https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/174101
>From 4c4344feefa0084aecfc61b22fa27c5b673e0849 Mon Sep 17 00:00:00 2001 From: Matt Arsenault <[email protected]> Date: Wed, 31 Dec 2025 12:13:26 +0100 Subject: [PATCH] InstCombine: Rudimentary support of shufflevector in SimplifyDemandedFPClass This should look more like the computeKnownFPClass handling, with knowledge of demanded vector elements. --- .../InstCombineSimplifyDemanded.cpp | 11 + ...simplify-demanded-fpclass-shufflevector.ll | 269 ++++++++++++++++++ 2 files changed, 280 insertions(+) create mode 100644 llvm/test/Transforms/InstCombine/simplify-demanded-fpclass-shufflevector.ll diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp index 9ae6ab25c6033..b71133a7b6470 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp @@ -2313,6 +2313,17 @@ Value *InstCombinerImpl::SimplifyDemandedUseFPClass(Value *V, Known = KnownVec | KnownInserted; break; } + case Instruction::ShuffleVector: { + KnownFPClass KnownLHS, KnownRHS; + if (SimplifyDemandedFPClass(I, 1, DemandedMask, KnownRHS, Depth + 1) || + SimplifyDemandedFPClass(I, 0, DemandedMask, KnownLHS, Depth + 1)) + return I; + + // TODO: This is overly conservative and should consider demanded elements, + // and splats. + Known = KnownLHS | KnownRHS; + break; + } default: Known = computeKnownFPClass(I, DemandedMask, CxtI, Depth + 1); break; diff --git a/llvm/test/Transforms/InstCombine/simplify-demanded-fpclass-shufflevector.ll b/llvm/test/Transforms/InstCombine/simplify-demanded-fpclass-shufflevector.ll new file mode 100644 index 0000000000000..042e77c3ee2a9 --- /dev/null +++ b/llvm/test/Transforms/InstCombine/simplify-demanded-fpclass-shufflevector.ll @@ -0,0 +1,269 @@ +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6 +; RUN: opt -S -passes=instcombine < %s | FileCheck %s + +declare nofpclass(inf zero sub norm) <4 x half> @returns_nan() +declare nofpclass(nan zero sub norm) <4 x half> @returns_inf() +declare nofpclass(inf zero sub norm) half @returns_nan_f16() + +define nofpclass(inf zero sub norm) <4 x half> @ret_only_nan__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(inf zero sub norm) <4 x half> @ret_only_nan__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[VEC0]], <4 x half> [[VEC1]], <4 x i32> <i32 6, i32 2, i32 3, i32 0> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(qnan inf zero sub norm) <4 x half> @ret_only_snan__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(qnan inf zero sub norm) <4 x half> @ret_only_snan__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[VEC0]], <4 x half> [[VEC1]], <4 x i32> <i32 6, i32 2, i32 3, i32 0> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(snan inf zero sub norm) <4 x half> @ret_only_qnan__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(snan inf zero sub norm) <4 x half> @ret_only_qnan__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[VEC0]], <4 x half> [[VEC1]], <4 x i32> <i32 6, i32 2, i32 3, i32 0> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan inf sub norm) <4 x half> @ret_only_zero__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(nan inf sub norm) <4 x half> @ret_only_zero__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[VEC0]], <4 x half> [[VEC1]], <4 x i32> <i32 6, i32 2, i32 3, i32 0> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan inf nzero sub norm) <4 x half> @ret_only_pzero__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(nan inf nzero sub norm) <4 x half> @ret_only_pzero__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: ret <4 x half> zeroinitializer +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan inf pzero sub norm) <4 x half> @ret_only_nzero__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(nan inf pzero sub norm) <4 x half> @ret_only_nzero__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: ret <4 x half> splat (half 0xH8000) +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan zero sub norm) <4 x half> @ret_only_inf__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(nan zero sub norm) <4 x