125 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-unknown-linux-gnu -mattr=+sse2 | FileCheck %s --check-prefixes=CHECK,SSE3; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-unknown-linux-gnu -mattr=+avx | FileCheck %s --check-prefixes=CHECK,AVX4; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-unknown-linux-gnu -mattr=+avx2 | FileCheck %s --check-prefixes=CHECK,AVX5 6define void @powof2div_uniform(ptr noalias nocapture %a, ptr noalias nocapture readonly %b, ptr noalias nocapture readonly %c){7; CHECK-LABEL: @powof2div_uniform(8; CHECK-NEXT: entry:9; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i32>, ptr [[B:%.*]], align 410; CHECK-NEXT: [[TMP3:%.*]] = load <4 x i32>, ptr [[C:%.*]], align 411; CHECK-NEXT: [[TMP4:%.*]] = add nsw <4 x i32> [[TMP3]], [[TMP1]]12; CHECK-NEXT: [[TMP5:%.*]] = sdiv <4 x i32> [[TMP4]], splat (i32 2)13; CHECK-NEXT: store <4 x i32> [[TMP5]], ptr [[A:%.*]], align 414; CHECK-NEXT: ret void15;16entry:17 %0 = load i32, ptr %b, align 418 %1 = load i32, ptr %c, align 419 %add = add nsw i32 %1, %020 %div = sdiv i32 %add, 221 store i32 %div, ptr %a, align 422 %arrayidx3 = getelementptr inbounds i32, ptr %b, i64 123 %2 = load i32, ptr %arrayidx3, align 424 %arrayidx4 = getelementptr inbounds i32, ptr %c, i64 125 %3 = load i32, ptr %arrayidx4, align 426 %add5 = add nsw i32 %3, %227 %div6 = sdiv i32 %add5, 228 %arrayidx7 = getelementptr inbounds i32, ptr %a, i64 129 store i32 %div6, ptr %arrayidx7, align 430 %arrayidx8 = getelementptr inbounds i32, ptr %b, i64 231 %4 = load i32, ptr %arrayidx8, align 432 %arrayidx9 = getelementptr inbounds i32, ptr %c, i64 233 %5 = load i32, ptr %arrayidx9, align 434 %add10 = add nsw i32 %5, %435 %div11 = sdiv i32 %add10, 236 %arrayidx12 = getelementptr inbounds i32, ptr %a, i64 237 store i32 %div11, ptr %arrayidx12, align 438 %arrayidx13 = getelementptr inbounds i32, ptr %b, i64 339 %6 = load i32, ptr %arrayidx13, align 440 %arrayidx14 = getelementptr inbounds i32, ptr %c, i64 341 %7 = load i32, ptr %arrayidx14, align 442 %add15 = add nsw i32 %7, %643 %div16 = sdiv i32 %add15, 244 %arrayidx17 = getelementptr inbounds i32, ptr %a, i64 345 store i32 %div16, ptr %arrayidx17, align 446 ret void47}48 49define void @powof2div_nonuniform(ptr noalias nocapture %a, ptr noalias nocapture readonly %b, ptr noalias nocapture readonly %c){50; SSE-LABEL: @powof2div_nonuniform(51; SSE-NEXT: entry:52; SSE-NEXT: [[TMP0:%.*]] = load i32, ptr [[B:%.*]], align 453; SSE-NEXT: [[TMP1:%.*]] = load i32, ptr [[C:%.*]], align 454; SSE-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP1]], [[TMP0]]55; SSE-NEXT: [[DIV:%.*]] = sdiv i32 [[ADD]], 256; SSE-NEXT: store i32 [[DIV]], ptr [[A:%.*]], align 457; SSE-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 158; SSE-NEXT: [[TMP2:%.*]] = load i32, ptr [[ARRAYIDX3]], align 459; SSE-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 160; SSE-NEXT: [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX4]], align 461; SSE-NEXT: [[ADD5:%.