[InstCombine] factorize left shifts of add/sub
We do similar factorization folds in SimplifyUsingDistributiveLaws,
but that drops no-wrap properties. Propagating those optimally may
help solve:
https://llvm.org/PR47430
The propagation is all-or-nothing for these patterns: when all
3 incoming ops have nsw or nuw, the 2 new ops should have the
same no-wrap property:
https://alive2.llvm.org/ce/z/Dv8wsU
This also solves:
https://llvm.org/PR47584
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 5ce32bc..c1cacdf 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -1199,6 +1199,42 @@
return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
}
+/// This is a specialization of a more general transform from
+/// SimplifyUsingDistributiveLaws. If that code can be made to work optimally
+/// for multi-use cases or propagating nsw/nuw, then we would not need this.
+static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
+ InstCombiner::BuilderTy &Builder) {
+ // TODO: Also handle mul by doubling the shift amount?
+ assert(I.getOpcode() == Instruction::Add ||
+ I.getOpcode() == Instruction::Sub && "Expected add/sub");
+ auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
+ auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
+ if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
+ return nullptr;
+
+ Value *X, *Y, *ShAmt;
+ if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
+ !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
+ return nullptr;
+
+ // No-wrap propagates only when all ops have no-wrap.
+ bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
+ Op1->hasNoSignedWrap();
+ bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
+ Op1->hasNoUnsignedWrap();
+
+ // add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
+ Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
+ if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
+ NewI->setHasNoSignedWrap(HasNSW);
+ NewI->setHasNoUnsignedWrap(HasNUW);
+ }
+ auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
+ NewShl->setHasNoSignedWrap(HasNSW);
+ NewShl->setHasNoUnsignedWrap(HasNUW);
+ return NewShl;
+}
+
Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
@@ -1215,6 +1251,9 @@
if (Value *V = SimplifyUsingDistributiveLaws(I))
return replaceInstUsesWith(I, V);
+ if (Instruction *R = factorizeMathWithShlOps(I, Builder))
+ return R;
+
if (Instruction *X = foldAddWithConstant(I))
return X;
@@ -1753,6 +1792,9 @@
if (Value *V = SimplifyUsingDistributiveLaws(I))
return replaceInstUsesWith(I, V);
+ if (Instruction *R = factorizeMathWithShlOps(I, Builder))
+ return R;
+
if (I.getType()->isIntOrIntVectorTy(1))
return BinaryOperator::CreateXor(Op0, Op1);
diff --git a/llvm/test/Transforms/InstCombine/shl-factor.ll b/llvm/test/Transforms/InstCombine/shl-factor.ll
index 274d6e3..37c54cc 100644
--- a/llvm/test/Transforms/InstCombine/shl-factor.ll
+++ b/llvm/test/Transforms/InstCombine/shl-factor.ll
@@ -5,9 +5,8 @@
define i6 @add_shl_same_amount(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @add_shl_same_amount(
-; CHECK-NEXT: [[XS:%.*]] = shl i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = add i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl i6 %x, %z
@@ -18,9 +17,8 @@
define <2 x i4> @add_shl_same_amount_nsw(<2 x i4> %x, <2 x i4> %y, <2 x i4> %z) {
; CHECK-LABEL: @add_shl_same_amount_nsw(
-; CHECK-NEXT: [[XS:%.*]] = shl nsw <2 x i4> [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nsw <2 x i4> [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = add nsw <2 x i4> [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add nsw <2 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl nsw <2 x i4> [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret <2 x i4> [[DIFF]]
;
%xs = shl nsw <2 x i4> %x, %z
@@ -31,9 +29,8 @@
define i64 @add_shl_same_amount_nuw(i64 %x, i64 %y, i64 %z) {
; CHECK-LABEL: @add_shl_same_amount_nuw(
-; CHECK-NEXT: [[XS:%.*]] = shl nuw i64 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nuw i64 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = add nuw i64 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add nuw i64 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl nuw i64 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i64 [[DIFF]]
;
%xs = shl nuw i64 %x, %z
@@ -46,8 +43,8 @@
; CHECK-LABEL: @add_shl_same_amount_nsw_extra_use1(
; CHECK-NEXT: [[XS:%.*]] = shl nuw nsw i8 [[X:%.*]], [[Z:%.*]]
; CHECK-NEXT: call void @use8(i8 [[XS]])
-; CHECK-NEXT: [[YS:%.*]] = shl nuw nsw i8 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = add nsw i8 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add nsw i8 [[X]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl nsw i8 [[TMP1]], [[Z]]
; CHECK-NEXT: ret i8 [[DIFF]]
;
%xs = shl nsw nuw i8 %x, %z
@@ -59,10 +56,10 @@
define i8 @add_shl_same_amount_nuw_extra_use2(i8 %x, i8 %y, i8 %z) {
; CHECK-LABEL: @add_shl_same_amount_nuw_extra_use2(
-; CHECK-NEXT: [[XS:%.*]] = shl nuw i8 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nuw nsw i8 [[Y:%.*]], [[Z]]
+; CHECK-NEXT: [[YS:%.*]] = shl nuw nsw i8 [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: call void @use8(i8 [[YS]])
-; CHECK-NEXT: [[DIFF:%.*]] = add nuw nsw i8 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add nuw i8 [[X:%.*]], [[Y]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl nuw i8 [[TMP1]], [[Z]]
; CHECK-NEXT: ret i8 [[DIFF]]
;
%xs = shl nuw i8 %x, %z
@@ -91,9 +88,8 @@
define i6 @add_shl_same_amount_partial_nsw1(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @add_shl_same_amount_partial_nsw1(
