From: Chandler Carruth <chandlerc@gmail.com>
Date: Thu, 7 Sep 2017 23:54:24 +0000 (+0000)
Subject: [x86] Extend the manual ISel of `add` and `sub` with both RMW memory
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=803b2c7e6939323250acfe1373f0a7ac6b4cb15e;p=llvm

[x86] Extend the manual ISel of `add` and `sub` with both RMW memory
operands and used flags to support matching immediate operands.

This is a bit trickier than register operands, and we still want to fall
back on a register operands even for things that appear to be
"immediates" when they won't actually select into the operation's
immediate operand. This also requires us to handle things like selecting
`sub` vs. `add` to minimize the number of bits needed to represent the
immediate, and picking the shortest immediate encoding. In order to
that, we in turn need to scan to make sure that CF isn't used as it will
get inverted.

The end result seems very nice though, and we're now generating
optimal instruction sequences for these patterns IMO.

A follow-up patch will further expand this to other operations with RMW
memory operands. But handing `add` and `sub` are useful starting points
to flesh out the machinery and make sure interesting and complex cases
can be handled.

Thanks to Craig Topper who provided a few fixes and improvements to this
patch in addition to the review!

Differential Revision: https://reviews.llvm.org/D37139

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@312764 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index 2fec38d7c40..17ee5ca2d4b 100644
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -1932,6 +1932,73 @@ static bool hasNoSignedComparisonUses(SDNode *N) {
   return true;
 }
 
+/// Test whether the given node which sets flags has any uses which require the
+/// CF flag to be accurate.
+static bool hasNoCarryFlagUses(SDNode *N) {
+  // Examine each user of the node.
+  for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); UI != UE;
+       ++UI) {
+    // Only check things that use the flags.
+    if (UI.getUse().getResNo() != 1)
+      continue;
+    // Only examine CopyToReg uses.
+    if (UI->getOpcode() != ISD::CopyToReg)
+      return false;
+    // Only examine CopyToReg uses that copy to EFLAGS.
+    if (cast<RegisterSDNode>(UI->getOperand(1))->getReg() != X86::EFLAGS)
+      return false;
+    // Examine each user of the CopyToReg use.
+    for (SDNode::use_iterator FlagUI = UI->use_begin(), FlagUE = UI->use_end();
+         FlagUI != FlagUE; ++FlagUI) {
+      // Only examine the Flag result.
+      if (FlagUI.getUse().getResNo() != 1)
+        continue;
+      // Anything unusual: assume conservatively.
+      if (!FlagUI->isMachineOpcode())
+        return false;
+      // Examine the opcode of the user.
+      switch (FlagUI->getMachineOpcode()) {
+      // Comparisons which don't examine the CF flag.
+      case X86::SETOr: case X86::SETNOr: case X86::SETEr: case X86::SETNEr:
+      case X86::SETSr: case X86::SETNSr: case X86::SETPr: case X86::SETNPr:
+      case X86::SETLr: case X86::SETGEr: case X86::SETLEr: case X86::SETGr:
+      case X86::JO_1: case X86::JNO_1: case X86::JE_1: case X86::JNE_1:
+      case X86::JS_1: case X86::JNS_1: case X86::JP_1: case X86::JNP_1:
+      case X86::JL_1: case X86::JGE_1: case X86::JLE_1: case X86::JG_1:
+      case X86::CMOVO16rr: case X86::CMOVO32rr: case X86::CMOVO64rr:
+      case X86::CMOVO16rm: case X86::CMOVO32rm: case X86::CMOVO64rm:
+      case X86::CMOVNO16rr: case X86::CMOVNO32rr: case X86::CMOVNO64rr:
+      case X86::CMOVNO16rm: case X86::CMOVNO32rm: case X86::CMOVNO64rm:
+      case X86::CMOVE16rr: case X86::CMOVE32rr: case X86::CMOVE64rr:
+      case X86::CMOVE16rm: case X86::CMOVE32rm: case X86::CMOVE64rm:
+      case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
+      case X86::CMOVNE16rm: case X86::CMOVNE32rm: case X86::CMOVNE64rm:
+      case X86::CMOVS16rr: case X86::CMOVS32rr: case X86::CMOVS64rr:
+      case X86::CMOVS16rm: case X86::CMOVS32rm: case X86::CMOVS64rm:
+      case X86::CMOVNS16rr: case X86::CMOVNS32rr: case X86::CMOVNS64rr:
+      case X86::CMOVNS16rm: case X86::CMOVNS32rm: case X86::CMOVNS64rm:
+      case X86::CMOVP16rr: case X86::CMOVP32rr: case X86::CMOVP64rr:
+      case X86::CMOVP16rm: case X86::CMOVP32rm: case X86::CMOVP64rm:
+      case X86::CMOVNP16rr: case X86::CMOVNP32rr: case X86::CMOVNP64rr:
+      case X86::CMOVNP16rm: case X86::CMOVNP32rm: case X86::CMOVNP64rm:
+      case X86::CMOVL16rr: case X86::CMOVL32rr: case X86::CMOVL64rr:
+      case X86::CMOVL16rm: case X86::CMOVL32rm: case X86::CMOVL64rm:
+      case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
+      case X86::CMOVGE16rm: case X86::CMOVGE32rm: case X86::CMOVGE64rm:
+      case X86::CMOVLE16rr: case X86::CMOVLE32rr: case X86::CMOVLE64rr:
+      case X86::CMOVLE16rm: case X86::CMOVLE32rm: case X86::CMOVLE64rm:
+      case X86::CMOVG16rr: case X86::CMOVG32rr: case X86::CMOVG64rr:
+      case X86::CMOVG16rm: case X86::CMOVG32rm: case X86::CMOVG64rm:
+        continue;
+      // Anything else: assume conservatively.
