[ARM] Add an extra fold for f32 extract(vdup(i32))

[dmgreen]

Summary

This adds another small fold for extract of a vdup, between a i32 and a f32, converting to a BITCAST. This allows some extra folding to happen, simplifying the resulting code.

[samtebbs]

Short and sweet, perfect.

llvm/lib/Target/ARM/ARMISelLowering.cpp

static SDValue PerformExtractEltCombine(SDNode *N,
 
   // extract (vdup x) -> x
   if (Op0->getOpcode() == ARMISD::VDUP) {
     SDValue X = Op0->getOperand(0);
     if (VT == MVT::f16 && X.getValueType() == MVT::i32)
       return DCI.DAG.getNode(ARMISD::VMOVhr, dl, VT, X);
     if (VT == MVT::i32 && X.getValueType() == MVT::f16)
       return DCI.DAG.getNode(ARMISD::VMOVrh, dl, VT, X);
+    if (VT == MVT::f32 && X.getValueType() == MVT::i32)
+      return DCI.DAG.getNode(ISD::BITCAST, dl, VT, X);
 
     while (X.getValueType() != VT && X->getOpcode() == ISD::BITCAST)
       X = X->getOperand(0);
     if (X.getValueType() == VT)
       return X;
   }
 
   // extract(bitcast(BUILD_VECTOR(VMOVDRR(a, b), ..))) -> a or b

llvm/test/CodeGen/Thumb2/mve-vst4.ll

entry:
   ret void
 }
 
 define void @vst4_v4i32_align1(<4 x i32> *%src, <16 x i32> *%dst) {
 ; CHECK-LABEL: vst4_v4i32_align1:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r7, lr}
 ; CHECK-NEXT:    push {r7, lr}
-; CHECK-NEXT:    .vsave {d8, d9}
-; CHECK-NEXT:    vpush {d8, d9}
+; CHECK-NEXT:    .vsave {d8, d9, d10}
+; CHECK-NEXT:    vpush {d8, d9, d10}
 ; CHECK-NEXT:    vldrw.u32 q0, [r0, #32]
-; CHECK-NEXT:    vldrw.u32 q3, [r0]
-; CHECK-NEXT:    vldrw.u32 q1, [r0, #16]
+; CHECK-NEXT:    vldrw.u32 q4, [r0]
 ; CHECK-NEXT:    vmov r2, r3, d1
 ; CHECK-NEXT:    vmov r12, lr, d0
-; CHECK-NEXT:    vdup.32 q4, r3
-; CHECK-NEXT:    vmov.f64 d0, d6
-; CHECK-NEXT:    vmov.f32 s1, s4
-; CHECK-NEXT:    vmov.f32 s4, s13
-; CHECK-NEXT:    vmov.f64 d4, d7
-; CHECK-NEXT:    vmov.f32 s12, s15
-; CHECK-NEXT:    vmov.f32 s13, s7
-; CHECK-NEXT:    vmov.f32 s14, s18
-; CHECK-NEXT:    vmov.f32 s15, s19
+; CHECK-NEXT:    vldrw.u32 q0, [r0, #16]
+; CHECK-NEXT:    vmov.f64 d2, d8
+; CHECK-NEXT:    vmov.f32 s5, s0
+; CHECK-NEXT:    vmov s10, r2
+; CHECK-NEXT:    vmov s14, r3
+; CHECK-NEXT:    vmov.f32 s8, s18
+; CHECK-NEXT:    vmov s20, lr
+; CHECK-NEXT:    vmov.f32 s9, s2
+; CHECK-NEXT:    vmov s6, r12
+; CHECK-NEXT:    vmov.f32 s0, s17
+; CHECK-NEXT:    vmov.f32 s12, s19
+; CHECK-NEXT:    vmov.f32 s13, s3
+; CHECK-NEXT:    vmov.f32 s2, s20
+; CHECK-NEXT:    vmov.f32 s15, s14
+; CHECK-NEXT:    vmov.f32 s11, s10
 ; CHECK-NEXT:    vstrb.8 q3, [r1, #48]
-; CHECK-NEXT:    vmov.f32 s9, s6
-; CHECK-NEXT:    vdup.32 q3, r2
-; CHECK-NEXT:    vmov.f32 s10, s14
-; CHECK-NEXT:    vmov.f32 s11, s15
+; CHECK-NEXT:    vmov.f32 s3, s20
 ; CHECK-NEXT:    vstrb.8 q2, [r1, #32]
-; CHECK-NEXT:    vdup.32 q2, lr
-; CHECK-NEXT:    vmov.f32 s6, s10
-; CHECK-NEXT:    vmov.f32 s7, s11
-; CHECK-NEXT:    vstrb.8 q1, [r1, #16]
-; CHECK-NEXT:    vdup.32 q1, r12
-; CHECK-NEXT:    vmov.f32 s2, s6
-; CHECK-NEXT:    vmov.f32 s3, s7
-; CHECK-NEXT:    vstrb.8 q0, [r1]
-; CHECK-NEXT:    vpop {d8, d9}
+; CHECK-NEXT:    vmov.f32 s7, s6
+; CHECK-NEXT:    vstrb.8 q0, [r1, #16]
+; CHECK-NEXT:    vstrb.8 q1, [r1]
+; CHECK-NEXT:    vpop {d8, d9, d10}
 ; CHECK-NEXT:    pop {r7, pc}
 entry:
   %s1 = getelementptr <4 x i32>, <4 x i32>* %src, i32 0
   %l1 = load <4 x i32>, <4 x i32>* %s1, align 4
   %s2 = getelementptr <4 x i32>, <4 x i32>* %src, i32 1
   %l2 = load <4 x i32>, <4 x i32>* %s2, align 4
   %s3 = getelementptr <4 x i32>, <4 x i32>* %src, i32 2
   %l3 = load <4 x i32>, <4 x i32>* %s3, align 4
entry:
   ret void
 }
 
