[FastISel][AArch64] Fix the immediate versions of the {s|u}{add|sub}.with.overflow...

ADDS and SUBS cannot encode negative immediates or immediates larger than 12bit.
This fix checks if the immediate version can be used under this constraints and
if we can convert ADDS to SUBS or vice versa to support negative immediates.

Also update the test cases to test the immediate versions.

llvm-svn: 214470

diff --git a/llvm/lib/Target/AArch64/AArch64FastISel.cpp b/llvm/lib/Target/AArch64/AArch64FastISel.cpp
index 8d7bddb..bfbf10a 100644 (file)
--- a/llvm/lib/Target/AArch64/AArch64FastISel.cpp
+++ b/llvm/lib/Target/AArch64/AArch64FastISel.cpp
@@ -1759,59 +1759,53 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) {
       return false;
     bool LHSIsKill = hasTrivialKill(LHS);
 
-    unsigned RHSReg = 0;
-    bool RHSIsKill = false;
-    bool UseImm = true;
-    if (!isa<ConstantInt>(RHS)) {
-      RHSReg = getRegForValue(RHS);
-      if (!RHSReg)
-        return false;
-      RHSIsKill = hasTrivialKill(RHS);
-      UseImm = false;
+    // Check if the immediate can be encoded in the instruction and if we should
+    // invert the instruction (adds -> subs) to handle negative immediates.
+    bool UseImm = false;
+    bool UseInverse = false;
+    uint64_t Imm = 0;
+    if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
+      if (C->isNegative()) {
+        UseInverse = true;
+        Imm = -(C->getSExtValue());
+      } else
+        Imm = C->getZExtValue();
+
+      if (isUInt<12>(Imm))
+        UseImm = true;
+
+      UseInverse = UseImm && UseInverse;
     }
 
+    static const unsigned OpcTable[2][2][2] = {
+      { {AArch64::ADDSWrr, AArch64::ADDSXrr},
+        {AArch64::ADDSWri, AArch64::ADDSXri} },
+      { {AArch64::SUBSWrr, AArch64::SUBSXrr},
+        {AArch64::SUBSWri, AArch64::SUBSXri} }
+    };
     unsigned Opc = 0;
     unsigned MulReg = 0;
+    unsigned RHSReg = 0;
+    bool RHSIsKill = false;
     AArch64CC::CondCode CC = AArch64CC::Invalid;
     bool Is64Bit = VT == MVT::i64;
     switch (II->getIntrinsicID()) {
     default: llvm_unreachable("Unexpected intrinsic!");
     case Intrinsic::sadd_with_overflow:
-      if (UseImm)
-        Opc = Is64Bit ? AArch64::ADDSXri : AArch64::ADDSWri;
-      else
-        Opc = Is64Bit ? AArch64::ADDSXrr : AArch64::ADDSWrr;
-      CC = AArch64CC::VS;
-      break;
+      Opc = OpcTable[UseInverse][UseImm][Is64Bit]; CC = AArch64CC::VS; break;
     case Intrinsic::uadd_with_overflow:
-      if (UseImm)
-        Opc = Is64Bit ? AArch64::ADDSXri : AArch64::ADDSWri;
-      else
-        Opc = Is64Bit ? AArch64::ADDSXrr : AArch64::ADDSWrr;
-      CC = AArch64CC::HS;
-      break;
+      Opc = OpcTable[UseInverse][UseImm][Is64Bit]; CC = AArch64CC::HS; break;
     case Intrinsic::ssub_with_overflow:
-      if (UseImm)
-        Opc = Is64Bit ? AArch64::SUBSXri : AArch64::SUBSWri;
-      else
-        Opc = Is64Bit ? AArch64::SUBSXrr : AArch64::SUBSWrr;
-      CC = AArch64CC::VS;
-      break;
+      Opc = OpcTable[!UseInverse][UseImm][Is64Bit]; CC = AArch64CC::VS; break;
     case Intrinsic::usub_with_overflow:
-      if (UseImm)
-        Opc = Is64Bit ? AArch64::SUBSXri : AArch64::SUBSWri;
-      else
-        Opc = Is64Bit ? AArch64::SUBSXrr : AArch64::SUBSWrr;
-      CC = AArch64CC::LO;
-      break;
+      Opc = OpcTable[!UseInverse][UseImm][Is64Bit]; CC = AArch64CC::LO; break;
     case Intrinsic::smul_with_overflow: {
       CC = AArch64CC::NE;
-      if (UseImm) {
-        RHSReg = getRegForValue(RHS);
-        if (!RHSReg)
-          return false;
-        RHSIsKill = hasTrivialKill(RHS);
-      }
+      RHSReg = getRegForValue(RHS);
+      if (!RHSReg)
+        return false;
+      RHSIsKill = hasTrivialKill(RHS);
+
       if (VT == MVT::i32) {
         MulReg = Emit_SMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
         unsigned ShiftReg = Emit_LSR_ri(MVT::i64, MulReg, false, 32);
@@ -1841,12 +1835,11 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) {
     }
     case Intrinsic::umul_with_overflow: {
       CC = AArch64CC::NE;
-      if (UseImm) {
-        RHSReg = getRegForValue(RHS);
-        if (!RHSReg)
-          return false;
-        RHSIsKill = hasTrivialKill(RHS);
-      }
+      RHSReg = getRegForValue(RHS);
+      if (!RHSReg)
+        return false;
+      RHSIsKill = hasTrivialKill(RHS);
+
       if (VT == MVT::i32) {
         MulReg = Emit_UMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
         unsigned CmpReg = createResultReg(TLI.getRegClassFor(MVT::i64));
@@ -1872,15 +1865,23 @@ bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) {
     }
     }
 