+      default:
+        return false;
+      }
+    }
+  }
+  return true;
+}
+
 /// Check whether or not the chain ending in StoreNode is suitable for doing
 /// the {load; op; store} to modify transformation.
 static bool isFusableLoadOpStorePattern(StoreSDNode *StoreNode,
@@ -2064,8 +2131,8 @@ bool X86DAGToDAGISel::foldLoadStoreIntoMemOperand(SDNode *Node) {
                   Segment))
     return false;
 
-  auto SelectOpcodeForSize = [&](unsigned Opc64, unsigned Opc32, unsigned Opc16,
-                                 unsigned Opc8) {
+  auto SelectOpcode = [&](unsigned Opc64, unsigned Opc32, unsigned Opc16,
+                          unsigned Opc8 = 0) {
     switch (MemVT.getSimpleVT().SimpleTy) {
     case MVT::i64:
       return Opc64;
@@ -2084,11 +2151,10 @@ bool X86DAGToDAGISel::foldLoadStoreIntoMemOperand(SDNode *Node) {
   switch (Opc) {
   case X86ISD::INC:
   case X86ISD::DEC: {
-    unsigned NewOpc = Opc == X86ISD::INC
-                          ? SelectOpcodeForSize(X86::INC64m, X86::INC32m,
-                                                X86::INC16m, X86::INC8m)
-                          : SelectOpcodeForSize(X86::DEC64m, X86::DEC32m,
-                                                X86::DEC16m, X86::DEC8m);
+    unsigned NewOpc =
+        Opc == X86ISD::INC
+            ? SelectOpcode(X86::INC64m, X86::INC32m, X86::INC16m, X86::INC8m)
+            : SelectOpcode(X86::DEC64m, X86::DEC32m, X86::DEC16m, X86::DEC8m);
     const SDValue Ops[] = {Base, Scale, Index, Disp, Segment, InputChain};
     Result =
         CurDAG->getMachineNode(NewOpc, SDLoc(Node), MVT::i32, MVT::Other, Ops);
@@ -2096,14 +2162,75 @@ bool X86DAGToDAGISel::foldLoadStoreIntoMemOperand(SDNode *Node) {
   }
   case X86ISD::ADD:
   case X86ISD::SUB: {
-    unsigned NewOpc = Opc == X86ISD::ADD
-                          ? SelectOpcodeForSize(X86::ADD64mr, X86::ADD32mr,
-                                                X86::ADD16mr, X86::ADD8mr)
-                          : SelectOpcodeForSize(X86::SUB64mr, X86::SUB32mr,
-                                                X86::SUB16mr, X86::SUB8mr);
-    const SDValue Ops[] = {Base,      Scale,   Index,
-                           Disp,      Segment, StoredVal->getOperand(1),
-                           InputChain};
+    auto SelectRegOpcode = [SelectOpcode](unsigned Opc) {
+      switch (Opc) {
+      case X86ISD::ADD:
+        return SelectOpcode(X86::ADD64mr, X86::ADD32mr, X86::ADD16mr,
+                            X86::ADD8mr);
+      case X86ISD::SUB:
+        return SelectOpcode(X86::SUB64mr, X86::SUB32mr, X86::SUB16mr,
+                            X86::SUB8mr);
+      default:
+        llvm_unreachable("Invalid opcode!");
+      }
+    };
+    auto SelectImm8Opcode = [SelectOpcode](unsigned Opc) {
+      switch (Opc) {
+      case X86ISD::ADD:
+        return SelectOpcode(X86::ADD64mi8, X86::ADD32mi8, X86::ADD16mi8);
+      case X86ISD::SUB:
+        return SelectOpcode(X86::SUB64mi8, X86::SUB32mi8, X86::SUB16mi8);
+      default:
+        llvm_unreachable("Invalid opcode!");
+      }
+    };
+    auto SelectImmOpcode = [SelectOpcode](unsigned Opc) {
+      switch (Opc) {
+      case X86ISD::ADD:
+        return SelectOpcode(X86::ADD64mi32, X86::ADD32mi, X86::ADD16mi,
+                            X86::ADD8mi);
+      case X86ISD::SUB:
+        return SelectOpcode(X86::SUB64mi32, X86::SUB32mi, X86::SUB16mi,
+                            X86::SUB8mi);
+      default:
+        llvm_unreachable("Invalid opcode!");
+      }
+    };
+
+    unsigned NewOpc = SelectRegOpcode(Opc);
+    SDValue Operand = StoredVal->getOperand(1);
+
+    // See if the operand is a constant that we can fold into an immediate
+    // operand.
