forked from OSchip/llvm-project
Revert "[InstCombine] add narrowing transform for low-masked binop with zext operand"
This reverts commit afa192cfb6.
This can cause an infinite loop as shown with an example in the
post-commit thread.
This commit is contained in:
Sanjay Patel
parent
60d6fbb621
commit
6fedc6a2b4
|
|
@ -1869,7 +1869,6 @@ Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {
|
|||
};
|
||||
BinaryOperator *BO;
|
||||
if (match(Op0, m_OneUse(m_BinOp(BO))) && isSuitableBinOpcode(BO)) {
|
||||
Instruction::BinaryOps BOpcode = BO->getOpcode();
|
||||
Value *X;
|
||||
const APInt *C1;
|
||||
// TODO: The one-use restrictions could be relaxed a little if the AND
|
||||
|
|
@ -1879,30 +1878,12 @@ Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {
|
|||
unsigned XWidth = X->getType()->getScalarSizeInBits();
|
||||
Constant *TruncC1 = ConstantInt::get(X->getType(), C1->trunc(XWidth));
|
||||
Value *BinOp = isa<ZExtInst>(BO->getOperand(0))
|
||||
? Builder.CreateBinOp(BOpcode, X, TruncC1)
|
||||
: Builder.CreateBinOp(BOpcode, TruncC1, X);
|
||||
? Builder.CreateBinOp(BO->getOpcode(), X, TruncC1)
|
||||
: Builder.CreateBinOp(BO->getOpcode(), TruncC1, X);
|
||||
Constant *TruncC = ConstantInt::get(X->getType(), C->trunc(XWidth));
|
||||
Value *And = Builder.CreateAnd(BinOp, TruncC);
|
||||
return new ZExtInst(And, Ty);
|
||||
}
|
||||
|
||||
if (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);
|
||||
}
|
||||
|
||||
if (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;
|
||||
|
|
|
|||
|
|
@ -744,9 +744,9 @@ define i64 @test39(i32 %X) {
|
|||
|
||||
define i32 @lowmask_add_zext(i8 %x, i32 %y) {
|
||||
; CHECK-LABEL: @lowmask_add_zext(
|
||||
; 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: [[ZX:%.*]] = zext i8 [[X:%.*]] to i32
|
||||
; CHECK-NEXT: [[BO:%.*]] = add i32 [[ZX]], [[Y:%.*]]
|
||||
; CHECK-NEXT: [[R:%.*]] = and i32 [[BO]], 255
|
||||
; CHECK-NEXT: ret i32 [[R]]
|
||||
;
|
||||
%zx = zext i8 %x to i32
|
||||
|
|
@ -758,9 +758,9 @@ define i32 @lowmask_add_zext(i8 %x, i32 %y) {
|
|||
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: [[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: [[ZX:%.*]] = zext i16 [[X:%.*]] to i32
|
||||
; CHECK-NEXT: [[BO:%.*]] = add i32 [[Y]], [[ZX]]
|
||||
; CHECK-NEXT: [[R:%.*]] = and i32 [[BO]], 65535
|
||||
; CHECK-NEXT: ret i32 [[R]]
|
||||
;
|
||||
%y = mul i32 %p, %p ; thwart complexity-based canonicalization
|
||||
|
|
@ -770,8 +770,6 @@ define i32 @lowmask_add_zext_commute(i16 %x, i32 %p) {
|
|||
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
|
||||
|
|
@ -785,8 +783,6 @@ define i32 @lowmask_add_zext_wrong_mask(i8 %x, i32 %y) {
|
|||
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
|
||||
|
|
@ -802,8 +798,6 @@ define i32 @lowmask_add_zext_use1(i8 %x, i32 %y) {
|
|||
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
|
||||
|
|
@ -819,13 +813,11 @@ define i32 @lowmask_add_zext_use2(i8 %x, i32 %y) {
|
|||
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: [[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: [[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: ret <2 x i32> [[R]]
|
||||
;
|
||||
%zx = zext <2 x i4> %x to <2 x i32>
|
||||
|
|
@ -834,13 +826,11 @@ define <2 x i32> @lowmask_sub_zext(<2 x i4> %x, <2 x i32> %y) {
|
|||
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: [[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: [[ZX:%.*]] = zext i5 [[X:%.*]] to i17
|
||||
; CHECK-NEXT: [[BO:%.*]] = sub i17 [[Y:%.*]], [[ZX]]
|
||||
; CHECK-NEXT: [[R:%.*]] = and i17 [[BO]], 31
|
||||
; CHECK-NEXT: ret i17 [[R]]
|
||||
;
|
||||
%zx = zext i5 %x to i17
|
||||
|
|
@ -851,9 +841,9 @@ define i17 @lowmask_sub_zext_commute(i5 %x, i17 %y) {
|
|||
|
||||
define i32 @lowmask_mul_zext(i8 %x, i32 %y) {
|
||||
; CHECK-LABEL: @lowmask_mul_zext(
|
||||
; 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: [[ZX:%.*]] = zext i8 [[X:%.*]] to i32
|
||||
; CHECK-NEXT: [[BO:%.*]] = mul i32 [[ZX]], [[Y:%.*]]
|
||||
; CHECK-NEXT: [[R:%.*]] = and i32 [[BO]], 255
|
||||
; CHECK-NEXT: ret i32 [[R]]
|
||||
;
|
||||
%zx = zext i8 %x to i32
|
||||
|
|
@ -862,8 +852,6 @@ define i32 @lowmask_mul_zext(i8 %x, i32 %y) {
|
|||
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]]
|
||||
|
|
@ -879,8 +867,6 @@ define i32 @lowmask_xor_zext_commute(i8 %x, i32 %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
|
||||
|
|
|
|||
Loading…
Reference in New Issue