half> @ret_only_inf__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[VEC0]], <4 x half> [[VEC1]], <4 x i32> <i32 6, i32 2, i32 3, i32 0> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan ninf zero sub norm) <4 x half> @ret_only_pinf__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(nan ninf zero sub norm) <4 x half> @ret_only_pinf__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: ret <4 x half> splat (half 0xH7C00) +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan pinf zero sub norm) <4 x half> @ret_only_ninf__shufflevector_unknown(<4 x half> %vec0, <4 x half> %vec1) { +; CHECK-LABEL: define nofpclass(nan pinf zero sub norm) <4 x half> @ret_only_ninf__shufflevector_unknown( +; CHECK-SAME: <4 x half> [[VEC0:%.*]], <4 x half> [[VEC1:%.*]]) { +; CHECK-NEXT: ret <4 x half> splat (half 0xHFC00) +; + %shuffle = shufflevector <4 x half> %vec0, <4 x half> %vec1, <4 x i32> <i32 6, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan) <4 x half> @ret_no_nan__shufflevector_one_input_select_nan(i1 %cond, <4 x half> %unknown) { +; CHECK-LABEL: define nofpclass(nan) <4 x half> @ret_no_nan__shufflevector_one_input_select_nan( +; CHECK-SAME: i1 [[COND:%.*]], <4 x half> [[UNKNOWN:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call <4 x half> @returns_nan() +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[UNKNOWN]], <4 x half> poison, <4 x i32> <i32 3, i32 1, i32 0, i32 0> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %nan = call <4 x half> @returns_nan() + %select = select i1 %cond, <4 x half> %nan, <4 x half> %unknown + %shuffle = shufflevector <4 x half> %select, <4 x half> poison, <4 x i32> <i32 3, i32 1, i32 0, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan) <4 x half> @ret_no_nan__shufflevector_two_input_select_nans(i1 %cond, <4 x half> %unknown0, <4 x half> %unknown1) { +; CHECK-LABEL: define nofpclass(nan) <4 x half> @ret_no_nan__shufflevector_two_input_select_nans( +; CHECK-SAME: i1 [[COND:%.*]], <4 x half> [[UNKNOWN0:%.*]], <4 x half> [[UNKNOWN1:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call <4 x half> @returns_nan() +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[UNKNOWN0]], <4 x half> [[UNKNOWN1]], <4 x i32> <i32 4, i32 7, i32 1, i32 3> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %nan = call <4 x half> @returns_nan() + %select0 = select i1 %cond, <4 x half> %nan, <4 x half> %unknown0 + %select1 = select i1 %cond, <4 x half> %unknown1, <4 x half> %nan + %shuffle = shufflevector <4 x half> %select0, <4 x half> %select1, <4 x i32> <i32 4, i32 7, i32 1, i32 3> + ret <4 x half> %shuffle +} + +define nofpclass(nan) <4 x half> @ret_no_nan__shufflevector_two_input_eliminate_vector0(<4 x half> %unknown) { +; CHECK-LABEL: define nofpclass(nan) <4 x half> @ret_no_nan__shufflevector_two_input_eliminate_vector0( +; CHECK-SAME: <4 x half> [[UNKNOWN:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call <4 x half> @returns_nan() +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[UNKNOWN]], <4 x half> poison, <4 x i32> <i32 0, i32 3, i32 poison, i32 poison> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %nan = call <4 x half> @returns_nan() + %shuffle = shufflevector <4 x half> %nan, <4 x half> %unknown, <4 x i32> <i32 4, i32 7, i32 1, i32 3> + ret <4 x half> %shuffle +} + +define nofpclass(nan) <4 x half> @ret_no_nan__shufflevector_two_input_eliminate_vector1(<4 x half> %unknown) { +; CHECK-LABEL: define nofpclass(nan) <4 x half> @ret_no_nan__shufflevector_two_input_eliminate_vector1( +; CHECK-SAME: <4 x half> [[UNKNOWN:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call <4 x half> @returns_nan() +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[UNKNOWN]], <4 x half> poison, <4 x i32> <i32 poison, i32 poison, i32 1, i32 3> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %nan = call <4 x half> @returns_nan() + %shuffle = shufflevector <4 x half> %unknown, <4 x half> %nan, <4 x i32> <i32 4, i32 7, i32 1, i32 3> + ret <4 x half> %shuffle +} + +; Cannot simplify +define nofpclass(inf nan) <4 x half> @ret_nonan_noinf___shuffle_not_nan_not_inf(i1 %cond, <4 x half> nofpclass(nan) %not.nan, <4 x half> %not.inf) { +; CHECK-LABEL: define nofpclass(nan inf) <4 x half> @ret_nonan_noinf___shuffle_not_nan_not_inf( +; CHECK-SAME: i1 [[COND:%.