[FastISel][AArch64] Fold 'AND' instruction during the address computation.

The 'AND' instruction could be used to mask out the lower 32 bits of a register.
If this is done inside an address computation we might be able to fold the
instruction into the memory instruction itself.

and  x1, x1, #0xffffffff   ---> ldrb x0, [x0, w1, uxtw]
ldrb x0, [x0, x1]

llvm-svn: 218030

2 files changed, 99 insertions, 2 deletions

diff --git a/llvm/lib/Target/AArch64/AArch64FastISel.cpp b/llvm/lib/Target/AArch64/AArch64FastISel.cpp
index d25956c7bc4..0e9ef990962 100644
--- a/llvm/lib/Target/AArch64/AArch64FastISel.cpp
+++ b/llvm/lib/Target/AArch64/AArch64FastISel.cpp
@@ -596,6 +596,29 @@ bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty)
         if (SE->getOperand(0)->getType()->isIntegerTy(32))
           Addr.setExtendType(AArch64_AM::SXTW);
 
+      if (const auto *AI = dyn_cast<BinaryOperator>(U))
+        if (AI->getOpcode() == Instruction::And) {
+          const Value *LHS = AI->getOperand(0);
+          const Value *RHS = AI->getOperand(1);
+
+          if (const auto *C = dyn_cast<ConstantInt>(LHS))
+            if (C->getValue() == 0xffffffff)
+              std::swap(LHS, RHS);
+
+          if (const auto *C = cast<ConstantInt>(RHS))
+            if (C->getValue() == 0xffffffff) {
+              Addr.setExtendType(AArch64_AM::UXTW);
+              unsigned Reg = getRegForValue(LHS);
+              if (!Reg)
+                return false;
+              bool RegIsKill = hasTrivialKill(LHS);
+              Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
+                                               AArch64::sub_32);
+              Addr.setOffsetReg(Reg);
+              return true;
+            }
+        }
+
       unsigned Reg = getRegForValue(U->getOperand(0));
       if (!Reg)
         return false;
@@ -660,6 +683,37 @@ bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty)
     Addr.setOffsetReg(Reg);
     return true;
   }
+  case Instruction::And: {
+    if (Addr.getOffsetReg())
+      break;
+
+    if (DL.getTypeSizeInBits(Ty) != 8)
+      break;
+
+    const Value *LHS = U->getOperand(0);
+    const Value *RHS = U->getOperand(1);
+
+    if (const auto *C = dyn_cast<ConstantInt>(LHS))
+      if (C->getValue() == 0xffffffff)
+        std::swap(LHS, RHS);
+
+    if (const auto *C = cast<ConstantInt>(RHS))
+      if (C->getValue() == 0xffffffff) {
+        Addr.setShift(0);
+        Addr.setExtendType(AArch64_AM::LSL);
+        Addr.setExtendType(AArch64_AM::UXTW);
+
+        unsigned Reg = getRegForValue(LHS);
+        if (!Reg)
+          return false;
+        bool RegIsKill = hasTrivialKill(LHS);
+        Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
+                                         AArch64::sub_32);
+        Addr.setOffsetReg(Reg);
+        return true;
+      }
+    break;
+  }
   } // end switch
 
   if (Addr.getReg()) {
diff --git a/llvm/test/CodeGen/AArch64/fast-isel-addressing-modes.ll b/llvm/test/CodeGen/AArch64/fast-isel-addressing-modes.ll
index 21fc6645009..b33c06ee278 100644
--- a/llvm/test/CodeGen/AArch64/fast-isel-addressing-modes.ll
+++ b/llvm/test/CodeGen/AArch64/fast-isel-addressing-modes.ll
@@ -1,5 +1,5 @@
-; RUN: llc -mtriple=aarch64-apple-darwin                      -verify-machineinstrs < %s | FileCheck %s --check-prefix=CHECK --check-prefix=SDAG
-; RUN: llc -mtriple=aarch64-apple-darwin -O0 -fast-isel-abort -verify-machineinstrs < %s | FileCheck %s --check-prefix=CHECK --check-prefix=FAST
+; RUN: llc -mtriple=aarch64-apple-darwin                             -verify-machineinstrs < %s | FileCheck %s --check-prefix=CHECK --check-prefix=SDAG
+; RUN: llc -mtriple=aarch64-apple-darwin -fast-isel -fast-isel-abort -verify-machineinstrs < %s | FileCheck %s --check-prefix=CHECK --check-prefix=FAST
 
 ; Load / Store Base Register only
 define zeroext i1 @load_breg_i1(i1* %a) {
@@ -425,6 +425,49 @@ define i32 @load_breg_mul_offreg_1(i64 %a, i64 %b) {
   ret i32 %4
 }
 
+define zeroext i8 @load_breg_and_offreg_1(i64 %a, i64 %b) {
+; CHECK-LABEL: load_breg_and_offreg_1
+; CHECK:       ldrb {{w[0-9]+}}, [x1, w0, uxtw]
+  %1 = and i64 %a, 4294967295
+  %2 = add i64 %1, %b
+  %3 = inttoptr i64 %2 to i8*
+  %4 = load i8* %3
+  ret i8 %4
+}
+
+define zeroext i16 @load_breg_and_offreg_2(i64 %a, i64 %b) {
+; CHECK-LABEL: load_breg_and_offreg_2
+; CHECK:       ldrh {{w[0-9]+}}, [x1, w0, uxtw #1]
+  %1 = and i64 %a, 4294967295
+  %2 = shl i64 %1, 1
+  %3 = add i64 %2, %b
+  %4 = inttoptr i64 %3 to i16*
+  %5 = load i16* %4
+  ret i16 %5
+}
+
+define i32 @load_breg_and_offreg_3(i64 %a, i64 %b) {
+; CHECK-LABEL: load_breg_and_offreg_3
+; CHECK:       ldr {{w[0-9]+}}, [x1, w0, uxtw #2]
+  %1 = and i64 %a, 4294967295
+  %2 = shl i64 %1, 2
+  %3 = add i64 %2, %b
+  %4 = inttoptr i64 %3 to i32*
+  %5 = load i32* %4
+  ret i32 %5
+}
+
+define i64 @load_breg_and_offreg_4(i64 %a, i64 %b) {
+; CHECK-LABEL: load_breg_and_offreg_4
+; CHECK:       ldr {{x[0-9]+}}, [x1, w0, uxtw #3]
+  %1 = and i64 %a, 4294967295
+  %2 = shl i64 %1, 3
+  %3 = add i64 %2, %b
+  %4 = inttoptr i64 %3 to i64*
+  %5 = load i64* %4
+  ret i64 %5
+}
+
 ; Load Base Register + Scaled Register Offset + Sign/Zero extension
 define i32 @load_breg_zext_shift_offreg_1(i32 %a, i64 %b) {
 ; CHECK-LABEL: load_breg_zext_shift_offreg_1