Use movw+movt in ARMFastISel::ARMMaterializeGV.

This eliminates a lot of constant pool entries for -O0 builds of code
with many global variable accesses.

This speeds up -O0 codegen of consumer-typeset by 2x because the
constant island pass no longer has to look at thousands of constant pool
entries.

<rdar://problem/10629774>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@147712 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 78 insertions, 43 deletions

diff --git a/lib/Target/ARM/ARMFastISel.cpp b/lib/Target/ARM/ARMFastISel.cpp
index 7ce5099..56e3d49 100644
--- a/lib/Target/ARM/ARMFastISel.cpp
+++ b/lib/Target/ARM/ARMFastISel.cpp
@@ -617,40 +617,61 @@ unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
   // TODO: Need more magic for ARM PIC.
   if (!isThumb2 && (RelocM == Reloc::PIC_)) return 0;
 
-  // MachineConstantPool wants an explicit alignment.
-  unsigned Align = TD.getPrefTypeAlignment(GV->getType());
-  if (Align == 0) {
-    // TODO: Figure out if this is correct.
-    Align = TD.getTypeAllocSize(GV->getType());
-  }
-
-  // Grab index.
-  unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
-  unsigned Id = AFI->createPICLabelUId();
-  ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
-                                                              ARMCP::CPValue,
-                                                              PCAdj);
-  unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
-
-  // Load value.
-  MachineInstrBuilder MIB;
   unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
-  if (isThumb2) {
-    unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
-    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
-          .addConstantPoolIndex(Idx);
-    if (RelocM == Reloc::PIC_)
-      MIB.addImm(Id);
+
+  // Use movw+movt when possible, it avoids constant pool entries.
+  if (Subtarget->isTargetDarwin() && Subtarget->useMovt()) {
+    unsigned Opc;
+    switch (RelocM) {
+    case Reloc::PIC_:
+      Opc = isThumb2 ? ARM::t2MOV_ga_pcrel : ARM::MOV_ga_pcrel;
+      break;
+    case Reloc::DynamicNoPIC:
+      Opc = isThumb2 ? ARM::t2MOV_ga_dyn : ARM::MOV_ga_dyn;
+      break;
+    default:
+      Opc = isThumb2 ? ARM::t2MOVi32imm : ARM::MOVi32imm;
+      break;
+    }
+    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
+                            DestReg).addGlobalAddress(GV));
   } else {
-    // The extra immediate is for addrmode2.
-    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
-                  DestReg)
-          .addConstantPoolIndex(Idx)
-          .addImm(0);
+    // MachineConstantPool wants an explicit alignment.
+    unsigned Align = TD.getPrefTypeAlignment(GV->getType());
+    if (Align == 0) {
+      // TODO: Figure out if this is correct.
+      Align = TD.getTypeAllocSize(GV->getType());
+    }
+
+    // Grab index.
+    unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 :
+      (Subtarget->isThumb() ? 4 : 8);
+    unsigned Id = AFI->createPICLabelUId();
+    ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
+                                                                ARMCP::CPValue,
+                                                                PCAdj);
+    unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
+
+    // Load value.
+    MachineInstrBuilder MIB;
+    if (isThumb2) {
+      unsigned Opc = (RelocM!=Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
+      MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
+        .addConstantPoolIndex(Idx);
+      if (RelocM == Reloc::PIC_)
+        MIB.addImm(Id);
+    } else {
+      // The extra immediate is for addrmode2.
+      MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
+                    DestReg)
+        .addConstantPoolIndex(Idx)
+        .addImm(0);
+    }
+    AddOptionalDefs(MIB);
   }
-  AddOptionalDefs(MIB);
 
   if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
+    MachineInstrBuilder MIB;
     unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
     if (isThumb2)
       MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
diff --git a/test/CodeGen/ARM/fast-isel-intrinsic.ll b/test/CodeGen/ARM/fast-isel-intrinsic.ll
index f2ba04f..5bc35ee 100644
--- a/test/CodeGen/ARM/fast-isel-intrinsic.ll
+++ b/test/CodeGen/ARM/fast-isel-intrinsic.ll
@@ -6,14 +6,16 @@
 
 define void @t1() nounwind ssp {
 ; ARM: t1
-; ARM: ldr r0, LCPI0_0
+; ARM: movw r0, :lower16:_message1
+; ARM: movt r0, :upper16:_message1
 ; ARM: add r0, r0, #5
 ; ARM: movw r1, #64
 ; ARM: movw r2, #10
 ; ARM: uxtb r1, r1
 ; ARM: bl _memset
 ; THUMB: t1
-; THUMB: ldr.n r0, LCPI0_0
+; THUMB: movw r0, :lower16:_message1
+; THUMB: movt r0, :upper16:_message1
 ; THUMB: adds r0, #5
 ; THUMB: movs r1, #64
 ; THUMB: movt r1, #0
@@ -29,7 +31,8 @@ declare void @llvm.memset.p0i8.i32(i8* nocapture, i8, i32, i32, i1) nounwind
 
