Skip to content

[InstCombine] Improve eq/ne by parts to handle ult/ugt equality pattern #69884

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 2 commits into from
Closed

[InstCombine] Improve eq/ne by parts to handle ult/ugt equality pattern #69884

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8208f91
[InstCombine] Add tests for new eq/ne patterns combining eq/ne by par…
goldsteinn Oct 22, 2023
1834ec7
[InstCombine] Improve eq/ne by parts to handle `ult/ugt` equality pat…
goldsteinn Oct 22, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
39 changes: 33 additions & 6 deletions llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1146,13 +1146,40 @@ Value *InstCombinerImpl::foldEqOfParts(ICmpInst *Cmp0, ICmpInst *Cmp1,
return nullptr;

CmpInst::Predicate Pred = IsAnd ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
if (Cmp0->getPredicate() != Pred || Cmp1->getPredicate() != Pred)
return nullptr;
auto GetMatchPart = [&](ICmpInst *Cmp,
unsigned OpNo) -> std::optional<IntPart> {
if (Pred == Cmp->getPredicate())
return matchIntPart(Cmp->getOperand(OpNo));

const APInt *C;
// (icmp eq (lshr x, C), (lshr y, C)) gets optimized to:
// (icmp ult (xor x, y), 1 << C) so also look for that.
if (Pred == CmpInst::ICMP_EQ && Cmp->getPredicate() == CmpInst::ICMP_ULT) {
if (!match(Cmp->getOperand(1), m_Power2(C)) ||
!match(Cmp->getOperand(0), m_Xor(m_Value(), m_Value())))
return std::nullopt;
}

// (icmp ne (lshr x, C), (lshr y, C)) gets optimized to:
// (icmp ugt (xor x, y), (1 << C) - 1) so also look for that.
else if (Pred == CmpInst::ICMP_NE &&
Cmp->getPredicate() == CmpInst::ICMP_UGT) {
if (!match(Cmp->getOperand(1), m_LowBitMask(C)) ||
!match(Cmp->getOperand(0), m_Xor(m_Value(), m_Value())))
return std::nullopt;
} else {
return std::nullopt;
}

unsigned From = Pred == CmpInst::ICMP_NE ? C->popcount() : C->countr_zero();
Instruction *I = cast<Instruction>(Cmp->getOperand(0));
return {{I->getOperand(OpNo), From, C->getBitWidth() - From}};
};

std::optional<IntPart> L0 = matchIntPart(Cmp0->getOperand(0));
std::optional<IntPart> R0 = matchIntPart(Cmp0->getOperand(1));
std::optional<IntPart> L1 = matchIntPart(Cmp1->getOperand(0));
std::optional<IntPart> R1 = matchIntPart(Cmp1->getOperand(1));
std::optional<IntPart> L0 = GetMatchPart(Cmp0, 0);
std::optional<IntPart> R0 = GetMatchPart(Cmp0, 1);
std::optional<IntPart> L1 = GetMatchPart(Cmp1, 0);
std::optional<IntPart> R1 = GetMatchPart(Cmp1, 1);
if (!L0 || !R0 || !L1 || !R1)
return nullptr;

Expand Down
123 changes: 123 additions & 0 deletions llvm/test/Transforms/InstCombine/eq-of-parts.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1333,3 +1333,126 @@ define i1 @ne_21_wrong_pred2(i32 %x, i32 %y) {
%c.210 = or i1 %c.2, %c.1
ret i1 %c.210
}

define i1 @eq_optimized_highbits_cmp(i32 %x, i32 %y) {
; CHECK-LABEL: @eq_optimized_highbits_cmp(
; CHECK-NEXT: [[R:%.*]] = icmp eq i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ult i32 %xor, 33554432
%tx = trunc i32 %x to i25
%ty = trunc i32 %y to i25
%cmp_lo = icmp eq i25 %tx, %ty
%r = and i1 %cmp_hi, %cmp_lo
ret i1 %r
}