+    if (auto *OperandC = dyn_cast<ConstantSDNode>(Operand)) {
+      auto OperandV = OperandC->getAPIntValue();
+
+      // Check if we can shrink the operand enough to fit in an immediate (or
+      // fit into a smaller immediate) by negating it and switching the
+      // operation.
+      if (((MemVT != MVT::i8 && OperandV.getMinSignedBits() > 8 &&
+            (-OperandV).getMinSignedBits() <= 8) ||
+           (MemVT == MVT::i64 && OperandV.getMinSignedBits() > 32 &&
+            (-OperandV).getMinSignedBits() <= 32)) &&
+          hasNoCarryFlagUses(StoredVal.getNode())) {
+        OperandV = -OperandV;
+        Opc = Opc == X86ISD::ADD ? X86ISD::SUB : X86ISD::ADD;
+      }
+
+      // First try to fit this into an Imm8 operand. If it doesn't fit, then try
+      // the larger immediate operand.
+      if (MemVT != MVT::i8 && OperandV.getMinSignedBits() <= 8) {
+        Operand = CurDAG->getTargetConstant(OperandV, SDLoc(Node), MemVT);
+        NewOpc = SelectImm8Opcode(Opc);
+      } else if (OperandV.getActiveBits() <= MemVT.getSizeInBits() &&
+                 (MemVT != MVT::i64 || OperandV.getMinSignedBits() <= 32)) {
+        Operand = CurDAG->getTargetConstant(OperandV, SDLoc(Node), MemVT);
+        NewOpc = SelectImmOpcode(Opc);
+      }
+    }
+
+    const SDValue Ops[] = {Base,    Scale,   Index,     Disp,
+                           Segment, Operand, InputChain};
     Result =
         CurDAG->getMachineNode(NewOpc, SDLoc(Node), MVT::i32, MVT::Other, Ops);
     break;
diff --git a/test/CodeGen/X86/add.ll b/test/CodeGen/X86/add.ll
index 28ababe29f7..634366bf05c 100644
--- a/test/CodeGen/X86/add.ll
+++ b/test/CodeGen/X86/add.ll
@@ -340,8 +340,7 @@ entry:
 define void @test12(i64* inreg %a) nounwind {
 ; X32-LABEL: test12:
 ; X32:       # BB#0: # %entry
-; X32-NEXT:    movl $-2147483648, %ecx # imm = 0x80000000
-; X32-NEXT:    addl %ecx, (%eax)
+; X32-NEXT:    addl $-2147483648, (%eax) # imm = 0x80000000
 ; X32-NEXT:    adcl $0, 4(%eax)
 ; X32-NEXT:    retl
 ;
@@ -364,8 +363,7 @@ entry:
 define void @test13(i64* inreg %a) nounwind {
 ; X32-LABEL: test13:
 ; X32:       # BB#0: # %entry
-; X32-NEXT:    movl $128, %ecx
-; X32-NEXT:    addl %ecx, (%eax)
+; X32-NEXT:    addl $128, (%eax)
 ; X32-NEXT:    adcl $0, 4(%eax)
 ; X32-NEXT:    retl
 ;
diff --git a/test/CodeGen/X86/fold-rmw-ops.ll b/test/CodeGen/X86/fold-rmw-ops.ll
index e9457bbf4a6..31554e4c1e0 100644
--- a/test/CodeGen/X86/fold-rmw-ops.ll
+++ b/test/CodeGen/X86/fold-rmw-ops.ll
@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -verify-machineinstrs -show-mc-encoding | FileCheck %s
 
 target triple = "x86_64-unknown-unknown"
 
@@ -11,19 +11,259 @@ target triple = "x86_64-unknown-unknown"
 declare void @a()
 declare void @b()
 
-define void @add64_imm_br() nounwind {
-; CHECK-LABEL: add64_imm_br:
+define void @add64_imm32_br() nounwind {
+; CHECK-LABEL: add64_imm32_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    movl $42, %eax
-; CHECK-NEXT:    addq %rax, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB0_1
+; CHECK-NEXT:    addq $16777214, {{.*}}(%rip) # encoding: [0x48,0x81,0x05,A,A,A,A,0xfe,0xff,0xff,0x00]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-8, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # imm = 0xFFFFFE
+; CHECK-NEXT:    js .LBB0_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB0_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  .LBB0_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i64, i64* @g64
-  %add = add nsw i64 %load1, 42
+  ; Add 0x00FFFFFE, a positive immediate requiring 24-bits.