 define void @t2() nounwind ssp {
 ; ARM: t2
-; ARM: ldr r0, LCPI1_0
+; ARM: movw r0, :lower16:L_temp$non_lazy_ptr
+; ARM: movt r0, :upper16:L_temp$non_lazy_ptr
 ; ARM: ldr r0, [r0]
 ; ARM: add r1, r0, #4
 ; ARM: add r0, r0, #16
@@ -39,7 +42,8 @@ define void @t2() nounwind ssp {
 ; ARM: ldr r1, [sp]                @ 4-byte Reload
 ; ARM: bl _memcpy
 ; THUMB: t2
-; THUMB: ldr.n r0, LCPI1_0
+; THUMB: movw r0, :lower16:L_temp$non_lazy_ptr
+; THUMB: movt r0, :upper16:L_temp$non_lazy_ptr
 ; THUMB: ldr r0, [r0]
 ; THUMB: adds r1, r0, #4
 ; THUMB: adds r0, #16
@@ -55,7 +59,8 @@ declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32,
 
 define void @t3() nounwind ssp {
 ; ARM: t3
-; ARM: ldr r0, LCPI2_0
+; ARM: movw r0, :lower16:L_temp$non_lazy_ptr
+; ARM: movt r0, :upper16:L_temp$non_lazy_ptr
 ; ARM: ldr r0, [r0]
 ; ARM: add r1, r0, #4
 ; ARM: add r0, r0, #16
@@ -63,7 +68,8 @@ define void @t3() nounwind ssp {
 ; ARM: mov r0, r1
 ; ARM: bl _memmove
 ; THUMB: t3
-; THUMB: ldr.n r0, LCPI2_0
+; THUMB: movw r0, :lower16:L_temp$non_lazy_ptr
+; THUMB: movt r0, :upper16:L_temp$non_lazy_ptr
 ; THUMB: ldr r0, [r0]
 ; THUMB: adds r1, r0, #4
 ; THUMB: adds r0, #16
@@ -77,9 +83,11 @@ define void @t3() nounwind ssp {
 
 define void @t4() nounwind ssp {
 ; ARM: t4
-; ARM: ldr r0, LCPI3_0
+; ARM: movw r0, :lower16:L_temp$non_lazy_ptr
+; ARM: movt r0, :upper16:L_temp$non_lazy_ptr
 ; ARM: ldr r0, [r0]
-; ARM: ldr r1, LCPI3_1
+; ARM: movw r1, :lower16:L_temp$non_lazy_ptr
+; ARM: movt r1, :upper16:L_temp$non_lazy_ptr
 ; ARM: ldr r1, [r1]
 ; ARM: ldr r2, [r1, #16]
 ; ARM: str r2, [r0, #4]
@@ -88,9 +96,11 @@ define void @t4() nounwind ssp {
 ; ARM: ldrh r1, [r1, #24]
 ; ARM: strh r1, [r0, #12]
 ; ARM: bx lr
-; THUMB: ldr.n r0, LCPI3_0
+; THUMB: movw r0, :lower16:L_temp$non_lazy_ptr
+; THUMB: movt r0, :upper16:L_temp$non_lazy_ptr
 ; THUMB: ldr r0, [r0]
-; THUMB: ldr.n r1, LCPI3_1
+; THUMB: movw r1, :lower16:L_temp$non_lazy_ptr
+; THUMB: movt r1, :upper16:L_temp$non_lazy_ptr
 ; THUMB: ldr r1, [r1]
 ; THUMB: ldr r2, [r1, #16]
 ; THUMB: str r2, [r0, #4]
diff --git a/test/CodeGen/ARM/fast-isel.ll b/test/CodeGen/ARM/fast-isel.ll
index 9f33aea..c8e0211 100644
--- a/test/CodeGen/ARM/fast-isel.ll
+++ b/test/CodeGen/ARM/fast-isel.ll
@@ -142,19 +142,23 @@ define void @test4() {
   store i32 %b, i32* @test4g
   ret void
 
-; THUMB: ldr.n r0, LCPI4_1
+; THUMB: movw r0, :lower16:L_test4g$non_lazy_ptr
+; THUMB: movt r0, :upper16:L_test4g$non_lazy_ptr
 ; THUMB: ldr r0, [r0]
 ; THUMB: ldr r0, [r0]
 ; THUMB: adds r0, #1
-; THUMB: ldr.n r1, LCPI4_0
+; THUMB: movw r1, :lower16:L_test4g$non_lazy_ptr
+; THUMB: movt r1, :upper16:L_test4g$non_lazy_ptr
 ; THUMB: ldr r1, [r1]
 ; THUMB: str r0, [r1]
 
-; ARM: ldr r0, LCPI4_1
+; ARM: movw r0, :lower16:L_test4g$non_lazy_ptr
+; ARM: movt r0, :upper16:L_test4g$non_lazy_ptr
 ; ARM: ldr r0, [r0]
 ; ARM: ldr r0, [r0]
 ; ARM: add r0, r0, #1
-; ARM: ldr r1, LCPI4_0
+; ARM: movw r1, :lower16:L_test4g$non_lazy_ptr
+; ARM: movt r1, :upper16:L_test4g$non_lazy_ptr
 ; ARM: ldr r1, [r1]
 ; ARM: str r0, [r1]
 }