[Bug target/92424] New: [aarch64] Broken code with -fpatchable-function-entry and BTI

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92424

            Bug ID: 92424
           Summary: [aarch64] Broken code with -fpatchable-function-entry
                    and BTI
           Product: gcc
           Version: 9.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ktkachov at gcc dot gnu.org
  Target Milestone: ---

-fpatchable-function-entry seems to interact badly with
-mbranch-protection=standard (used to insert BTI markers)

void g(void);

void f(void)
{ g(); g(); }

aarch64-linux-gnu-gcc  -mbranch-protection=standard
-fpatchable-function-entry=2,1 -O3 -c foo.c -save-temps


Produces:
        .arch armv8-a
        .file   "foo.c"
        .text
        .align  2
        .p2align 3,,7
        .global f
        .section        __patchable_function_entries,"aw",@progbits
        .8byte  .LPFE1
        .text
.LPFE1:
        nop
        .type   f, %function
f:
        nop  // <------- Function entry has NOP rather than PACIASP landing pad
.LFB0:
        .cfi_startproc
        hint    25 // paciasp
        .cfi_window_save
        stp     x29, x30, [sp, -16]!
        .cfi_def_cfa_offset 16
        .cfi_offset 29, -16
        .cfi_offset 30, -8
        mov     x29, sp
        bl      g
        ldp     x29, x30, [sp], 16
        .cfi_restore 30
        .cfi_restore 29
        .cfi_def_cfa_offset 0
        hint    29 // autiasp
        .cfi_window_save
        b       g
        .cfi_endproc
.LFE0:
        .size   f, .-f


Could a possible solution be to emit ofr -mbranch-protection={standard or bti}
-fpatchable-function-entry=N,M:

f-4*M:
        .rept M
        nop
        .endr
f:
        bti c
        .rept N-M
        nop
        .endr
        // paciasp may be used from here on in

Unless we fix the codegen we should perhaps error out when the two options are
combined rather than producing silently wrong code?