// ((C1 OP zext(X)) & C2) -> zext((C1 OP X) & C2) if C2 fits in the
// bitwidth of X and OP behaves well when given trunc(C1) and X.
- auto isSuitableBinOpcode = [](BinaryOperator *B) {
+ auto isNarrowableBinOpcode = [](BinaryOperator *B) {
switch (B->getOpcode()) {
case Instruction::Xor:
case Instruction::Or:
}
};
BinaryOperator *BO;
- if (match(Op0, m_OneUse(m_BinOp(BO))) && isSuitableBinOpcode(BO)) {
+ if (match(Op0, m_OneUse(m_BinOp(BO))) && isNarrowableBinOpcode(BO)) {
+ Instruction::BinaryOps BOpcode = BO->getOpcode();
Value *X;
const APInt *C1;
// TODO: The one-use restrictions could be relaxed a little if the AND
// is going to be removed.
+ // Try to narrow the 'and' and a binop with constant operand:
+ // and (bo (zext X), C1), C --> zext (and (bo X, TruncC1), TruncC)
if (match(BO, m_c_BinOp(m_OneUse(m_ZExt(m_Value(X))), m_APInt(C1))) &&
C->isIntN(X->getType()->getScalarSizeInBits())) {
unsigned XWidth = X->getType()->getScalarSizeInBits();
Constant *TruncC1 = ConstantInt::get(X->getType(), C1->trunc(XWidth));
Value *BinOp = isa<ZExtInst>(BO->getOperand(0))
- ? Builder.CreateBinOp(BO->getOpcode(), X, TruncC1)
- : Builder.CreateBinOp(BO->getOpcode(), TruncC1, X);
+ ? Builder.CreateBinOp(BOpcode, X, TruncC1)
+ : Builder.CreateBinOp(BOpcode, TruncC1, X);
Constant *TruncC = ConstantInt::get(X->getType(), C->trunc(XWidth));
Value *And = Builder.CreateAnd(BinOp, TruncC);
return new ZExtInst(And, Ty);
}
+
+ // Similar to above: if the mask matches the zext input width, then the
+ // 'and' can be eliminated, so we can truncate the other variable op:
+ // and (bo (zext X), Y), C --> zext (bo X, (trunc Y))
+ if (isa<Instruction>(BO->getOperand(0)) &&
+ match(BO->getOperand(0), m_OneUse(m_ZExt(m_Value(X)))) &&
+ C->isMask(X->getType()->getScalarSizeInBits())) {
+ Y = BO->getOperand(1);
+ Value *TrY = Builder.CreateTrunc(Y, X->getType(), Y->getName() + ".tr");
+ Value *NewBO =
+ Builder.CreateBinOp(BOpcode, X, TrY, BO->getName() + ".narrow");
+ return new ZExtInst(NewBO, Ty);
+ }
+ // and (bo Y, (zext X)), C --> zext (bo (trunc Y), X)
+ if (isa<Instruction>(BO->getOperand(1)) &&
+ match(BO->getOperand(1), m_OneUse(m_ZExt(m_Value(X)))) &&
+ C->isMask(X->getType()->getScalarSizeInBits())) {
+ Y = BO->getOperand(0);
+ Value *TrY = Builder.CreateTrunc(Y, X->getType(), Y->getName() + ".tr");
+ Value *NewBO =
+ Builder.CreateBinOp(BOpcode, TrY, X, BO->getName() + ".narrow");
+ return new ZExtInst(NewBO, Ty);
+ }
}
Constant *C1, *C2;
define i32 @lowmask_add_zext(i8 %x, i32 %y) {
; CHECK-LABEL: @lowmask_add_zext(
-; CHECK-NEXT: [[ZX:%.*]] = zext i8 [[X:%.*]] to i32
-; CHECK-NEXT: [[BO:%.*]] = add i32 [[ZX]], [[Y:%.*]]
-; CHECK-NEXT: [[R:%.*]] = and i32 [[BO]], 255
+; CHECK-NEXT: [[Y_TR:%.*]] = trunc i32 [[Y:%.*]] to i8
+; CHECK-NEXT: [[BO_NARROW:%.*]] = add i8 [[Y_TR]], [[X:%.*]]
+; CHECK-NEXT: [[R:%.*]] = zext i8 [[BO_NARROW]] to i32
; CHECK-NEXT: ret i32 [[R]]
;
%zx = zext i8 %x to i32
define i32 @lowmask_add_zext_commute(i16 %x, i32 %p) {
; CHECK-LABEL: @lowmask_add_zext_commute(
