[PATCH] hppa: Implement PA 2.0 symbolic relocations for long displacements

[PATCH] hppa: Implement PA 2.0 symbolic relocations for long displacements

[John David Anglin]

[-- Attachment #1: Type: text/plain, Size: 4445 bytes --]

Tested on hppa-unknown-linux-gnu.  There are no regressions in
all the binutils suites.  I also tested with a modified version
of gcc which allowed symbolic floating point loads and stores.
There were no regressions in the gcc tests.

Okay for trunk?

Dave
---

hppa: Implement PA 2.0 symbolic relocations for long displacements

The PA 2.0 architecture introduced several new load and store
instructions with long displacements.  These include floating
point loads and stores for word mode, and integer and floating
point loads and stores for double words.  Currently, ld does
not correctly support symbolic relocations for these instructions.

If these are used, ld applies the standard R_PARISC_DPREL14R
relocation and corrupts the instruction.  This change uses
bfd_hppa_insn2fmt to determine the correct relocation format.

We need to check the computed displacement as the immediate
value used in these instruction must be a multiple of 4 or 8
depending on whether the access is for a word or double word.

A misaligned offset can potentially occur if the symbol is not
properly aligned or if $global$ (the global pointer) is not
double word aligned.  $global$ is provided as a .data section
start symbol.  The patch adjusts elf.sc and hppalinux.sh to
align .data to a 8-byte boundary in non-shared and non-pie
links.

2024-04-01  John David Anglin  <danglin@gcc.gnu.org>

	PR ld/31503

bfd/ChangeLog:

	* elf32-hppa.c (final_link_relocate): Output

ld/ChangeLog:

	* emulparams/hppalinux.sh (DATA_SECTION_ALIGNMENT): Define.
	* scripttempl/elf.sc: Align .data section to DATA_SECTION_ALIGNMENT
	when relocating.

diff --git a/bfd/elf32-hppa.c b/bfd/elf32-hppa.c
index bb4bfdc00b8..e9b3ce24359 100644
--- a/bfd/elf32-hppa.c
+++ b/bfd/elf32-hppa.c
@@ -3110,7 +3110,7 @@ final_link_relocate (asection *input_section,
   unsigned int r_type = ELF32_R_TYPE (rela->r_info);
   unsigned int orig_r_type = r_type;
   reloc_howto_type *howto = elf_hppa_howto_table + r_type;
-  int r_format = howto->bitsize;
+  int r_format;
   enum hppa_reloc_field_selector_type_alt r_field;
   bfd *input_bfd = input_section->owner;
   bfd_vma offset = rela->r_offset;
@@ -3432,6 +3432,50 @@ final_link_relocate (asection *input_section,
       break;
     }
 
+  r_format = bfd_hppa_insn2fmt (input_bfd, insn);
+  switch (r_format)
+    {
+    case 10:
+    case -10:
+      if (val & 7)
+	{
+	  _bfd_error_handler
+	    /* xgettext:c-format */
+	    (_("%pB(%pA+%#" PRIx64 "): displacement %#x for insn %#x "
+	       "is not a multiple of 8 (gp %#x)"),
+	     input_bfd,
+	     input_section,
+	     (uint64_t) offset,
+	     val,
+	     insn,
+	     (unsigned int) elf_gp (input_section->output_section->owner));
+	  bfd_set_error (bfd_error_bad_value);
+	  return bfd_reloc_notsupported;
+	}
+      break;
+
+    case -11:
+    case -16:
+      if (val & 3)
+	{
+	  _bfd_error_handler
+	    /* xgettext:c-format */
+	    (_("%pB(%pA+%#" PRIx64 "): displacement %#x for insn %#x "
+	       "is not a multiple of 4 (gp %#x)"),
+	     input_bfd,
+	     input_section,
+	     (uint64_t) offset,
+	     val,
+	     insn,
+	     (unsigned int) elf_gp (input_section->output_section->owner));
+	  bfd_set_error (bfd_error_bad_value);
+	  return bfd_reloc_notsupported;
+	}
+      break;
+
+    default:
+      break;
+    }
   insn = hppa_rebuild_insn (insn, val, r_format);
 
   /* Update the instruction word.  */
diff --git a/ld/emulparams/hppalinux.sh b/ld/emulparams/hppalinux.sh
index 7892df9130d..5a462119742 100644
--- a/ld/emulparams/hppalinux.sh
+++ b/ld/emulparams/hppalinux.sh
@@ -19,6 +19,7 @@ else
   DATA_SEGMENT_END=". = DATA_SEGMENT_END (.);"
   DATA_SEGMENT_RELRO_END=". = DATA_SEGMENT_RELRO_END (${SEPARATE_GOTPLT-0}, .);"
 fi
+DATA_SECTION_ALIGNMENT="${CREATE_SHLIB-${CREATE_PIE-ALIGN(8)}}"
 ARCH=hppa
 MACHINE=hppa1.1    # We use 1.1 specific features.
 NOP=0x08000240
diff --git a/ld/scripttempl/elf.sc b/ld/scripttempl/elf.sc
index fae7c2ad71c..d5022fa502f 100644
--- a/ld/scripttempl/elf.sc
+++ b/ld/scripttempl/elf.sc
@@ -669,7 +669,7 @@ cat <<EOF
 
   ${DATA_PLT+${PLT_BEFORE_GOT-${PLT}}}
 
-  .data         ${RELOCATING-0} :
+  .data         ${RELOCATING-0}${RELOCATING+${DATA_SECTION_ALIGNMENT}} :
   {
     ${RELOCATING+${DATA_START_SYMBOLS}}
     *(.data${RELOCATING+ .data.* .gnu.linkonce.d.*})

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: [PATCH] hppa: Implement PA 2.0 symbolic relocations for long displacements

[Alan Modra]

On Mon, Apr 01, 2024 at 06:10:38PM +0000, John David Anglin wrote:
> Tested on hppa-unknown-linux-gnu.  There are no regressions in
> all the binutils suites.  I also tested with a modified version
> of gcc which allowed symbolic floating point loads and stores.
> There were no regressions in the gcc tests.
> 
> Okay for trunk?

OK.  (You didn't really need to ask.)

-- 
Alan Modra
Australia Development Lab, IBM