113 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -S %s -passes=scalarize-masked-mem-intrin -mtriple=aarch64-linux-gnu | FileCheck -check-prefixes=CHECK-LE %s3; RUN: opt -S %s -passes=scalarize-masked-mem-intrin -mtriple=aarch64-linux-gnu -mattr=+sve | FileCheck -check-prefixes=CHECK-SVE-LE %s4; RUN: opt -S %s -passes=scalarize-masked-mem-intrin -mtriple=aarch64_be-linux-gnu -data-layout="E-m:o-i64:64-i128:128-n32:64-S128" | FileCheck -check-prefixes=CHECK-BE %s5 6define void @scalarize_v2i64(ptr %p, <2 x i1> %mask, <2 x i64> %data) {7; CHECK-LE-LABEL: @scalarize_v2i64(8; CHECK-LE-NEXT: [[SCALAR_MASK:%.*]] = bitcast <2 x i1> [[MASK:%.*]] to i29; CHECK-LE-NEXT: [[TMP1:%.*]] = and i2 [[SCALAR_MASK]], 110; CHECK-LE-NEXT: [[TMP2:%.*]] = icmp ne i2 [[TMP1]], 011; CHECK-LE-NEXT: br i1 [[TMP2]], label [[COND_STORE:%.*]], label [[ELSE:%.*]]12; CHECK-LE: cond.store:13; CHECK-LE-NEXT: [[TMP3:%.*]] = extractelement <2 x i64> [[DATA:%.*]], i64 014; CHECK-LE-NEXT: [[TMP4:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i32 015; CHECK-LE-NEXT: store i64 [[TMP3]], ptr [[TMP4]], align 816; CHECK-LE-NEXT: br label [[ELSE]]17; CHECK-LE: else:18; CHECK-LE-NEXT: [[TMP5:%.*]] = and i2 [[SCALAR_MASK]], -219; CHECK-LE-NEXT: [[TMP6:%.*]] = icmp ne i2 [[TMP5]], 020; CHECK-LE-NEXT: br i1 [[TMP6]], label [[COND_STORE1:%.*]], label [[ELSE2:%.*]]21; CHECK-LE: cond.store1:22; CHECK-LE-NEXT: [[TMP7:%.*]] = extractelement <2 x i64> [[DATA]], i64 123; CHECK-LE-NEXT: [[TMP8:%.*]] = getelementptr inbounds i64, ptr [[P]], i32 124; CHECK-LE-NEXT: store i64 [[TMP7]], ptr [[TMP8]], align 825; CHECK-LE-NEXT: br label [[ELSE2]]26; CHECK-LE: else2:27; CHECK-LE-NEXT: ret void28;29; CHECK-SVE-LE-LABEL: @scalarize_v2i64(30; CHECK-SVE-LE-NEXT: call void @llvm.masked.store.v2i64.p0(<2 x i64> [[DATA:%.*]], ptr align 128 [[P:%.*]], <2 x i1> [[MASK:%.*]])31; CHECK-SVE-LE-NEXT: ret void32;33; CHECK-BE-LABEL: @scalarize_v2i64(34; CHECK-BE-NEXT: [[SCALAR_MASK:%.*]] = bitcast <2 x i1> [[MASK:%.*]] to i235; CHECK-BE-NEXT: [[TMP1:%.*]] = and i2 [[SCALAR_MASK]], -236; CHECK-BE-NEXT: [[TMP2:%.*]] = icmp ne i2 [[TMP1]], 037; CHECK-BE-NEXT: br i1 [[TMP2]], label [[COND_STORE:%.*]], label [[ELSE:%.*]]38; CHECK-BE: cond.store:39; CHECK-BE-NEXT: [[TMP3:%.*]] = extractelement <2 x i64> [[DATA:%.*]], i64 040; CHECK-BE-NEXT: [[TMP4:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i32 041; CHECK-BE-NEXT: store i64 [[TMP3]], ptr [[TMP4]], align 842; CHECK-BE-NEXT: br label [[ELSE]]43; CHECK-BE: else:44; CHECK-BE-NEXT: [[TMP5:%.*]] = and i2 [[SCALAR_MASK]], 145; CHECK-BE-NEXT: [[TMP6:%.*]] = icmp ne i2 [[TMP5]], 046; CHECK-BE-NEXT: br i1 [[TMP6]], label [[COND_STORE1:%.*]], label [[ELSE2:%.*]]47; CHECK-BE: cond.store1:48; CHECK-BE-NEXT: [[TMP7:%.