IR, X86: Understand !absolute_symbol metadata on global variables.

Summary:
Attaching !absolute_symbol to a global variable does two things:
1) Marks it as an absolute symbol reference.
2) Specifies the value range of that symbol's address.
Teach the X86 backend to allow absolute symbols to appear in place of
immediates by extending the relocImm and mov64imm32 matchers. Start using
relocImm in more places where it is legal.

As previously proposed on llvm-dev:
http://lists.llvm.org/pipermail/llvm-dev/2016-October/105800.html

Differential Revision: https://reviews.llvm.org/D25878

llvm-svn: 289087

llvm/docs/LangRef.rst

@@ -4589,6 +4589,25 @@ Examples:
     !2 = !{ i8 0, i8 2, i8 3, i8 6 }
     !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
 
+'``absolute_symbol``' Metadata
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+``absolute_symbol`` metadata may be attached to a global variable
+declaration. It marks the declaration as a reference to an absolute symbol,
+which causes the backend to use absolute relocations for the symbol even
+in position independent code, and expresses the possible ranges that the
+global variable's *address* (not its value) is in, in the same format as
+``range`` metadata.
+
+Example:
+
+.. code-block:: llvm
+
+      @a = external global i8, !absolute_symbol !0 ; Absolute symbol in range [0,256)
+
+    ...
+    !0 = !{ i64 0, i64 256 }
+
 '``unpredictable``' Metadata
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 

llvm/include/llvm/IR/GlobalValue.h

@@ -33,6 +33,7 @@
 namespace llvm {
 
 class Comdat;
+class ConstantRange;
 class Error;
 class GlobalObject;
 class Module;
@@ -511,6 +512,13 @@ public:
   }
   GlobalObject *getBaseObject();
 
+  /// Returns whether this is a reference to an absolute symbol.
+  bool isAbsoluteSymbolRef() const;
+
+  /// If this is an absolute symbol reference, returns the range of the symbol,
+  /// otherwise returns None.
+  Optional<ConstantRange> getAbsoluteSymbolRange() const;
+
   /// This method unlinks 'this' from the containing module, but does not delete
   /// it.
   virtual void removeFromParent() = 0;

llvm/include/llvm/IR/LLVMContext.h

@@ -77,6 +77,7 @@ public:
     MD_loop = 18,                     // "llvm.loop"
     MD_type = 19,                     // "type"
     MD_section_prefix = 20,           // "section_prefix"
+    MD_absolute_symbol = 21,          // "absolute_symbol"
   };
 
   /// Known operand bundle tag IDs, which always have the same value.  All

llvm/lib/Analysis/ValueTracking.cpp

@@ -3345,11 +3345,11 @@ bool llvm::isKnownNonNull(const Value *V) {
   if (const Argument *A = dyn_cast<Argument>(V))
     return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr();
 
-  // A global variable in address space 0 is non null unless extern weak.
-  // Other address spaces may have null as a valid address for a global,
-  // so we can't assume anything.
+  // A global variable in address space 0 is non null unless extern weak
+  // or an absolute symbol reference. Other address spaces may have null as a
+  // valid address for a global, so we can't assume anything.
   if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
-    return !GV->hasExternalWeakLinkage() &&
+    return !GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
            GV->getType()->getAddressSpace() == 0;
 
   // A Load tagged with nonnull metadata is never null.

llvm/lib/IR/Globals.cpp

@@ -15,6 +15,7 @@
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/Triple.h"
 #include "llvm/IR/Constants.h"
+#include "llvm/IR/ConstantRange.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/GlobalAlias.h"
 #include "llvm/IR/GlobalValue.h"
@@ -222,6 +223,26 @@ GlobalObject *GlobalValue::getBaseObject() {
   return nullptr;
 }
 
+bool GlobalValue::isAbsoluteSymbolRef() const {
+  auto *GO = dyn_cast<GlobalObject>(this);
+  if (!GO)
+    return false;
+
+  return GO->getMetadata(LLVMContext::MD_absolute_symbol);
+}
+
+Optional<ConstantRange> GlobalValue::getAbsoluteSymbolRange() const {
+  auto *GO = dyn_cast<GlobalObject>(this);
+  if (!GO)
+    return None;
+
+  MDNode *MD = GO->getMetadata(LLVMContext::MD_absolute_symbol);
+  if (!MD)
+    return None;
+
+  return getConstantRangeFromMetadata(*MD);
+}
+
 //===----------------------------------------------------------------------===//
 // GlobalVariable Implementation
 //===----------------------------------------------------------------------===//