+    if (!UseImm) {
+      RHSReg = getRegForValue(RHS);
+      if (!RHSReg)
+        return false;
+      RHSIsKill = hasTrivialKill(RHS);
+    }
+
     unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
     if (Opc) {
       MachineInstrBuilder MIB;
       MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc),
                     ResultReg)
               .addReg(LHSReg, getKillRegState(LHSIsKill));
-      if (UseImm)
-        MIB.addImm(cast<ConstantInt>(RHS)->getZExtValue());
-      else
+      if (UseImm) {
+        MIB.addImm(Imm);
+        MIB.addImm(0);
+      } else
         MIB.addReg(RHSReg, getKillRegState(RHSIsKill));
     }
     else

diff --git a/llvm/test/CodeGen/AArch64/arm64-xaluo.ll b/llvm/test/CodeGen/AArch64/arm64-xaluo.ll
index efc63db..fe81d8d 100644 (file)
--- a/llvm/test/CodeGen/AArch64/arm64-xaluo.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-xaluo.ll
@@ -4,9 +4,9 @@
 ;
 ; Get the actual value of the overflow bit.
 ;
-define zeroext i1 @saddo.i32(i32 %v1, i32 %v2, i32* %res) {
+define zeroext i1 @saddo1.i32(i32 %v1, i32 %v2, i32* %res) {
 entry:
-; CHECK-LABEL:  saddo.i32
+; CHECK-LABEL:  saddo1.i32
 ; CHECK:        adds {{w[0-9]+}}, w0, w1
 ; CHECK-NEXT:   cset {{w[0-9]+}}, vs
   %t = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
@@ -16,9 +16,48 @@ entry:
   ret i1 %obit
 }
 
