[DAG] visitORCommutative - fold build_pair(not(x),not(y)) -> not(build_pair(x,y)) style patterns

(Sorry, not an actual build_pair node just a similar pattern).

For cases where we're concatenating 2 integers into a double width integer, see if both integer sources are NOT patterns.

We could take this further and handle all logic ops with a constant operands, but I just wanted to handle the case reported on #89533 initially.

Fixes #89533

Author: RKSimon

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -7620,6 +7620,7 @@ SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, const SDLoc &DL) {
 static SDValue visitORCommutative(SelectionDAG &DAG, SDValue N0, SDValue N1,
                                   SDNode *N) {
   EVT VT = N0.getValueType();
+  unsigned BW = VT.getScalarSizeInBits();
   SDLoc DL(N);
 
   auto peekThroughResize = [](SDValue V) {
@@ -7689,6 +7690,26 @@ static SDValue visitORCommutative(SelectionDAG &DAG, SDValue N0, SDValue N1,
       peekThroughZext(N0.getOperand(2)) == peekThroughZext(N1.getOperand(1)))
     return N0;
 
+  // Attempt to match a legalized build_pair-esque pattern:
+  // or(shl(aext(X),BW/2),zext(Y))
+  SDValue Lo, Hi;
+  if (sd_match(N0,
+               m_OneUse(m_Shl(m_AnyExt(m_Value(Hi)), m_SpecificInt(BW / 2)))) &&
+      sd_match(N1, m_ZExt(m_Value(Lo))) &&
+      Lo.getScalarValueSizeInBits() == (BW / 2) &&
+      Lo.getValueType() == Hi.getValueType()) {
+    // Fold build_pair(not(Lo),not(Hi)) -> not(build_pair(Lo,Hi)).
+    SDValue NotLo, NotHi;
+    if (sd_match(Lo, m_OneUse(m_Not(m_Value(NotLo)))) &&
+        sd_match(Hi, m_OneUse(m_Not(m_Value(NotHi))))) {
+      Lo = DAG.getZExtOrTrunc(NotLo, DL, VT);
+      Hi = DAG.getZExtOrTrunc(NotHi, DL, VT);
+      Hi = DAG.getNode(ISD::SHL, DL, VT, Hi,
+                       DAG.getShiftAmountConstant(BW / 2, VT, DL));
+      return DAG.getNOT(DL, DAG.getNode(ISD::OR, DL, VT, Lo, Hi), VT);
+    }
+  }
+
   return SDValue();
 }
 

