1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -S --passes=vector-combine -mtriple=x86_64-- -mcpu=x86-64 | FileCheck %s --check-prefixes=CHECK,X64 3; RUN: opt < %s -S --passes=vector-combine -mtriple=x86_64-- -mcpu=x86-64-v2 | FileCheck %s --check-prefixes=CHECK,X64 4; RUN: opt < %s -S --passes=vector-combine -mtriple=x86_64-- -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=CHECK,X64 5; RUN: opt < %s -S --passes=vector-combine -mtriple=x86_64-- -mcpu=x86-64-v4 | FileCheck %s --check-prefixes=CHECK,AVX512 6 7; 8; Fold reduce(trunc(X)) -> trunc(reduce(X)) if more cost efficient 9; 10 11; Cheap AVX512 v8i64 -> v8i32 truncation 12define i32 @reduce_add_trunc_v8i64_i32(<8 x i64> %a0) { 13; X64-LABEL: @reduce_add_trunc_v8i64_i32( 14; X64-NEXT: [[TMP1:%.*]] = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> [[A0:%.*]]) 15; X64-NEXT: [[RED:%.*]] = trunc i64 [[TMP1]] to i32 16; X64-NEXT: ret i32 [[RED]] 17; 18; AVX512-LABEL: @reduce_add_trunc_v8i64_i32( 19; AVX512-NEXT: [[TR:%.*]] = trunc <8 x i64> [[A0:%.*]] to <8 x i32> 20; AVX512-NEXT: [[RED:%.*]] = tail call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TR]]) 21; AVX512-NEXT: ret i32 [[RED]] 22; 23 %tr = trunc <8 x i64> %a0 to <8 x i32> 24 %red = tail call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> %tr) 25 ret i32 %red 26} 27declare i32 @llvm.vector.reduce.add.v8i32(<8 x i32>) 28 29; No legal vXi8 multiplication so vXi16 is always cheaper 30define i8 @reduce_mul_trunc_v16i16_i8(<16 x i16> %a0) { 31; CHECK-LABEL: @reduce_mul_trunc_v16i16_i8( 32; CHECK-NEXT: [[TMP1:%.*]] = call i16 @llvm.vector.reduce.mul.v16i16(<16 x i16> [[A0:%.*]]) 33; CHECK-NEXT: [[RED:%.*]] = trunc i16 [[TMP1]] to i8 34; CHECK-NEXT: ret i8 [[RED]] 35; 36 %tr = trunc <16 x i16> %a0 to <16 x i8> 37 %red = tail call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> %tr) 38 ret i8 %red 39} 40declare i8 @llvm.vector.reduce.mul.v16i8(<16 x i8>) 41 42define i8 @reduce_or_trunc_v8i32_i8(<8 x i32> %a0) { 43; CHECK-LABEL: @reduce_or_trunc_v8i32_i8( 44; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.vector.reduce.or.v8i32(<8 x i32> [[A0:%.*]]) 45; CHECK-NEXT: [[RED:%.*]] = trunc i32 [[TMP1]] to i8 46; CHECK-NEXT: ret i8 [[RED]] 47; 48 %tr = trunc <8 x i32> %a0 to <8 x i8> 49 %red = tail call i8 @llvm.vector.reduce.or.v8i32(<8 x i8> %tr) 50 ret i8 %red 51} 52declare i8 @llvm.vector.reduce.or.v8i8(<8 x i8>) 53 54define i8 @reduce_xor_trunc_v16i64_i8(<16 x i64> %a0) { 55; CHECK-LABEL: @reduce_xor_trunc_v16i64_i8( 56; CHECK-NEXT: [[TMP1:%.*]] = call i64 @llvm.vector.reduce.xor.v16i64(<16 x i64> [[A0:%.*]]) 57; CHECK-NEXT: [[RED:%.