[InstCombine] try multi-use demanded bits fold for 'sub'

This is similar to D133788 / 73919a8, but for sub
the transform is valid only for low zeros in operand 1.

https://alive2.llvm.org/ce/z/EmRsXC

Author: rotateright

llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

@@ -1080,6 +1080,18 @@ Value *InstCombinerImpl::SimplifyMultipleUseDemandedBits(
 
     break;
   }
+  case Instruction::Sub: {
+    unsigned NLZ = DemandedMask.countLeadingZeros();
+    APInt DemandedFromOps = APInt::getLowBitsSet(BitWidth, BitWidth - NLZ);
+
+    // If an operand subtracts zeros from every bit below the highest demanded
+    // bit, that operand doesn't change the result. Return the other side.
+    computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
+    if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
+      return I->getOperand(0);
+
+    break;
+  }
   case Instruction::AShr: {
     // Compute the Known bits to simplify things downstream.
     computeKnownBits(I, Known, Depth, CxtI);

llvm/test/Transforms/InstCombine/sub.ll

@@ -2074,7 +2074,7 @@ define i5 @demand_low_bits_uses(i8 %x, i8 %y) {
 ; CHECK-NEXT:    [[M:%.*]] = and i8 [[X:%.*]], 96
 ; CHECK-NEXT:    [[A:%.*]] = sub i8 [[Y:%.*]], [[M]]
 ; CHECK-NEXT:    call void @use8(i8 [[A]])
-; CHECK-NEXT:    [[R:%.*]] = trunc i8 [[A]] to i5
+; CHECK-NEXT:    [[R:%.*]] = trunc i8 [[Y]] to i5
 ; CHECK-NEXT:    ret i5 [[R]]
 ;
   %m = and i8 %x, 96 ; 0x60
@@ -2101,6 +2101,8 @@ define i6 @demand_low_bits_uses_extra_bit(i8 %x, i8 %y) {
   ret i6 %r
 }
 
+; negative test - must be operand 1
+
 define i8 @demand_low_bits_uses_commute(i8 %x, i8 %y, i8 %z) {
 ; CHECK-LABEL: @demand_low_bits_uses_commute(
 ; CHECK-NEXT:    [[M:%.*]] = and i8 [[X:%.*]], -64