 define void @vst4_v4f32_align1(<4 x float> *%src, <16 x float> *%dst) {
 ; CHECK-LABEL: vst4_v4f32_align1:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r7, lr}
 ; CHECK-NEXT:    push {r7, lr}
-; CHECK-NEXT:    .vsave {d8, d9, d10, d11}
-; CHECK-NEXT:    vpush {d8, d9, d10, d11}
-; CHECK-NEXT:    vldrw.u32 q4, [r0]
+; CHECK-NEXT:    .vsave {d8, d9, d10, d11, d12}
+; CHECK-NEXT:    vpush {d8, d9, d10, d11, d12}
 ; CHECK-NEXT:    vldrw.u32 q0, [r0, #32]
-; CHECK-NEXT:    vldrw.u32 q2, [r0, #16]
-; CHECK-NEXT:    vmov.f64 d2, d8
+; CHECK-NEXT:    vldrw.u32 q5, [r0]
+; CHECK-NEXT:    vldrw.u32 q1, [r0, #16]
 ; CHECK-NEXT:    vmov r2, r3, d1
-; CHECK-NEXT:    vmov.f32 s5, s8
-; CHECK-NEXT:    vdup.32 q5, r3
-; CHECK-NEXT:    vmov.f32 s8, s17
-; CHECK-NEXT:    vmov.f64 d6, d9
-; CHECK-NEXT:    vmov.f32 s16, s19
-; CHECK-NEXT:    vmov.f32 s17, s11
-; CHECK-NEXT:    vmov.f32 s18, s22
-; CHECK-NEXT:    vmov.f32 s19, s23
-; CHECK-NEXT:    vstrb.8 q4, [r1, #48]
-; CHECK-NEXT:    vmov.f32 s13, s10
-; CHECK-NEXT:    vdup.32 q4, r2
 ; CHECK-NEXT:    vmov r12, lr, d0
-; CHECK-NEXT:    vmov.f32 s14, s18
+; CHECK-NEXT:    vmov.f64 d4, d10
+; CHECK-NEXT:    vmov.f32 s9, s4
+; CHECK-NEXT:    vmov s14, r2
+; CHECK-NEXT:    vmov s18, r3
+; CHECK-NEXT:    vmov.f32 s12, s22
+; CHECK-NEXT:    vmov s24, lr
+; CHECK-NEXT:    vmov.f32 s13, s6
+; CHECK-NEXT:    vmov.f32 s4, s21
+; CHECK-NEXT:    vmov.f32 s16, s23
+; CHECK-NEXT:    vmov.f32 s17, s7
+; CHECK-NEXT:    vmov s10, r12
+; CHECK-NEXT:    vmov.f32 s6, s24
+; CHECK-NEXT:    vmov.f32 s19, s18
 ; CHECK-NEXT:    vmov.f32 s15, s2
+; CHECK-NEXT:    vstrb.8 q4, [r1, #48]
+; CHECK-NEXT:    vmov.f32 s7, s24
 ; CHECK-NEXT:    vstrb.8 q3, [r1, #32]
-; CHECK-NEXT:    vdup.32 q3, lr
-; CHECK-NEXT:    vmov.f32 s10, s14
-; CHECK-NEXT:    vmov.f32 s11, s15
-; CHECK-NEXT:    vstrb.8 q2, [r1, #16]
-; CHECK-NEXT:    vdup.32 q2, r12
-; CHECK-NEXT:    vmov.f32 s6, s10
-; CHECK-NEXT:    vmov.f32 s7, s0
-; CHECK-NEXT:    vstrb.8 q1, [r1]
-; CHECK-NEXT:    vpop {d8, d9, d10, d11}
+; CHECK-NEXT:    vmov.f32 s11, s0
+; CHECK-NEXT:    vstrb.8 q1, [r1, #16]
+; CHECK-NEXT:    vstrb.8 q2, [r1]
+; CHECK-NEXT:    vpop {d8, d9, d10, d11, d12}
 ; CHECK-NEXT:    pop {r7, pc}
 entry:
   %s1 = getelementptr <4 x float>, <4 x float>* %src, i32 0
   %l1 = load <4 x float>, <4 x float>* %s1, align 4
   %s2 = getelementptr <4 x float>, <4 x float>* %src, i32 1
   %l2 = load <4 x float>, <4 x float>* %s2, align 4
   %s3 = getelementptr <4 x float>, <4 x float>* %src, i32 2
   %l3 = load <4 x float>, <4 x float>* %s3, align 4