[PowerPC] [Constant Hoisting] Enable constant hoisting on PPC

Implements the various TTI functions to enable constant hoisting on PPC. The
only significant test-suite change is this:

MultiSource/Benchmarks/VersaBench/bmm/bmm - 20% speedup
(which essentially reverses the slowdown from r206120).

llvm-svn: 206141

lib/Target/PowerPC/PPCTargetTransformInfo.cpp

@@ -18,11 +18,15 @@
 #include "PPC.h"
 #include "PPCTargetMachine.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Target/CostTable.h"
 #include "llvm/Target/TargetLowering.h"
 using namespace llvm;
 
+static cl::opt<bool> DisablePPCConstHoist("disable-ppc-constant-hoisting",
+cl::desc("disable constant hoisting on PPC"), cl::init(false), cl::Hidden);
+
 // Declare the pass initialization routine locally as target-specific passes
 // don't havve a target-wide initialization entry point, and so we rely on the
 // pass constructor initialization.
@@ -67,6 +71,13 @@ public:
 
   /// \name Scalar TTI Implementations
   /// @{
+  unsigned getIntImmCost(const APInt &Imm, Type *Ty) const override;
+
+  unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
+                         Type *Ty) const override;
+  unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
+                         Type *Ty) const override;
+
   virtual PopcntSupportKind
   getPopcntSupport(unsigned TyWidth) const override;
   virtual void getUnrollingPreferences(
@@ -123,6 +134,142 @@ PPCTTI::PopcntSupportKind PPCTTI::getPopcntSupport(unsigned TyWidth) const {
   return PSK_Software;
 }
 
+unsigned PPCTTI::getIntImmCost(const APInt &Imm, Type *Ty) const {
+  if (DisablePPCConstHoist)
+    return TargetTransformInfo::getIntImmCost(Imm, Ty);
+
+  assert(Ty->isIntegerTy());
+
+  unsigned BitSize = Ty->getPrimitiveSizeInBits();
+  if (BitSize == 0)
+    return ~0U;
+
+  if (Imm == 0)
+    return TCC_Free;
+
+  if (Imm.getBitWidth() <= 64) {
+    if (isInt<16>(Imm.getSExtValue()))
+      return TCC_Basic;
+
+    if (isInt<32>(Imm.getSExtValue())) {
+      // A constant that can be materialized using lis.
+      if ((Imm.getZExtValue() & 0xFFFF) == 0)
+        return TCC_Basic;
+
+      return 2 * TCC_Basic;
+    }
+  }
+
+  return 4 * TCC_Basic;
+}
+
+unsigned PPCTTI::getIntImmCost(Intrinsic::ID IID, unsigned Idx,
+                               const APInt &Imm, Type *Ty) const {
+  if (DisablePPCConstHoist)
+    return TargetTransformInfo::getIntImmCost(IID, Idx, Imm, Ty);
+
+  assert(Ty->isIntegerTy());
+
+  unsigned BitSize = Ty->getPrimitiveSizeInBits();
+  if (BitSize == 0)
+    return ~0U;
+
+  switch (IID) {
+  default: return TCC_Free;
+  case Intrinsic::sadd_with_overflow:
+  case Intrinsic::uadd_with_overflow:
+  case Intrinsic::ssub_with_overflow:
+  case Intrinsic::usub_with_overflow:
+    if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<16>(Imm.getSExtValue()))
+      return TCC_Free;
+    break;
+  }
+  return PPCTTI::getIntImmCost(Imm, Ty);
+}
+
+unsigned PPCTTI::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
+                               Type *Ty) const {
+  if (DisablePPCConstHoist)
+    return TargetTransformInfo::getIntImmCost(Opcode, Idx, Imm, Ty);
+
+  assert(Ty->isIntegerTy());
+
+  unsigned BitSize = Ty->getPrimitiveSizeInBits();
+  if (BitSize == 0)
+    return ~0U;
+
+  unsigned ImmIdx = ~0U;
+  bool ShiftedFree = false, RunFree = false, UnsignedFree = false,
+       ZeroFree = false;
+  switch (Opcode) {
+  default: return TCC_Free;
+  case Instruction::GetElementPtr:
+    // Always hoist the base address of a GetElementPtr. This prevents the
+    // creation of new constants for every base constant that gets constant
+    // folded with the offset.
+    if (Idx == 0)
+      return 2 * TCC_Basic;
+    return TCC_Free;
+  case Instruction::And:
+    RunFree = true; // (for the rotate-and-mask instructions)
+    // Fallthrough...
+  case Instruction::Add:
+  case Instruction::Or:
+  case Instruction::Xor:
+    ShiftedFree = true;
+    // Fallthrough...
+  case Instruction::Sub:
+  case Instruction::Mul:
+  case Instruction::Shl:
+  case Instruction::LShr:
+  case Instruction::AShr:
+    ImmIdx = 1;
+    break;
+  case Instruction::ICmp:
+    UnsignedFree = true;
+    ImmIdx = 1;
+    // Fallthrough... (zero comparisons can use record-form instructions)
+  case Instruction::Select:
+    ZeroFree = true;
+    break;
+  case Instruction::PHI:
+  case Instruction::Call:
+  case Instruction::Ret:
+  case Instruction::Load:
+  case Instruction::Store:
+    break;
+  }
+
+  if (ZeroFree && Imm == 0)
+    return TCC_Free;
+
+  if (Idx == ImmIdx && Imm.getBitWidth() <= 64) {
+    if (isInt<16>(Imm.getSExtValue()))
+      return TCC_Free;
+
+    if (RunFree) {
+      if (Imm.getBitWidth() <= 32 &&
+          (isShiftedMask_32(Imm.getZExtValue()) ||
+           isShiftedMask_32(~Imm.getZExtValue())))
+        return TCC_Free;
+
+
+      if (ST->isPPC64() &&
+          (isShiftedMask_64(Imm.getZExtValue()) ||
+           isShiftedMask_64(~Imm.getZExtValue())))
+        return TCC_Free;
+    }
+
+    if (UnsignedFree && isUInt<16>(Imm.getZExtValue()))
+      return TCC_Free;
+
+    if (ShiftedFree && (Imm.getZExtValue() & 0xFFFF) == 0)
+      return TCC_Free;
+  }
+
+  return PPCTTI::getIntImmCost(Imm, Ty);
+}
+
 void PPCTTI::getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const {
   if (ST->getDarwinDirective() == PPC::DIR_A2) {
     // The A2 is in-order with a deep pipeline, and concatenation unrolling

