Add some DAGCombines for (adde 0, 0, glue), which are useful to optimize legalized code for large integer arithmetic.

1. Inform users of ADDEs with two 0 operands that it never sets carry
2. Fold other ADDs or ADDCs into the ADDE if possible

It would be neat if we could do the same thing for SETCC+ADD eventually, but we can't do that in target independent code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@126557 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -1290,6 +1290,16 @@ SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
   return SDValue();
 }
 
+/// isCarryMaterialization - Returns true if V is an ADDE node that is known to
+/// return 0 or 1 depending on the carry flag.
+static bool isCarryMaterialization(SDValue V) {
+  if (V.getOpcode() != ISD::ADDE)
+    return false;
+
+  ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(0));
+  return C && C->isNullValue() && V.getOperand(0) == V.getOperand(1);
+}
+
 SDValue DAGCombiner::visitADD(SDNode *N) {
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
@@ -1453,6 +1463,18 @@ SDValue DAGCombiner::visitADD(SDNode *N) {
     return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
   }
 
+  // add (adde 0, 0, glue), X -> adde X, 0, glue
+  if (N0->hasOneUse() && isCarryMaterialization(N0))
+    return DAG.getNode(ISD::ADDE, N->getDebugLoc(),
+                       DAG.getVTList(VT, MVT::Glue), N1, N0.getOperand(0),
+                       N0.getOperand(2));
+
+  // add X, (adde 0, 0, glue) -> adde X, 0, glue
+  if (N1->hasOneUse() && isCarryMaterialization(N1))
+    return DAG.getNode(ISD::ADDE, N->getDebugLoc(),
+                       DAG.getVTList(VT, MVT::Glue), N0, N1.getOperand(0),
+                       N1.getOperand(2));
+
   return SDValue();
 }
 
@@ -1496,6 +1518,16 @@ SDValue DAGCombiner::visitADDC(SDNode *N) {
                                    N->getDebugLoc(), MVT::Glue));
   }
 
+  // addc (adde 0, 0, glue), X -> adde X, 0, glue
+  if (N0->hasOneUse() && isCarryMaterialization(N0))
+    return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(), N1,
+                       DAG.getConstant(0, VT), N0.getOperand(2));
+
+  // addc X, (adde 0, 0, glue) -> adde X, 0, glue
+  if (N1->hasOneUse() && isCarryMaterialization(N1))
+    return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(), N0,
+                       DAG.getConstant(0, VT), N1.getOperand(2));
+
   return SDValue();
 }
 
@@ -1506,6 +1538,12 @@ SDValue DAGCombiner::visitADDE(SDNode *N) {
   ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
   ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
 
+  // If both operands are null we know that carry out will always be false.
+  if (N0C && N0C->isNullValue() && N0 == N1)
+    DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), DAG.getNode(ISD::CARRY_FALSE,
+                                                             N->getDebugLoc(),
+                                                             MVT::Glue));
+
   // canonicalize constant to RHS
   if (N0C && !N1C)
     return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),

lib/Target/README.txt

@@ -1780,43 +1780,6 @@ case it choses instead to keep the max operation obvious.
 
 //===---------------------------------------------------------------------===//
 
-Take the following testcase on x86-64 (similar testcases exist for all targets
-with addc/adde):
-
-define void @a(i64* nocapture %s, i64* nocapture %t, i64 %a, i64 %b,
-i64 %c) nounwind {
-entry:
- %0 = zext i64 %a to i128                        ; <i128> [#uses=1]
- %1 = zext i64 %b to i128                        ; <i128> [#uses=1]
- %2 = add i128 %1, %0                            ; <i128> [#uses=2]
- %3 = zext i64 %c to i128                        ; <i128> [#uses=1]
- %4 = shl i128 %3, 64                            ; <i128> [#uses=1]
- %5 = add i128 %4, %2                            ; <i128> [#uses=1]
- %6 = lshr i128 %5, 64                           ; <i128> [#uses=1]
- %7 = trunc i128 %6 to i64                       ; <i64> [#uses=1]
- store i64 %7, i64* %s, align 8
- %8 = trunc i128 %2 to i64                       ; <i64> [#uses=1]
- store i64 %8, i64* %t, align 8
- ret void
-}
-
-Generated code:
-        addq	%rcx, %rdx
-        sbbq	%rax, %rax
-        subq	%rax, %r8
-        movq	%r8, (%rdi)
-        movq	%rdx, (%rsi)
-        ret
-
-Expected code:
-       addq    %rcx, %rdx
-       adcq    $0, %r8
-       movq    %r8, (%rdi)
-       movq    %rdx, (%rsi)
-       ret
-
-//===---------------------------------------------------------------------===//
-
 Switch lowering generates less than ideal code for the following switch:
 define void @a(i32 %x) nounwind {
 entry:

test/CodeGen/X86/adde-carry.ll

@@ -0,0 +1,26 @@
+; RUN: llc -march=x86-64 < %s | FileCheck %s -check-prefix=CHECK-64
+; RUN: llc -march=x86 < %s | FileCheck %s -check-prefix=CHECK-32
+
+define void @a(i64* nocapture %s, i64* nocapture %t, i64 %a, i64 %b, i64 %c) nounwind {
+entry:
+ %0 = zext i64 %a to i128
+ %1 = zext i64 %b to i128
+ %2 = add i128 %1, %0
+ %3 = zext i64 %c to i128
+ %4 = shl i128 %3, 64
+ %5 = add i128 %4, %2
+ %6 = lshr i128 %5, 64
+ %7 = trunc i128 %6 to i64
+ store i64 %7, i64* %s, align 8
+ %8 = trunc i128 %2 to i64
+ store i64 %8, i64* %t, align 8
+ ret void
+
+; CHECK-32: addl
+; CHECK-32: adcl
+; CHECK-32: adcl $0
+; CHECK-32: adcl $0
+
+; CHECK-64: addq
+; CHECK-64: adcq $0
+}