[AArch64] Implement TLS relaxations

lib/Target/AArch64/AArch64Relocator.cpp

@@ -4,7 +4,6 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //===----------------------------------------------------------------------===//
 
-
 #include "AArch64Relocator.h"
 #include "AArch64InsnHelpers.h"
 #include "AArch64PLT.h"
@@ -52,8 +51,7 @@ static uint64_t getSigningSchema(const Relocation &pReloc) {
 /// helper_DynRel - Get an relocation entry in .rela.dyn
 Relocation *helper_DynRel_init(ELFObjectFile *Obj, Relocation *R,
                                ResolveInfo *pSym, Fragment *F, uint32_t pOffset,
-                               Relocator::Type pType,
-                               AArch64LDBackend &B) {
+                               Relocator::Type pType, AArch64LDBackend &B) {
   Relocation *rela_entry = nullptr;
 
   if (pType == R_AARCH64_TLSDESC)
@@ -190,9 +188,8 @@ bool AArch64Relocator::relocNeedsDynRel(Relocation &pReloc) const {
                     pReloc.type() == llvm::ELF::R_AARCH64_ABS32 ||
                     pReloc.type() == llvm::ELF::R_AARCH64_ABS16 ||
                     pReloc.type() == llvm::ELF::R_AARCH64_AUTH_ABS64;
-  return getTarget().symbolNeedsDynRel(
-                   *rsym, (rsym->reserved() & ReservePLT),
-                   isAbsReloc);
+  return getTarget().symbolNeedsDynRel(*rsym, (rsym->reserved() & ReservePLT),
+                                       isAbsReloc);
 }
 
 Relocator::Result AArch64Relocator::applyRelocation(Relocation &pRelocation) {
@@ -252,12 +249,10 @@ void AArch64Relocator::scanLocalReloc(InputFile &pInput, Relocation &pReloc,
       rsym->setReserved(rsym->reserved() | ReserveRel);
       getTarget().checkAndSetHasTextRel(pSection);
       // set up the dyn rel directly
-      Relocation::Type relType =
-        isAuthAbs ? llvm::ELF::R_AARCH64_AUTH_RELATIVE
-                  : llvm::ELF::R_AARCH64_RELATIVE;
+      Relocation::Type relType = isAuthAbs ? llvm::ELF::R_AARCH64_AUTH_RELATIVE
+                                           : llvm::ELF::R_AARCH64_RELATIVE;
       helper_DynRel_init(Obj, &pReloc, rsym, pReloc.targetRef()->frag(),
-                         pReloc.targetRef()->offset(), relType,
-                         m_Target);
+                         pReloc.targetRef()->offset(), relType, m_Target);
     }
   }
     return;
@@ -330,6 +325,14 @@ void AArch64Relocator::scanLocalReloc(InputFile &pInput, Relocation &pReloc,
       G->setValueType(GOT::TLSStaticSymbolValue);
       return;
     }
+    if (config().options().isPIE()) {
+      // PIE executable: relax TLSDESC to IE
+      AArch64GOT *G = m_Target.createGOT(GOT::TLS_IE, Obj, rsym);
+      helper_DynRel_init(Obj, &pReloc, rsym, G, 0x0,
+                         llvm::ELF::R_AARCH64_TLS_TPREL64, m_Target);
+      rsym->setReserved(rsym->reserved() | ReserveGOT);
+      return;
+    }
     AArch64GOT *G = m_Target.createGOT(GOT::TLS_DESC, Obj, rsym);
     helper_DynRel_init(Obj, &pReloc, rsym, G->getFirst(), 0x0,
                        llvm::ELF::R_AARCH64_TLSDESC, m_Target);
@@ -382,8 +385,8 @@ void AArch64Relocator::scanGlobalReloc(InputFile &pInput, Relocation &pReloc,
         // for signed pointers"
         if (isAuthAbs) {
           config().raise(Diag::non_pic_relocation)
-            << getName(pReloc.type()) << pReloc.symInfo()->name()
-            << pReloc.getSourcePath(config().options());
+              << getName(pReloc.type()) << pReloc.symInfo()->name()
+              << pReloc.getSourcePath(config().options());
           m_Target.getModule().setFailure(true);
           return;
         }
@@ -409,10 +412,8 @@ void AArch64Relocator::scanGlobalReloc(InputFile &pInput, Relocation &pReloc,
           relType = isAuthAbs ? llvm::ELF::R_AARCH64_AUTH_RELATIVE
                               : llvm::ELF::R_AARCH64_RELATIVE;
         }
-        helper_DynRel_init(
-            Obj, &pReloc, rsym, pReloc.targetRef()->frag(),
-            pReloc.targetRef()->offset(),
-            relType, m_Target);
+        helper_DynRel_init(Obj, &pReloc, rsym, pReloc.targetRef()->frag(),
+                           pReloc.targetRef()->offset(), relType, m_Target);
       }
     }
   }
@@ -562,6 +563,14 @@ void AArch64Relocator::scanGlobalReloc(InputFile &pInput, Relocation &pReloc,
       G->setValueType(GOT::TLSStaticSymbolValue);
       return;
     }
+    if (config().options().isPIE()) {
+      // PIE executable: relax TLSDESC to IE
+      AArch64GOT *G = m_Target.createGOT(GOT::TLS_IE, Obj, rsym);
+      helper_DynRel_init(Obj, &pReloc, rsym, G, 0x0,
+                         llvm::ELF::R_AARCH64_TLS_TPREL64, m_Target);
+      rsym->setReserved(rsym->reserved() | ReserveGOT);
+      return;
+    }
     AArch64GOT *G = m_Target.createGOT(GOT::TLS_DESC, Obj, rsym);
     helper_DynRel_init(Obj, &pReloc, rsym, G->getFirst(), 0x0,
                        llvm::ELF::R_AARCH64_TLSDESC, m_Target);
@@ -1023,7 +1032,8 @@ Relocator::Result ld64_got_lo12(Relocation &pReloc, AArch64Relocator &pParent) {
 }
 
