ELF,SystemZ: Don't sort relocations for TLS GD/LD optimization; support CREL

The General dynamic and Local Dynamic TLS models call
`__tls_get_offset`, which is eliminated after optimization. To avoid an
unneeded PLT entry and its .dynsym entry, llvm#75643 sorted the relocations.

However, sorting relocations does not play well with CREL. Since GNU ld
has a false dependency on `__tls_get_offset` as well, let's remove the
relocation sorting. Fix llvm#149511

Pull Request: llvm#149640

Author: MaskRay

lld/ELF/Arch/SystemZ.cpp

@@ -23,7 +23,6 @@ namespace {
 class SystemZ : public TargetInfo {
 public:
   SystemZ(Ctx &);
-  int getTlsGdRelaxSkip(RelType type) const override;
   RelExpr getRelExpr(RelType type, const Symbol &s,
                      const uint8_t *loc) const override;
   RelType getDynRel(RelType type) const override;
@@ -277,29 +276,6 @@ RelExpr SystemZ::adjustTlsExpr(RelType type, RelExpr expr) const {
   return expr;
 }
 
-int SystemZ::getTlsGdRelaxSkip(RelType type) const {
-  // A __tls_get_offset call instruction is marked with 2 relocations:
-  //
-  //   R_390_TLS_GDCALL / R_390_TLS_LDCALL: marker relocation
-  //   R_390_PLT32DBL: __tls_get_offset
-  //
-  // After the relaxation we no longer call __tls_get_offset and should skip
-  // both relocations to not create a false dependence on __tls_get_offset
-  // being defined.
-  //
-  // Note that this mechanism only works correctly if the R_390_TLS_[GL]DCALL
-  // is seen immediately *before* the R_390_PLT32DBL.  Unfortunately, current
-  // compilers on the platform will typically generate the inverse sequence.
-  // To fix this, we sort relocations by offset in RelocationScanner::scan;
-  // this ensures the correct sequence as the R_390_TLS_[GL]DCALL applies to
-  // the first byte of the brasl instruction, while the R_390_PLT32DBL applies
-  // to its third byte (the relative displacement).
-
-  if (type == R_390_TLS_GDCALL || type == R_390_TLS_LDCALL)
-    return 2;
-  return 1;
-}
-
 void SystemZ::relaxTlsGdToIe(uint8_t *loc, const Relocation &rel,
                              uint64_t val) const {
   // The general-dynamic code sequence for a global `x`:
@@ -320,6 +296,9 @@ void SystemZ::relaxTlsGdToIe(uint8_t *loc, const Relocation &rel,
   // Relaxing to initial-exec entails:
   // 1) Replacing the call by a load from the GOT.
   // 2) Replacing the relocation on the constant LC0 by R_390_TLS_GOTIE64.
+  //
+  // While we no longer call __tls_get_offset after optimization, we still
+  // generate an unused PLT entry. This simple behavior matches GNU ld.
 
   switch (rel.type) {
   case R_390_TLS_GDCALL:

lld/ELF/Relocations.cpp

@@ -1654,10 +1654,8 @@ void RelocationScanner::scan(Relocs<RelTy> rels) {
   // For EhInputSection, OffsetGetter expects the relocations to be sorted by
   // r_offset. In rare cases (.eh_frame pieces are reordered by a linker
   // script), the relocations may be unordered.
-  // On SystemZ, all sections need to be sorted by r_offset, to allow TLS
-  // relaxation to be handled correctly - see SystemZ::getTlsGdRelaxSkip.
   SmallVector<RelTy, 0> storage;
-  if (isa<EhInputSection>(sec) || ctx.arg.emachine == EM_S390)
+  if (isa<EhInputSection>(sec))
     rels = sortRels(rels, storage);
 
   if constexpr (RelTy::IsCrel) {

lld/test/ELF/systemz-plt.s

@@ -3,7 +3,7 @@
 
 # RUN: llvm-mc -filetype=obj -triple=s390x-unknown-linux %t1.s -o %t1.o
 # RUN: ld.lld -shared %t1.o -soname=t1.so -o %t1.so
-# RUN: llvm-mc -filetype=obj -triple=s390x-unknown-linux %s -o %t.o
+# RUN: llvm-mc -filetype=obj -triple=s390x-unknown-linux --crel %s -o %t.o
 # RUN: ld.lld %t.o %t1.so -z separate-code -o %t
 # RUN: llvm-readelf -S -s -r -x .got.plt %t | FileCheck %s
 # RUN: llvm-objdump -d %t | FileCheck --check-prefixes=DIS %s

lld/test/ELF/systemz-tls-gd.s

@@ -1,6 +1,7 @@
 # REQUIRES: systemz
 # RUN: llvm-mc -filetype=obj -triple=s390x-unknown-linux %s -o %t.o
-# RUN: echo '.tbss; .globl b, c; b: .zero 4; c:' | llvm-mc -filetype=obj -triple=s390x-unknown-linux - -o %t1.o
+# RUN: echo '.globl __tls_get_offset; __tls_get_offset:; .tbss; .globl b, c; b: .zero 4; c:' | \
+# RUN:   llvm-mc -filetype=obj -triple=s390x-unknown-linux - -o %t1.o
 # RUN: ld.lld -shared -soname=t1.so %t1.o -o %t1.so
 
