diff --git a/llvm/lib/Analysis/InstructionSimplify.cpp b/llvm/lib/Analysis/InstructionSimplify.cpp --- a/llvm/lib/Analysis/InstructionSimplify.cpp +++ b/llvm/lib/Analysis/InstructionSimplify.cpp @@ -1803,6 +1803,27 @@ return nullptr; } +/// Try to simplify and/or of icmp with ctpop intrinsic. +static Value *simplifyAndOrOfICmpsWithCtpop(ICmpInst *Cmp0, ICmpInst *Cmp1, + bool IsAnd) { + ICmpInst::Predicate Pred0, Pred1; + Value *X; + const APInt *C; + if (!match(Cmp0, m_ICmp(Pred0, m_Intrinsic(m_Value(X)), + m_APInt(C))) || + !match(Cmp1, m_ICmp(Pred1, m_Specific(X), m_ZeroInt())) || C->isZero()) + return nullptr; + + // (ctpop(X) == C) || (X != 0) --> X != 0 where C > 0 + if (!IsAnd && Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_NE) + return Cmp1; + // (ctpop(X) != C) && (X == 0) --> X == 0 where C > 0 + if (IsAnd && Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_EQ) + return Cmp1; + + return nullptr; +} + static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1, const SimplifyQuery &Q) { if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true, Q)) @@ -1824,6 +1845,11 @@ if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true)) return X; + if (Value *X = simplifyAndOrOfICmpsWithCtpop(Op0, Op1, true)) + return X; + if (Value *X = simplifyAndOrOfICmpsWithCtpop(Op1, Op0, true)) + return X; + if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, Q.IIQ)) return X; if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, Q.IIQ)) @@ -1900,6 +1926,11 @@ if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false)) return X; + if (Value *X = simplifyAndOrOfICmpsWithCtpop(Op0, Op1, false)) + return X; + if (Value *X = simplifyAndOrOfICmpsWithCtpop(Op1, Op0, false)) + return X; + if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, Q.IIQ)) return X; if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, Q.IIQ)) diff --git a/llvm/test/Transforms/InstCombine/ispow2.ll b/llvm/test/Transforms/InstCombine/ispow2.ll --- a/llvm/test/Transforms/InstCombine/ispow2.ll +++ b/llvm/test/Transforms/InstCombine/ispow2.ll @@ -899,11 +899,8 @@ define i1 @is_pow2or0_ctpop_wrong_pred2_logical(i32 %x) { ; CHECK-LABEL: @is_pow2or0_ctpop_wrong_pred2_logical( -; CHECK-NEXT: [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]] -; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[T0]], 1 -; CHECK-NEXT: [[ISZERO:%.*]] = icmp ne i32 [[X]], 0 -; CHECK-NEXT: [[R:%.*]] = or i1 [[ISZERO]], [[CMP]] -; CHECK-NEXT: ret i1 [[R]] +; CHECK-NEXT: [[ISZERO:%.*]] = icmp ne i32 [[X:%.*]], 0 +; CHECK-NEXT: ret i1 [[ISZERO]] ; %t0 = tail call i32 @llvm.ctpop.i32(i32 %x) %cmp = icmp eq i32 %t0, 1 @@ -1091,11 +1088,8 @@ define i1 @isnot_pow2nor0_ctpop_wrong_pred2_logical(i32 %x) { ; CHECK-LABEL: @isnot_pow2nor0_ctpop_wrong_pred2_logical( -; CHECK-NEXT: [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]] -; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[T0]], 1 -; CHECK-NEXT: [[NOTZERO:%.*]] = icmp eq i32 [[X]], 0 -; CHECK-NEXT: [[R:%.*]] = and i1 [[NOTZERO]], [[CMP]] -; CHECK-NEXT: ret i1 [[R]] +; CHECK-NEXT: [[NOTZERO:%.*]] = icmp eq i32 [[X:%.