 // R_AARCH64_LD64_GOTPAGE_LO15: G(GDAT(S)) - Page(GOT)
-Relocator::Result ld64_gotpage_lo15(Relocation &pReloc, AArch64Relocator &pParent) {
+Relocator::Result ld64_gotpage_lo15(Relocation &pReloc,
+                                    AArch64Relocator &pParent) {
   if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT)) {
     return Relocator::BadReloc;
   }
@@ -1148,7 +1158,8 @@ Relocator::Result tls_gottprel_page(Relocation &pReloc,
   DiagnosticEngine *DiagEngine = pParent.config().getDiagEngine();
   Relocator::DWord A = pReloc.addend();
 
-  if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT)) {
+  if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT) ||
+      pParent.config().isCodeStatic()) {
     Relocator::DWord X =
         pParent.getSymValue(&pReloc) + AArch64LDBackend::getStaticTCBSize();
     // Convert to movz
@@ -1173,7 +1184,8 @@ Relocator::Result tls_gottprel_lo(Relocation &pReloc,
                                   AArch64Relocator &pParent) {
   Relocator::DWord A = pReloc.addend();
 
-  if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT)) {
+  if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT) ||
+      pParent.config().isCodeStatic()) {
     Relocator::DWord X =
         pParent.getSymValue(&pReloc) + AArch64LDBackend::getStaticTCBSize();
     // Convert to movk
@@ -1216,7 +1228,8 @@ Relocator::Result tls_tlsdesc_page(Relocation &pReloc,
                                    AArch64Relocator &pParent) {
   Relocator::DWord A = pReloc.addend();
 
-  if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT)) {
+  if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT) ||
+      pParent.config().isCodeStatic()) {
     Relocator::DWord X =
         pParent.getSymValue(&pReloc) + AArch64LDBackend::getStaticTCBSize();
     // Convert to movz
@@ -1241,34 +1254,36 @@ Relocator::Result tls_tlsdesc_page(Relocation &pReloc,
 Relocator::Result tls_tlsdesc_lo(Relocation &pReloc,
                                  AArch64Relocator &pParent) {
   Relocator::DWord A = pReloc.addend();
-
-  if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT)) {
+  if (!(pReloc.symInfo()->reserved() & Relocator::ReserveGOT) ||
+      pParent.config().isCodeStatic()) {
     Relocator::DWord X =
         pParent.getSymValue(&pReloc) + AArch64LDBackend::getStaticTCBSize();
-    // Convert to movk, save to x0
     uint32_t movk = 0xF2800000;
     pReloc.target() = helper_reencode_movzk_imm(movk, X);
     return Relocator::OK;
   }
-
+  if (pParent.config().options().isPIE()) {
+    uint32_t insn = 0xF9400000; // ldr x0, [x0, #0]
+    Relocator::Address GOT_S = pParent.getTarget()
+                                   .findEntryInGOT(pReloc.symInfo())
+                                   ->getAddr(pParent.config().getDiagEngine());
+    Relocator::DWord GX = helper_get_page_offset(GOT_S + A);
+    pReloc.target() = helper_reencode_ldst_pos_imm(insn, GX >> 3);
+    return Relocator::OK;
+  }
   Relocator::Address GOT_S = pParent.getTarget()
                                  .findEntryInGOT(pReloc.symInfo())
                                  ->getAddr(pParent.config().getDiagEngine());
   Relocator::DWord GX = helper_get_page_offset(GOT_S + A);
   pReloc.target() = helper_reencode_ldst_pos_imm(pReloc.target(), GX >> 3);
-
-  // Convert Rt to X0 if static
-  if (pParent.config().isCodeStatic())
-    pReloc.target() = pReloc.target() & ~0x1F;
-
   return Relocator::OK;
 }
 
