[LLD][ARM] Handle .ARM.exidx sections at non-zero output sec offset

[peter.smith]

Summary

Embedded systems that do not use an ELF loader locate the .ARM.exidx exception table via linker defined exidx_start and exidx_end rather than use the PT_ARM_EXIDX program header. This means that some linker scripts such as the picolibc C library's linker script, do not have the .ARM.exidx sections at offset 0 in the OutputSection. For example from https://github.com/picolibc/picolibc/blob/main/picolibc.ld.in#L126:

.except_unordered : {
    . = ALIGN(8);
    PROVIDE(__exidx_start = .);
    *(.ARM.exidx*)
    PROVIDE(__exidx_end = .);
} >flash AT>flash :text

This is within the specification of Arm exception tables, and is handled correctly by ld.bfd. The specification even permits the .ARM.exidx sections to be placed within the .rodata section, but that is very uncommon.

This patch has 2 parts. The first updates the writing of the data of the .ARM.exidx SyntheticSection to account for a non-zero OutputSection offset. The second part makes the PT_ARM_EXIDX program header generation a special case so that it covers only the SyntheticSection and not the parent OutputSection. While not strictly necessary for programs locating the exception tables via the symbols it may cause ELF utilities that locate the exception tables via the PT_ARM_EXIDX program header to fail. This does not seem to be the
case for GNU and LLVM readelf which seems to look for the SHT_ARM_EXIDX section.

An alternative simpler way of handling the PT_ARM_EXIDX is to write a PT_NULL program header if the output section offset of the SyntheticSection is non 0. I don't think that we can remove the program header as the size of the table has already been calculated.

[peter.smith]

Update comments, better description of expected results and add missing full stops.

[MaskRay]

Thanks for the patch. I prefer the current approach than changing PT_ARM_EXIDX to PT_NULL. The code change looks good.

This does not seem to be the case for GNU and LLVM readelf which seems to look for the SHT_ARM_EXIDX section.

Confirmed that they just check the section type SHT_ARM_EXIDX, not the program header type. Perhaps readelf -D -u can be made to check PT_ARM_EXIDX, but the change is likely over-engineering..

lld/ELF/Writer.cpp

2631

Perhaps use if (part.armExidx && p->p_type == PT_ARM_EXIDX). The two conditions are equivalent, but checking part.armExidx is safer in case part.armExidx is null.

lld/test/ELF/arm-exidx-nonzero-offset.s

10

Shall we use llvm-readelf --unwind output as well? The output can be more readable, though it doesn't use the GNU readelf style...

29

42

.exceptions has different alignment than the previous line.

[peter.smith]

Address review comments:

• Use part.armExidx instead of config->emachine
• Tidy up white space in test.
• Add llvm-readelf --unwind output to test.

I noticed that llvm-readelf --unwind can't decode function names with the Thumb bit set although arm-none-eabi-readelf can. I've used CHECK rather than CHECK-NEXT just in case llvm-readelf --unwind output is improved.

[MaskRay]

Looks great!

Optional: consider adding a note to lld/docs/ReleaseNotes.rst. There are some examples in previous releases, e.g. git show origin/release/15.x:lld/docs/ReleaseNotes.rst

[peter.smith]

Add entry in release notes under Fixes.

lld/ELF/SyntheticSections.cpp

     }
     executableSections = std::move(selectedSections);
   }
-
+  // offset is within the SyntheticSection.
   size_t offset = 0;
   size = 0;
   for (InputSection *isec : executableSections) {

            dataOffset += 4)
         write32(buf + offset + dataOffset,
                 read32(d->content().data() + dataOffset));
-      target->relocateAlloc(*d, buf + d->outSecOff);
+      // Recalculate outSecOff as finalizeAddressDependentContent()
+      // may have altered syntheticSection outSecOff.
+      d->outSecOff = offset + outSecOff;
+      target->relocateAlloc(*d, buf + offset);
       offset += d->getSize();
     } else {
       // A Linker generated CANTUNWIND section.

lld/ELF/Writer.cpp

     OutputSection *first = p->firstSec;
     OutputSection *last = p->lastSec;
 
