[InstCombine] Support (sub (sext x), (sext y)) --> (sext (sub x, y)) and (sub (zext x), (zext y)) --> (zext (sub x, y))

[craig.topper]

Summary

If the sub doesn't overflow in the original type we can move it above the sext/zext.

This is similar to what we do for add. The overflow checking for sub is currently weaker than add, so the test cases are constructed for what is supported.

[spatel]

Logic seems right, but we already had a big pile of repeated code.
I think we can handle sub the same way as add/mul after:
rL342292
The only extra work will be swapping the constant operand to LHS for sub?

[craig.topper]

Rebase to use the code from r342292.

[spatel]

LGTM - but please add at least one negative test where the math actually could overflow, so we verify that the value tracking is in place.

lib/Transforms/InstCombine/InstCombineAddSub.cpp

if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
     // --> (A < 0) ? -A : A
     Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
     // Copy the nuw/nsw flags from the sub to the negate.
     Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
                                    I.hasNoSignedWrap());
     return SelectInst::Create(Cmp, Neg, A);
   }
 
+  if (Instruction *Ext = narrowMathIfNoOverflow(I))
+    return Ext;
+
   bool Changed = false;
   if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
     Changed = true;
     I.setHasNoSignedWrap(true);
   }
   if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
     Changed = true;
     I.setHasNoUnsignedWrap(true);

lib/Transforms/InstCombine/InstructionCombining.cpp

Instruction *InstCombiner::foldShuffledBinop(BinaryOperator &Inst) {
   return nullptr;
 }
 
 /// Try to narrow the width of a binop if at least 1 operand is an extend of
 /// of a value. This requires a potentially expensive known bits check to make
 /// sure the narrow op does not overflow.
 Instruction *InstCombiner::narrowMathIfNoOverflow(BinaryOperator &BO) {
   // We need at least one extended operand.
-  Value *LHS = BO.getOperand(0), *RHS = BO.getOperand(1);
+  Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1);
+
+  // If this is a sub, we swap the operands since we always want an extension
+  // on the RHS. The LHS can be an extension or a constant.
+  if (BO.getOpcode() == Instruction::Sub)
+    std::swap(Op0, Op1);
+
   Value *X;
-  bool IsSext = match(LHS, m_SExt(m_Value(X)));
-  if (!IsSext && !match(LHS, m_ZExt(m_Value(X))))
+  bool IsSext = match(Op0, m_SExt(m_Value(X)));
+  if (!IsSext && !match(Op0, m_ZExt(m_Value(X))))
     return nullptr;
 
   // If both operands are the same extension from the same source type and we
   // can eliminate at least one (hasOneUse), this might work.
   CastInst::CastOps CastOpc = IsSext ? Instruction::SExt : Instruction::ZExt;
   Value *Y;
-  if (!(match(RHS, m_ZExtOrSExt(m_Value(Y))) && X->getType() == Y->getType() &&
-        cast<Operator>(RHS)->getOpcode() == CastOpc &&
-        (LHS->hasOneUse() || RHS->hasOneUse()))) {
+  if (!(match(Op1, m_ZExtOrSExt(m_Value(Y))) && X->getType() == Y->getType() &&
+        cast<Operator>(Op1)->getOpcode() == CastOpc &&
+        (Op0->hasOneUse() || Op1->hasOneUse()))) {
     // If that did not match, see if we have a suitable constant operand.
     // Truncating and extending must produce the same constant.
     Constant *WideC;
-    if (!LHS->hasOneUse() || !match(RHS, m_Constant(WideC)))
+    if (!Op0->hasOneUse() || !match(Op1, m_Constant(WideC)))
       return nullptr;
     Constant *NarrowC = ConstantExpr::getTrunc(WideC, X->getType());
     if (ConstantExpr::getCast(CastOpc, NarrowC, BO.getType()) != WideC)
       return nullptr;
     Y = NarrowC;
   }
+
+  // Swap back now that we found our operands.
+  if (BO.getOpcode() == Instruction::Sub)
+    std::swap(X, Y);
+
   // Both operands have narrow versions. Last step: the math must not overflow
   // in the narrow width.
   if (!willNotOverflow(BO.getOpcode(), X, Y, BO, IsSext))
     return nullptr;
 
   // bo (ext X), (ext Y) --> ext (bo X, Y)
   // bo (ext X), C       --> ext (bo X, C')
   Value *NarrowBO = Builder.CreateBinOp(BO.getOpcode(), X, Y, "narrow");

test/Transforms/InstCombine/narrow-math.ll

;
   %call1 = call i32 @callee(), !range !1
   %call2 = call i32 @callee(), !range !1
   %sext1 = sext i32 %call1 to i64
   %sext2 = sext i32 %call2 to i64
   %add = mul i64 %sext1, %sext2
   ret i64 %add
 }
 
