[PowerPC] Exploit vabsd for absolute difference of vectors on P9

[Esme]

Summary

Exploit vabsd for some cases on power9, for example:

void foo (char *restrict p, char *restrict q, char *restrict t)
{
  for (int i = 0; i < 16; i++)
     t[i] = abs (p[i] - q[i]);
}

this case should be matched to the HW instruction vabsdub.

[Esme]

Fix whitespace error.

[qiucf]

minor comment: would it be better to make 'some cases' in title more accurate, like 'absolute difference' (how ISA names the instr)?

[steven.zhang]

LGTM. Thank you for adding this.

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

     // Check if CRBits did anything and return that if it did.
     if (SDValue CRTruncValue = DAGCombineTruncBoolExt(N, DCI))
       return CRTruncValue;
   }
 
   SDLoc dl(N);
   SDValue Op0 = N->getOperand(0);
 
+  // fold (truncate (abs (sub (zext a), (zext b)))) -> (vabsd a, b)
+  if (Subtarget.hasP9Altivec() && Op0.getOpcode() == ISD::ABS) {
+    EVT VT = N->getValueType(0);
+    if (VT != MVT::v4i32 && VT != MVT::v8i16 && VT != MVT::v16i8)
+      return SDValue();
+    SDValue Sub = Op0.getOperand(0);
+    if (Sub.getOpcode() == ISD::SUB) {
+      SDValue SubOp0 = Sub.getOperand(0);
+      SDValue SubOp1 = Sub.getOperand(1);
+      if ((SubOp0.getOpcode() == ISD::ZERO_EXTEND) &&
+          (SubOp1.getOpcode() == ISD::ZERO_EXTEND)) {
+        return DCI.DAG.getNode(PPCISD::VABSD, dl, VT, SubOp0.getOperand(0),
+                               SubOp1.getOperand(0),
+                               DCI.DAG.getTargetConstant(0, dl, MVT::i32));
+      }
+    }
+  }
+  
   // Looking for a truncate of i128 to i64.
   if (Op0.getValueType() != MVT::i128 || N->getValueType(0) != MVT::i64)
     return SDValue();
 
   int EltToExtract = DCI.DAG.getDataLayout().isBigEndian() ? 1 : 0;
 
   // SRL feeding TRUNCATE.
   if (Op0.getOpcode() == ISD::SRL) {

llvm/test/CodeGen/PowerPC/vec_absd.ll

 define <4 x i32> @test_word(<4 x i32> %a, <4 x i32> %b) {
 entry:
   %res = tail call <4 x i32> @llvm.ppc.altivec.vabsduw(<4 x i32> %a, <4 x i32> %b)
   ret <4 x i32> %res
 ; CHECK-LABEL: @test_word
 ; CHECK: vabsduw 2, 2, 3
 ; CHECK: blr
 }
+
+define <16 x i8> @test_vabsdub(<16 x i8> %0, <16 x i8> %1) {
+entry:
+  %2 = zext <16 x i8> %0 to <16 x i32>
+  %3 = zext <16 x i8> %1 to <16 x i32>
+  %4 = sub nsw <16 x i32> %2, %3
+  %5 = icmp slt <16 x i32> %4, zeroinitializer
+  %6 = sub nsw <16 x i32> zeroinitializer, %4
+  %7 = select <16 x i1> %5, <16 x i32> %6, <16 x i32> %4
+  %8 = trunc <16 x i32> %7 to <16 x i8>
+  ret <16 x i8> %8
+; CHECK-LABEL: @test_vabsdub
+; CHECK: vabsdub 2, 2, 3
+; CHECK: blr
+}
+
+define <8 x i16> @test_vabsduh(<8 x i16> %0, <8 x i16> %1) {
+entry:
+  %2 = zext <8 x i16> %0 to <8 x i32>
+  %3 = zext <8 x i16> %1 to <8 x i32>
+  %4 = sub nsw <8 x i32> %2, %3
+  %5 = icmp slt <8 x i32> %4, zeroinitializer
+  %6 = sub nsw <8 x i32> zeroinitializer, %4
+  %7 = select <8 x i1> %5, <8 x i32> %6, <8 x i32> %4
+  %8 = trunc <8 x i32> %7 to <8 x i16>
+  ret <8 x i16> %8
+; CHECK-LABEL: @test_vabsduh
+; CHECK: vabsduh 2, 2, 3
+; CHECK: blr
+}
+
+define <4 x i32> @test_vabsduw(<4 x i32> %0, <4 x i32> %1) {
+entry:
+  %2 = sub nsw <4 x i32> %0, %1
+  %3 = icmp slt <4 x i32> %2, zeroinitializer
+  %4 = sub nsw <4 x i32> zeroinitializer, %2
+  %5 = select <4 x i1> %3, <4 x i32> %4, <4 x i32> %2
+  ret <4 x i32> %5
+; CHECK-LABEL: @test_vabsduw
+; CHECK: vabsduw 2, 2, 3
+; CHECK: blr
+}