+    // .ARM.exidx sections may not be within a single .ARM.exidx
+    // output section. We always want to describe just the
+    // SyntheticSection.
+    if (part.armExidx && p->p_type == PT_ARM_EXIDX) {
+      p->p_filesz = part.armExidx->getSize();
+      p->p_memsz = part.armExidx->getSize();
+      p->p_offset = first->offset + part.armExidx->outSecOff;
+      p->p_vaddr = first->addr + part.armExidx->outSecOff;
+      p->p_align = part.armExidx->addralign;
+      if (part.elfHeader)
+        p->p_offset -= part.elfHeader->getParent()->offset;
+
+      if (!p->hasLMA)
+        p->p_paddr = first->getLMA() + part.armExidx->outSecOff;
+      return;
+    }
+
     if (first) {
       p->p_filesz = last->offset - first->offset;
       if (last->type != SHT_NOBITS)

lld/docs/ReleaseNotes.rst

 WebAssembly Improvements
 ------------------------
 
+Fixes
+#####
+
+* Arm exception index tables (.ARM.exidx sections) are now ouptut
+  correctly when they are at a non zero offset within their output
+  section. (`D148033 <https://reviews.llvm.org/D148033>`_)

lld/test/ELF/arm-exidx-nonzero-offset.s

+// REQUIRES: arm
+// RUN: rm -rf %t && split-file %s %t
+
+// RUN: llvm-mc -filetype=obj -triple=armv7a-none-linux-gnueabi --arm-add-build-attributes %t/a.s -o %t/a.o
+// RUN: ld.lld %t/a.o -T %t/exidx-non-zero-offset.t -o %t/non-zero
+// RUN: llvm-readelf --program-headers --unwind --symbols -x .exceptions %t/non-zero | FileCheck %s
+// RUN: ld.lld %t/a.o -T %t/exidx-zero-offset.t -o %t/zero
+// RUN: llvm-readelf --program-headers --unwind --symbols -x .exceptions %t/zero | FileCheck %s
+
+/// On platforms that load ELF files directly the ARM.exidx sections
+/// are located with the PT_ARM_EXIDX program header. This requires
+/// all .ARM.exidx input sections to be placed in a single
+/// .ARM.exidx output section. Embedded systems that do not load
+/// from an ELF file use the linker defined symbols __exidx_start and
+/// __exidx_stop. There is no requirement to place the .ARM.exidx
+/// input sections in their own output section. This test case checks
+/// that a .ARM.exidx synthetic section that isn't at a zero offset
+/// within the output section gets the correct offsets. We check
+/// this by checking that equal amounts of alignment padding
+/// inserted before and after the section start produce the same
+/// results.
+
+/// For the two linker scripts, one with the .ARM.exidx table
+/// at a zero offset within its output section and one with a
+/// non-zero offset, but both at the same file and address;
+/// the PT_ARM_EXIDX, symbols and table contents should be the
+/// same.
+// CHECK:      EXIDX 0x010080 0x00000080 0x00000080 0x00028 0x00028 R   0x4
+
+/// unwind entries starting from the right address are identical.
+/// llvm-readelf does not seem to be able to detect Thumb functionNames
+/// whereas arm-none-eabi-readelf can. Used CHECK rather than CHECK-NEXT
+/// to cope with possible improvements llvm-readelf.
+// CHECK:        FunctionAddress: 0x0
+// CHECK-NEXT:   FunctionName: f1
+// CHECK-NEXT:   Model: Compact (Inline)
+// CHECK-NEXT:   PersonalityIndex: 0
+// CHECK-NEXT:   Opcodes [
+// CHECK-NEXT:     0x97      ; vsp = r7
+// CHECK-NEXT:     0x84 0x08 ; pop {r7, lr}
+// CHECK-NEXT:   ]
+// CHECK-NEXT: }
+// CHECK-NEXT: Entry {
+// CHECK-NEXT:   FunctionAddress: 0x8