 // R_AARCH64_TLSDESC_ADD_LO12_NC : G(GTLSDESC(S+A))
 Relocator::Result tls_tlsdesc_add(Relocation &pReloc,
                                   AArch64Relocator &pParent) {
   Relocator::DWord A = pReloc.addend();
-  if (pParent.config().isCodeStatic()) {
+  if (pParent.config().isCodeStatic() || pParent.config().options().isPIE()) {
     // Convert to nop
     pReloc.target() = 0xD503201F;
     return Relocator::OK;
@@ -1285,7 +1300,7 @@ Relocator::Result tls_tlsdesc_add(Relocation &pReloc,
 
 // R_AARCH64_TLSDESC_CALL
 Relocator::Result tls_call(Relocation &pReloc, AArch64Relocator &pParent) {
-  if (pParent.config().isCodeStatic()) {
+  if (pParent.config().isCodeStatic() || pParent.config().options().isPIE()) {
     // Convert to nop
     pReloc.target() = 0xD503201F;
     return Relocator::OK;

test/AArch64/standalone/TLS_DESC/DESC.test

@@ -1,10 +1,14 @@
 RUN: %clangas %clangasopts -filetype obj -target-cpu generic -target-feature +neon -mrelax-all %p/Inputs/t.s -o %t.o
-RUN: %link %linkopts  -march aarch64 -static %t.o -z max-page-size=0x1000 -o %t.out
+RUN: %link %linkopts -march aarch64 -static -z max-page-size=0x1000 %t.o -o %t.out
 RUN: llvm-objdump -d %t.out | %filecheck %s
 
-#CHECK: {{.*}} <_test_tls_desc>:
-#CHECK: {{.*}} adrp    x0, 0x1000
-#CHECK: {{.*}} ldr     x0, [x0]
-#CHECK: {{.*}} <_test_tls_desc_local>:
-#CHECK: {{.*}} adrp    x0, 0x1000
-#CHECK: {{.*}} ldr     x0, [x0, #0x8]
+CHECK: _test_tls_desc
+CHECK: d2a00000       movz
+CHECK: f2800200       movk
+CHECK: d503201f       nop
+CHECK: d503201f       nop
+CHECK: _test_tls_desc_local
+CHECK: d2a00000       movz
+CHECK: f2800280       movk
+CHECK: d503201f       nop
+CHECK: d503201f       nop

test/AArch64/standalone/TLS_DESC/TLS_DESC.test

@@ -0,0 +1,14 @@
+RUN: %clangas %clangasopts -filetype obj -target-cpu generic -target-feature +neon -mrelax-all %p/Inputs/t.s -o %t.o
+RUN: %link %linkopts -march aarch64 -static -z max-page-size=0x1000 %t.o -o %t.out
+RUN: llvm-objdump -d %t.out | %filecheck %s
+
+CHECK: _test_tls_desc
+CHECK: d2a00000       movz
+CHECK: f2800200       movk
+CHECK: d503201f       nop
+CHECK: d503201f       nop
+CHECK: _test_tls_desc_local
+CHECK: d2a00000       movz
+CHECK: f2800280       movk
+CHECK: d503201f       nop
+CHECK: d503201f       nop

test/AArch64/standalone/TLS_DESC/TLS_DESC_PIE.test

@@ -0,0 +1,14 @@
+RUN: %clangas %clangasopts -filetype obj -target-cpu generic -target-feature +neon -mrelax-all %p/Inputs/t.s -o %t.o
+RUN: %link %linkopts -march aarch64 -pie -z max-page-size=0x1000 %t.o -o %t.out
+RUN: llvm-objdump -d %t.out | %filecheck %s
+
+CHECK: _test_tls_desc
+CHECK: b0000000       adrp
+CHECK: f9408000       ldr
+CHECK: d503201f       nop
+CHECK: d503201f       nop
+CHECK: _test_tls_desc_local
+CHECK: b0000000       adrp
+CHECK: f9408400       ldr
+CHECK: d503201f       nop
+CHECK: d503201f       nop

test/AArch64/standalone/TLS_IE/IE.test

@@ -8,8 +8,8 @@ RUN: %clang %clangopts -target aarch64 %p/Inputs/f.c -c -o %t3.o
 RUN: %link %linkopts -static -march aarch64 %t2.o %t3.o -z max-page-size=0x1000 -o %t2.out
 