*]], <4 x half> nofpclass(nan) [[NOT_NAN:%.*]], <4 x half> [[NOT_INF:%.*]]) { +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[NOT_NAN]], <4 x half> [[NOT_INF]], <4 x i32> <i32 4, i32 2, i32 3, i32 0> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %shuffle = shufflevector <4 x half> %not.nan, <4 x half> %not.inf, <4 x i32> <i32 4, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +; poison +define nofpclass(inf nan) <4 x half> @ret_nonan_noinf__shufflevector_nan_inf(i1 %cond) { +; CHECK-LABEL: define nofpclass(nan inf) <4 x half> @ret_nonan_noinf__shufflevector_nan_inf( +; CHECK-SAME: i1 [[COND:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call <4 x half> @returns_nan() +; CHECK-NEXT: [[INF:%.*]] = call <4 x half> @returns_inf() +; CHECK-NEXT: ret <4 x half> <half undef, half poison, half poison, half poison> +; + %nan = call <4 x half> @returns_nan() + %inf = call <4 x half> @returns_inf() + %shuffle = shufflevector <4 x half> %inf, <4 x half> %nan, <4 x i32> <i32 4, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(ninf nnorm nsub nzero) <4 x half> @ret_positives___shuffle_fabs_positive_or_nans(i1 %cond, <4 x half> nofpclass(ninf nnorm nsub nzero) %known.positive0, <4 x half> nofpclass(ninf nnorm nsub nzero) %known.positive1) { +; CHECK-LABEL: define nofpclass(ninf nzero nsub nnorm) <4 x half> @ret_positives___shuffle_fabs_positive_or_nans( +; CHECK-SAME: i1 [[COND:%.*]], <4 x half> nofpclass(ninf nzero nsub nnorm) [[KNOWN_POSITIVE0:%.*]], <4 x half> nofpclass(ninf nzero nsub nnorm) [[KNOWN_POSITIVE1:%.*]]) { +; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x half> [[KNOWN_POSITIVE0]], <4 x half> [[KNOWN_POSITIVE1]], <4 x i32> <i32 4, i32 2, i32 3, i32 0> +; CHECK-NEXT: [[SHUFFLE:%.*]] = call <4 x half> @llvm.fabs.v4f16(<4 x half> [[TMP1]]) +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %fabs.0 = call <4 x half> @llvm.fabs.v4f16(<4 x half> %known.positive0) + %fabs.1 = call <4 x half> @llvm.fabs.v4f16(<4 x half> %known.positive1) + %shuffle = shufflevector <4 x half> %fabs.0, <4 x half> %fabs.1, <4 x i32> <i32 4, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(ninf nnorm nsub nzero) <4 x half> @ret_positives___shuffle_fabs_positives(i1 %cond, <4 x half> nofpclass(nan ninf nnorm nsub nzero) %known.positive0, <4 x half> nofpclass(nan ninf nnorm nsub nzero) %known.positive1) { +; CHECK-LABEL: define nofpclass(ninf nzero nsub nnorm) <4 x half> @ret_positives___shuffle_fabs_positives( +; CHECK-SAME: i1 [[COND:%.*]], <4 x half> nofpclass(nan ninf nzero nsub nnorm) [[KNOWN_POSITIVE0:%.*]], <4 x half> nofpclass(nan ninf nzero nsub nnorm) [[KNOWN_POSITIVE1:%.*]]) { +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[KNOWN_POSITIVE0]], <4 x half> [[KNOWN_POSITIVE1]], <4 x i32> <i32 4, i32 2, i32 3, i32 0> +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %fabs.0 = call <4 x half> @llvm.fabs.v4f16(<4 x half> %known.positive0) + %fabs.1 = call <4 x half> @llvm.fabs.v4f16(<4 x half> %known.positive1) + %shuffle = shufflevector <4 x half> %fabs.0, <4 x half> %fabs.1, <4 x i32> <i32 4, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +; TODO: Should be able to eliminate fabs with no-nans from return +define nofpclass(nan ninf nnorm nsub nzero) <4 x half> @ret_positives_non_nan___shuffle_fabs(i1 %cond, <4 x half> nofpclass(ninf nnorm nsub nzero) %known.positive0, <4 x half> nofpclass(ninf nnorm nsub nzero) %known.positive1) { +; CHECK-LABEL: define nofpclass(nan ninf nzero nsub nnorm) <4 x half> @ret_positives_non_nan___shuffle_fabs( +; CHECK-SAME: i1 [[COND:%.*]], <4 x half> nofpclass(ninf nzero nsub nnorm) [[KNOWN_POSITIVE0:%.*]], <4 x half> nofpclass(ninf nzero nsub nnorm) [[KNOWN_POSITIVE1:%.*]]) { +; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x half> [[KNOWN_POSITIVE0]], <4 x half> [[KNOWN_POSITIVE1]], <4 x i32> <i32 4, i32 2, i32 3, i32 0> +; CHECK-NEXT: [[SHUFFLE:%.