maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[InstCombine] fold adds of constants separated by sext/zext 

This is part of a transform that may be done in the backend:
D13757
...but it should always be beneficial to fold this sooner in IR
for all targets.

https://rise4fun.com/Alive/vaiW

  Name: sext add nsw
  %add = add nsw i8 %i, C0
  %ext = sext i8 %add to i32
  %r = add i32 %ext, C1
  =>
  %s = sext i8 %i to i32
  %r = add i32 %s, sext(C0)+C1

  Name: zext add nuw
  %add = add nuw i8 %i, C0
  %ext = zext i8 %add to i16
  %r = add i16 %ext, C1
  =>
  %s = zext i8 %i to i16
  %r = add i16 %s, zext(C0)+C1

llvm-svn: 355118
This commit is contained in:
Sanjay Patel 2019-02-28 19:05:26 +00:00
parent 8b1703fc1d
commit 4a47f5f550
2 changed files with 56 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

52
llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
View File

@ -822,6 +822,47 @@ static Value *checkForNegativeOperand(BinaryOperator &I,
return nullptr;
}
/// Wrapping flags may allow combining constants separated by an extend.
static Instruction *foldNoWrapAdd(BinaryOperator &Add,
InstCombiner::BuilderTy &Builder) {
Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
Type *Ty = Add.getType();
Constant *Op1C;
if (!match(Op1, m_Constant(Op1C)))
return nullptr;
// Try this match first because it results in an add in the narrow type.
// (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
Value *X;
const APInt *C1, *C2;
if (match(Op1, m_APInt(C1)) &&
match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
Constant *NewC =
ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
}
// More general combining of constants in the wide type.
// (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
Constant *NarrowC;
if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
Value *WideX = Builder.CreateSExt(X, Ty);
return BinaryOperator::CreateAdd(WideX, NewC);
}
// (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
Value *WideX = Builder.CreateZExt(X, Ty);
return BinaryOperator::CreateAdd(WideX, NewC);
}
return nullptr;
}
Instruction *InstCombiner::foldAddWithConstant(BinaryOperator &Add) {
Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
Constant *Op1C;
@ -870,14 +911,6 @@ Instruction *InstCombiner::foldAddWithConstant(BinaryOperator &Add) {
C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
return CastInst::Create(Instruction::SExt, X, Ty);
// (add (zext (add nuw X, C2)), C) --> (zext (add nuw X, C2 + C))
if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
C->isNegative() && C->sge(-C2->sext(C->getBitWidth()))) {
Constant *NewC =
ConstantInt::get(X->getType(), *C2 + C->trunc(C2->getBitWidth()));
return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
}
if (C->isOneValue() && Op0->hasOneUse()) {
// add (sext i1 X), 1 --> zext (not X)
// TODO: The smallest IR representation is (select X, 0, 1), and that would
@ -1050,6 +1083,9 @@ Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
if (Instruction *X = foldAddWithConstant(I))
return X;
if (Instruction *X = foldNoWrapAdd(I, Builder))
return X;
// FIXME: This should be moved into the above helper function to allow these
// transforms for general constant or constant splat vectors.
Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);

