diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
--- a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
@@ -549,6 +549,16 @@
         SimplifyDemandedBits(I, 0, DemandedFromOps, LHSKnown, Depth + 1))
       return disableWrapFlagsBasedOnUnusedHighBits(I, NLZ);
 
+    // If low order bits are not demanded and are known to be zero in RHS,
+    // then we don't need to demand them from LHS, since they can't cause a
+    // borrow from any bits that are demanded in the result.
+    unsigned NTZ = (~DemandedMask & RHSKnown.Zero).countTrailingOnes();
+    APInt DemandedFromLHS = DemandedFromOps;
+    DemandedFromLHS.clearLowBits(NTZ);
+    if (ShrinkDemandedConstant(I, 0, DemandedFromLHS) ||
+        SimplifyDemandedBits(I, 0, DemandedFromLHS, LHSKnown, Depth + 1))
+      return I;
+
     // If we are known to be subtracting zeros from every bit below
     // the highest demanded bit, we just return the other side.
     if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
diff --git a/llvm/test/Transforms/InstCombine/sub.ll b/llvm/test/Transforms/InstCombine/sub.ll
--- a/llvm/test/Transforms/InstCombine/sub.ll
+++ b/llvm/test/Transforms/InstCombine/sub.ll
@@ -2120,12 +2120,12 @@
   ret i8 %r
 }
 
+; sub becomes negate and combines with shl
+
 define i8 @shrink_sub_from_constant_lowbits(i8 %x) {
 ; CHECK-LABEL: @shrink_sub_from_constant_lowbits(
-; CHECK-NEXT:    [[X000:%.*]] = shl i8 [[X:%.*]], 3
-; CHECK-NEXT:    [[SUB:%.*]] = sub i8 7, [[X000]]
-; CHECK-NEXT:    [[R:%.*]] = and i8 [[SUB]], -8
-; CHECK-NEXT:    ret i8 [[R]]
+; CHECK-NEXT:    [[X000_NEG:%.*]] = mul i8 [[X:%.*]], -8
+; CHECK-NEXT:    ret i8 [[X000_NEG]]
 ;
   %x000 = shl i8 %x, 3   ; 3 low bits are known zero
   %sub = sub i8 7, %x000
@@ -2133,6 +2133,8 @@
   ret i8 %r
 }
 
+; negative test - extra use prevents shrinking '7'
+
 define i8 @shrink_sub_from_constant_lowbits_uses(i8 %x) {
 ; CHECK-LABEL: @shrink_sub_from_constant_lowbits_uses(
 ; CHECK-NEXT:    [[X000:%.*]] = shl i8 [[X:%.*]], 3
@@ -2148,10 +2150,12 @@
   ret i8 %r
 }
 
+; safe to clear 3 low bits (2 higher bits remain set)
+
 define i8 @shrink_sub_from_constant_lowbits2(i8 %x) {
 ; CHECK-LABEL: @shrink_sub_from_constant_lowbits2(
 ; CHECK-NEXT:    [[X000:%.*]] = and i8 [[X:%.*]], -8
-; CHECK-NEXT:    [[SUB:%.*]] = sub nsw i8 30, [[X000]]
+; CHECK-NEXT:    [[SUB:%.*]] = sub nsw i8 24, [[X000]]
 ; CHECK-NEXT:    [[R:%.*]] = and i8 [[SUB]], -16
 ; CHECK-NEXT:    ret i8 [[R]]
 ;
@@ -2161,11 +2165,13 @@
   ret i8 %r
 }
 
+; safe to clear 3 low bits (2 higher bits remain set)
+
 define <2 x i8> @shrink_sub_from_constant_lowbits3(<2 x i8> %x) {
 ; CHECK-LABEL: @shrink_sub_from_constant_lowbits3(
 ; CHECK-NEXT:    [[X0000:%.*]] = shl <2 x i8> [[X:%.*]], <i8 4, i8 4>
-; CHECK-NEXT:    [[SUB:%.*]] = sub nuw <2 x i8> <i8 31, i8 31>, [[X0000]]
-; CHECK-NEXT:    [[R:%.*]] = lshr <2 x i8> [[SUB]], <i8 3, i8 3>
+; CHECK-NEXT:    [[SUB:%.*]] = sub nuw <2 x i8> <i8 24, i8 24>, [[X0000]]
+; CHECK-NEXT:    [[R:%.*]] = lshr exact <2 x i8> [[SUB]], <i8 3, i8 3>
 ; CHECK-NEXT:    ret <2 x i8> [[R]]
 ;
   %x0000 = shl <2 x i8> %x, <i8 4, i8 4>     ; 4 low bits are known zero
@@ -2174,12 +2180,14 @@
   ret <2 x i8> %r
 }
 
+; eliminate the mask of y or the mask of the result
+
 define i8 @demand_sub_from_variable_lowbits(i8 %x, i8 %y) {
 ; CHECK-LABEL: @demand_sub_from_variable_lowbits(
 ; CHECK-NEXT:    [[X000:%.*]] = shl i8 [[X:%.*]], 3
-; CHECK-NEXT:    [[Y000:%.*]] = and i8 [[Y:%.*]], -8
-; CHECK-NEXT:    [[SUB:%.*]] = sub i8 [[Y000]], [[X000]]
-; CHECK-NEXT:    ret i8 [[SUB]]
+; CHECK-NEXT:    [[SUB:%.*]] = sub i8 [[Y:%.*]], [[X000]]
+; CHECK-NEXT:    [[R:%.*]] = and i8 [[SUB]], -8
+; CHECK-NEXT:    ret i8 [[R]]
 ;
   %x000 = shl i8 %x, 3   ; 3 low bits are known zero
   %y000 = and i8 %y, -8
@@ -2188,11 +2196,12 @@
   ret i8 %r
 }
 
+; setting the low 3 bits of y doesn't change anything
+
 define i8 @demand_sub_from_variable_lowbits2(i8 %x, i8 %y) {
 ; CHECK-LABEL: @demand_sub_from_variable_lowbits2(
 ; CHECK-NEXT:    [[X0000:%.*]] = shl i8 [[X:%.*]], 4
-; CHECK-NEXT:    [[Y111:%.*]] = or i8 [[Y:%.*]], 7
-; CHECK-NEXT:    [[SUB:%.*]] = sub nuw nsw i8 [[Y111]], [[X0000]]
+; CHECK-NEXT:    [[SUB:%.*]] = sub nuw nsw i8 [[Y:%.*]], [[X0000]]
 ; CHECK-NEXT:    [[R:%.*]] = lshr i8 [[SUB]], 4
 ; CHECK-NEXT:    ret i8 [[R]]
 ;
@@ -2203,6 +2212,8 @@
   ret i8 %r
 }
 
+; negative test - the mask of y removes an extra bit, so that instruction is needed
+
 define i8 @demand_sub_from_variable_lowbits3(i8 %x, i8 %y) {
 ; CHECK-LABEL: @demand_sub_from_variable_lowbits3(
 ; CHECK-NEXT:    [[X0000:%.*]] = shl i8 [[X:%.*]], 4