AArch64: correct CodeGen of MOVZ/MOVK combinations.

According to the AArch64 ELF specification (4.6.8), it's the
assembler's responsibility to make sure the shift amount is correct in
relocated MOVZ/MOVK instructions.

This wasn't being obeyed by either the MCJIT CodeGen or RuntimeDyldELF
(which happened to work out well for JIT tests). This commit should
make us compliant in this area.

llvm-svn: 185360

3 files changed, 22 insertions, 17 deletions

diff --git a/llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp b/llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp
index 722ed10f733..cb3b5129f23 100644
--- a/llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp
+++ b/llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp
@@ -334,8 +334,8 @@ void RuntimeDyldELF::resolveAArch64Relocation(const SectionEntry &Section,
     *TargetPtr &= 0xff80001fU;
     // Immediate goes in bits 20:5 of MOVZ/MOVK instruction
     *TargetPtr |= Result >> (48 - 5);
-    // Shift is "lsl #48", in bits 22:21
-    *TargetPtr |= 3 << 21;
+    // Shift must be "lsl #48", in bits 22:21
+    assert((*TargetPtr >> 21 & 0x3) == 3 && "invalid shift for relocation");
     break;
   }
   case ELF::R_AARCH64_MOVW_UABS_G2_NC: {
@@ -347,8 +347,8 @@ void RuntimeDyldELF::resolveAArch64Relocation(const SectionEntry &Section,
     *TargetPtr &= 0xff80001fU;
     // Immediate goes in bits 20:5 of MOVZ/MOVK instruction
     *TargetPtr |= ((Result & 0xffff00000000ULL) >> (32 - 5));
-    // Shift is "lsl #32", in bits 22:21
-    *TargetPtr |= 2 << 21;
+    // Shift must be "lsl #32", in bits 22:21
+    assert((*TargetPtr >> 21 & 0x3) == 2 && "invalid shift for relocation");
     break;
   }
   case ELF::R_AARCH64_MOVW_UABS_G1_NC: {
@@ -359,8 +359,8 @@ void RuntimeDyldELF::resolveAArch64Relocation(const SectionEntry &Section,
     *TargetPtr &= 0xff80001fU;
     // Immediate goes in bits 20:5 of MOVZ/MOVK instruction
     *TargetPtr |= ((Result & 0xffff0000U) >> (16 - 5));
-    // Shift is "lsl #16", in bits 22:21
-    *TargetPtr |= 1 << 21;
+    // Shift must be "lsl #16", in bits 22:2
+    assert((*TargetPtr >> 21 & 0x3) == 1 && "invalid shift for relocation");
     break;
   }
   case ELF::R_AARCH64_MOVW_UABS_G0_NC: {
@@ -371,7 +371,8 @@ void RuntimeDyldELF::resolveAArch64Relocation(const SectionEntry &Section,
     *TargetPtr &= 0xff80001fU;
     // Immediate goes in bits 20:5 of MOVZ/MOVK instruction
     *TargetPtr |= ((Result & 0xffffU) << 5);
-    // Shift is "lsl #0", in bits 22:21. No action needed.
+    // Shift must be "lsl #0", in bits 22:21.
+    assert((*TargetPtr >> 21 & 0x3) == 0 && "invalid shift for relocation");
     break;
   }
   }
diff --git a/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp b/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
index 2e37cb41f9b..f258a966948 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
@@ -70,10 +70,11 @@ public:
 
   /// Used for pre-lowered address-reference nodes, so we already know
   /// the fields match. This operand's job is simply to add an
-  /// appropriate shift operand (i.e. 0) to the MOVZ/MOVK instruction.
+  /// appropriate shift operand to the MOVZ/MOVK instruction.
+  template<unsigned LogShift>
   bool SelectMOVWAddressRef(SDValue N, SDValue &Imm, SDValue &Shift) {
     Imm = N;
-    Shift = CurDAG->getTargetConstant(0, MVT::i32);
+    Shift = CurDAG->getTargetConstant(LogShift, MVT::i32);
     return true;
   }
 
diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.td b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
index d2cfc7db223..725a12164be 100644
--- a/llvm/lib/Target/AArch64/AArch64InstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
@@ -3974,14 +3974,17 @@ def : movalias<MOVZxii, GPR64, movz64_movimm>;
 def : movalias<MOVNwii, GPR32, movn32_movimm>;
 def : movalias<MOVNxii, GPR64, movn64_movimm>;
 
-def movw_addressref : ComplexPattern<i64, 2, "SelectMOVWAddressRef">;
-
-def : Pat<(A64WrapperLarge movw_addressref:$G3, movw_addressref:$G2,
-                           movw_addressref:$G1, movw_addressref:$G0),
-          (MOVKxii (MOVKxii (MOVKxii (MOVZxii movw_addressref:$G3),
-                                     movw_addressref:$G2),
-                            movw_addressref:$G1),
-                   movw_addressref:$G0)>;
+def movw_addressref_g0 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<0>">;
+def movw_addressref_g1 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<1>">;
+def movw_addressref_g2 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<2>">;
+def movw_addressref_g3 : ComplexPattern<i64, 2, "SelectMOVWAddressRef<3>">;
+
+def : Pat<(A64WrapperLarge movw_addressref_g3:$G3, movw_addressref_g2:$G2,
+                           movw_addressref_g1:$G1, movw_addressref_g0:$G0),
+          (MOVKxii (MOVKxii (MOVKxii (MOVZxii movw_addressref_g3:$G3),
+                                     movw_addressref_g2:$G2),
+                            movw_addressref_g1:$G1),
+                   movw_addressref_g0:$G0)>;
 
 //===----------------------------------------------------------------------===//
 // PC-relative addressing instructions