+  %add = add i64 %load1, 16777214
+  store i64 %add, i64* @g64
+  %cond = icmp slt i64 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add64_sext_imm32_br() nounwind {
+; CHECK-LABEL: add64_sext_imm32_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    addq $-2147483648, {{.*}}(%rip) # encoding: [0x48,0x81,0x05,A,A,A,A,0x00,0x00,0x00,0x80]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-8, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # imm = 0x80000000
+; CHECK-NEXT:    js .LBB1_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB1_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB1_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  ; Add -0x80000000, which requires sign-extended 32 bits.
+  %add = add i64 %load1, -2147483648
+  store i64 %add, i64* @g64
+  %cond = icmp slt i64 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add64_imm32_via_sub_br() nounwind {
+; CHECK-LABEL: add64_imm32_via_sub_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    subq $-2147483648, {{.*}}(%rip) # encoding: [0x48,0x81,0x2d,A,A,A,A,0x00,0x00,0x00,0x80]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-8, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # imm = 0x80000000
+; CHECK-NEXT:    js .LBB2_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB2_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB2_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  ; Add 0x80000000, which cannot fit in a sign extended 32-bit immediate. This
+  ; get's folded because we can instead subtract -0x80000000.
+  %add = add i64 %load1, 2147483648
+  store i64 %add, i64* @g64
+  %cond = icmp slt i64 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add64_no_imm32_via_sub_due_to_cf_br() nounwind {
+; CHECK-LABEL: add64_no_imm32_via_sub_due_to_cf_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    movl $2147483648, %eax # encoding: [0xb8,0x00,0x00,0x00,0x80]
+; CHECK-NEXT:    # imm = 0x80000000
+; CHECK-NEXT:    addq %rax, {{.*}}(%rip) # encoding: [0x48,0x01,0x05,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    jae .LBB3_2 # encoding: [0x73,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB3_2-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB3_2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  ; Add 0x80000000, which cannot fit in a sign extended 32-bit immediate, but
+  ; could in theory be folded into an immediate operand of a sub. However, we
+  ; use the CF flag here and so shouldn't make that transformation.
+  %add = add i64 %load1, 2147483648
+  store i64 %add, i64* @g64
+  %cond = icmp ult i64 %add, 2147483648
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add64_too_large_imm32_br() nounwind {
+; CHECK-LABEL: add64_too_large_imm32_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    movl $2147483649, %eax # encoding: [0xb8,0x01,0x00,0x00,0x80]
+; CHECK-NEXT:    # imm = 0x80000001
+; CHECK-NEXT:    addq %rax, {{.*}}(%rip) # encoding: [0x48,0x01,0x05,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB4_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB4_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB4_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  ; Add 0x80000001, which cannot fit in a sign extended 32-bit immediate. This
+  ; should not get folded into an immediate.
+  %add = add i64 %load1, 2147483649
+  store i64 %add, i64* @g64
+  %cond = icmp slt i64 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add64_imm8_via_sub_br() nounwind {
+; CHECK-LABEL: add64_imm8_via_sub_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    subq $-128, {{.*}}(%rip) # encoding: [0x48,0x83,0x2d,A,A,A,A,0x80]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB5_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB5_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB5_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  ; Add 0x80 which can't quite fit into an imm8 because it would be sign
+  ; extended, but which can fit if we convert to a sub and negate the value.
+  %add = add i64 %load1, 128
+  store i64 %add, i64* @g64
+  %cond = icmp slt i64 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add64_imm8_br() nounwind {
+; CHECK-LABEL: add64_imm8_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    addq $42, {{.*}}(%rip) # encoding: [0x48,0x83,0x05,A,A,A,A,0x2a]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB6_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB6_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB6_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  %add = add i64 %load1, 42
+  store i64 %add, i64* @g64
+  %cond = icmp slt i64 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add64_imm8_neg_br() nounwind {
+; CHECK-LABEL: add64_imm8_neg_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    addq $-42, {{.*}}(%rip) # encoding: [0x48,0x83,0x05,A,A,A,A,0xd6]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB7_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB7_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB7_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  %add = add i64 %load1, -42
   store i64 %add, i64* @g64
   %cond = icmp slt i64 %add, 0
   br i1 %cond, label %a, label %b
@@ -40,16 +280,85 @@ b:
 define void @add32_imm_br() nounwind {
 ; CHECK-LABEL: add32_imm_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    movl $42, %eax
-; CHECK-NEXT:    addl %eax, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB1_1
+; CHECK-NEXT:    addl $-2147483648, {{.*}}(%rip) # encoding: [0x81,0x05,A,A,A,A,0x00,0x00,0x00,0x80]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g32-8, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # imm = 0x80000000
+; CHECK-NEXT:    js .LBB8_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB8_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB1_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB8_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i32, i32* @g32
-  %add = add nsw i32 %load1, 42
+  ; Add 0x80000000, a positive number requiring 32 bits of immediate.