*]] = trunc i64 [[TMP1]] to i8 58; CHECK-NEXT: ret i8 [[RED]] 59; 60 %tr = trunc <16 x i64> %a0 to <16 x i8> 61 %red = tail call i8 @llvm.vector.reduce.xor.v16i8(<16 x i8> %tr) 62 ret i8 %red 63} 64declare i8 @llvm.vector.reduce.xor.v16i8(<16 x i8>) 65 66; Truncation source has other uses - OK to truncate reduction 67define i16 @reduce_and_trunc_v16i64_i16(<16 x i64> %a0) { 68; CHECK-LABEL: @reduce_and_trunc_v16i64_i16( 69; CHECK-NEXT: [[TMP1:%.*]] = call i64 @llvm.vector.reduce.and.v16i64(<16 x i64> [[A0:%.*]]) 70; CHECK-NEXT: [[RED:%.*]] = trunc i64 [[TMP1]] to i16 71; CHECK-NEXT: call void @use_v16i64(<16 x i64> [[A0]]) 72; CHECK-NEXT: ret i16 [[RED]] 73; 74 %tr = trunc <16 x i64> %a0 to <16 x i16> 75 %red = tail call i16 @llvm.vector.reduce.and.v16i16(<16 x i16> %tr) 76 call void @use_v16i64(<16 x i64> %a0) 77 ret i16 %red 78} 79declare i16 @llvm.vector.reduce.and.v16i16(<16 x i16>) 80 81; Negative Test: vXi16 multiply is much cheaper than vXi64 82define i16 @reduce_mul_trunc_v8i64_i16(<8 x i64> %a0) { 83; CHECK-LABEL: @reduce_mul_trunc_v8i64_i16( 84; CHECK-NEXT: [[TR:%.*]] = trunc <8 x i64> [[A0:%.*]] to <8 x i16> 85; CHECK-NEXT: [[RED:%.*]] = tail call i16 @llvm.vector.reduce.mul.v8i16(<8 x i16> [[TR]]) 86; CHECK-NEXT: ret i16 [[RED]] 87; 88 %tr = trunc <8 x i64> %a0 to <8 x i16> 89 %red = tail call i16 @llvm.vector.reduce.mul.v8i16(<8 x i16> %tr) 90 ret i16 %red 91} 92declare i16 @llvm.vector.reduce.mul.v8i16(<8 x i16>) 93 94; Negative Test: min/max reductions can't use pre-truncated types. 95define i8 @reduce_smin_trunc_v16i16_i8(<16 x i16> %a0) { 96; CHECK-LABEL: @reduce_smin_trunc_v16i16_i8( 97; CHECK-NEXT: [[TR:%.*]] = trunc <16 x i16> [[A0:%.*]] to <16 x i8> 98; CHECK-NEXT: [[RED:%.*]] = tail call i8 @llvm.vector.reduce.smin.v16i8(<16 x i8> [[TR]]) 99; CHECK-NEXT: ret i8 [[RED]] 100; 101 %tr = trunc <16 x i16> %a0 to <16 x i8> 102 %red = tail call i8 @llvm.vector.reduce.smin.v16i8(<16 x i8> %tr) 103 ret i8 %red 104} 105declare i8 @llvm.vector.reduce.smin.v16i8(<16 x i8>) 106 107; Negative Test: Truncation has other uses. 108define i16 @reduce_and_trunc_v16i64_i16_multiuse(<16 x i64> %a0) { 109; CHECK-LABEL: @reduce_and_trunc_v16i64_i16_multiuse( 110; CHECK-NEXT: [[TR:%.*]] = trunc <16 x i64> [[A0:%.*]] to <16 x i16> 111; CHECK-NEXT: [[RED:%.*]] = tail call i16 @llvm.vector.reduce.and.v16i16(<16 x i16> [[TR]]) 112; CHECK-NEXT: call void @use_v16i16(<16 x i16> [[TR]]) 113; CHECK-NEXT: ret i16 [[RED]] 114; 115 %tr = trunc <16 x i64> %a0 to <16 x i16> 116 %red = tail call i16 @llvm.vector.reduce.and.v16i16(<16 x i16> %tr) 117 call void @use_v16i16(<16 x i16> %tr) 118 ret i16 %red 119} 120 121declare void @use_v16i64(<16 x i64>) 122declare void @use_v16i16(<16 x i16>) 123 124