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

[InstCombine] reduce checks for power-of-2-or-zero using ctpop 

This follows up the transform from rL363956 to use the ctpop intrinsic when checking for power-of-2-or-zero.

This is matching the isPowerOf2() patterns used in PR42314:
https://bugs.llvm.org/show_bug.cgi?id=42314

But there's at least 1 instcombine follow-up needed to match the alternate form:

(v & (v - 1)) == 0;

We should have all of the backend expansions handled with:
rL364319
(x86-specific changes still needed for optimal code based on subtarget)

And the larger patterns to exclude zero as a power-of-2 are joining with this change after:
rL364153 ( D63660 )
rL364246

Differential Revision: https://reviews.llvm.org/D63777

llvm-svn: 364341
This commit is contained in:
Sanjay Patel 2019-06-25 18:51:44 +00:00
parent 30519a68d5
commit fcfa056ceb
2 changed files with 38 additions and 59 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
View File

@ -3829,23 +3829,21 @@ Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
return new ICmpInst(Pred, A, B);
// Canonicalize checking for a power-of-2-or-zero value:
// (A & -A) == A --> (A & (A - 1)) == 0
// (-A & A) == A --> (A & (A - 1)) == 0
// A == (A & -A) --> (A & (A - 1)) == 0
// A == (-A & A) --> (A & (A - 1)) == 0
// TODO: This could be reduced by using the ctpop intrinsic.
// (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
// (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
A = nullptr;
if (match(Op0, m_OneUse(m_c_And(m_OneUse(m_Neg(m_Specific(Op1))),
m_Specific(Op1)))))
if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
A = Op1;
else if (match(Op1, m_OneUse(m_c_And(m_OneUse(m_Neg(m_Specific(Op0))),
m_Specific(Op0)))))
else if (match(Op1,
m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
A = Op0;
if (A) {
Type *Ty = A->getType();
Value *Dec = Builder.CreateAdd(A, ConstantInt::getAllOnesValue(Ty));
Value *And = Builder.CreateAnd(A, Dec);
return new ICmpInst(Pred, And, ConstantInt::getNullValue(Ty));
CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
return Pred == ICmpInst::ICMP_EQ
? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
: new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
}
return nullptr;

