[lld] [lld] Merge GOT entries for symbols that have been ICFed (PR #131630)

[Peter Smith via llvm-commits]

================
@@ -0,0 +1,1057 @@
+// REQUIRES: aarch64
+
+// RUN: llvm-mc -filetype=obj -triple=aarch64 %s -o %t
+// RUN: ld.lld %t -o %t2 --icf=all
+// RUN: llvm-objdump --section-headers %t2 | FileCheck %s
+
+// CHECK: {{.*}}.got 00000008{{.*}}
+
+.addrsig
+
+callee:
+ret
+
+.section .rodata.dummy1,"a", at progbits
+sym1:
+.long 111
+.long 122
+.byte 123
+
+.section .rodata.dummy2,"a", at progbits
+sym2:
+.long 111
+.long 122
+sym3:
+.byte 123
+
+.macro f, index
+
+.section .text.f1_\index,"ax", at progbits
+f1_\index:
+adrp x0, :got:g\index
+mov x1, #\index
+b f2_\index
+
+.section .text.f2_\index,"ax", at progbits
+f2_\index:
+ldr x0, [x0, :got_lo12:g\index] 
+b callee
+
+.section .rodata.g\index,"a", at progbits
+g_\index:
+.long 111
+.long 122
+
+g\index:
+.byte 123
+
+.section .text._start,"ax", at progbits
+bl f1_\index
+
+.endm
+
+.section .text._start,"ax", at progbits
+.globl _start
+_start:
+
+f 0
----------------
smithp35 wrote:

In this particular case I don't think it is necessary to show that ADRP may end up in a different page. I think the underlying problem is that we have two relocations that must target the same GOT entry are now different. If the fix didn't remove the uneccessary GOT entries for the folded symbols, but instead made the calculations of the LDR refer to the correct entry then we'd need them, and we'd need a test for the disassembly.

LLD tends to prefer the smallest possible test case as it is normally easier to understand and update, ultimately it would be MaskRay's opinion that counts.

https://github.com/llvm/llvm-project/pull/131630