+  %add = add i32 %load1, 2147483648
+  store i32 %add, i32* @g32
+  %cond = icmp slt i32 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add32_imm8_br() nounwind {
+; CHECK-LABEL: add32_imm8_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    addl $42, {{.*}}(%rip) # encoding: [0x83,0x05,A,A,A,A,0x2a]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g32-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB9_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB9_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB9_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i32, i32* @g32
+  %add = add i32 %load1, 42
+  store i32 %add, i32* @g32
+  %cond = icmp slt i32 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add32_imm8_neg_br() nounwind {
+; CHECK-LABEL: add32_imm8_neg_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    addl $-42, {{.*}}(%rip) # encoding: [0x83,0x05,A,A,A,A,0xd6]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g32-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB10_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB10_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB10_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i32, i32* @g32
+  %add = add i32 %load1, -42
   store i32 %add, i32* @g32
   %cond = icmp slt i32 %add, 0
   br i1 %cond, label %a, label %b
@@ -66,16 +375,85 @@ b:
 define void @add16_imm_br() nounwind {
 ; CHECK-LABEL: add16_imm_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    movw $42, %ax
-; CHECK-NEXT:    addw %ax, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB2_1
+; CHECK-NEXT:    addw $-32768, {{.*}}(%rip) # encoding: [0x66,0x81,0x05,A,A,A,A,0x00,0x80]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g16-6, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # imm = 0x8000
+; CHECK-NEXT:    js .LBB11_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB11_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB2_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB11_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i16, i16* @g16
+  ; Add 0x8000, a positive number requiring 16 bits of immediate.
+  %add = add i16 %load1, 32768
+  store i16 %add, i16* @g16
+  %cond = icmp slt i16 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add16_imm8_br() nounwind {
+; CHECK-LABEL: add16_imm8_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    addw $42, {{.*}}(%rip) # encoding: [0x66,0x83,0x05,A,A,A,A,0x2a]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g16-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB12_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB12_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB12_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i16, i16* @g16
-  %add = add nsw i16 %load1, 42
+  %add = add i16 %load1, 42
+  store i16 %add, i16* @g16
+  %cond = icmp slt i16 %add, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @add16_imm8_neg_br() nounwind {
+; CHECK-LABEL: add16_imm8_neg_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    addw $-42, {{.*}}(%rip) # encoding: [0x66,0x83,0x05,A,A,A,A,0xd6]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g16-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB13_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB13_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB13_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i16, i16* @g16
+  %add = add i16 %load1, -42
   store i16 %add, i16* @g16
   %cond = icmp slt i16 %add, 0
   br i1 %cond, label %a, label %b
@@ -92,16 +470,21 @@ b:
 define void @add8_imm_br() nounwind {
 ; CHECK-LABEL: add8_imm_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    movb $42, %al
-; CHECK-NEXT:    addb %al, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB3_1
+; CHECK-NEXT:    addb $-2, {{.*}}(%rip) # encoding: [0x80,0x05,A,A,A,A,0xfe]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g8-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB14_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB14_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB3_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB14_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i8, i8* @g8
-  %add = add nsw i8 %load1, 42
+  %add = add i8 %load1, -2
   store i8 %add, i8* @g8
   %cond = icmp slt i8 %add, 0
   br i1 %cond, label %a, label %b
@@ -118,15 +501,21 @@ b:
 define void @add64_reg_br(i64 %arg) nounwind {
 ; CHECK-LABEL: add64_reg_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    addq %rdi, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB4_1
+; CHECK-NEXT:    addq %rdi, {{.*}}(%rip) # encoding: [0x48,0x01,0x3d,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB15_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB15_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB4_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB15_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i64, i64* @g64
-  %add = add nsw i64 %load1, %arg
+  %add = add i64 %load1, %arg
   store i64 %add, i64* @g64
   %cond = icmp slt i64 %add, 0
   br i1 %cond, label %a, label %b
@@ -143,15 +532,21 @@ b:
 define void @add32_reg_br(i32 %arg) nounwind {
 ; CHECK-LABEL: add32_reg_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    addl %edi, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB5_1
+; CHECK-NEXT:    addl %edi, {{.*}}(%rip) # encoding: [0x01,0x3d,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g32-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB16_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB16_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB5_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB16_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i32, i32* @g32
-  %add = add nsw i32 %load1, %arg
+  %add = add i32 %load1, %arg
   store i32 %add, i32* @g32
   %cond = icmp slt i32 %add, 0
   br i1 %cond, label %a, label %b