llvm/lib/IR/LLVMContext.cpp

@@ -57,6 +57,7 @@ LLVMContext::LLVMContext() : pImpl(new LLVMContextImpl(*this)) {
     {MD_loop, "llvm.loop"},
     {MD_type, "type"},
     {MD_section_prefix, "section_prefix"},
+    {MD_absolute_symbol, "absolute_symbol"},
   };
 
   for (auto &MDKind : MDKinds) {

llvm/lib/Target/X86/X86ISelDAGToDAG.cpp

@@ -24,6 +24,7 @@
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/SelectionDAGISel.h"
+#include "llvm/IR/ConstantRange.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Intrinsics.h"
@@ -1571,7 +1572,15 @@ bool X86DAGToDAGISel::selectMOV64Imm32(SDValue N, SDValue &Imm) {
     return false;
 
   Imm = N;
-  return TM.getCodeModel() == CodeModel::Small;
+  if (N->getOpcode() != ISD::TargetGlobalAddress)
+    return TM.getCodeModel() == CodeModel::Small;
+
+  Optional<ConstantRange> CR =
+      cast<GlobalAddressSDNode>(N)->getGlobal()->getAbsoluteSymbolRange();
+  if (!CR)
+    return TM.getCodeModel() == CodeModel::Small;
+
+  return CR->getUnsignedMax().ult(1ull << 32);
 }
 
 bool X86DAGToDAGISel::selectLEA64_32Addr(SDValue N, SDValue &Base,
@@ -1710,10 +1719,39 @@ bool X86DAGToDAGISel::selectRelocImm(SDValue N, SDValue &Op) {
     return true;
   }
 
+  // Keep track of the original value type and whether this value was
+  // truncated. If we see a truncation from pointer type to VT that truncates
+  // bits that are known to be zero, we can use a narrow reference.
+  EVT VT = N.getValueType();
+  bool WasTruncated = false;
+  if (N.getOpcode() == ISD::TRUNCATE) {
+    WasTruncated = true;
+    N = N.getOperand(0);
+  }
+
   if (N.getOpcode() != X86ISD::Wrapper)
     return false;
 
-  Op = N.getOperand(0);
+  // We can only use non-GlobalValues as immediates if they were not truncated,
+  // as we do not have any range information. If we have a GlobalValue and the
+  // address was not truncated, we can select it as an operand directly.
+  unsigned Opc = N.getOperand(0)->getOpcode();
+  if (Opc != ISD::TargetGlobalAddress || !WasTruncated) {
+    Op = N.getOperand(0);
+    // We can only select the operand directly if we didn't have to look past a
+    // truncate.
+    return !WasTruncated;
+  }
+
+  // Check that the global's range fits into VT.
+  auto *GA = cast<GlobalAddressSDNode>(N.getOperand(0));
+  Optional<ConstantRange> CR = GA->getGlobal()->getAbsoluteSymbolRange();
+  if (!CR || CR->getUnsignedMax().uge(1ull << VT.getSizeInBits()))
+    return false;
+
+  // Okay, we can use a narrow reference.
+  Op = CurDAG->getTargetGlobalAddress(GA->getGlobal(), SDLoc(N), VT,
+                                      GA->getOffset(), GA->getTargetFlags());
   return true;
 }
 

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -13676,7 +13676,11 @@ static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget,
 }
 
 // Returns the appropriate wrapper opcode for a global reference.
-unsigned X86TargetLowering::getGlobalWrapperKind() const {
+unsigned X86TargetLowering::getGlobalWrapperKind(const GlobalValue *GV) const {
+  // References to absolute symbols are never PC-relative.
+  if (GV && GV->isAbsoluteSymbolRef())
+    return X86ISD::Wrapper;
+
   CodeModel::Model M = getTargetMachine().getCodeModel();
   if (Subtarget.isPICStyleRIPRel() &&
       (M == CodeModel::Small || M == CodeModel::Kernel))
@@ -13805,7 +13809,7 @@ SDValue X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV,
     Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
   }
 
-  Result = DAG.getNode(getGlobalWrapperKind(), dl, PtrVT, Result);
+  Result = DAG.getNode(getGlobalWrapperKind(GV), dl, PtrVT, Result);
 
   // With PIC, the address is actually $g + Offset.
   if (isGlobalRelativeToPICBase(OpFlags)) {

llvm/lib/Target/X86/X86ISelLowering.h

@@ -1115,7 +1115,7 @@ namespace llvm {
     SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
 
