ARM64: use GOT for weak symbols & PIC.

Weak symbols cannot use the small code model's usual ADRP sequences since the
instruction simply may not be able to encode a value of 0.

This redirects them to use the GOT, which hopefully linkers are able to cope
with even in the static relocation model.

llvm-svn: 205426

lib/Target/ARM64/ARM64Subtarget.cpp

@@ -50,16 +50,33 @@ ARM64Subtarget::ClassifyGlobalReference(const GlobalValue *GV,
   if (GV->isDeclaration() && !GV->isMaterializable())
     isDecl = true;
 
-  // If symbol visibility is hidden, the extra load is not needed if
-  // the symbol is definitely defined in the current translation unit.
-  if (TM.getRelocationModel() != Reloc::Static && GV->hasDefaultVisibility() &&
-      (isDecl || GV->isWeakForLinker()))
-    return ARM64II::MO_GOT;
-
+  // MachO large model always goes via a GOT, simply to get a single 8-byte
+  // absolute relocation on all global addresses.
   if (TM.getCodeModel() == CodeModel::Large && isTargetMachO())
     return ARM64II::MO_GOT;
 
-  // FIXME: this will fail on static ELF for weak symbols.
+  // The small code mode's direct accesses use ADRP, which cannot necessarily
+  // produce the value 0 (if the code is above 4GB). Therefore they must use the
+  // GOT.
+  if (TM.getCodeModel() == CodeModel::Small && GV->isWeakForLinker() && isDecl)
+    return ARM64II::MO_GOT;
+
+  // If symbol visibility is hidden, the extra load is not needed if
+  // the symbol is definitely defined in the current translation unit.
+
+  // The handling of non-hidden symbols in PIC mode is rather target-dependent:
+  //   + On MachO, if the symbol is defined in this module the GOT can be
+  //     skipped.
+  //   + On ELF, the R_AARCH64_COPY relocation means that even symbols actually
+  //     defined could end up in unexpected places. Use a GOT.
+  if (TM.getRelocationModel() != Reloc::Static && GV->hasDefaultVisibility()) {
+    if (isTargetMachO())
+      return (isDecl || GV->isWeakForLinker()) ? ARM64II::MO_GOT
+                                               : ARM64II::MO_NO_FLAG;
+    else
+      return ARM64II::MO_GOT;
+  }
+
   return ARM64II::MO_NO_FLAG;
 }
 

test/CodeGen/ARM64/basic-pic.ll

@@ -0,0 +1,54 @@
+; RUN: llc -mtriple=arm64-none-linux-gnu -verify-machineinstrs -relocation-model=pic %s -o - | FileCheck %s
+
+@var = global i32 0
+
+define i32 @get_globalvar() {
+; CHECK-LABEL: get_globalvar:
+
+  %val = load i32* @var
+; CHECK: adrp x[[GOTHI:[0-9]+]], :got:var
+; CHECK: ldr x[[GOTLOC:[0-9]+]], [x[[GOTHI]], :got_lo12:var]
+; CHECK: ldr w0, [x[[GOTLOC]]]
+
+  ret i32 %val
+}
+
+define i32* @get_globalvaraddr() {
+; CHECK-LABEL: get_globalvaraddr:
+
+  %val = load i32* @var
+; CHECK: adrp x[[GOTHI:[0-9]+]], :got:var
+; CHECK: ldr x0, [x[[GOTHI]], :got_lo12:var]
+
+  ret i32* @var
+}
+
+@hiddenvar = hidden global i32 0
+
+define i32 @get_hiddenvar() {
+; CHECK-LABEL: get_hiddenvar:
+
+  %val = load i32* @hiddenvar
+; CHECK: adrp x[[HI:[0-9]+]], hiddenvar
+; CHECK: ldr w0, [x[[HI]], :lo12:hiddenvar]
+
+  ret i32 %val
+}
+
+define i32* @get_hiddenvaraddr() {
+; CHECK-LABEL: get_hiddenvaraddr:
+
+  %val = load i32* @hiddenvar
+; CHECK: adrp [[HI:x[0-9]+]], hiddenvar
+; CHECK: add x0, [[HI]], :lo12:hiddenvar
+
+  ret i32* @hiddenvar
+}
+
+define void()* @get_func() {
+; CHECK-LABEL: get_func:
+
+  ret void()* bitcast(void()*()* @get_func to void()*)
+; CHECK: adrp x[[GOTHI:[0-9]+]], :got:get_func
+; CHECK: ldr x0, [x[[GOTHI]], :got_lo12:get_func]
+}

test/CodeGen/ARM64/extern-weak.ll

@@ -0,0 +1,51 @@
+; RUN: llc -mtriple=arm64-none-linux-gnu -o - < %s | FileCheck %s
+; RUN: llc -mtriple=arm64-none-linux-gnu -code-model=large -o - < %s | FileCheck --check-prefix=CHECK-LARGE %s
+
+declare extern_weak i32 @var()
+
+define i32()* @foo() {
+; The usual ADRP/ADD pair can't be used for a weak reference because it must
+; evaluate to 0 if the symbol is undefined. We use a litpool entry.
+  ret i32()* @var
+
+; CHECK: adrp x[[VAR:[0-9]+]], :got:var
+; CHECK: ldr x0, [x[[VAR]], :got_lo12:var]
+
+  ; In the large model, the usual relocations are absolute and can
+  ; materialise 0.
+; CHECK-LARGE: movz x0, #:abs_g3:var
+; CHECK-LARGE: movk x0, #:abs_g2_nc:var
+; CHECK-LARGE: movk x0, #:abs_g1_nc:var
+; CHECK-LARGE: movk x0, #:abs_g0_nc:var
+}
+
+
+@arr_var = extern_weak global [10 x i32]
+
+define i32* @bar() {
+  %addr = getelementptr [10 x i32]* @arr_var, i32 0, i32 5
+; CHECK: adrp x[[ARR_VAR_HI:[0-9]+]], :got:arr_var
+; CHECK: ldr [[ARR_VAR:x[0-9]+]], [x[[ARR_VAR_HI]], :got_lo12:arr_var]
+; CHECK: add x0, [[ARR_VAR]], #20
+  ret i32* %addr
+
+  ; In the large model, the usual relocations are absolute and can
+  ; materialise 0.
+; CHECK-LARGE: movz [[ARR_VAR:x[0-9]+]], #:abs_g3:arr_var
+; CHECK-LARGE: movk [[ARR_VAR]], #:abs_g2_nc:arr_var
+; CHECK-LARGE: movk [[ARR_VAR]], #:abs_g1_nc:arr_var
+; CHECK-LARGE: movk [[ARR_VAR]], #:abs_g0_nc:arr_var
+}
+
+@defined_weak_var = internal unnamed_addr global i32 0
+
+define i32* @wibble() {
+  ret i32* @defined_weak_var
+; CHECK: adrp [[BASE:x[0-9]+]], defined_weak_var
+; CHECK: add x0, [[BASE]], :lo12:defined_weak_var
+
+; CHECK-LARGE: movz x0, #:abs_g3:defined_weak_var
+; CHECK-LARGE: movk x0, #:abs_g2_nc:defined_weak_var
+; CHECK-LARGE: movk x0, #:abs_g1_nc:defined_weak_var
+; CHECK-LARGE: movk x0, #:abs_g0_nc:defined_weak_var
+}