*]], 0 +; CHECK-NEXT: ret i1 [[NOTZERO]] ; %t0 = tail call i32 @llvm.ctpop.i32(i32 %x) %cmp = icmp ne i32 %t0, 1 diff --git a/llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll b/llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll new file mode 100644 --- /dev/null +++ b/llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll @@ -0,0 +1,102 @@ +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py +; RUN: opt < %s -passes=instsimplify -S | FileCheck %s + +;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; +; +; (ctpop(X) == N) || (X != 0) --> X != 0 +; where N > 0 +; +;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + +declare i32 @llvm.ctpop.i32(i32) + +define i1 @eq_or_non_0(i32 %x) { +; CHECK-LABEL: @eq_or_non_0( +; CHECK-NEXT: [[NOTZERO:%.*]] = icmp ne i32 [[X:%.*]], 0 +; CHECK-NEXT: ret i1 [[NOTZERO]] +; + %t0 = tail call i32 @llvm.ctpop.i32(i32 %x) + %cmp = icmp eq i32 %t0, 10 + %notzero = icmp ne i32 %x, 0 + %r = or i1 %notzero, %cmp + ret i1 %r +} + +declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32>) + +define <2 x i1> @eq_or_non_0_commute(<2 x i32> %x) { +; CHECK-LABEL: @eq_or_non_0_commute( +; CHECK-NEXT: [[NOTZERO:%.*]] = icmp ne <2 x i32> [[X:%.*]], zeroinitializer +; CHECK-NEXT: ret <2 x i1> [[NOTZERO]] +; + %t0 = tail call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> %x) + %cmp = icmp eq <2 x i32> %t0, + %notzero = icmp ne <2 x i32> %x, + %r = or <2 x i1> %cmp, %notzero + ret <2 x i1> %r +} + +; Negative test - wrong predicate + +define i1 @eq_or_non_0_wrong_pred1(i32 %x) { +; CHECK-LABEL: @eq_or_non_0_wrong_pred1( +; CHECK-NEXT: [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]) +; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[T0]], 10 +; CHECK-NEXT: [[NOTZERO:%.*]] = icmp ne i32 [[X]], 0 +; CHECK-NEXT: [[R:%.*]] = or i1 [[NOTZERO]], [[CMP]] +; CHECK-NEXT: ret i1 [[R]] +; + %t0 = tail call i32 @llvm.ctpop.i32(i32 %x) + %cmp = icmp ne i32 %t0, 10 + %notzero = icmp ne i32 %x, 0 + %r = or i1 %notzero, %cmp + ret i1 %r +} + +;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; +; +; (ctpop(X) != N) && (X == 0) --> X == 0 +; where N > 0 +; +;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + +define i1 @ne_and_is_0(i32 %x) { +; CHECK-LABEL: @ne_and_is_0( +; CHECK-NEXT: [[ISZERO:%.*]] = icmp eq i32 [[X:%.*]], 0 +; CHECK-NEXT: ret i1 [[ISZERO]] +; + %t0 = tail call i32 @llvm.ctpop.i32(i32 %x) + %cmp = icmp ne i32 %t0, 10 + %iszero = icmp eq i32 %x, 0 + %r = and i1 %iszero, %cmp + ret i1 %r +} + +define <2 x i1> @ne_and_is_0_commute(<2 x i32> %x) { +; CHECK-LABEL: @ne_and_is_0_commute( +; CHECK-NEXT: [[ISZERO:%.*]] = icmp eq <2 x i32> [[X:%.*]], zeroinitializer +; CHECK-NEXT: ret <2 x i1> [[ISZERO]] +; + %t0 = tail call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> %x) + %cmp = icmp ne <2 x i32> %t0, + %iszero = icmp eq <2 x i32> %x, + %r = and <2 x i1> %cmp, %iszero + ret <2 x i1> %r +} + +; Negative test - wrong predicate + +define i1 @ne_and_is_0_wrong_pred1(i32 %x) { +; CHECK-LABEL: @ne_and_is_0_wrong_pred1( +; CHECK-NEXT: [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]) +; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[T0]], 10 +; CHECK-NEXT: [[ISZERO:%.*]] = icmp eq i32 [[X]], 0 +; CHECK-NEXT: [[R:%.*]] = or i1 [[ISZERO]], [[CMP]] +; CHECK-NEXT: ret i1 [[R]] +; + %t0 = tail call i32 @llvm.ctpop.i32(i32 %x) + %cmp = icmp ne i32 %t0, 10 + %iszero = icmp eq i32 %x, 0 + %r = or i1 %iszero, %cmp + ret i1 %r +}