llvm/test/CodeGen/X86/subvectorwise-store-of-vector-splat.ll

@@ -3060,12 +3060,7 @@ define void @vec384_v3i32(ptr %in.subvec.ptr, ptr %out.subvec.ptr, ptr %out.vec.
 ; SCALAR:       # %bb.0:
 ; SCALAR-NEXT:    movl 8(%rdi), %eax
 ; SCALAR-NEXT:    movq (%rdi), %rcx
-; SCALAR-NEXT:    movq %rcx, %rdi
-; SCALAR-NEXT:    shrq $32, %rdi
-; SCALAR-NEXT:    notl %edi
-; SCALAR-NEXT:    shlq $32, %rdi
-; SCALAR-NEXT:    notl %ecx
-; SCALAR-NEXT:    orq %rdi, %rcx
+; SCALAR-NEXT:    notq %rcx
 ; SCALAR-NEXT:    notl %eax
 ; SCALAR-NEXT:    movl %eax, 8(%rsi)
 ; SCALAR-NEXT:    movq %rcx, (%rsi)
@@ -3196,12 +3191,7 @@ define void @vec384_v3f32(ptr %in.subvec.ptr, ptr %out.subvec.ptr, ptr %out.vec.
 ; SCALAR:       # %bb.0:
 ; SCALAR-NEXT:    movl 8(%rdi), %eax
 ; SCALAR-NEXT:    movq (%rdi), %rcx
-; SCALAR-NEXT:    movq %rcx, %rdi
-; SCALAR-NEXT:    shrq $32, %rdi
-; SCALAR-NEXT:    notl %edi
-; SCALAR-NEXT:    shlq $32, %rdi
-; SCALAR-NEXT:    notl %ecx
-; SCALAR-NEXT:    orq %rdi, %rcx
+; SCALAR-NEXT:    notq %rcx
 ; SCALAR-NEXT:    notl %eax
 ; SCALAR-NEXT:    movl %eax, 8(%rsi)
 ; SCALAR-NEXT:    movq %rcx, (%rsi)
@@ -4216,25 +4206,10 @@ define void @vec384_v6i32(ptr %in.subvec.ptr, ptr %out.subvec.ptr, ptr %out.vec.
 ; SCALAR:       # %bb.0:
 ; SCALAR-NEXT:    movq (%rdi), %rax
 ; SCALAR-NEXT:    movq 8(%rdi), %rcx
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shrq $32, %r8
-; SCALAR-NEXT:    movq %rcx, %r9
-; SCALAR-NEXT:    shrq $32, %r9
 ; SCALAR-NEXT:    movq 16(%rdi), %rdi
-; SCALAR-NEXT:    movq %rdi, %r10
-; SCALAR-NEXT:    shrq $32, %r10
-; SCALAR-NEXT:    notl %r10d
-; SCALAR-NEXT:    shlq $32, %r10
-; SCALAR-NEXT:    notl %edi
-; SCALAR-NEXT:    orq %r10, %rdi
-; SCALAR-NEXT:    notl %r9d
-; SCALAR-NEXT:    shlq $32, %r9
-; SCALAR-NEXT:    notl %ecx
-; SCALAR-NEXT:    orq %r9, %rcx
-; SCALAR-NEXT:    notl %r8d
-; SCALAR-NEXT:    shlq $32, %r8
-; SCALAR-NEXT:    notl %eax
-; SCALAR-NEXT:    orq %r8, %rax
+; SCALAR-NEXT:    notq %rdi
+; SCALAR-NEXT:    notq %rcx
+; SCALAR-NEXT:    notq %rax
 ; SCALAR-NEXT:    movq %rax, (%rsi)
 ; SCALAR-NEXT:    movq %rcx, 8(%rsi)
 ; SCALAR-NEXT:    movq %rdi, 16(%rsi)
@@ -4303,25 +4278,10 @@ define void @vec384_v6f32(ptr %in.subvec.ptr, ptr %out.subvec.ptr, ptr %out.vec.
 ; SCALAR:       # %bb.0:
 ; SCALAR-NEXT:    movq (%rdi), %rax
 ; SCALAR-NEXT:    movq 8(%rdi), %rcx
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shrq $32, %r8
-; SCALAR-NEXT:    movq %rcx, %r9
-; SCALAR-NEXT:    shrq $32, %r9
 ; SCALAR-NEXT:    movq 16(%rdi), %rdi
-; SCALAR-NEXT:    movq %rdi, %r10
-; SCALAR-NEXT:    shrq $32, %r10
-; SCALAR-NEXT:    notl %r10d
-; SCALAR-NEXT:    shlq $32, %r10
-; SCALAR-NEXT:    notl %edi
-; SCALAR-NEXT:    orq %r10, %rdi
-; SCALAR-NEXT:    notl %r9d
-; SCALAR-NEXT:    shlq $32, %r9
-; SCALAR-NEXT:    notl %ecx
-; SCALAR-NEXT:    orq %r9, %rcx
-; SCALAR-NEXT:    notl %r8d
-; SCALAR-NEXT:    shlq $32, %r8
-; SCALAR-NEXT:    notl %eax
-; SCALAR-NEXT:    orq %r8, %rax
+; SCALAR-NEXT:    notq %rdi
+; SCALAR-NEXT:    notq %rcx
+; SCALAR-NEXT:    notq %rax
 ; SCALAR-NEXT:    movq %rax, (%rsi)
 ; SCALAR-NEXT:    movq %rcx, 8(%rsi)
 ; SCALAR-NEXT:    movq %rdi, 16(%rsi)