[DAGCombiner] Push truncate through adde when the carry isn't used.

[deadalnix]

Summary

As per title.

[arsenm]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

7784

Extra whitespace change

7874–7876

Needs a legality check for the resulting adde?

[deadalnix]

Check for adde's validity before combining.

[arsenm]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

7877

Usually this is done as !LegalizeDAG || TLI.isOperationLegal..

7878

SDLoc SL(N)

test/CodeGen/X86/adde-carry.ll

165–166

Can you add tests with more types?

[deadalnix]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

7877

There is no wrapper for this ?

[deadalnix]

Add test cases for various sizes.
Use SDLoc SL(N) .

[RKSimon]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

7875

clang-format

[deadalnix]

Rebase, clang-format

[RKSimon]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

7876

This should probably be (!LegalOperations || TLI.isOperationLegal(ISD::ADDE, VT)) - otherwise you might get into trouble if custom lowering results in the truncate being regenerated.

7879

Typically we don't use auto for these cases - just use SDvalue

[deadalnix]

Use (!LegalOperations || TLI.isOperationLegal(ISD::ADDE, VT))

[arsenm]

LGTM

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
   // Currently we only perform this optimization on scalars because vectors
   // may have different active low bits.
   if (!VT.isVector()) {
     if (SDValue Shorter =
             GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
                                                      VT.getSizeInBits())))
       return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
   }
 
   // fold (truncate (load x)) -> (smaller load x)
   // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
   if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
     if (SDValue Reduced = ReduceLoadWidth(N))
       return Reduced;
 
     // Handle the case where the load remains an extending load even
     // after truncation.
if (N0.getOpcode() == ISD::BITCAST && !VT.isVector()) {
     }
   }
 
   // Simplify the operands using demanded-bits information.
   if (!VT.isVector() &&
       SimplifyDemandedBits(SDValue(N, 0)))
     return SDValue(N, 0);
 
+  // (trunc adde(X, Y, Carry)) -> (adde trunc(X), trunc(Y), Carry)
+  // When the adde's carry is not used.
+  if (N0.getOpcode() == ISD::ADDE && N0.hasOneUse() &&
+      !N0.getNode()->hasAnyUseOfValue(1) &&
+      TLI.isOperationLegalOrCustom(ISD::ADDE, VT)) {
+    SDLoc SL(N);
+    auto X = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(0));
+    auto Y = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
+    return DAG.getNode(ISD::ADDE, SL, DAG.getVTList(VT, MVT::Glue),
+                       X, Y, N0.getOperand(2));
+  }
+
   return SDValue();
 }
 
 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
   SDValue Elt = N->getOperand(i);
   if (Elt.getOpcode() != ISD::MERGE_VALUES)
     return Elt.getNode();
   return Elt.getOperand(Elt.getResNo()).getNode();

test/CodeGen/X86/adde-carry.ll

  store i64 %8, i64* %t, align 8
  ret void
 }
 
 define void @b(i32* nocapture %r, i64 %a, i64 %b, i32 %c) nounwind {
 ; CHECK-LABEL: b:
 ; CHECK:       # BB#0: # %entry
 ; CHECK-NEXT:    addq %rdx, %rsi
-; CHECK-NEXT:    sbbq %rax, %rax
-; CHECK-NEXT:    subl %eax, %ecx
+; CHECK-NEXT:    adcl $0, %ecx
 ; CHECK-NEXT:    movl %ecx, (%rdi)
 ; CHECK-NEXT:    retq
 entry:
  %0 = zext i64 %a to i128
  %1 = zext i64 %b to i128
  %2 = zext i32 %c to i128
  %3 = add i128 %1, %0
  %4 = lshr i128 %3, 64
  %5 = add i128 %4, %2
  %6 = trunc i128 %5 to i32
  store i32 %6, i32* %r, align 4
  ret void
 }
 
 define void @c(i16* nocapture %r, i64 %a, i64 %b, i16 %c) nounwind {
 ; CHECK-LABEL: c:
 ; CHECK:       # BB#0: # %entry
 ; CHECK-NEXT:    addq %rdx, %rsi
-; CHECK-NEXT:    sbbq %rax, %rax
-; CHECK-NEXT:    subl %eax, %ecx
+; CHECK-NEXT:    adcw $0, %cx
 ; CHECK-NEXT:    movw %cx, (%rdi)
 ; CHECK-NEXT:    retq
 entry:
  %0 = zext i64 %a to i128
  %1 = zext i64 %b to i128
  %2 = zext i16 %c to i128
  %3 = add i128 %1, %0
  %4 = lshr i128 %3, 64
  %5 = add i128 %4, %2
  %6 = trunc i128 %5 to i16
  store i16 %6, i16* %r, align 4
  ret void
 }
 
 define void @d(i8* nocapture %r, i64 %a, i64 %b, i8 %c) nounwind {
 ; CHECK-LABEL: d:
 ; CHECK:       # BB#0: # %entry
 ; CHECK-NEXT:    addq %rdx, %rsi
-; CHECK-NEXT:    sbbq %rax, %rax
-; CHECK-NEXT:    subl %eax, %ecx
+; CHECK-NEXT:    adcb $0, %cl
 ; CHECK-NEXT:    movb %cl, (%rdi)
 ; CHECK-NEXT:    retq
 entry:
  %0 = zext i64 %a to i128
  %1 = zext i64 %b to i128
  %2 = zext i8 %c to i128
  %3 = add i128 %1, %0
  %4 = lshr i128 %3, 64

 ; CHECK-LABEL: muladd:
 ; CHECK:       # BB#0: # %entry
 ; CHECK-NEXT:    movq %rdx, %rax
 ; CHECK-NEXT:    mulq %rsi
 ; CHECK-NEXT:    addq (%rdi), %rax
 ; CHECK-NEXT:    adcq $0, %rdx
 ; CHECK-NEXT:    movq %rax, (%rdi)
 ; CHECK-NEXT:    addq 8(%rdi), %rdx
-; CHECK-NEXT:    sbbq %rax, %rax
 ; CHECK-NEXT:    movq %rdx, 8(%rdi)
-; CHECK-NEXT:    subl %eax, 16(%rdi)
+; CHECK-NEXT:    adcl $0, 16(%rdi)
 ; CHECK-NEXT:    retq
 entry:
   %0 = zext i64 %arg.a to i128
   %1 = zext i64 %arg.b to i128
   %2 = mul nuw i128 %1, %0
   %3 = getelementptr inbounds %accumulator, %accumulator* %this, i64 0, i32 0
   %4 = load i64, i64* %3, align 8
   %5 = zext i64 %4 to i128
   %6 = add i128 %5, %2