75
llvm/test/Transforms/InstCombine/ispow2.ll
View File

@ -3,9 +3,8 @@
define i1 @is_pow2or0_negate_op(i32 %x) {
; CHECK-LABEL: @is_pow2or0_negate_op(
; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[X:%.*]], -1
; CHECK-NEXT: [[TMP2:%.*]] = and i32 [[TMP1]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[TMP2]], 0
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range !0
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[TMP1]], 2
; CHECK-NEXT: ret i1 [[CMP]]
;
%neg = sub i32 0, %x
@ -16,9 +15,8 @@ define i1 @is_pow2or0_negate_op(i32 %x) {
define <2 x i1> @is_pow2or0_negate_op_vec(<2 x i32> %x) {
; CHECK-LABEL: @is_pow2or0_negate_op_vec(
; CHECK-NEXT: [[TMP1:%.*]] = add <2 x i32> [[X:%.*]], <i32 -1, i32 -1>
; CHECK-NEXT: [[TMP2:%.*]] = and <2 x i32> [[TMP1]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp eq <2 x i32> [[TMP2]], zeroinitializer
; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> [[X:%.*]])
; CHECK-NEXT: [[CMP:%.*]] = icmp ult <2 x i32> [[TMP1]], <i32 2, i32 2>
; CHECK-NEXT: ret <2 x i1> [[CMP]]
;
%neg = sub <2 x i32> zeroinitializer, %x
@ -55,9 +53,8 @@ define <2 x i1> @is_pow2or0_decrement_op_vec(<2 x i8> %x) {
define i1 @isnot_pow2or0_negate_op(i32 %x) {
; CHECK-LABEL: @isnot_pow2or0_negate_op(
; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[X:%.*]], -1
; CHECK-NEXT: [[TMP2:%.*]] = and i32 [[TMP1]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[TMP2]], 0
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range !0
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP1]], 1
; CHECK-NEXT: ret i1 [[CMP]]
;
%neg = sub i32 0, %x
@ -68,9 +65,8 @@ define i1 @isnot_pow2or0_negate_op(i32 %x) {
define <2 x i1> @isnot_pow2or0_negate_op_vec(<2 x i32> %x) {
; CHECK-LABEL: @isnot_pow2or0_negate_op_vec(
; CHECK-NEXT: [[TMP1:%.*]] = add <2 x i32> [[X:%.*]], <i32 -1, i32 -1>
; CHECK-NEXT: [[TMP2:%.*]] = and <2 x i32> [[TMP1]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp ne <2 x i32> [[TMP2]], zeroinitializer
; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> [[X:%.*]])
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt <2 x i32> [[TMP1]], <i32 1, i32 1>
; CHECK-NEXT: ret <2 x i1> [[CMP]]
;
%neg = sub <2 x i32> zeroinitializer, %x
@ -108,9 +104,8 @@ define <2 x i1> @isnot_pow2or0_decrement_op_vec(<2 x i8> %x) {
define i1 @is_pow2or0_negate_op_commute1(i32 %p) {
; CHECK-LABEL: @is_pow2or0_negate_op_commute1(
; CHECK-NEXT: [[X:%.*]] = srem i32 42, [[P:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = add nsw i32 [[X]], -1
; CHECK-NEXT: [[TMP2:%.*]] = and i32 [[X]], [[TMP1]]
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[TMP2]], 0
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range !1
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[TMP1]], 2
; CHECK-NEXT: ret i1 [[CMP]]
;
%x = srem i32 42, %p ; thwart complexity-based canonicalization
@ -125,9 +120,8 @@ define i1 @is_pow2or0_negate_op_commute1(i32 %p) {
define i1 @isnot_pow2or0_negate_op_commute2(i32 %p) {
; CHECK-LABEL: @isnot_pow2or0_negate_op_commute2(
; CHECK-NEXT: [[X:%.*]] = urem i32 42, [[P:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = add nsw i32 [[X]], -1
; CHECK-NEXT: [[TMP2:%.*]] = and i32 [[X]], [[TMP1]]
; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[TMP2]], 0
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range !2
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP1]], 1
; CHECK-NEXT: ret i1 [[CMP]]
;
%x = urem i32 42, %p ; thwart complexity-based canonicalization
@ -140,9 +134,8 @@ define i1 @isnot_pow2or0_negate_op_commute2(i32 %p) {
define i1 @isnot_pow2or0_negate_op_commute3(i32 %p) {
; CHECK-LABEL: @isnot_pow2or0_negate_op_commute3(
; CHECK-NEXT: [[X:%.*]] = urem i32 42, [[P:%.*]]
