7.10.10 .cfi_def_cfa_register register

.cfi_def_cfa_register modifies a rule for computing CFA. From now on register will be used instead of the old one. Offset remains the same.

7.10.11 .cfi_def_cfa_offset offset

.cfi_def_cfa_offset modifies a rule for computing CFA. Register remains the same, but offset is new. Note that it is the absolute offset that will be added to a defined register to compute CFA address.

7.10.12 .cfi_adjust_cfa_offset offset

Same as .cfi_def_cfa_offset but offset is a relative value that is added/substracted from the previous offset.

7.10.13 .cfi_offset register, offset

Previous value of register is saved at offset offset from CFA.

7.10.14 .cfi_rel_offset register, offset

Previous value of register is saved at offset offset from the current CFA register. This is transformed to .cfi_offset using the known displacement of the CFA register from the CFA. This is often easier to use, because the number will match the code it’s annotating.

7.10.15 .cfi_register register1, register2

Previous value of register1 is saved in register register2.

7.10.16 .cfi_restore register

.cfi_restore says that the rule for register is now the same as it was at the beginning of the function, after all initial instruction added by .cfi_startproc were executed.

7.10.17 .cfi_undefined register

From now on the previous value of register can’t be restored anymore.

7.10.18 .cfi_same_value register

Current value of register is the same like in the previous frame, i.e. no restoration needed.

7.10.19 .cfi_remember_state and .cfi_restore_state

.cfi_remember_state pushes the set of rules for every register onto an implicit stack, while .cfi_restore_state pops them off the stack and places them in the current row. This is useful for situations where you have multiple .cfi_* directives that need to be undone due to the control flow of the program. For example, we could have something like this (assuming the CFA is the value of rbp):

        je label
        popq %rbx
        .cfi_restore %rbx
        popq %r12
        .cfi_restore %r12
        popq %rbp
        .cfi_restore %rbp
        .cfi_def_cfa %rsp, 8
        ret
label:
        /* Do something else */

Here, we want the .cfi directives to affect only the rows corresponding to the instructions before label. This means we’d have to add multiple .cfi directives after label to recreate the original save locations of the registers, as well as setting the CFA back to the value of rbp. This would be clumsy, and result in a larger binary size. Instead, we can write:

        je label
        popq %rbx
        .cfi_remember_state
        .cfi_restore %rbx
        popq %r12
        .cfi_restore %r12
        popq %rbp
        .cfi_restore %rbp
        .cfi_def_cfa %rsp, 8
        ret
label:
        .cfi_restore_state
        /* Do something else */

That way, the rules for the instructions after label will be the same as before the first .cfi_restore without having to use multiple .cfi directives.

7.10.20 .cfi_return_column register

Change return column register, i.e. the return address is either directly in register or can be accessed by rules for register.

7.10.21 .cfi_signal_frame

Mark current function as signal trampoline.

7.10.22 .cfi_window_save

SPARC register window has been saved.

7.10.23 .cfi_escape expression[, …]

Allows the user to add arbitrary bytes to the unwind info. One might use this to add OS-specific CFI opcodes, or generic CFI opcodes that GAS does not yet support.

7.10.24 .cfi_val_encoded_addr register, encoding, label

The current value of register is label. The value of label will be encoded in the output file according to encoding; see the description of .cfi_personality for details on this encoding.

The usefulness of equating a register to a fixed label is probably limited to the return address register. Here, it can be useful to mark a code segment that has only one return address which is reached by a direct branch and no copy of the return address exists in memory or another register.