[ELF][test] Add testing for dynamic TLS relocations in .debug_info

[MaskRay]

Please place this in one of the x86-64-tls-ld*.s files. For ppc64 we use ppc64-dtprel.s

lld/test/ELF/relocation-dyn-tls-debug-info.s

31 ↗

(On Diff #378269)

Two values should be sufficient.

You can use .quad 0 or .space to add a gap if you think 0 as a value is too special.

[MaskRay]

lld/test/ELF/relocation-dyn-tls-debug-info.s

10 ↗

(On Diff #378269)

Drop --gc-sections from this test. It doesn't do anything special.

The property is tested in debug-dead-reloc* files.

[andrewng]

In D111436#3051813, @MaskRay wrote:

Please place this in one of the x86-64-tls-ld*.s files. For ppc64 we use ppc64-dtprel.s

This test is testing the patching behaviour of DTPREL in non-alloc sections, specifically .debug_info because that's the usual use case, so I'm not sure it really fits with the x86-64-tls-ld*.s tests. Do you want me to rename this test to x86-64-dtprel.s or do you have another suggestion?

lld/test/ELF/relocation-dyn-tls-debug-info.s

10 ↗

(On Diff #378269)

The focus of this testing is the following code from InputSection::relocateNonAlloc in InputSection.cpp:

if (tombstone ||
    (isDebug && (type == target->symbolicRel || expr == R_DTPREL))) {

This is why I added --gc-sections because although the actual behaviour now is not special there is this specific code path. However, thanks for spotting debug-dead-reloc-tls.s which already provides coverage. Not sure how I managed to miss this in my search for existing coverage. I will remove this --gc-sections test case.

31 ↗

(On Diff #378269)

Yes, two values are sufficient, the third value was really for the --gc-sections test case which is no longer required.

[MaskRay]

In D111436#3053359, @andrewng wrote:

In D111436#3051813, @MaskRay wrote:

Please place this in one of the x86-64-tls-ld*.s files. For ppc64 we use ppc64-dtprel.s

This test is testing the patching behaviour of DTPREL in non-alloc sections, specifically .debug_info because that's the usual use case, so I'm not sure it really fits with the x86-64-tls-ld*.s tests. Do you want me to rename this test to x86-64-dtprel.s or do you have another suggestion?

DTPREL relocations are used in the local-dynamic TLS model. The non-alloc section context is slightly different, but I think just reusing one file is fine (in this case I think fewer tests are easier to organize).

[andrewng]

Address review comments and moved testing to x86-64-tls-ld-local.s.

[MaskRay]

LGTM.

lld/test/ELF/x86-64-tls-ld-local.s

 // REQUIRES: x86
 // RUN: llvm-mc %s -o %t.o -filetype=obj -triple=x86_64-pc-linux
 // RUN: ld.lld %t.o -o %t.so -shared
-// RUN: llvm-readobj -r -S %t.so | FileCheck %s
+// RUN: llvm-readobj -r -S --hex-dump=.debug_info %t.so | FileCheck %s
 
 // CHECK:      Relocations [
 // CHECK-NEXT:   Section ({{.*}}) .rela.dyn {
 // CHECK-NEXT:     R_X86_64_DTPMOD64 - 0x0
 // CHECK-NEXT:     R_X86_64_DTPMOD64 - 0x0
 // CHECK-NEXT:   }
 // CHECK-NEXT:   Section ({{.*}}) .rela.plt {
 // CHECK-NEXT:     R_X86_64_JUMP_SLOT __tls_get_addr 0x0
 // CHECK-NEXT:   }
 // CHECK-NEXT: ]
 
+// CHECK:      Hex dump of section '.debug_info':
+// CHECK-NEXT: 0x00000000 04000000 00000000 00000000 00000000
+
   data16
   leaq  bar@TLSGD(%rip), %rdi
   data16
   data16
   rex64
   callq __tls_get_addr@PLT
 
   leaq  bar@TLSLD(%rip), %rdi
   callq __tls_get_addr@PLT
   leaq  bar@DTPOFF(%rax), %rax
 
   .section  .tdata,"awT",@progbits
 bar:
   .long 42
+foo:
+  .long 43
+
+  .section  .debug_info,"",@progbits
+  .quad foo@DTPOFF
+  .quad bar@DTPOFF