-    unsigned getGlobalWrapperKind() const;
+    unsigned getGlobalWrapperKind(const GlobalValue *GV = nullptr) const;
     SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerGlobalAddress(const GlobalValue *GV, const SDLoc &dl,

llvm/lib/Target/X86/X86InstrInfo.td

@@ -948,10 +948,10 @@ def i64immSExt32 : ImmLeaf<i64, [{ return isInt<32>(Imm); }]>;
 // Eventually, it would be nice to allow ConstantHoisting to merge constants
 // globally for potentially added savings.
 //
-def imm8_su : PatLeaf<(i8 imm), [{
+def imm8_su : PatLeaf<(i8 relocImm), [{
     return !shouldAvoidImmediateInstFormsForSize(N);
 }]>;
-def imm16_su : PatLeaf<(i16 imm), [{
+def imm16_su : PatLeaf<(i16 relocImm), [{
     return !shouldAvoidImmediateInstFormsForSize(N);
 }]>;
 def imm32_su : PatLeaf<(i32 relocImm), [{

llvm/lib/Target/X86/X86InstrShiftRotate.td

@@ -591,19 +591,20 @@ def ROR64rCL : RI<0xD3, MRM1r, (outs GR64:$dst), (ins GR64:$src1),
 
 def ROR8ri   : Ii8<0xC0, MRM1r, (outs GR8 :$dst), (ins GR8 :$src1, u8imm:$src2),
                    "ror{b}\t{$src2, $dst|$dst, $src2}",
-                   [(set GR8:$dst, (rotr GR8:$src1, (i8 imm:$src2)))], IIC_SR>;
+                   [(set GR8:$dst, (rotr GR8:$src1, (i8 relocImm:$src2)))],
+                   IIC_SR>;
 def ROR16ri  : Ii8<0xC1, MRM1r, (outs GR16:$dst), (ins GR16:$src1, u8imm:$src2),
                    "ror{w}\t{$src2, $dst|$dst, $src2}",
-                   [(set GR16:$dst, (rotr GR16:$src1, (i8 imm:$src2)))],
+                   [(set GR16:$dst, (rotr GR16:$src1, (i8 relocImm:$src2)))],
                    IIC_SR>, OpSize16;
 def ROR32ri  : Ii8<0xC1, MRM1r, (outs GR32:$dst), (ins GR32:$src1, u8imm:$src2),
                    "ror{l}\t{$src2, $dst|$dst, $src2}",
-                   [(set GR32:$dst, (rotr GR32:$src1, (i8 imm:$src2)))],
+                   [(set GR32:$dst, (rotr GR32:$src1, (i8 relocImm:$src2)))],
                    IIC_SR>, OpSize32;
 def ROR64ri  : RIi8<0xC1, MRM1r, (outs GR64:$dst),
                     (ins GR64:$src1, u8imm:$src2),
                     "ror{q}\t{$src2, $dst|$dst, $src2}",
-                    [(set GR64:$dst, (rotr GR64:$src1, (i8 imm:$src2)))],
+                    [(set GR64:$dst, (rotr GR64:$src1, (i8 relocImm:$src2)))],
                     IIC_SR>;
 
 // Rotate by 1

llvm/lib/Target/X86/X86Subtarget.cpp

@@ -92,6 +92,10 @@ unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV,
   if (TM.getCodeModel() == CodeModel::Large)
     return X86II::MO_NO_FLAG;
 
+  // Absolute symbols can be referenced directly.
+  if (GV && GV->isAbsoluteSymbolRef())
+    return X86II::MO_NO_FLAG;
+
   if (TM.shouldAssumeDSOLocal(M, GV))
     return classifyLocalReference(GV);
 

llvm/test/CodeGen/X86/absolute-bit-mask.ll

@@ -0,0 +1,61 @@
+; RUN: llc < %s | FileCheck %s
+; RUN: llc -relocation-model=pic < %s | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@bit_mask8 = external hidden global i8, !absolute_symbol !0
+@bit_mask32 = external hidden global i8, !absolute_symbol !1
+@bit_mask64 = external hidden global i8, !absolute_symbol !2
+
+declare void @f()
+
+define void @foo8(i8* %ptr) {
+  %load = load i8, i8* %ptr
+  ; CHECK: testb $bit_mask8, (%rdi)
+  %and = and i8 %load, ptrtoint (i8* @bit_mask8 to i8)
+  %icmp = icmp eq i8 %and, 0
+  br i1 %icmp, label %t, label %f
+
+t:
+  call void @f()
+  ret void
+
+f:
+  ret void
+}
+
+define void @foo32(i32* %ptr) {
+  %load = load i32, i32* %ptr
+  ; CHECK: testl $bit_mask32, (%rdi)
+  %and = and i32 %load, ptrtoint (i8* @bit_mask32 to i32)
+  %icmp = icmp eq i32 %and, 0
+  br i1 %icmp, label %t, label %f
+
+t:
+  call void @f()
+  ret void
+
+f:
+  ret void
+}
+
+define void @foo64(i64* %ptr) {
+  %load = load i64, i64* %ptr
+  ; CHECK: movabsq $bit_mask64, %rax
+  ; CHECK: testq (%rdi), %rax
+  %and = and i64 %load, ptrtoint (i8* @bit_mask64 to i64)
+  %icmp = icmp eq i64 %and, 0
+  br i1 %icmp, label %t, label %f
+
+t:
+  call void @f()
+  ret void
+
+f:
+  ret void
+}
+
+!0 = !{i64 0, i64 256}
+!1 = !{i64 0, i64 4294967296}
+!2 = !{i64 -1, i64 -1}