@@ -168,15 +563,21 @@ b:
 define void @add16_reg_br(i16 %arg) nounwind {
 ; CHECK-LABEL: add16_reg_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    addw %di, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB6_1
+; CHECK-NEXT:    addw %di, {{.*}}(%rip) # encoding: [0x66,0x01,0x3d,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g16-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB17_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB17_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB6_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB17_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i16, i16* @g16
-  %add = add nsw i16 %load1, %arg
+  %add = add i16 %load1, %arg
   store i16 %add, i16* @g16
   %cond = icmp slt i16 %add, 0
   br i1 %cond, label %a, label %b
@@ -193,15 +594,21 @@ b:
 define void @add8_reg_br(i8 %arg) nounwind {
 ; CHECK-LABEL: add8_reg_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    addb %dil, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB7_1
+; CHECK-NEXT:    addb %dil, {{.*}}(%rip) # encoding: [0x40,0x00,0x3d,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g8-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB18_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB18_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB7_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB18_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i8, i8* @g8
-  %add = add nsw i8 %load1, %arg
+  %add = add i8 %load1, %arg
   store i8 %add, i8* @g8
   %cond = icmp slt i8 %add, 0
   br i1 %cond, label %a, label %b
@@ -215,19 +622,95 @@ b:
   ret void
 }
 
-define void @sub64_imm_br() nounwind {
-; CHECK-LABEL: sub64_imm_br:
+define void @sub64_imm32_br() nounwind {
+; CHECK-LABEL: sub64_imm32_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    movq $-42, %rax
-; CHECK-NEXT:    addq %rax, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB8_1
+; CHECK-NEXT:    subq $-2147483648, {{.*}}(%rip) # encoding: [0x48,0x81,0x2d,A,A,A,A,0x00,0x00,0x00,0x80]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-8, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # imm = 0x80000000
+; CHECK-NEXT:    js .LBB19_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB19_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB8_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB19_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i64, i64* @g64
-  %sub = sub nsw i64 %load1, 42
+  ; Subtract -0x80000000, which can't be negated into a sign-extended 32-bit
+  ; immediate, so that we have to select sub here.
+  %sub = sub i64 %load1, -2147483648
+  store i64 %sub, i64* @g64
+  %cond = icmp slt i64 %sub, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @sub64_too_large_imm32_br() nounwind {
+; CHECK-LABEL: sub64_too_large_imm32_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    movabsq $-4294967295, %rax # encoding: [0x48,0xb8,0x01,0x00,0x00,0x00,0xff,0xff,0xff,0xff]
+; CHECK-NEXT:    # imm = 0xFFFFFFFF00000001
+; CHECK-NEXT:    addq %rax, {{.*}}(%rip) # encoding: [0x48,0x01,0x05,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB20_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB20_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB20_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  ; Subtract 0xFFFFFFFF, which cannot fit in a sign extended 32-bit immediate,
+  ; even if negated and sign extended as an add.
+  %sub = sub i64 %load1, 4294967295
+  store i64 %sub, i64* @g64
+  %cond = icmp slt i64 %sub, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @sub64_imm8_br() nounwind {
+; CHECK-LABEL: sub64_imm8_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    subq $-128, {{.*}}(%rip) # encoding: [0x48,0x83,0x2d,A,A,A,A,0x80]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB21_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB21_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB21_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i64, i64* @g64
+  ; Subtract -0x80, which can be done with an 8-bit immediate but only as
+  ; a subtract where that immediate can be negative.
+  %sub = sub i64 %load1, -128
   store i64 %sub, i64* @g64
   %cond = icmp slt i64 %sub, 0
   br i1 %cond, label %a, label %b
@@ -244,16 +727,57 @@ b:
 define void @sub32_imm_br() nounwind {
 ; CHECK-LABEL: sub32_imm_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    movl $-42, %eax
-; CHECK-NEXT:    addl %eax, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB9_1
+; CHECK-NEXT:    addl $-2147483648, {{.*}}(%rip) # encoding: [0x81,0x05,A,A,A,A,0x00,0x00,0x00,0x80]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g32-8, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # imm = 0x80000000
+; CHECK-NEXT:    js .LBB22_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB22_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB9_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB22_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i32, i32* @g32
-  %sub = sub nsw i32 %load1, 42
+  ; Subtract -0x80000000, which requires 32 bits of immediate but still gets
+  ; lowered as an add.
+  %sub = sub i32 %load1, -2147483648
+  store i32 %sub, i32* @g32
+  %cond = icmp slt i32 %sub, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @sub32_imm8_br() nounwind {
+; CHECK-LABEL: sub32_imm8_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    subl $-128, {{.*}}(%rip) # encoding: [0x83,0x2d,A,A,A,A,0x80]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g32-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB23_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB23_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB23_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i32, i32* @g32
+  ; Subtract -0x80, which can be done with an 8-bit immediate but only as
+  ; a subtract where that immediate can be negative.