+define i64 @test13(i32 %V) {
+; CHECK-LABEL: @test13(
+; CHECK-NEXT:    [[CALL1:%.*]] = call i32 @callee(), !range !2
+; CHECK-NEXT:    [[CALL2:%.*]] = call i32 @callee(), !range !3
+; CHECK-NEXT:    [[SUBCONV:%.*]] = sub nsw i32 [[CALL1]], [[CALL2]]
+; CHECK-NEXT:    [[SUB:%.*]] = sext i32 [[SUBCONV]] to i64
+; CHECK-NEXT:    ret i64 [[SUB]]
+;
+  %call1 = call i32 @callee(), !range !2
+  %call2 = call i32 @callee(), !range !3
+  %sext1 = sext i32 %call1 to i64
+  %sext2 = sext i32 %call2 to i64
+  %sub = sub i64 %sext1, %sext2
+  ret i64 %sub
+}
+
+define i64 @test14(i32 %V) {
+; CHECK-LABEL: @test14(
+; CHECK-NEXT:    [[CALL1:%.*]] = call i32 @callee(), !range !2
+; CHECK-NEXT:    [[CALL2:%.*]] = call i32 @callee(), !range !0
+; CHECK-NEXT:    [[SUBCONV:%.*]] = sub nuw nsw i32 [[CALL1]], [[CALL2]]
+; CHECK-NEXT:    [[SUB:%.*]] = zext i32 [[SUBCONV]] to i64
+; CHECK-NEXT:    ret i64 [[SUB]]
+;
+  %call1 = call i32 @callee(), !range !2
+  %call2 = call i32 @callee(), !range !0
+  %zext1 = zext i32 %call1 to i64
+  %zext2 = zext i32 %call2 to i64
+  %sub = sub i64 %zext1, %zext2
+  ret i64 %sub
+}
+
+define i64 @test15(i32 %V) {
+; CHECK-LABEL: @test15(
+; CHECK-NEXT:    [[ASHR:%.*]] = ashr i32 [[V:%.*]], 1
+; CHECK-NEXT:    [[SUBCONV:%.*]] = sub nsw i32 8, [[ASHR]]
+; CHECK-NEXT:    [[SUB:%.*]] = sext i32 [[SUBCONV]] to i64
+; CHECK-NEXT:    ret i64 [[SUB]]
+;
+  %ashr = ashr i32 %V, 1
+  %sext = sext i32 %ashr to i64
+  %sub = sub i64 8, %sext
+  ret i64 %sub
+}
+
+define <2 x i64> @test15vec(<2 x i32> %V) {
+; CHECK-LABEL: @test15vec(
+; CHECK-NEXT:    [[ASHR:%.*]] = ashr <2 x i32> [[V:%.*]], <i32 1, i32 1>
+; CHECK-NEXT:    [[SUBCONV:%.*]] = sub nsw <2 x i32> <i32 8, i32 8>, [[ASHR]]
+; CHECK-NEXT:    [[SUB:%.*]] = sext <2 x i32> [[SUBCONV]] to <2 x i64>
+; CHECK-NEXT:    ret <2 x i64> [[SUB]]
+;
+  %ashr = ashr <2 x i32> %V, <i32 1, i32 1>
+  %sext = sext <2 x i32> %ashr to <2 x i64>
+  %sub = sub <2 x i64> <i64 8, i64 8>, %sext
+  ret <2 x i64> %sub
+}
+
+define i64 @test16(i32 %V) {
+; CHECK-LABEL: @test16(
+; CHECK-NEXT:    [[LSHR:%.*]] = lshr i32 [[V:%.*]], 1
+; CHECK-NEXT:    [[SUBCONV:%.*]] = sub nuw i32 -2, [[LSHR]]
+; CHECK-NEXT:    [[SUB:%.*]] = zext i32 [[SUBCONV]] to i64
+; CHECK-NEXT:    ret i64 [[SUB]]
+;
+  %lshr = lshr i32 %V, 1
+  %zext = zext i32 %lshr to i64
+  %sub = sub i64 4294967294, %zext
+  ret i64 %sub
+}
+
+define <2 x i64> @test16vec(<2 x i32> %V) {
+; CHECK-LABEL: @test16vec(
+; CHECK-NEXT:    [[LSHR:%.*]] = lshr <2 x i32> [[V:%.*]], <i32 1, i32 1>
+; CHECK-NEXT:    [[SUBCONV:%.*]] = sub nuw <2 x i32> <i32 -2, i32 -2>, [[LSHR]]
+; CHECK-NEXT:    [[SUB:%.*]] = zext <2 x i32> [[SUBCONV]] to <2 x i64>
+; CHECK-NEXT:    ret <2 x i64> [[SUB]]
+;
+  %lshr = lshr <2 x i32> %V, <i32 1, i32 1>
+  %zext = zext <2 x i32> %lshr to <2 x i64>
+  %sub = sub <2 x i64> <i64 4294967294, i64 4294967294>, %zext
+  ret <2 x i64> %sub
+}
+
 !0 = !{ i32 0, i32 2000 }
 !1 = !{ i32 -2000, i32 0 }
+!2 = !{ i32 -512, i32 -255 }
+!3 = !{ i32 -128, i32 0 }