define i1 @eq_optimized_highbits_cmp_todo_overlapping(i32 %x, i32 %y) {
; CHECK-LABEL: @eq_optimized_highbits_cmp_todo_overlapping(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ult i32 [[XOR]], 16777216
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i25
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i25
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp eq i25 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = and i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ult i32 %xor, 16777216
%tx = trunc i32 %x to i25
%ty = trunc i32 %y to i25
%cmp_lo = icmp eq i25 %tx, %ty
%r = and i1 %cmp_hi, %cmp_lo
ret i1 %r
}

define i1 @eq_optimized_highbits_cmp_fail_not_pow2(i32 %x, i32 %y) {
; CHECK-LABEL: @eq_optimized_highbits_cmp_fail_not_pow2(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ult i32 [[XOR]], 16777215
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i24
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i24
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp eq i24 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = and i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ult i32 %xor, 16777215
%tx = trunc i32 %x to i24
%ty = trunc i32 %y to i24
%cmp_lo = icmp eq i24 %tx, %ty
%r = and i1 %cmp_hi, %cmp_lo
ret i1 %r
}

define i1 @ne_optimized_highbits_cmp(i32 %x, i32 %y) {
; CHECK-LABEL: @ne_optimized_highbits_cmp(
; CHECK-NEXT: [[R:%.*]] = icmp ne i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ugt i32 %xor, 16777215
%tx = trunc i32 %x to i24
%ty = trunc i32 %y to i24
%cmp_lo = icmp ne i24 %tx, %ty
%r = or i1 %cmp_hi, %cmp_lo
ret i1 %r
}

define i1 @ne_optimized_highbits_cmp_fail_not_mask(i32 %x, i32 %y) {
; CHECK-LABEL: @ne_optimized_highbits_cmp_fail_not_mask(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ugt i32 [[XOR]], 16777216
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i24
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i24
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp ne i24 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = or i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ugt i32 %xor, 16777216
%tx = trunc i32 %x to i24
%ty = trunc i32 %y to i24
%cmp_lo = icmp ne i24 %tx, %ty
%r = or i1 %cmp_hi, %cmp_lo
ret i1 %r
}

define i1 @ne_optimized_highbits_cmp_fail_no_combined_int(i32 %x, i32 %y) {
; CHECK-LABEL: @ne_optimized_highbits_cmp_fail_no_combined_int(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ugt i32 [[XOR]], 16777215
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i23
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i23
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp ne i23 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = or i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ugt i32 %xor, 16777215
%tx = trunc i32 %x to i23
%ty = trunc i32 %y to i23
%cmp_lo = icmp ne i23 %tx, %ty
%r = or i1 %cmp_hi, %cmp_lo
ret i1 %r
}

define i1 @ne_optimized_highbits_cmp_todo_overlapping(i32 %x, i32 %y) {
; CHECK-LABEL: @ne_optimized_highbits_cmp_todo_overlapping(
; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[CMP_HI:%.*]] = icmp ugt i32 [[XOR]], 8388607
; CHECK-NEXT: [[TX:%.*]] = trunc i32 [[X]] to i24
; CHECK-NEXT: [[TY:%.*]] = trunc i32 [[Y]] to i24
; CHECK-NEXT: [[CMP_LO:%.*]] = icmp ne i24 [[TX]], [[TY]]
; CHECK-NEXT: [[R:%.*]] = or i1 [[CMP_HI]], [[CMP_LO]]
; CHECK-NEXT: ret i1 [[R]]
;
%xor = xor i32 %y, %x
%cmp_hi = icmp ugt i32 %xor, 8388607
%tx = trunc i32 %x to i24
%ty = trunc i32 %y to i24
%cmp_lo = icmp ne i24 %tx, %ty
%r = or i1 %cmp_hi, %cmp_lo
ret i1 %r
}