+  %sub = sub i32 %load1, -128
   store i32 %sub, i32* @g32
   %cond = icmp slt i32 %sub, 0
   br i1 %cond, label %a, label %b
@@ -270,16 +794,57 @@ b:
 define void @sub16_imm_br() nounwind {
 ; CHECK-LABEL: sub16_imm_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    movw $-42, %ax
-; CHECK-NEXT:    addw %ax, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB10_1
+; CHECK-NEXT:    addw $-32768, {{.*}}(%rip) # encoding: [0x66,0x81,0x05,A,A,A,A,0x00,0x80]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g16-6, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # imm = 0x8000
+; CHECK-NEXT:    js .LBB24_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB24_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB10_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB24_1: # %a
+; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
+entry:
+  %load1 = load i16, i16* @g16
+  ; Subtract -0x8000, which requires a 16 bits of immediate but still gets
+  ; lowered as an add.
+  %sub = sub i16 %load1, -32768
+  store i16 %sub, i16* @g16
+  %cond = icmp slt i16 %sub, 0
+  br i1 %cond, label %a, label %b
+
+a:
+  tail call void @a()
+  ret void
+
+b:
+  tail call void @b()
+  ret void
+}
+
+define void @sub16_imm8_br() nounwind {
+; CHECK-LABEL: sub16_imm8_br:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    subw $-128, {{.*}}(%rip) # encoding: [0x66,0x83,0x2d,A,A,A,A,0x80]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g16-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB25_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB25_1-1, kind: FK_PCRel_1
+; CHECK-NEXT:  # BB#2: # %b
+; CHECK-NEXT:    jmp b # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB25_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i16, i16* @g16
-  %sub = sub nsw i16 %load1, 42
+  ; Subtract -0x80, which can be done with an 8-bit immediate but only as
+  ; a subtract where that immediate can be negative.
+  %sub = sub i16 %load1, -128
   store i16 %sub, i16* @g16
   %cond = icmp slt i16 %sub, 0
   br i1 %cond, label %a, label %b
@@ -296,16 +861,23 @@ b:
 define void @sub8_imm_br() nounwind {
 ; CHECK-LABEL: sub8_imm_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    movb $-42, %al
-; CHECK-NEXT:    addb %al, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB11_1
+; CHECK-NEXT:    addb $-128, {{.*}}(%rip) # encoding: [0x80,0x05,A,A,A,A,0x80]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g8-5, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB26_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB26_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB11_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB26_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i8, i8* @g8
-  %sub = sub nsw i8 %load1, 42
+  ; Subtract -0x80, which requires an 8-bit immediate but still gets lowered as
+  ; an add.
+  %sub = sub i8 %load1, -128
   store i8 %sub, i8* @g8
   %cond = icmp slt i8 %sub, 0
   br i1 %cond, label %a, label %b
@@ -322,15 +894,21 @@ b:
 define void @sub64_reg_br(i64 %arg) nounwind {
 ; CHECK-LABEL: sub64_reg_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    subq %rdi, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB12_1
+; CHECK-NEXT:    subq %rdi, {{.*}}(%rip) # encoding: [0x48,0x29,0x3d,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g64-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB27_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB27_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB12_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB27_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i64, i64* @g64
-  %sub = sub nsw i64 %load1, %arg
+  %sub = sub i64 %load1, %arg
   store i64 %sub, i64* @g64
   %cond = icmp slt i64 %sub, 0
   br i1 %cond, label %a, label %b
@@ -347,15 +925,21 @@ b:
 define void @sub32_reg_br(i32 %arg) nounwind {
 ; CHECK-LABEL: sub32_reg_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    subl %edi, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB13_1
+; CHECK-NEXT:    subl %edi, {{.*}}(%rip) # encoding: [0x29,0x3d,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 2, value: g32-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB28_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB28_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB13_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB28_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i32, i32* @g32
