[DAGCombiner] Allow promoted constants when lowering vector SDIV exacts

Author: SavchenkoValeriy

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -6345,7 +6345,6 @@ static SDValue BuildExactSDIV(const TargetLowering &TLI, SDNode *N,
   SDValue Op0 = N->getOperand(0);
   SDValue Op1 = N->getOperand(1);
   EVT VT = N->getValueType(0);
-  EVT SVT = VT.getScalarType();
   EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
   EVT ShSVT = ShVT.getScalarType();
 
@@ -6355,6 +6354,8 @@ static SDValue BuildExactSDIV(const TargetLowering &TLI, SDNode *N,
   auto BuildSDIVPattern = [&](ConstantSDNode *C) {
     if (C->isZero())
       return false;
+
+    EVT CT = C->getValueType(0);
     APInt Divisor = C->getAPIntValue();
     unsigned Shift = Divisor.countr_zero();
     if (Shift) {
@@ -6363,12 +6364,13 @@ static SDValue BuildExactSDIV(const TargetLowering &TLI, SDNode *N,
     }
     APInt Factor = Divisor.multiplicativeInverse();
     Shifts.push_back(DAG.getConstant(Shift, dl, ShSVT));
-    Factors.push_back(DAG.getConstant(Factor, dl, SVT));
+    Factors.push_back(DAG.getConstant(Factor, dl, CT));
     return true;
   };
 
   // Collect all magic values from the build vector.
-  if (!ISD::matchUnaryPredicate(Op1, BuildSDIVPattern))
+  if (!ISD::matchUnaryPredicate(Op1, BuildSDIVPattern, /*AllowUndefs=*/false,
+                                /*AllowTruncation=*/true))
     return SDValue();
 
   SDValue Shift, Factor;