*]] = call <4 x half> @llvm.fabs.v4f16(<4 x half> [[TMP1]]) +; CHECK-NEXT: ret <4 x half> [[SHUFFLE]] +; + %fabs.0 = call <4 x half> @llvm.fabs.v4f16(<4 x half> %known.positive0) + %fabs.1 = call <4 x half> @llvm.fabs.v4f16(<4 x half> %known.positive1) + %shuffle = shufflevector <4 x half> %fabs.0, <4 x half> %fabs.1, <4 x i32> <i32 4, i32 2, i32 3, i32 0> + ret <4 x half> %shuffle +} + +define nofpclass(nan) <4 x half> @ret_nonan_splat(i1 %cond, half %unknown) { +; CHECK-LABEL: define nofpclass(nan) <4 x half> @ret_nonan_splat( +; CHECK-SAME: i1 [[COND:%.*]], half [[UNKNOWN:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call half @returns_nan_f16() +; CHECK-NEXT: [[UNKNOWN_VEC:%.*]] = insertelement <4 x half> poison, half [[UNKNOWN]], i64 0 +; CHECK-NEXT: [[SPLAT_UNKNOWN:%.*]] = shufflevector <4 x half> [[UNKNOWN_VEC]], <4 x half> poison, <4 x i32> zeroinitializer +; CHECK-NEXT: ret <4 x half> [[SPLAT_UNKNOWN]] +; + %nan = call half @returns_nan_f16() + %nan.vec = insertelement <4 x half> poison, half %nan, i32 0 + %unknown.vec = insertelement <4 x half> poison, half %unknown, i32 0 + %splat.nan = shufflevector <4 x half> %nan.vec, <4 x half> poison, <4 x i32> zeroinitializer + %splat.unknown = shufflevector <4 x half> %unknown.vec, <4 x half> poison, <4 x i32> zeroinitializer + %select = select i1 %cond, <4 x half> %splat.nan, <4 x half> %splat.unknown + ret <4 x half> %select +} + +define nofpclass(nan) <4 x half> @ret_nonan_splat_select_unknown_or_nan(i1 %cond, half %unknown) { +; CHECK-LABEL: define nofpclass(nan) <4 x half> @ret_nonan_splat_select_unknown_or_nan( +; CHECK-SAME: i1 [[COND:%.*]], half [[UNKNOWN:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call half @returns_nan_f16() +; CHECK-NEXT: [[SELECT_VEC:%.*]] = insertelement <4 x half> poison, half [[UNKNOWN]], i64 0 +; CHECK-NEXT: [[SPLAT:%.*]] = shufflevector <4 x half> [[SELECT_VEC]], <4 x half> poison, <4 x i32> zeroinitializer +; CHECK-NEXT: ret <4 x half> [[SPLAT]] +; + %nan = call half @returns_nan_f16() + %select = select i1 %cond, half %nan, half %unknown + %select.vec = insertelement <4 x half> poison, half %select, i32 0 + %splat = shufflevector <4 x half> %select.vec, <4 x half> poison, <4 x i32> zeroinitializer + ret <4 x half> %splat +} + +; Do not rewrite exp into fadd +define nofpclass(snan) <4 x half> @ret_nonnan_rhs_taints_known_nan_result(i1 %cond, <4 x half> %unknown) { +; CHECK-LABEL: define nofpclass(snan) <4 x half> @ret_nonnan_rhs_taints_known_nan_result( +; CHECK-SAME: i1 [[COND:%.*]], <4 x half> [[UNKNOWN:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call <4 x half> @returns_nan() +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[NAN]], <4 x half> [[UNKNOWN]], <4 x i32> <i32 5, i32 0, i32 1, i32 3> +; CHECK-NEXT: [[EXP:%.*]] = call <4 x half> @llvm.exp.v4f16(<4 x half> [[SHUFFLE]]) +; CHECK-NEXT: ret <4 x half> [[EXP]] +; + %nan = call <4 x half> @returns_nan() + %shuffle = shufflevector <4 x half> %nan, <4 x half> %unknown, <4 x i32> <i32 5, i32 0, i32 1, i32 3> + %exp = call <4 x half> @llvm.exp.v4f16(<4 x half> %shuffle) + ret <4 x half> %exp +} + +; Do not rewrite exp into fadd +define nofpclass(snan) <4 x half> @ret_nonnan_lhs_taints_known_nan_result(i1 %cond, <4 x half> %unknown) { +; CHECK-LABEL: define nofpclass(snan) <4 x half> @ret_nonnan_lhs_taints_known_nan_result( +; CHECK-SAME: i1 [[COND:%.*]], <4 x half> [[UNKNOWN:%.*]]) { +; CHECK-NEXT: [[NAN:%.*]] = call <4 x half> @returns_nan() +; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <4 x half> [[UNKNOWN]], <4 x half> [[NAN]], <4 x i32> <i32 5, i32 0, i32 1, i32 3> +; CHECK-NEXT: [[EXP:%.*]] = call <4 x half> @llvm.exp.v4f16(<4 x half> [[SHUFFLE]]) +; CHECK-NEXT: ret <4 x half> [[EXP]] +; + %nan = call <4 x half> @returns_nan() + %shuffle = shufflevector <4 x half> %unknown, <4 x half> %nan, <4 x i32> <i32 5, i32 0, i32 1, i32 3> + %exp = call <4 x half> @llvm.exp.v4f16(<4 x half> %shuffle) + ret <4 x half> %exp +} _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