-  %sub = sub nsw i32 %load1, %arg
+  %sub = sub i32 %load1, %arg
   store i32 %sub, i32* @g32
   %cond = icmp slt i32 %sub, 0
   br i1 %cond, label %a, label %b
@@ -372,15 +956,21 @@ b:
 define void @sub16_reg_br(i16 %arg) nounwind {
 ; CHECK-LABEL: sub16_reg_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    subw %di, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB14_1
+; CHECK-NEXT:    subw %di, {{.*}}(%rip) # encoding: [0x66,0x29,0x3d,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g16-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB29_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB29_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB14_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB29_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i16, i16* @g16
-  %sub = sub nsw i16 %load1, %arg
+  %sub = sub i16 %load1, %arg
   store i16 %sub, i16* @g16
   %cond = icmp slt i16 %sub, 0
   br i1 %cond, label %a, label %b
@@ -397,15 +987,21 @@ b:
 define void @sub8_reg_br(i8 %arg) nounwind {
 ; CHECK-LABEL: sub8_reg_br:
 ; CHECK:       # BB#0: # %entry
-; CHECK-NEXT:    subb %dil, {{.*}}(%rip)
-; CHECK-NEXT:    js .LBB15_1
+; CHECK-NEXT:    subb %dil, {{.*}}(%rip) # encoding: [0x40,0x28,0x3d,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 3, value: g8-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    js .LBB30_1 # encoding: [0x78,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: .LBB30_1-1, kind: FK_PCRel_1
 ; CHECK-NEXT:  # BB#2: # %b
 ; CHECK-NEXT:    jmp b # TAILCALL
-; CHECK-NEXT:  .LBB15_1: # %a
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: b-1, kind: FK_PCRel_1
+; CHECK-NEXT:  .LBB30_1: # %a
 ; CHECK-NEXT:    jmp a # TAILCALL
+; CHECK-NEXT:    # encoding: [0xeb,A]
+; CHECK-NEXT:    # fixup A - offset: 1, value: a-1, kind: FK_PCRel_1
 entry:
   %load1 = load i8, i8* @g8
-  %sub = sub nsw i8 %load1, %arg
+  %sub = sub i8 %load1, %arg
   store i8 %sub, i8* @g8
   %cond = icmp slt i8 %sub, 0
   br i1 %cond, label %a, label %b
diff --git a/test/CodeGen/X86/peephole-na-phys-copy-folding.ll b/test/CodeGen/X86/peephole-na-phys-copy-folding.ll
index 6a83fe01b09..e2f28e6ba4c 100644
--- a/test/CodeGen/X86/peephole-na-phys-copy-folding.ll
+++ b/test/CodeGen/X86/peephole-na-phys-copy-folding.ll
@@ -66,8 +66,7 @@ define i1 @plus_forty_two() nounwind {
 ; CHECK32-LABEL: plus_forty_two:
 ; CHECK32:       # BB#0: # %entry
 ; CHECK32-NEXT:    movb M, %al
-; CHECK32-NEXT:    movl $42, %ecx
-; CHECK32-NEXT:    addl %ecx, L
+; CHECK32-NEXT:    addl $42, L
 ; CHECK32-NEXT:    jne .LBB1_2
 ; CHECK32-NEXT:  # BB#1: # %entry
 ; CHECK32-NEXT:    andb $8, %al
@@ -82,8 +81,7 @@ define i1 @plus_forty_two() nounwind {
 ; CHECK64-LABEL: plus_forty_two:
 ; CHECK64:       # BB#0: # %entry
 ; CHECK64-NEXT:    movb {{.*}}(%rip), %al
-; CHECK64-NEXT:    movl $42, %ecx
-; CHECK64-NEXT:    addl %ecx, {{.*}}(%rip)
+; CHECK64-NEXT:    addl $42, {{.*}}(%rip)
 ; CHECK64-NEXT:    jne .LBB1_2
 ; CHECK64-NEXT:  # BB#1: # %entry
 ; CHECK64-NEXT:    andb $8, %al
@@ -164,8 +162,7 @@ define i1 @minus_forty_two() nounwind {
 ; CHECK32-LABEL: minus_forty_two:
 ; CHECK32:       # BB#0: # %entry
 ; CHECK32-NEXT:    movb M, %al
-; CHECK32-NEXT:    movl $-42, %ecx
-; CHECK32-NEXT:    addl %ecx, L
+; CHECK32-NEXT:    addl $-42, L
 ; CHECK32-NEXT:    jne .LBB3_2
 ; CHECK32-NEXT:  # BB#1: # %entry
 ; CHECK32-NEXT:    andb $8, %al
@@ -180,8 +177,7 @@ define i1 @minus_forty_two() nounwind {
 ; CHECK64-LABEL: minus_forty_two:
 ; CHECK64:       # BB#0: # %entry
 ; CHECK64-NEXT:    movb {{.*}}(%rip), %al
-; CHECK64-NEXT:    movl $-42, %ecx
-; CHECK64-NEXT:    addl %ecx, {{.*}}(%rip)
+; CHECK64-NEXT:    addl $-42, {{.*}}(%rip)
 ; CHECK64-NEXT:    jne .LBB3_2
 ; CHECK64-NEXT:  # BB#1: # %entry
 ; CHECK64-NEXT:    andb $8, %al
diff --git a/test/CodeGen/X86/pr32659.ll b/test/CodeGen/X86/pr32659.ll
index dd50745da32..b74d99f1015 100644
--- a/test/CodeGen/X86/pr32659.ll
+++ b/test/CodeGen/X86/pr32659.ll
@@ -50,9 +50,7 @@ define void @fn2() nounwind optsize {
 ; CHECK-NEXT:    sarl $31, %eax
 ; CHECK-NEXT:    andl %eax, e+4
 ; CHECK-NEXT:    decl g
-; CHECK-NEXT:    xorl %eax, %eax
-; CHECK-NEXT:    incl %eax
-; CHECK-NEXT:    addl %eax, f
+; CHECK-NEXT:    addl $1, f
 ; CHECK-NEXT:    adcl $0, f+4
 ; CHECK-NEXT:    addl $8, %esp
 ; CHECK-NEXT:    popl %ebx