 CHECK: _test_tls_IE
-CHECK: b0000000       adrp
-CHECK: f9400000       ldr
+CHECK: d2a00000       movz
+CHECK: f2800200       movk
 CHECK: _test_tls_IE_local
 CHECK: d2a00000       movz
 CHECK: f2800280       movk

test/AArch64/standalone/TLS_IE_PIE/IE_PIE.test

@@ -0,0 +1,12 @@
+# Verify that IE TLS relocations remain as GOT-indirect adrp+ldr
+# in PIE executables (IE stays as IE, no relaxation to LE).
+RUN: %clangas %clangasopts -filetype obj -target-cpu generic -target-feature +neon -mrelax-all %p/Inputs/ie.s -o %t.o
+RUN: %link %linkopts -march aarch64 -pie -z max-page-size=0x1000 %t.o -o %t.out
+RUN: llvm-objdump -d %t.out | %filecheck %s
+
+CHECK: _test_tls_IE
+CHECK: b0000000       adrp
+CHECK: f9408000       ldr
+CHECK: _test_tls_IE_local
+CHECK: b0000000       adrp
+CHECK: f9408400       ldr

test/AArch64/standalone/TLS_IE_PIE/Inputs/ie.s

@@ -0,0 +1,29 @@
+        .global tlsievar
+        .section        .tbss,"awT",%nobits
+        .align  2
+        .type   tlsievar, %object
+        .size   tlsievar, 4
+tlsievar:
+        .zero   4
+
+       .align  2
+        .type   l_tlsievar, %object
+        .size   l_tlsievar, 4
+l_tlsievar:
+        .zero   4
+
+.text
+_test_tls_IE:
+
+        // R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21 tlsievar
+        adrp x0, :gottprel:tlsievar
+        // R_AARCH64_TLSIE_GOTTPREL_LO12_NC    tlsievar
+        ldr  x0, [x0, :gottprel_lo12:tlsievar]
+
+_test_tls_IE_local:
+
+        // R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21 l_tlsievar
+        adrp x0, :gottprel:l_tlsievar
+        // R_AARCH64_TLSIE_GOTTPREL_LO12_NC    l_tlsievar
+        ldr  x0, [x0, :gottprel_lo12:l_tlsievar]
+

test/AArch64/standalone/TLS_Relaxation_Verify/Inputs/tls_test.c

@@ -0,0 +1,10 @@
+/* Verify TLS relaxations for AArch64.
+ * Global-dynamic (TLSDESC) should be relaxed to:
+ * - Local Exec (movz+movk) for static executables
+ * - Initial Exec (adrp+ldr) for PIE executables
+ */
+__thread int tls_global = 42;
+static __thread int tls_local = 100;
+
+int get_global() { return tls_global; }
+int get_local()  { return tls_local; }

test/AArch64/standalone/TLS_Relaxation_Verify/TLS_Relaxation_Verify.test

@@ -0,0 +1,34 @@
+RUN: %clang %clangopts -target aarch64-unknown-none-elf \
+RUN:   -ftls-model=global-dynamic -fPIC \
+RUN:   -c %p/Inputs/tls_test.c -o %t.o
+RUN: %link %linkopts -march aarch64 -static \
+RUN:   -z max-page-size=0x1000 %t.o -o %t.static.out
+RUN: llvm-objdump -d %t.static.out | %filecheck %s --check-prefix=STATIC
+RUN: %link %linkopts -march aarch64 -pie \
+RUN:   -z max-page-size=0x1000 %t.o -o %t.pie.out
+RUN: llvm-objdump -d %t.pie.out | %filecheck %s --check-prefix=PIE
+RUN: %readelf -r %t.pie.out | %filecheck %s --check-prefix=PIE-RELOC
+
+STATIC: <get_global>:
+STATIC: d2a00000       movz
+STATIC: f2800200       movk
+STATIC: d503201f       nop
+STATIC: d503201f       nop
+STATIC: <get_local>:
+STATIC: d2a00000       movz
+STATIC: f2800280       movk
+STATIC: d503201f       nop
+STATIC: d503201f       nop
+
+PIE: <get_global>:
+PIE: b0000000          adrp
+PIE: f9408400          ldr
+PIE: d503201f          nop
+PIE: d503201f          nop
+PIE: <get_local>:
+PIE: b0000000          adrp
+PIE: f9408800          ldr
+PIE: d503201f          nop
+PIE: d503201f          nop
+
+PIE-RELOC: R_AARCH64_TLS_TPREL64