test/Transforms/ConstantHoisting/PowerPC/const-base-addr.ll

@@ -0,0 +1,23 @@
+; RUN: opt -S -consthoist < %s | FileCheck %s
+target datalayout = "E-m:e-i64:64-n32:64"
+target triple = "powerpc64-unknown-linux-gnu"
+
+%T = type { i32, i32, i32, i32 }
+
+; Test if even cheap base addresses are hoisted.
+define i32 @test1() nounwind {
+; CHECK-LABEL:  @test1
+; CHECK:        %const = bitcast i32 12345678 to i32
+; CHECK:        %1 = inttoptr i32 %const to %T*
+; CHECK:        %addr1 = getelementptr %T* %1, i32 0, i32 1
+  %addr1 = getelementptr %T* inttoptr (i32 12345678 to %T*), i32 0, i32 1
+  %tmp1 = load i32* %addr1
+  %addr2 = getelementptr %T* inttoptr (i32 12345678 to %T*), i32 0, i32 2
+  %tmp2 = load i32* %addr2
+  %addr3 = getelementptr %T* inttoptr (i32 12345678 to %T*), i32 0, i32 3
+  %tmp3 = load i32* %addr3
+  %tmp4 = add i32 %tmp1, %tmp2
+  %tmp5 = add i32 %tmp3, %tmp4
+  ret i32 %tmp5
+}
+

test/Transforms/ConstantHoisting/PowerPC/lit.local.cfg

@@ -0,0 +1,4 @@
+targets = set(config.root.targets_to_build.split())
+if not 'PowerPC' in targets:
+    config.unsupported = True
+

test/Transforms/ConstantHoisting/PowerPC/masks.ll

@@ -0,0 +1,66 @@
+; RUN: opt -S -consthoist < %s | FileCheck %s
+target datalayout = "E-m:e-i64:64-n32:64"
+target triple = "powerpc64-unknown-linux-gnu"
+
+; Here the masks are all contiguous, and should not be hoisted.
+define i32 @test1() nounwind {
+entry:
+; CHECK-LABEL:  @test1
+; CHECK-NOT: bitcast i32 65535 to i32
+; CHECK: and i32 undef, 65535
+  %conv121 = and i32 undef, 65535
+  br i1 undef, label %if.then152, label %if.end167
+
+if.then152:
+; CHECK: and i32 undef, 65535
+  %conv153 = and i32 undef, 65535
+  br i1 undef, label %if.end167, label %end2
+
+if.end167:
+; CHECK: and i32 {{.*}}, 32768
+  %shl161 = shl nuw nsw i32 %conv121, 15
+  %0 = load i8* undef, align 1
+  %conv169 = zext i8 %0 to i32
+  %shl170 = shl nuw nsw i32 %conv169, 7
+  %shl161.masked = and i32 %shl161, 32768
+  %conv174 = or i32 %shl170, %shl161.masked
+  %cmp178 = icmp ugt i32 %conv174, 32767
+  br i1 %cmp178, label %end1, label %end2
+
+end1:
+  unreachable
+
+end2:
+  unreachable
+}
+
+; Here the masks are not contiguous, and should be hoisted.
+define i32 @test2() nounwind {
+entry:
+; CHECK-LABEL: @test2
+; CHECK: bitcast i32 65531 to i32
+  %conv121 = and i32 undef, 65531
+  br i1 undef, label %if.then152, label %if.end167
+
+if.then152:
+  %conv153 = and i32 undef, 65531
+  br i1 undef, label %if.end167, label %end2
+
+if.end167:
+; CHECK: add i32 {{.*}}, -32758
+  %shl161 = shl nuw nsw i32 %conv121, 15
+  %0 = load i8* undef, align 1
+  %conv169 = zext i8 %0 to i32
+  %shl170 = shl nuw nsw i32 %conv169, 7
+  %shl161.masked = and i32 %shl161, 32773
+  %conv174 = or i32 %shl170, %shl161.masked
+  %cmp178 = icmp ugt i32 %conv174, 32767
+  br i1 %cmp178, label %end1, label %end2
+
+end1:
+  unreachable
+
+end2:
+  unreachable
+}
+