-; CHECK-NEXT: [[XS:%.*]] = shl nsw i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nsw i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = add i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl nsw i6 %x, %z
@@ -104,9 +100,8 @@
define i6 @add_shl_same_amount_partial_nsw2(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @add_shl_same_amount_partial_nsw2(
-; CHECK-NEXT: [[XS:%.*]] = shl i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nsw i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = add nsw i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl i6 %x, %z
@@ -117,9 +112,8 @@
define i6 @add_shl_same_amount_partial_nuw1(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @add_shl_same_amount_partial_nuw1(
-; CHECK-NEXT: [[XS:%.*]] = shl nuw i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nuw i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = add i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl nuw i6 %x, %z
@@ -130,9 +124,8 @@
define i6 @add_shl_same_amount_partial_nuw2(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @add_shl_same_amount_partial_nuw2(
-; CHECK-NEXT: [[XS:%.*]] = shl nuw i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = add nuw i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = add i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl nuw i6 %x, %z
@@ -143,9 +136,8 @@
define i6 @sub_shl_same_amount(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @sub_shl_same_amount(
-; CHECK-NEXT: [[XS:%.*]] = shl i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = sub i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl i6 %x, %z
@@ -156,9 +148,8 @@
define <2 x i4> @sub_shl_same_amount_nsw(<2 x i4> %x, <2 x i4> %y, <2 x i4> %z) {
; CHECK-LABEL: @sub_shl_same_amount_nsw(
-; CHECK-NEXT: [[XS:%.*]] = shl nsw <2 x i4> [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nsw <2 x i4> [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = sub nsw <2 x i4> [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub nsw <2 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl nsw <2 x i4> [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret <2 x i4> [[DIFF]]
;
%xs = shl nsw <2 x i4> %x, %z
@@ -169,9 +160,8 @@
define i64 @sub_shl_same_amount_nuw(i64 %x, i64 %y, i64 %z) {
; CHECK-LABEL: @sub_shl_same_amount_nuw(
-; CHECK-NEXT: [[XS:%.*]] = shl nuw i64 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nuw i64 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = sub nuw i64 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub nuw i64 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl nuw i64 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i64 [[DIFF]]
;
%xs = shl nuw i64 %x, %z
@@ -184,8 +174,8 @@
; CHECK-LABEL: @sub_shl_same_amount_nsw_extra_use1(
; CHECK-NEXT: [[XS:%.*]] = shl nuw nsw i8 [[X:%.*]], [[Z:%.*]]
; CHECK-NEXT: call void @use8(i8 [[XS]])
-; CHECK-NEXT: [[YS:%.*]] = shl nuw nsw i8 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = sub nsw i8 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub nsw i8 [[X]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl nsw i8 [[TMP1]], [[Z]]
; CHECK-NEXT: ret i8 [[DIFF]]
;
%xs = shl nsw nuw i8 %x, %z
@@ -197,10 +187,10 @@
define i8 @sub_shl_same_amount_nuw_extra_use2(i8 %x, i8 %y, i8 %z) {
; CHECK-LABEL: @sub_shl_same_amount_nuw_extra_use2(
-; CHECK-NEXT: [[XS:%.*]] = shl nuw i8 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nuw nsw i8 [[Y:%.*]], [[Z]]
+; CHECK-NEXT: [[YS:%.*]] = shl nuw nsw i8 [[Y:%.*]], [[Z:%.*]]
; CHECK-NEXT: call void @use8(i8 [[YS]])
-; CHECK-NEXT: [[DIFF:%.*]] = sub nuw nsw i8 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub nuw i8 [[X:%.*]], [[Y]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl nuw i8 [[TMP1]], [[Z]]
; CHECK-NEXT: ret i8 [[DIFF]]
;
%xs = shl nuw i8 %x, %z
@@ -229,9 +219,8 @@
define i6 @sub_shl_same_amount_partial_nsw1(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @sub_shl_same_amount_partial_nsw1(
-; CHECK-NEXT: [[XS:%.*]] = shl nsw i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nsw i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = sub i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl nsw i6 %x, %z
@@ -242,9 +231,8 @@
define i6 @sub_shl_same_amount_partial_nsw2(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @sub_shl_same_amount_partial_nsw2(
-; CHECK-NEXT: [[XS:%.*]] = shl i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nsw i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = sub nsw i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl i6 %x, %z
@@ -255,9 +243,8 @@
define i6 @sub_shl_same_amount_partial_nuw1(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @sub_shl_same_amount_partial_nuw1(
-; CHECK-NEXT: [[XS:%.*]] = shl nuw i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl nuw i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = sub i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl nuw i6 %x, %z
@@ -268,9 +255,8 @@
define i6 @sub_shl_same_amount_partial_nuw2(i6 %x, i6 %y, i6 %z) {
; CHECK-LABEL: @sub_shl_same_amount_partial_nuw2(
-; CHECK-NEXT: [[XS:%.*]] = shl nuw i6 [[X:%.*]], [[Z:%.*]]
-; CHECK-NEXT: [[YS:%.*]] = shl i6 [[Y:%.*]], [[Z]]
-; CHECK-NEXT: [[DIFF:%.*]] = sub nuw i6 [[XS]], [[YS]]
+; CHECK-NEXT: [[TMP1:%.*]] = sub i6 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[DIFF:%.*]] = shl i6 [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret i6 [[DIFF]]
;
%xs = shl nuw i6 %x, %z