; CHECK-NEXT: [[TMP1:%.*]] = add nsw i32 [[X]], -1
; CHECK-NEXT: [[TMP2:%.*]] = and i32 [[X]], [[TMP1]]
; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[TMP2]], 0
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range !2
; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP1]], 1
; CHECK-NEXT: ret i1 [[CMP]]
;
%x = urem i32 42, %p ; thwart complexity-based canonicalization
@ -158,8 +151,8 @@ define i1 @is_pow2or0_negate_op_extra_use1(i32 %x) {
; CHECK-LABEL: @is_pow2or0_negate_op_extra_use1(
; CHECK-NEXT: [[NEG:%.*]] = sub i32 0, [[X:%.*]]
; CHECK-NEXT: call void @use(i32 [[NEG]])
; CHECK-NEXT: [[AND:%.*]] = and i32 [[NEG]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[AND]], [[X]]
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range !0
; CHECK-NEXT: [[CMP:%.*]] = icmp ult i32 [[TMP1]], 2
; CHECK-NEXT: ret i1 [[CMP]]
;
%neg = sub i32 0, %x
@ -428,12 +421,9 @@ define i1 @isnot_pow2_ctpop_wrong_pred2(i32 %x) {
define i1 @is_pow2_negate_op(i32 %x) {
; CHECK-LABEL: @is_pow2_negate_op(
; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[X:%.*]], -1
; CHECK-NEXT: [[TMP2:%.*]] = and i32 [[TMP1]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[TMP2]], 0
; CHECK-NEXT: [[NOTZERO:%.*]] = icmp ne i32 [[X]], 0
; CHECK-NEXT: [[R:%.*]] = and i1 [[NOTZERO]], [[CMP]]
; CHECK-NEXT: ret i1 [[R]]
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range !0
; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i32 [[TMP1]], 1
; CHECK-NEXT: ret i1 [[TMP2]]
;
%neg = sub i32 0, %x
%and = and i32 %neg, %x
@ -445,12 +435,9 @@ define i1 @is_pow2_negate_op(i32 %x) {
define <2 x i1> @is_pow2_negate_op_vec(<2 x i32> %x) {
; CHECK-LABEL: @is_pow2_negate_op_vec(
; CHECK-NEXT: [[TMP1:%.*]] = add <2 x i32> [[X:%.*]], <i32 -1, i32 -1>
; CHECK-NEXT: [[TMP2:%.*]] = and <2 x i32> [[TMP1]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp eq <2 x i32> [[TMP2]], zeroinitializer
; CHECK-NEXT: [[NOTZERO:%.*]] = icmp ne <2 x i32> [[X]], zeroinitializer
; CHECK-NEXT: [[R:%.*]] = and <2 x i1> [[CMP]], [[NOTZERO]]
; CHECK-NEXT: ret <2 x i1> [[R]]
; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> [[X:%.*]])
; CHECK-NEXT: [[TMP2:%.*]] = icmp eq <2 x i32> [[TMP1]], <i32 1, i32 1>
; CHECK-NEXT: ret <2 x i1> [[TMP2]]
;
%neg = sub <2 x i32> zeroinitializer, %x
%and = and <2 x i32> %neg, %x
@ -496,12 +483,9 @@ define <2 x i1> @is_pow2_decrement_op_vec(<2 x i8> %x) {
define i1 @isnot_pow2_negate_op(i32 %x) {
; CHECK-LABEL: @isnot_pow2_negate_op(
; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[X:%.*]], -1
; CHECK-NEXT: [[TMP2:%.*]] = and i32 [[TMP1]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[TMP2]], 0
; CHECK-NEXT: [[ISZERO:%.*]] = icmp eq i32 [[X]], 0
; CHECK-NEXT: [[R:%.*]] = or i1 [[CMP]], [[ISZERO]]
; CHECK-NEXT: ret i1 [[R]]
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range !0
; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 1
; CHECK-NEXT: ret i1 [[TMP2]]
;
%neg = sub i32 0, %x
%and = and i32 %neg, %x
@ -513,12 +497,9 @@ define i1 @isnot_pow2_negate_op(i32 %x) {
define <2 x i1> @isnot_pow2_negate_op_vec(<2 x i32> %x) {
; CHECK-LABEL: @isnot_pow2_negate_op_vec(
; CHECK-NEXT: [[TMP1:%.*]] = add <2 x i32> [[X:%.*]], <i32 -1, i32 -1>
; CHECK-NEXT: [[TMP2:%.*]] = and <2 x i32> [[TMP1]], [[X]]
; CHECK-NEXT: [[CMP:%.*]] = icmp ne <2 x i32> [[TMP2]], zeroinitializer
; CHECK-NEXT: [[ISZERO:%.*]] = icmp eq <2 x i32> [[X]], zeroinitializer
; CHECK-NEXT: [[R:%.*]] = or <2 x i1> [[ISZERO]], [[CMP]]
; CHECK-NEXT: ret <2 x i1> [[R]]
; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> [[X:%.*]])
; CHECK-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[TMP1]], <i32 1, i32 1>
; CHECK-NEXT: ret <2 x i1> [[TMP2]]
;
%neg = sub <2 x i32> zeroinitializer, %x
%and = and <2 x i32> %neg, %x