 # RUN: ld.lld -shared %t.o %t1.o -o %t.so
@@ -19,12 +20,12 @@
 # RUN: llvm-objdump --section .data.rel.ro --full-contents %t.ie | FileCheck --check-prefix=IE-DATA %s
 
 # GD-REL: Relocation section '.rela.dyn' at offset {{.*}} contains 6 entries:
-# GD-REL:      0000000000002570 0000000200000036 R_390_TLS_DTPMOD 0000000000000008 a + 0
-# GD-REL-NEXT: 0000000000002578 0000000200000037 R_390_TLS_DTPOFF 0000000000000008 a + 0
-# GD-REL-NEXT: 0000000000002580 0000000300000036 R_390_TLS_DTPMOD 000000000000000c b + 0
-# GD-REL-NEXT: 0000000000002588 0000000300000037 R_390_TLS_DTPOFF 000000000000000c b + 0
-# GD-REL-NEXT: 0000000000002590 0000000400000036 R_390_TLS_DTPMOD 0000000000000010 c + 0
-# GD-REL-NEXT: 0000000000002598 0000000400000037 R_390_TLS_DTPOFF 0000000000000010 c + 0
+# GD-REL:      0000000000002570 {{.*}}           R_390_TLS_DTPMOD 0000000000000008 a + 0
+# GD-REL-NEXT: 0000000000002578 {{.*}}           R_390_TLS_DTPOFF 0000000000000008 a + 0
+# GD-REL-NEXT: 0000000000002580 {{.*}}           R_390_TLS_DTPMOD 000000000000000c b + 0
+# GD-REL-NEXT: 0000000000002588 {{.*}}           R_390_TLS_DTPOFF 000000000000000c b + 0
+# GD-REL-NEXT: 0000000000002590 {{.*}}           R_390_TLS_DTPMOD 0000000000000010 c + 0
+# GD-REL-NEXT: 0000000000002598 {{.*}}           R_390_TLS_DTPOFF 0000000000000010 c + 0
 
 ## _GLOBAL_OFFSET_TABLE is at 0x2558
 # GD:      larl    %r12, 0x2558
@@ -80,33 +81,36 @@
 
 
 # IE-REL: Relocation section '.rela.dyn' at offset {{.*}} contains 2 entries:
-# IE-REL:      0000000001002430 0000000200000038 R_390_TLS_TPOFF 0000000000000000 b + 0
-# IE-REL-NEXT: 0000000001002438 0000000300000038 R_390_TLS_TPOFF 0000000000000000 c + 0
+# IE-REL:      0000000001002500 {{.*}}           R_390_TLS_TPOFF 0000000000000000 b + 0
+# IE-REL-NEXT: 0000000001002508 {{.*}}           R_390_TLS_TPOFF 0000000000000000 c + 0
+## Benign false dependency on __tls_get_offset
+# IE-REL: Relocation section '.rela.plt' at offset {{.*}} contains 1
+# IE-REL:                                        R_390_JMP_SLOT  0000000000000000 __tls_get_offset
 
-## _GLOBAL_OFFSET_TABLE is at 0x1002418
-# IE:      larl    %r12, 0x1002418
+## _GLOBAL_OFFSET_TABLE
+# IE:      larl    %r12, 0x10024e8
 
-## TP offset of a is at 0x1002340
-# IE-NEXT: lgrl    %r2, 0x1002340
+## TP offset of a
+# IE-NEXT: lgrl    %r2, 0x10023d0
 # IE-NEXT: jgnop
 # IE-NEXT: lgf     %r2, 0(%r2,%r7)
 
-## GOT offset of the TP offset for b is at 0x1002348
-# IE-NEXT: lgrl    %r2, 0x1002348
+## GOT offset of the TP offset for b
+# IE-NEXT: lgrl    %r2, 0x10023d8
 # IE-NEXT: lg      %r2, 0(%r2,%r12)
 # IE-NEXT: lgf     %r2, 0(%r2,%r7)
 
-## GOT offset of the TP offset for c is at 0x1002350
-# IE-NEXT: lgrl    %r2, 0x1002350
+## GOT offset of the TP offset for c
+# IE-NEXT: lgrl    %r2, 0x10023e0
 # IE-NEXT: lg      %r2, 0(%r2,%r12)
 # IE-NEXT: lgf     %r2, 0(%r2,%r7)
 
 ## TP offsets (a) / GOT offset of TP offsets (b, c)
 # a: -4
-# b: 0x1002430 / 0x18
-# c: 0x1002438 / 0x20
-# IE-DATA:      1002340 ffffffff fffffffc 00000000 00000018
-# IE-DATA-NEXT: 1002350 00000000 00000020
+# b: 0x10023d0 / 0x18
+# c: 0x10023e0 / 0x20
+# IE-DATA:      10023d0 ffffffff fffffffc 00000000 00000018
+# IE-DATA-NEXT: 10023e0 00000000 00000020
 
 
 ear     %r7,%a0

lld/test/ELF/systemz-tls-ld.s

@@ -91,6 +91,9 @@ lgf     %r1,0(%r1,%r2)
 lgrl    %r1, .LC3
 lgf     %r1,0(%r1,%r2)
 
+.globl __tls_get_offset
+__tls_get_offset:
+
         .section        .data.rel.ro,"aw"
         .align  8
 .LC0: