[DAGCombine] (x - C) - y -> (x - y) - C fold. Try 3

Summary:
Again only vectors affected. Frustrating. Let me take a look into that..

https://rise4fun.com/Alive/AAq

This is a recommit, originally committed in rL361852, but reverted
to investigate test-suite compile-time hangs, and then reverted in
rL362109 to fix missing constant folds that were causing
endless combine loops.

Reviewers: RKSimon, craig.topper, spatel

Reviewed By: RKSimon

Subscribers: javed.absar, JDevlieghere, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D62294

llvm-svn: 362145

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -2990,6 +2990,13 @@ SDValue DAGCombiner::visitSUB(SDNode *N) {
     SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(0));
     return DAG.getNode(ISD::SUB, DL, VT, Sub, N1.getOperand(1));
   }
+  // (x - C) - y  ->  (x - y) - C
+  // This is necessary because SUB(X,C) -> ADD(X,-C) doesn't work for vectors.
+  if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB &&
+      isConstantOrConstantVector(N0.getOperand(1))) {
+    SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N1);
+    return DAG.getNode(ISD::SUB, DL, VT, Sub, N0.getOperand(1));
+  }
 
   // If the target's bool is represented as 0/-1, prefer to make this 'add 0/-1'
   // rather than 'sub 0/1' (the sext should get folded).

llvm/test/CodeGen/AArch64/sink-addsub-of-const.ll

@@ -314,8 +314,8 @@ define <4 x i32> @vec_sink_sub_of_const_to_sub(<4 x i32> %a, <4 x i32> %b, <4 x
 ; CHECK-NEXT:    adrp x8, .LCPI20_0
 ; CHECK-NEXT:    ldr q3, [x8, :lo12:.LCPI20_0]
 ; CHECK-NEXT:    sub v0.4s, v0.4s, v1.4s
-; CHECK-NEXT:    sub v0.4s, v0.4s, v3.4s
 ; CHECK-NEXT:    sub v0.4s, v0.4s, v2.4s
+; CHECK-NEXT:    sub v0.4s, v0.4s, v3.4s
 ; CHECK-NEXT:    ret
   %t0 = sub <4 x i32> %a, %b
   %t1 = sub <4 x i32> %t0, <i32 42, i32 24, i32 undef, i32 46>

llvm/test/CodeGen/X86/sink-addsub-of-const.ll

@@ -460,15 +460,15 @@ define <4 x i32> @vec_sink_sub_of_const_to_sub(<4 x i32> %a, <4 x i32> %b, <4 x
 ; X32-LABEL: vec_sink_sub_of_const_to_sub:
 ; X32:       # %bb.0:
 ; X32-NEXT:    psubd %xmm1, %xmm0
-; X32-NEXT:    psubd {{\.LCPI.*}}, %xmm0
 ; X32-NEXT:    psubd %xmm2, %xmm0
+; X32-NEXT:    psubd {{\.LCPI.*}}, %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: vec_sink_sub_of_const_to_sub:
 ; X64:       # %bb.0:
 ; X64-NEXT:    psubd %xmm1, %xmm0
-; X64-NEXT:    psubd {{.*}}(%rip), %xmm0
 ; X64-NEXT:    psubd %xmm2, %xmm0
+; X64-NEXT:    psubd {{.*}}(%rip), %xmm0
 ; X64-NEXT:    retq
   %t0 = sub <4 x i32> %a, %b
   %t1 = sub <4 x i32> %t0, <i32 42, i32 24, i32 undef, i32 46>