+// CHECK-NEXT:   FunctionName: f2
+// CHECK-NEXT:   Model: CantUnwind
+// CHECK-NEXT: }
+// CHECK-NEXT: Entry {
+// CHECK-NEXT:   FunctionAddress: 0xE
+// CHECK:        ExceptionHandlingTable: .ARM.extab
+// CHECK-NEXT:   TableEntryAddress: 0xA8
+// CHECK-NEXT:   Model: Generic
+// CHECK-NEXT:   PersonalityRoutineAddress: 0x12
+// CHECK-NEXT: }
+// CHECK-NEXT: Entry {
+// CHECK-NEXT:   FunctionAddress: 0x10
+// CHECK:        Model: CantUnwind
+// CHECK-NEXT: }
+// CHECK-NEXT: Entry {
+// CHECK-NEXT:   FunctionAddress: 0x14
+// CHECK-NEXT:   FunctionName: __ARMv7ABSLongThunk_f3
+// CHECK-NEXT:   Model: CantUnwind
+// CHECK-NEXT: }
+
+// CHECK:      {{[0-9]+}}: 00000080 {{.*}} __exidx_start
+// CHECK-NEXT: {{[0-9]+}}: 000000a8 {{.*}} __exidx_end
+
+// CHECK:      0x00000080 80ffff7f 08849780 80ffff7f 01000000
+// CHECK-NEXT: 0x00000090 7effff7f 14000000 78ffff7f 01000000
+// CHECK-NEXT: 0x000000a0 74ffff7f 01000000
+
+//--- exidx-non-zero-offset.t
+SECTIONS {
+  .text : { *(.text .text.*) }
+  /* Addition of thunk in .text changes alignment padding */
+  .exceptions : {
+  /* Alignment padding within .exceptions */
+  . = ALIGN(128);
+  /* Embedded C libraries find exceptions via linker defined
+     symbols */
+  __exidx_start = .;
+  *(.ARM.exidx) ;
+  __exidx_end = .;
+  }
+  .ARM.extab : { *(.ARM.extab .ARM.extab.*) }
+}
+
+//--- exidx-zero-offset.t
+SECTIONS {
+  .text : { *(.text .text.*) }
+  /* Addition of thunk in .text changes alignment padding */
+  /* alignment padding before .exceptions starts */
+  .exceptions : ALIGN(128) {
+  /* Embedded C libraries find exceptions via linker defined
+     symbols */
+  __exidx_start = .;
+  *(.ARM.exidx) ;
+  __exidx_end = .;
+  }
+  .ARM.extab : { *(.ARM.extab .ARM.extab.*) }
+}
+
+//--- a.s
+.syntax unified
+
+/// Expect inline unwind instructions.
+.section .text.01, "ax", %progbits
+.arm
+.balign 4
+.global f1
+.type f1, %function
+f1:
+.fnstart
+/// provoke an interworking thunk.
+b f3
+bx lr
+.save {r7, lr}
+.setfp r7, sp, #0
+.fnend
+
+/// Expect no unwind information from assembler. The linker must
+/// synthesise an EXIDX_CANTUNWIND entry to prevent an exception
+/// thrown through f2 from matching against the unwind instructions
+/// for f1.
+.section .text.02, "ax", %progbits
+.global f2
+.type f2, %function
+.balign 4
+f2:
+bx lr
+
+/// Expect 1 EXIDX_CANTUNWIND entry that can be merged into the linker
+/// generated EXIDX_CANTUNWIND as if the assembler had generated it.
+.section .text.03, "ax",%progbits
+.global f3
+.type f3, %function
+.thumb
+.balign 2
+f3:
+.fnstart
+bx lr
+.cantunwind
+.fnend
+
+/// Expect a section with a reference to an .ARM.extab.
+.section .text.04, "ax",%progbits
+.global f4
+.balign 2
+.type f4, %function
+f4:
+.fnstart
+bx lr
+.personality __gxx_personality_v0
+.handlerdata
+.long 0
+.fnend
+
+
+/// Dummy implementation of personality routines to satisfy reference
+/// from exception tables, linker will generate EXIDX_CANTUNWIND.
+.section .text.__aeabi_unwind_cpp_pr0, "ax", %progbits
+.global __aeabi_unwind_cpp_pr0
+.type __aeabi_unwind_cpp_pr0, %function
+.balign 2
+__aeabi_unwind_cpp_pr0:
+bx lr
+
+.section .text.__gcc_personality_v0, "ax", %progbits
+.global __gxx_personality_v0
+.type __gcc_personality_v0, %function
+.balign 2
+__gxx_personality_v0:
+bx lr