; CHECK-NEXT: [[Y:%.*]] = mul i32 [[P:%.*]], [[P]]
-; CHECK-NEXT: [[ZX:%.*]] = zext i16 [[X:%.*]] to i32
-; CHECK-NEXT: [[BO:%.*]] = add i32 [[Y]], [[ZX]]
-; CHECK-NEXT: [[R:%.*]] = and i32 [[BO]], 65535
+; CHECK-NEXT: [[Y_TR:%.*]] = trunc i32 [[Y]] to i16
+; CHECK-NEXT: [[BO_NARROW:%.*]] = add i16 [[Y_TR]], [[X:%.*]]
+; CHECK-NEXT: [[R:%.*]] = zext i16 [[BO_NARROW]] to i32
; CHECK-NEXT: ret i32 [[R]]
;
%y = mul i32 %p, %p ; thwart complexity-based canonicalization
ret i32 %r
}
+; negative test - the mask constant must match the zext source type
+
define i32 @lowmask_add_zext_wrong_mask(i8 %x, i32 %y) {
; CHECK-LABEL: @lowmask_add_zext_wrong_mask(
; CHECK-NEXT: [[ZX:%.*]] = zext i8 [[X:%.*]] to i32
ret i32 %r
}
+; negative test - extra use
+
define i32 @lowmask_add_zext_use1(i8 %x, i32 %y) {
; CHECK-LABEL: @lowmask_add_zext_use1(
; CHECK-NEXT: [[ZX:%.*]] = zext i8 [[X:%.*]] to i32
ret i32 %r
}
+; negative test - extra use
+
define i32 @lowmask_add_zext_use2(i8 %x, i32 %y) {
; CHECK-LABEL: @lowmask_add_zext_use2(
; CHECK-NEXT: [[ZX:%.*]] = zext i8 [[X:%.*]] to i32
ret i32 %r
}
+; vector splats work too
+
define <2 x i32> @lowmask_sub_zext(<2 x i4> %x, <2 x i32> %y) {
; CHECK-LABEL: @lowmask_sub_zext(
-; CHECK-NEXT: [[ZX:%.*]] = zext <2 x i4> [[X:%.*]] to <2 x i32>
-; CHECK-NEXT: [[BO:%.*]] = sub <2 x i32> [[ZX]], [[Y:%.*]]
-; CHECK-NEXT: [[R:%.*]] = and <2 x i32> [[BO]], <i32 15, i32 15>
+; CHECK-NEXT: [[Y_TR:%.*]] = trunc <2 x i32> [[Y:%.*]] to <2 x i4>
+; CHECK-NEXT: [[BO_NARROW:%.*]] = sub <2 x i4> [[X:%.*]], [[Y_TR]]
+; CHECK-NEXT: [[R:%.*]] = zext <2 x i4> [[BO_NARROW]] to <2 x i32>
; CHECK-NEXT: ret <2 x i32> [[R]]
;
%zx = zext <2 x i4> %x to <2 x i32>
ret <2 x i32> %r
}
+; weird types are allowed
+
define i17 @lowmask_sub_zext_commute(i5 %x, i17 %y) {
; CHECK-LABEL: @lowmask_sub_zext_commute(
-; CHECK-NEXT: [[ZX:%.*]] = zext i5 [[X:%.*]] to i17
-; CHECK-NEXT: [[BO:%.*]] = sub i17 [[Y:%.*]], [[ZX]]
-; CHECK-NEXT: [[R:%.*]] = and i17 [[BO]], 31
+; CHECK-NEXT: [[Y_TR:%.*]] = trunc i17 [[Y:%.*]] to i5
+; CHECK-NEXT: [[BO_NARROW:%.*]] = sub i5 [[Y_TR]], [[X:%.*]]
+; CHECK-NEXT: [[R:%.*]] = zext i5 [[BO_NARROW]] to i17
; CHECK-NEXT: ret i17 [[R]]
;
%zx = zext i5 %x to i17
define i32 @lowmask_mul_zext(i8 %x, i32 %y) {
; CHECK-LABEL: @lowmask_mul_zext(
-; CHECK-NEXT: [[ZX:%.*]] = zext i8 [[X:%.*]] to i32
-; CHECK-NEXT: [[BO:%.*]] = mul i32 [[ZX]], [[Y:%.*]]
-; CHECK-NEXT: [[R:%.*]] = and i32 [[BO]], 255
+; CHECK-NEXT: [[Y_TR:%.*]] = trunc i32 [[Y:%.*]] to i8
+; CHECK-NEXT: [[BO_NARROW:%.*]] = mul i8 [[Y_TR]], [[X:%.*]]
+; CHECK-NEXT: [[R:%.*]] = zext i8 [[BO_NARROW]] to i32
; CHECK-NEXT: ret i32 [[R]]
;
%zx = zext i8 %x to i32
ret i32 %r
}
+; TODO: we could have narrowed the xor
+
define i32 @lowmask_xor_zext_commute(i8 %x, i32 %p) {
; CHECK-LABEL: @lowmask_xor_zext_commute(
; CHECK-NEXT: [[Y:%.*]] = mul i32 [[P:%.*]], [[P]]
ret i32 %r
}
+; TODO: we could have narrowed the or
+
define i24 @lowmask_or_zext_commute(i16 %x, i24 %y) {
; CHECK-LABEL: @lowmask_or_zext_commute(
; CHECK-NEXT: [[ZX:%.*]] = zext i16 [[X:%.*]] to i24