ELF: Do not relax ADRP/LDR -> ADRP/ADD for absolute symbols in PIC.

[pcc]

Summary

GOT references to absolute symbols can't be relaxed to use ADRP/ADD in
position-independent code because these instructions produce a relative
address.

[MaskRay]

Thanks for catching this case. For absolute symbols we can choose not relaxing in -no-pie mode. Do you prefer relaxing?

lld/test/ELF/aarch64-adrp-ldr-got-symbols.s

42

can be applied in -no-pie mode?

[pcc]

In D128492#3606984, @MaskRay wrote:

Thanks for catching this case. For absolute symbols we can choose not relaxing in -no-pie mode. Do you prefer relaxing?

Either way is fine with me.

[pcc]

Comment

[peter.smith]

Thanks for the fix. I'll see about getting a condition added to the ABI to mention Absolute addresses in a PI context.

[peter.smith]

For the ABI https://github.com/ARM-software/abi-aa/pull/155

lld/ELF/Arch/AArch64.cpp

bool AArch64Relaxer::tryRelaxAdrpLdr(const Relocation &adrpRel,
   uint32_t adrpDestReg = adrpInstr & 0x1f;
   uint32_t ldrDestReg = ldrInstr & 0x1f;
   uint32_t ldrSrcReg = (ldrInstr >> 5) & 0x1f;
   // Check if ADPR and LDR use the same register.
   if (adrpDestReg != ldrDestReg || adrpDestReg != ldrSrcReg)
     return false;
 
   Symbol &sym = *adrpRel.sym;
+  // GOT references to absolute symbols can't be relaxed to use ADRP/ADD in
+  // position-independent code because these instructions produce a relative
+  // address.
+  if (config->isPic && !cast<Defined>(sym).section)
+    return false;
   // Check if the address difference is within 4GB range.
   int64_t val =
       getAArch64Page(sym.getVA()) - getAArch64Page(secAddr + adrpRel.offset);
   if (val != llvm::SignExtend64(val, 33))
     return false;
 
   Relocation adrpSymRel = {R_AARCH64_PAGE_PC, R_AARCH64_ADR_PREL_PG_HI21,
                            adrpRel.offset, /*addend=*/0, &sym};

lld/test/ELF/aarch64-adrp-ldr-got-symbols.s

 ## This test verifies that the pair adrp + ldr is relaxed/not relaxed
 ## depending on the target symbol properties.
 
 # REQUIRES: aarch64
 # RUN: rm -rf %t && split-file %s %t
 
 # RUN: llvm-mc -filetype=obj -triple=aarch64 %t/symbols.s -o %t/symbols.o
+# RUN: llvm-mc -filetype=obj -triple=aarch64 %t/abs.s -o %t/abs.o
 
-# RUN: ld.lld -shared -T %t/linker.t %t/symbols.o -o %t/symbols.so
+# RUN: ld.lld -shared -T %t/linker.t %t/symbols.o %t/abs.o -o %t/symbols.so
 # RUN: llvm-objdump --no-show-raw-insn -d %t/symbols.so | \
 # RUN:   FileCheck --check-prefix=LIB %s
 
 ## Symbol 'hidden_sym' is nonpreemptible, the relaxation should be applied.
 LIB:      adrp   x0
 LIB-NEXT: add    x0
 
 ## Symbol 'global_sym' is preemptible, no relaxations should be applied.
 LIB-NEXT: adrp   x1
 LIB-NEXT: ldr    x1
 
 ## Symbol 'undefined_sym' is undefined, no relaxations should be applied.
 LIB-NEXT: adrp   x2
 LIB-NEXT: ldr    x2
 
 ## Symbol 'ifunc_sym' is STT_GNU_IFUNC, no relaxations should be applied.
 LIB-NEXT: adrp   x3
 LIB-NEXT: ldr    x3
 
-# RUN: ld.lld -T %t/linker.t -z undefs %t/symbols.o -o %t/symbols
+## Symbol 'abs_sym' is absolute, no relaxations should be applied.
+LIB-NEXT: adrp   x4
+LIB-NEXT: ldr    x4
+
+# RUN: ld.lld -T %t/linker.t -z undefs %t/symbols.o %t/abs.o -o %t/symbols
 # RUN: llvm-objdump --no-show-raw-insn -d %t/symbols | \
 # RUN:   FileCheck --check-prefix=EXE %s
 
 ## Symbol 'global_sym' is nonpreemptible, the relaxation should be applied.
 EXE:      adrp   x1
 EXE-NEXT: add    x1
 
+## Symbol 'abs_sym' is absolute, relaxations should be applied.
+EXE:      adrp   x4
+EXE-NEXT: add    x4
+
 ## The linker script ensures that .rodata and .text are sufficiently (>1MB)
 ## far apart so that the adrp + ldr pair cannot be relaxed to adr + nop.
 #--- linker.t
 SECTIONS {
  .rodata 0x1000: { *(.rodata) }
  .text   0x300100: { *(.text) }
 }
 
+# This symbol is defined in a separate file to prevent the definition from
+# being folded into the instructions that reference it.
+#--- abs.s
+.global abs_sym
+.hidden abs_sym
+abs_sym = 0x1000
+
 #--- symbols.s
 .rodata
 .hidden hidden_sym
 hidden_sym:
 .word 10
 
 .global global_sym
 global_sym:
_start:
   adrp    x0, :got:hidden_sym
   ldr     x0, [x0, #:got_lo12:hidden_sym]
   adrp    x1, :got:global_sym
   ldr     x1, [x1, #:got_lo12:global_sym]
   adrp    x2, :got:undefined_sym
   ldr     x2, [x2, #:got_lo12:undefined_sym]
   adrp    x3, :got:ifunc_sym
   ldr     x3, [x3, #:got_lo12:ifunc_sym]
+  adrp    x4, :got:abs_sym
+  ldr     x4, [x4, #:got_lo12:abs_sym]