llvm/test/CodeGen/X86/absolute-bt.ll

@@ -0,0 +1,51 @@
+; RUN: llc < %s | FileCheck %s
+; RUN: llc -relocation-model=pic < %s | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@bit_mask8 = external hidden global i8, !absolute_symbol !0
+@bit_mask32 = external hidden global i8, !absolute_symbol !1
+@bit_mask64 = external hidden global i8, !absolute_symbol !2
+
+declare void @f()
+
+define void @foo32(i32* %ptr) {
+  %load = load i32, i32* %ptr
+  %and = and i32 %load, 31
+  %shl = shl i32 1, %and
+  %and2 = and i32 %shl, ptrtoint (i8* @bit_mask32 to i32)
+  ; CHECK: movl $bit_mask32, %eax
+  ; CHECK: btl %ecx, %eax
+  %icmp = icmp eq i32 %and2, 0
+  br i1 %icmp, label %t, label %f
+
+t:
+  call void @f()
+  ret void
+
+f:
+  ret void
+}
+
+define void @foo64(i64* %ptr) {
+  %load = load i64, i64* %ptr
+  %and = and i64 %load, 63
+  %shl = shl i64 1, %and
+  %and2 = and i64 %shl, ptrtoint (i8* @bit_mask64 to i64)
+  ; CHECK: movabsq $bit_mask64, %rax
+  ; CHECK: btq %rcx, %rax
+  %icmp = icmp eq i64 %and2, 0
+  br i1 %icmp, label %t, label %f
+
+t:
+  call void @f()
+  ret void
+
+f:
+  ret void
+}
+
+!0 = !{i64 0, i64 256}
+!1 = !{i64 0, i64 4294967296}
+!2 = !{i64 -1, i64 -1}

llvm/test/CodeGen/X86/absolute-constant.ll

@@ -0,0 +1,28 @@
+; RUN: llc < %s | FileCheck %s
+; RUN: llc -relocation-model=pic < %s | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@foo = external global i8, align 1, !absolute_symbol !0
+
+define void @bar(i8* %x) {
+entry:
+  %0 = load i8, i8* %x, align 1
+  %conv = sext i8 %0 to i32
+  ; CHECK: testb $foo, (%rdi)
+  %and = and i32 %conv, sext (i8 ptrtoint (i8* @foo to i8) to i32)
+  %tobool = icmp eq i32 %and, 0
+  br i1 %tobool, label %if.end, label %if.then
+
+if.then:                                          ; preds = %entry
+  tail call void (...) @xf()
+  br label %if.end
+
+if.end:                                           ; preds = %entry, %if.then
+  ret void
+}
+
+declare void @xf(...)
+
+!0 = !{i32 0, i32 256}

llvm/test/CodeGen/X86/absolute-rotate.ll

@@ -0,0 +1,27 @@
+; RUN: llc < %s | FileCheck %s
+; RUN: llc -relocation-model=pic < %s | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@align = external hidden global i8, !absolute_symbol !0
+
+declare void @f()
+
+define void @foo(i64 %val) {
+  %shr = lshr i64 %val, zext (i8 ptrtoint (i8* @align to i8) to i64)
+  %shl = shl i64 %val, zext (i8 sub (i8 64, i8 ptrtoint (i8* @align to i8)) to i64)
+  ; CHECK: rorq $align, %rdi
+  %ror = or i64 %shr, %shl
+  %cmp = icmp ult i64 %ror, 109
+  br i1 %cmp, label %t, label %f
+
+t:
+  call void @f()
+  ret void
+
+f:
+  ret void
+}
+
+!0 = !{i64 0, i64 256}

llvm/test/Transforms/FunctionAttrs/nonnull-global.ll

@@ -0,0 +1,10 @@
+; RUN: opt -S -functionattrs %s | FileCheck %s
+
+@a = external global i8, !absolute_symbol !0
+
+; CHECK-NOT: define nonnull
+define i8* @foo() {
+  ret i8* @a
+}
+
+!0 = !{i64 0, i64 256}