llvm/test/CodeGen/AArch64/sdiv-by-const-promoted-ops.ll

@@ -49,122 +49,9 @@ define <8 x i16> @sdiv_exact_v8i16_by_255(<8 x i16> %x) {
 define <16 x i16> @sdiv_exact_v16i16_by_255(<16 x i16> %x) {
 ; CHECK-LABEL: sdiv_exact_v16i16_by_255:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    smov x11, v0.h[1]
-; CHECK-NEXT:    smov x10, v0.h[0]
-; CHECK-NEXT:    mov x8, #-32639 // =0xffffffffffff8081
-; CHECK-NEXT:    smov x13, v0.h[3]
-; CHECK-NEXT:    smov x14, v1.h[1]
-; CHECK-NEXT:    movk x8, #32896, lsl #16
-; CHECK-NEXT:    smov x16, v1.h[0]
-; CHECK-NEXT:    smov w12, v0.h[1]
-; CHECK-NEXT:    smov w15, v0.h[0]
-; CHECK-NEXT:    smov x18, v1.h[2]
-; CHECK-NEXT:    smov w0, v0.h[3]
-; CHECK-NEXT:    smov w1, v1.h[1]
-; CHECK-NEXT:    smull x11, w11, w8
-; CHECK-NEXT:    smov w2, v1.h[0]
-; CHECK-NEXT:    smov x9, v0.h[2]
-; CHECK-NEXT:    smull x10, w10, w8
-; CHECK-NEXT:    smov w17, v0.h[2]
-; CHECK-NEXT:    smov w3, v1.h[2]
-; CHECK-NEXT:    smull x13, w13, w8
-; CHECK-NEXT:    smull x14, w14, w8
-; CHECK-NEXT:    add x12, x12, x11, lsr #32
-; CHECK-NEXT:    smull x16, w16, w8
-; CHECK-NEXT:    add x10, x15, x10, lsr #32
-; CHECK-NEXT:    smull x15, w18, w8
-; CHECK-NEXT:    add x11, x0, x13, lsr #32
-; CHECK-NEXT:    smov x0, v0.h[4]
-; CHECK-NEXT:    add x13, x1, x14, lsr #32
-; CHECK-NEXT:    asr w18, w10, #7
-; CHECK-NEXT:    smull x9, w9, w8
-; CHECK-NEXT:    add x14, x2, x16, lsr #32
-; CHECK-NEXT:    asr w16, w12, #7
-; CHECK-NEXT:    smov x2, v1.h[3]
-; CHECK-NEXT:    add w18, w18, w10, lsr #31
-; CHECK-NEXT:    add x15, x3, x15, lsr #32
-; CHECK-NEXT:    smov w10, v0.h[5]
-; CHECK-NEXT:    add w12, w16, w12, lsr #31
-; CHECK-NEXT:    asr w16, w14, #7
-; CHECK-NEXT:    add x9, x17, x9, lsr #32
-; CHECK-NEXT:    fmov s2, w18
-; CHECK-NEXT:    smov w17, v0.h[4]
-; CHECK-NEXT:    smull x0, w0, w8
-; CHECK-NEXT:    add w14, w16, w14, lsr #31
-; CHECK-NEXT:    asr w16, w13, #7
-; CHECK-NEXT:    asr w1, w9, #7
-; CHECK-NEXT:    smov x18, v0.h[5]
-; CHECK-NEXT:    fmov s3, w14
-; CHECK-NEXT:    mov v2.h[1], w12
-; CHECK-NEXT:    add w12, w16, w13, lsr #31
-; CHECK-NEXT:    smov w13, v1.h[3]
-; CHECK-NEXT:    smov x14, v1.h[4]
-; CHECK-NEXT:    smull x16, w2, w8
-; CHECK-NEXT:    add w1, w1, w9, lsr #31
-; CHECK-NEXT:    add x17, x17, x0, lsr #32
-; CHECK-NEXT:    asr w0, w15, #7
-; CHECK-NEXT:    mov v3.h[1], w12
-; CHECK-NEXT:    smov w12, v1.h[4]
-; CHECK-NEXT:    smull x18, w18, w8
-; CHECK-NEXT:    mov v2.h[2], w1
-; CHECK-NEXT:    asr w1, w11, #7
-; CHECK-NEXT:    add w15, w0, w15, lsr #31
-; CHECK-NEXT:    add x13, x13, x16, lsr #32
-; CHECK-NEXT:    smov x16, v1.h[5]
-; CHECK-NEXT:    smull x14, w14, w8
-; CHECK-NEXT:    add w11, w1, w11, lsr #31
-; CHECK-NEXT:    smov x0, v0.h[6]
-; CHECK-NEXT:    add x10, x10, x18, lsr #32
-; CHECK-NEXT:    asr w1, w13, #7
-; CHECK-NEXT:    mov v3.h[2], w15
-; CHECK-NEXT:    smov w15, v1.h[5]
-; CHECK-NEXT:    add x12, x12, x14, lsr #32
-; CHECK-NEXT:    mov v2.h[3], w11
-; CHECK-NEXT:    asr w11, w17, #7
-; CHECK-NEXT:    add w13, w1, w13, lsr #31
-; CHECK-NEXT:    smull x16, w16, w8
-; CHECK-NEXT:    smov x14, v1.h[6]
-; CHECK-NEXT:    asr w18, w12, #7
-; CHECK-NEXT:    add w11, w11, w17, lsr #31
-; CHECK-NEXT:    smov w9, v0.h[6]
-; CHECK-NEXT:    mov v3.h[3], w13
-; CHECK-NEXT:    smull x17, w0, w8
-; CHECK-NEXT:    smov x0, v1.h[7]
-; CHECK-NEXT:    add x13, x15, x16, lsr #32
-; CHECK-NEXT:    add w12, w18, w12, lsr #31
-; CHECK-NEXT:    smov w16, v1.h[6]
-; CHECK-NEXT:    mov v2.h[4], w11
-; CHECK-NEXT:    smov x11, v0.h[7]
-; CHECK-NEXT:    smull x14, w14, w8
-; CHECK-NEXT:    asr w15, w10, #7
-; CHECK-NEXT:    asr w18, w13, #7
-; CHECK-NEXT:    smov w1, v0.h[7]
-; CHECK-NEXT:    mov v3.h[4], w12
-; CHECK-NEXT:    add x9, x9, x17, lsr #32
-; CHECK-NEXT:    add w10, w15, w10, lsr #31
-; CHECK-NEXT:    add w12, w18, w13, lsr #31
-; CHECK-NEXT:    add x13, x16, x14, lsr #32
-; CHECK-NEXT:    smov w14, v1.h[7]
-; CHECK-NEXT:    smull x11, w11, w8
-; CHECK-NEXT:    smull x8, w0, w8
-; CHECK-NEXT:    mov v2.h[5], w10
-; CHECK-NEXT:    asr w10, w9, #7
-; CHECK-NEXT:    mov v3.h[5], w12
-; CHECK-NEXT:    asr w12, w13, #7
-; CHECK-NEXT:    add w9, w10, w9, lsr #31
-; CHECK-NEXT:    add x10, x1, x11, lsr #32
-; CHECK-NEXT:    add w11, w12, w13, lsr #31
-; CHECK-NEXT:    add x8, x14, x8, lsr #32
-; CHECK-NEXT:    mov v2.h[6], w9
-; CHECK-NEXT:    asr w9, w10, #7
-; CHECK-NEXT:    mov v3.h[6], w11
-; CHECK-NEXT:    asr w11, w8, #7
-; CHECK-NEXT:    add w9, w9, w10, lsr #31
-; CHECK-NEXT:    add w8, w11, w8, lsr #31
-; CHECK-NEXT:    mov v2.h[7], w9
-; CHECK-NEXT:    mov v3.h[7], w8
-; CHECK-NEXT:    mov v0.16b, v2.16b
-; CHECK-NEXT:    mov v1.16b, v3.16b
+; CHECK-NEXT:    mvni v2.8h, #1, lsl #8
+; CHECK-NEXT:    mul v0.8h, v0.8h, v2.8h
+; CHECK-NEXT:    mul v1.8h, v1.8h, v2.8h
 ; CHECK-NEXT:    ret
   %div = sdiv exact <16 x i16> %x, splat (i16 255)
   ret <16 x i16> %div