*]] = add nsw i32 [[TMP3]], [[TMP2]]62; SSE-NEXT: [[DIV6:%.*]] = sdiv i32 [[ADD5]], 463; SSE-NEXT: [[ARRAYIDX7:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 164; SSE-NEXT: store i32 [[DIV6]], ptr [[ARRAYIDX7]], align 465; SSE-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 266; SSE-NEXT: [[TMP4:%.*]] = load i32, ptr [[ARRAYIDX8]], align 467; SSE-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 268; SSE-NEXT: [[TMP5:%.*]] = load i32, ptr [[ARRAYIDX9]], align 469; SSE-NEXT: [[ADD10:%.*]] = add nsw i32 [[TMP5]], [[TMP4]]70; SSE-NEXT: [[DIV11:%.*]] = sdiv i32 [[ADD10]], 871; SSE-NEXT: [[ARRAYIDX12:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 272; SSE-NEXT: store i32 [[DIV11]], ptr [[ARRAYIDX12]], align 473; SSE-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 374; SSE-NEXT: [[TMP6:%.*]] = load i32, ptr [[ARRAYIDX13]], align 475; SSE-NEXT: [[ARRAYIDX14:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 376; SSE-NEXT: [[TMP7:%.*]] = load i32, ptr [[ARRAYIDX14]], align 477; SSE-NEXT: [[ADD15:%.*]] = add nsw i32 [[TMP7]], [[TMP6]]78; SSE-NEXT: [[DIV16:%.*]] = sdiv i32 [[ADD15]], 1679; SSE-NEXT: [[ARRAYIDX17:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 380; SSE-NEXT: store i32 [[DIV16]], ptr [[ARRAYIDX17]], align 481; SSE-NEXT: ret void82;83; AVX-LABEL: @powof2div_nonuniform(84; AVX-NEXT: entry:85; AVX-NEXT: [[TMP1:%.*]] = load <4 x i32>, ptr [[B:%.*]], align 486; AVX-NEXT: [[TMP3:%.*]] = load <4 x i32>, ptr [[C:%.*]], align 487; AVX-NEXT: [[TMP4:%.*]] = add nsw <4 x i32> [[TMP3]], [[TMP1]]88; AVX-NEXT: [[TMP5:%.*]] = sdiv <4 x i32> [[TMP4]], <i32 2, i32 4, i32 8, i32 16>89; AVX-NEXT: store <4 x i32> [[TMP5]], ptr [[A:%.*]], align 490; AVX-NEXT: ret void91;92entry:93 %0 = load i32, ptr %b, align 494 %1 = load i32, ptr %c, align 495 %add = add nsw i32 %1, %096 %div = sdiv i32 %add, 297 store i32 %div, ptr %a, align 498 %arrayidx3 = getelementptr inbounds i32, ptr %b, i64 199 %2 = load i32, ptr %arrayidx3, align 4100 %arrayidx4 = getelementptr inbounds i32, ptr %c, i64 1101 %3 = load i32, ptr %arrayidx4, align 4102 %add5 = add nsw i32 %3, %2103 %div6 = sdiv i32 %add5, 4104 %arrayidx7 = getelementptr inbounds i32, ptr %a, i64 1105 store i32 %div6, ptr %arrayidx7, align 4106 %arrayidx8 = getelementptr inbounds i32, ptr %b, i64 2107 %4 = load i32, ptr %arrayidx8, align 4108 %arrayidx9 = getelementptr inbounds i32, ptr %c, i64 2109 %5 = load i32, ptr %arrayidx9, align 4110 %add10 = add nsw i32 %5, %4111 %div11 = sdiv i32 %add10, 8112 %arrayidx12 = getelementptr inbounds i32, ptr %a, i64 2113 store i32 %div11, ptr %arrayidx12, align 4114 %arrayidx13 = getelementptr inbounds i32, ptr %b, i64 3115 %6 = load i32, ptr %arrayidx13, align 4116 %arrayidx14 = getelementptr inbounds i32, ptr %c, i64 3117 %7 = load i32, ptr %arrayidx14, align 4118 %add15 = add nsw i32 %7, %6119 %div16 = sdiv i32 %add15, 16120 %arrayidx17 = getelementptr inbounds i32, ptr %a, i64 3121 store i32 %div16, ptr %arrayidx17, align 4122 ret void123}124 125