22
llvm/test/Transforms/InstCombine/add.ll
View File

@ -396,9 +396,8 @@ define i8 @add_nuw_signbit(i8 %x) {
define i32 @add_nsw_sext_add(i8 %x) {
; CHECK-LABEL: @add_nsw_sext_add(
; CHECK-NEXT: [[ADD:%.*]] = add nsw i8 [[X:%.*]], 42
; CHECK-NEXT: [[EXT:%.*]] = sext i8 [[ADD]] to i32
; CHECK-NEXT: [[R:%.*]] = add nsw i32 [[EXT]], 356
; CHECK-NEXT: [[TMP1:%.*]] = sext i8 [[X:%.*]] to i32
; CHECK-NEXT: [[R:%.*]] = add nsw i32 [[TMP1]], 398
; CHECK-NEXT: ret i32 [[R]]
;
%add = add nsw i8 %x, 42
@ -407,6 +406,8 @@ define i32 @add_nsw_sext_add(i8 %x) {
ret i32 %r
}
; Negative test - extra use of the sext means increase of instructions.
define i32 @add_nsw_sext_add_extra_use_1(i8 %x, i32* %p) {
; CHECK-LABEL: @add_nsw_sext_add_extra_use_1(
; CHECK-NEXT: [[ADD:%.*]] = add nsw i8 [[X:%.*]], 42
@ -426,8 +427,8 @@ define <2 x i32> @add_nsw_sext_add_vec_extra_use_2(<2 x i8> %x, <2 x i8>* %p) {
; CHECK-LABEL: @add_nsw_sext_add_vec_extra_use_2(
; CHECK-NEXT: [[ADD:%.*]] = add nsw <2 x i8> [[X:%.*]], <i8 42, i8 -5>
; CHECK-NEXT: store <2 x i8> [[ADD]], <2 x i8>* [[P:%.*]], align 2
; CHECK-NEXT: [[EXT:%.*]] = sext <2 x i8> [[ADD]] to <2 x i32>
; CHECK-NEXT: [[R:%.*]] = add nsw <2 x i32> [[EXT]], <i32 356, i32 12>
; CHECK-NEXT: [[TMP1:%.*]] = sext <2 x i8> [[X]] to <2 x i32>
; CHECK-NEXT: [[R:%.*]] = add nsw <2 x i32> [[TMP1]], <i32 398, i32 7>
; CHECK-NEXT: ret <2 x i32> [[R]]
;
%add = add nsw <2 x i8> %x, <i8 42, i8 -5>
@ -439,9 +440,8 @@ define <2 x i32> @add_nsw_sext_add_vec_extra_use_2(<2 x i8> %x, <2 x i8>* %p) {
define <2 x i32> @add_nuw_zext_add_vec(<2 x i16> %x) {
; CHECK-LABEL: @add_nuw_zext_add_vec(
; CHECK-NEXT: [[ADD:%.*]] = add nuw <2 x i16> [[X:%.*]], <i16 -42, i16 5>
; CHECK-NEXT: [[EXT:%.*]] = zext <2 x i16> [[ADD]] to <2 x i32>
; CHECK-NEXT: [[R:%.*]] = add nsw <2 x i32> [[EXT]], <i32 356, i32 -12>
; CHECK-NEXT: [[TMP1:%.*]] = zext <2 x i16> [[X:%.*]] to <2 x i32>
; CHECK-NEXT: [[R:%.*]] = add nsw <2 x i32> [[TMP1]], <i32 65850, i32 -7>
; CHECK-NEXT: ret <2 x i32> [[R]]
;
%add = add nuw <2 x i16> %x, <i16 -42, i16 5>
@ -450,6 +450,8 @@ define <2 x i32> @add_nuw_zext_add_vec(<2 x i16> %x) {
ret <2 x i32> %r
}
; Negative test - extra use of the zext means increase of instructions.
define i64 @add_nuw_zext_add_extra_use_1(i8 %x, i64* %p) {
; CHECK-LABEL: @add_nuw_zext_add_extra_use_1(
; CHECK-NEXT: [[ADD:%.*]] = add nuw i8 [[X:%.*]], 42
@ -469,8 +471,8 @@ define i64 @add_nuw_zext_add_extra_use_2(i8 %x, i8* %p) {
; CHECK-LABEL: @add_nuw_zext_add_extra_use_2(
; CHECK-NEXT: [[ADD:%.*]] = add nuw i8 [[X:%.*]], 42
; CHECK-NEXT: store i8 [[ADD]], i8* [[P:%.*]], align 1
; CHECK-NEXT: [[EXT:%.*]] = zext i8 [[ADD]] to i64
; CHECK-NEXT: [[R:%.*]] = add nuw nsw i64 [[EXT]], -356
; CHECK-NEXT: [[TMP1:%.*]] = zext i8 [[X]] to i64
; CHECK-NEXT: [[R:%.*]] = add nuw nsw i64 [[TMP1]], -314
; CHECK-NEXT: ret i64 [[R]]
;
%add = add nuw i8 %x, 42