[AArch64] Unconditionally use DW_EH_PE_indirect|DW_EH_PE_pcrel personality/lsda/ttype...
authorFangrui Song <i@maskray.me>
Sun, 5 Feb 2023 18:46:43 +0000 (10:46 -0800)
committerFangrui Song <i@maskray.me>
Sun, 5 Feb 2023 18:46:43 +0000 (10:46 -0800)
For -fno-pic, without DW_EH_PE_indirect, the personality routine pointer in a
CIE needs an R_AARCH64_ABS64 relocation. In common configurations that
`__gcc_personality_v0` is defined in a shared object, this will lead to a
discouraged canonical PLT entry, or, if `ld.lld -z notext` (betwen D122459 and
D143136), a dynamic R_AARCH64_ABS64 relocation with an incorrect offset:
https://github.com/llvm/llvm-project/issues/60392

Since GCC uses DW_EH_PE_indirect for -fno-pic code (the behavior hasn't changed
since the initial port in 2012), let's follow suit by simplifying the code.
(
For tiny and small code models, we use DW_EH_PE_sdata8 instead of GCC's
DW_EH_PE_sdata4. This is a deliberate choice to support personality-.eh_frame
offset > 2GiB. This is unneeded for small code model since "Max text segment
size < 2GiB" but making `-fno-pic -mcmodel={tiny,small}` different seems
unnecessary: the scenarios that uses both -fno-pic and C++ exceptions have been
increasingly rare now, so there is little advantage optimizing for the little
size saving with code complexity.
)

---

Two clang/test/Interpreter tests would fail without 6747fc07d1aa94e22622e278e5a02ba70675ac9b
([ORC] Use JITLink as the default linker for LLJIT on Linux/arm64.)

Reviewed By: MatzeB

Differential Revision: https://reviews.llvm.org/D143039

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp
llvm/test/CodeGen/AArch64/arm64-big-endian-eh.ll
llvm/test/CodeGen/AArch64/pic-eh-stubs.ll
llvm/test/DebugInfo/AArch64/eh_frame_personality.ll
llvm/test/Transforms/CodeGenPrepare/AArch64/large-offset-gep.ll

index 39f7ee7..bdb328f 100644 (file)
@@ -182,26 +182,14 @@ void TargetLoweringObjectFileELF::Initialize(MCContext &Ctx,
     // The small model guarantees static code/data size < 4GB, but not where it
     // will be in memory. Most of these could end up >2GB away so even a signed
     // pc-relative 32-bit address is insufficient, theoretically.
-    if (isPositionIndependent()) {
-      // ILP32 uses sdata4 instead of sdata8
-      if (TgtM.getTargetTriple().getEnvironment() == Triple::GNUILP32) {
-        PersonalityEncoding = dwarf::DW_EH_PE_indirect | dwarf::DW_EH_PE_pcrel |
-                              dwarf::DW_EH_PE_sdata4;
-        LSDAEncoding = dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata4;
-        TTypeEncoding = dwarf::DW_EH_PE_indirect | dwarf::DW_EH_PE_pcrel |
-                        dwarf::DW_EH_PE_sdata4;
-      } else {
-        PersonalityEncoding = dwarf::DW_EH_PE_indirect | dwarf::DW_EH_PE_pcrel |
-                              dwarf::DW_EH_PE_sdata8;
-        LSDAEncoding = dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata8;
-        TTypeEncoding = dwarf::DW_EH_PE_indirect | dwarf::DW_EH_PE_pcrel |
-                        dwarf::DW_EH_PE_sdata8;
-      }
-    } else {
-      PersonalityEncoding = dwarf::DW_EH_PE_absptr;
-      LSDAEncoding = dwarf::DW_EH_PE_absptr;
-      TTypeEncoding = dwarf::DW_EH_PE_absptr;
-    }
+    //
+    // Use DW_EH_PE_indirect even for -fno-pic to avoid copy relocations.
+    LSDAEncoding = dwarf::DW_EH_PE_pcrel |
+                   (TgtM.getTargetTriple().getEnvironment() == Triple::GNUILP32
+                        ? dwarf::DW_EH_PE_sdata4
+                        : dwarf::DW_EH_PE_sdata8);
+    PersonalityEncoding = LSDAEncoding | dwarf::DW_EH_PE_indirect;
+    TTypeEncoding = LSDAEncoding | dwarf::DW_EH_PE_indirect;
     break;
   case Triple::lanai:
     LSDAEncoding = dwarf::DW_EH_PE_absptr;
index c6f955f..6c54cec 100644 (file)
@@ -71,5 +71,6 @@ declare void @_ZSt9terminatev()
 ; CHECK-LABEL: Contents of section .eh_frame:
 ; CHECK-NEXT: {{^ 0000}}
 ; CHECK-NEXT: {{^ 0010}}
-; CHECK-NEXT: 0000 0000001c
+; CHECK-NEXT: 0020 0000000c 00440e10 9e040000 0000001c .....D..........
+; CHECK-NEXT: 0030 00000000 017a504c 5200017c 1e0b9c00 .....zPLR..|....
 
index 3d072bc..bdb557e 100644 (file)
@@ -1,3 +1,4 @@
+; RUN: llc -mtriple=aarch64-none-linux-gnu -relocation-model=static -simplifycfg-require-and-preserve-domtree=1 -o - %s | FileCheck %s
 ; RUN: llc -mtriple=aarch64-none-linux-gnu -relocation-model=pic -simplifycfg-require-and-preserve-domtree=1 -o - %s | FileCheck %s
 ; RUN: llc -mtriple=aarch64_be-none-linux-gnu -relocation-model=pic -simplifycfg-require-and-preserve-domtree=1 -o - %s | FileCheck %s
 
index 912ebf1..2ee65a8 100644 (file)
@@ -17,7 +17,7 @@ clean:
 
 ; CHECK: Contents of section .eh_frame:
 ; CHECK: 0000 1c000000 00000000 017a504c 5200017c  .........zPLR..|
-; CHECK: 0010 1e0b0000 00000000 00000000 1b0c1f00  ................
+; CHECK: 0010 1e0b9c00 00000000 0000001c 1b0c1f00 ................
 
 ; Don't really care about the rest:
 
index 9d8ef90..301bf66 100644 (file)
@@ -138,8 +138,8 @@ define void @test4(i32 %n) uwtable personality ptr @__FrameHandler {
 ; CHECK-LABEL: test4:
 ; CHECK:       .Lfunc_begin0:
 ; CHECK-NEXT:    .cfi_startproc
-; CHECK-NEXT:    .cfi_personality 0, __FrameHandler
-; CHECK-NEXT:    .cfi_lsda 0, .Lexception0
+; CHECK-NEXT:    .cfi_personality 156, DW.ref.__FrameHandler
+; CHECK-NEXT:    .cfi_lsda 28, .Lexception0
 ; CHECK-NEXT:  // %bb.0: // %entry
 ; CHECK-NEXT:    stp x30, x21, [sp, #-32]! // 16-byte Folded Spill
 ; CHECK-NEXT:    .cfi_def_cfa_offset 32