*]] = extractelement <2 x i64> [[DATA]], i64 149; CHECK-BE-NEXT: [[TMP8:%.*]] = getelementptr inbounds i64, ptr [[P]], i32 150; CHECK-BE-NEXT: store i64 [[TMP7]], ptr [[TMP8]], align 851; CHECK-BE-NEXT: br label [[ELSE2]]52; CHECK-BE: else2:53; CHECK-BE-NEXT: ret void54;55 call void @llvm.masked.store.v2i64.p0(<2 x i64> %data, ptr %p, i32 128, <2 x i1> %mask)56 ret void57}58 59define void @scalarize_v2i64_ones_mask(ptr %p, <2 x i64> %data) {60; CHECK-LE-LABEL: @scalarize_v2i64_ones_mask(61; CHECK-LE-NEXT: store <2 x i64> [[DATA:%.*]], ptr [[P:%.*]], align 862; CHECK-LE-NEXT: ret void63;64; CHECK-SVE-LE-LABEL: @scalarize_v2i64_ones_mask(65; CHECK-SVE-LE-NEXT: call void @llvm.masked.store.v2i64.p0(<2 x i64> [[DATA:%.*]], ptr align 8 [[P:%.*]], <2 x i1> splat (i1 true))66; CHECK-SVE-LE-NEXT: ret void67;68; CHECK-BE-LABEL: @scalarize_v2i64_ones_mask(69; CHECK-BE-NEXT: store <2 x i64> [[DATA:%.*]], ptr [[P:%.*]], align 870; CHECK-BE-NEXT: ret void71;72 call void @llvm.masked.store.v2i64.p0(<2 x i64> %data, ptr %p, i32 8, <2 x i1> <i1 true, i1 true>)73 ret void74}75 76define void @scalarize_v2i64_zero_mask(ptr %p, <2 x i64> %data) {77; CHECK-LE-LABEL: @scalarize_v2i64_zero_mask(78; CHECK-LE-NEXT: ret void79;80; CHECK-SVE-LE-LABEL: @scalarize_v2i64_zero_mask(81; CHECK-SVE-LE-NEXT: call void @llvm.masked.store.v2i64.p0(<2 x i64> [[DATA:%.*]], ptr align 8 [[P:%.*]], <2 x i1> zeroinitializer)82; CHECK-SVE-LE-NEXT: ret void83;84; CHECK-BE-LABEL: @scalarize_v2i64_zero_mask(85; CHECK-BE-NEXT: ret void86;87 call void @llvm.masked.store.v2i64.p0(<2 x i64> %data, ptr %p, i32 8, <2 x i1> <i1 false, i1 false>)88 ret void89}90 91define void @scalarize_v2i64_const_mask(ptr %p, <2 x i64> %data) {92; CHECK-LE-LABEL: @scalarize_v2i64_const_mask(93; CHECK-LE-NEXT: [[TMP1:%.*]] = extractelement <2 x i64> [[DATA:%.*]], i64 194; CHECK-LE-NEXT: [[TMP2:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i32 195; CHECK-LE-NEXT: store i64 [[TMP1]], ptr [[TMP2]], align 896; CHECK-LE-NEXT: ret void97;98; CHECK-SVE-LE-LABEL: @scalarize_v2i64_const_mask(99; CHECK-SVE-LE-NEXT: call void @llvm.masked.store.v2i64.p0(<2 x i64> [[DATA:%.*]], ptr align 8 [[P:%.*]], <2 x i1> <i1 false, i1 true>)100; CHECK-SVE-LE-NEXT: ret void101;102; CHECK-BE-LABEL: @scalarize_v2i64_const_mask(103; CHECK-BE-NEXT: [[TMP1:%.*]] = extractelement <2 x i64> [[DATA:%.*]], i64 1104; CHECK-BE-NEXT: [[TMP2:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i32 1105; CHECK-BE-NEXT: store i64 [[TMP1]], ptr [[TMP2]], align 8106; CHECK-BE-NEXT: ret void107;108 call void @llvm.masked.store.v2i64.p0(<2 x i64> %data, ptr %p, i32 8, <2 x i1> <i1 false, i1 true>)109 ret void110}111 112declare void @llvm.masked.store.v2i64.p0(<2 x i64>, ptr, i32, <2 x i1>)113