-define zeroext i1 @saddo.i64(i64 %v1, i64 %v2, i64* %res) {
+; Test the immediate version.
+define zeroext i1 @saddo2.i32(i32 %v1, i32* %res) {
 entry:
-; CHECK-LABEL:  saddo.i64
+; CHECK-LABEL:  saddo2.i32
+; CHECK:        adds {{w[0-9]+}}, w0, #4
+; CHECK-NEXT:   cset {{w[0-9]+}}, vs
+  %t = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %v1, i32 4)
+  %val = extractvalue {i32, i1} %t, 0
+  %obit = extractvalue {i32, i1} %t, 1
+  store i32 %val, i32* %res
+  ret i1 %obit
+}
+
+; Test negative immediates.
+define zeroext i1 @saddo3.i32(i32 %v1, i32* %res) {
+entry:
+; CHECK-LABEL:  saddo3.i32
+; CHECK:        subs {{w[0-9]+}}, w0, #4
+; CHECK-NEXT:   cset {{w[0-9]+}}, vs
+  %t = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %v1, i32 -4)
+  %val = extractvalue {i32, i1} %t, 0
+  %obit = extractvalue {i32, i1} %t, 1
+  store i32 %val, i32* %res
+  ret i1 %obit
+}
+
+; Test immediates that are too large to be encoded.
+define zeroext i1 @saddo4.i32(i32 %v1, i32* %res) {
+entry:
+; CHECK-LABEL:  saddo4.i32
+; CHECK:        adds {{w[0-9]+}}, w0, {{w[0-9]+}}
+; CHECK-NEXT:   cset {{w[0-9]+}}, vs
+  %t = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %v1, i32 16777215)
+  %val = extractvalue {i32, i1} %t, 0
+  %obit = extractvalue {i32, i1} %t, 1
+  store i32 %val, i32* %res
+  ret i1 %obit
+}
+
+define zeroext i1 @saddo1.i64(i64 %v1, i64 %v2, i64* %res) {
+entry:
+; CHECK-LABEL:  saddo1.i64
 ; CHECK:        adds {{x[0-9]+}}, x0, x1
 ; CHECK-NEXT:   cset {{w[0-9]+}}, vs
   %t = call {i64, i1} @llvm.sadd.with.overflow.i64(i64 %v1, i64 %v2)
@@ -28,6 +67,30 @@ entry:
   ret i1 %obit
 }
 
+define zeroext i1 @saddo2.i64(i64 %v1, i64* %res) {
+entry:
+; CHECK-LABEL:  saddo2.i64
+; CHECK:        adds {{x[0-9]+}}, x0, #4
+; CHECK-NEXT:   cset {{w[0-9]+}}, vs
+  %t = call {i64, i1} @llvm.sadd.with.overflow.i64(i64 %v1, i64 4)
+  %val = extractvalue {i64, i1} %t, 0
+  %obit = extractvalue {i64, i1} %t, 1
+  store i64 %val, i64* %res
+  ret i1 %obit
+}
+
+define zeroext i1 @saddo3.i64(i64 %v1, i64* %res) {
+entry:
+; CHECK-LABEL:  saddo3.i64
+; CHECK:        subs {{x[0-9]+}}, x0, #4
+; CHECK-NEXT:   cset {{w[0-9]+}}, vs
+  %t = call {i64, i1} @llvm.sadd.with.overflow.i64(i64 %v1, i64 -4)
+  %val = extractvalue {i64, i1} %t, 0
+  %obit = extractvalue {i64, i1} %t, 1
+  store i64 %val, i64* %res
+  ret i1 %obit
+}
+
 define zeroext i1 @uaddo.i32(i32 %v1, i32 %v2, i32* %res) {
 entry:
 ; CHECK-LABEL:  uaddo.i32
@@ -52,9 +115,9 @@ entry:
   ret i1 %obit
 }
 
-define zeroext i1 @ssubo.i32(i32 %v1, i32 %v2, i32* %res) {
+define zeroext i1 @ssubo1.i32(i32 %v1, i32 %v2, i32* %res) {
 entry:
-; CHECK-LABEL:  ssubo.i32
+; CHECK-LABEL:  ssubo1.i32
 ; CHECK:        subs {{w[0-9]+}}, w0, w1
 ; CHECK-NEXT:   cset {{w[0-9]+}}, vs
   %t = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %v1, i32 %v2)
@@ -64,6 +127,18 @@ entry:
   ret i1 %obit
 }
 
+define zeroext i1 @ssubo2.i32(i32 %v1, i32* %res) {
+entry:
+; CHECK-LABEL:  ssubo2.i32
+; CHECK:        adds {{w[0-9]+}}, w0, #4
+; CHECK-NEXT:   cset {{w[0-9]+}}, vs
+  %t = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %v1, i32 -4)
+  %val = extractvalue {i32, i1} %t, 0
+  %obit = extractvalue {i32, i1} %t, 1
+  store i32 %val, i32* %res
+  ret i1 %obit
+}
+
 define zeroext i1 @ssubo.i64(i64 %v1, i64 %v2, i64* %res) {
 entry:
 ; CHECK-LABEL:  ssubo.i64