1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -mattr=sse2 -slp-vectorizer -S | FileCheck %s --check-prefix=SSE 3; RUN: opt < %s -mattr=avx2 -slp-vectorizer -S | FileCheck %s --check-prefix=AVX 4 5; TODO: 6; With AVX, we are able to vectorize the 1st 4 elements as 256-bit vector ops, 7; but the final 2 elements remain scalar. They should get vectorized using 8; 128-bit ops identically to what happens with SSE. 9 10target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128" 11target triple = "x86_64-unknown-linux-gnu" 12 13define void @PR28457(double* noalias nocapture align 32 %q, double* noalias nocapture readonly align 32 %p) { 14; SSE-LABEL: @PR28457( 15; SSE-NEXT: [[P0:%.*]] = getelementptr inbounds double, double* [[P:%.*]], i64 0 16; SSE-NEXT: [[P1:%.*]] = getelementptr inbounds double, double* [[P]], i64 1 17; SSE-NEXT: [[P2:%.*]] = getelementptr inbounds double, double* [[P]], i64 2 18; SSE-NEXT: [[P3:%.*]] = getelementptr inbounds double, double* [[P]], i64 3 19; SSE-NEXT: [[P4:%.*]] = getelementptr inbounds double, double* [[P]], i64 4 20; SSE-NEXT: [[P5:%.*]] = getelementptr inbounds double, double* [[P]], i64 5 21; SSE-NEXT: [[Q0:%.*]] = getelementptr inbounds double, double* [[Q:%.*]], i64 0 22; SSE-NEXT: [[Q1:%.*]] = getelementptr inbounds double, double* [[Q]], i64 1 23; SSE-NEXT: [[Q2:%.*]] = getelementptr inbounds double, double* [[Q]], i64 2 24; SSE-NEXT: [[Q3:%.*]] = getelementptr inbounds double, double* [[Q]], i64 3 25; SSE-NEXT: [[Q4:%.*]] = getelementptr inbounds double, double* [[Q]], i64 4 26; SSE-NEXT: [[Q5:%.*]] = getelementptr inbounds double, double* [[Q]], i64 5 27; SSE-NEXT: [[TMP1:%.*]] = bitcast double* [[P0]] to <2 x double>* 28; SSE-NEXT: [[TMP2:%.*]] = load <2 x double>, <2 x double>* [[TMP1]], align 8 29; SSE-NEXT: [[TMP3:%.*]] = bitcast double* [[P2]] to <2 x double>* 30; SSE-NEXT: [[TMP4:%.*]] = load <2 x double>, <2 x double>* [[TMP3]], align 8 31; SSE-NEXT: [[TMP5:%.*]] = bitcast double* [[P4]] to <2 x double>* 32; SSE-NEXT: [[TMP6:%.*]] = load <2 x double>, <2 x double>* [[TMP5]], align 8 33; SSE-NEXT: [[TMP7:%.*]] = fadd <2 x double> [[TMP2]], <double 1.000000e+00, double 1.000000e+00> 34; SSE-NEXT: [[TMP8:%.*]] = fadd <2 x double> [[TMP4]], <double 1.000000e+00, double 1.000000e+00> 35; SSE-NEXT: [[TMP9:%.*]] = fadd <2 x double> [[TMP6]], <double 1.000000e+00, double 1.000000e+00> 36; SSE-NEXT: [[TMP10:%.*]] = bitcast double* [[Q0]] to <2 x double>* 37; SSE-NEXT: store <2 x double> [[TMP7]], <2 x double>* [[TMP10]], align 8 38; SSE-NEXT: [[TMP11:%.*]] = bitcast double* [[Q2]] to <2 x double>* 39; SSE-NEXT: store <2 x double> [[TMP8]], <2 x double>* [[TMP11]], align 8 40; SSE-NEXT: [[TMP12:%.*]] = bitcast double* [[Q4]] to <2 x double>* 41; SSE-NEXT: store <2 x double> [[TMP9]], <2 x double>* [[TMP12]], align 8 42; SSE-NEXT: ret void 43; 44; AVX-LABEL: @PR28457( 45; AVX-NEXT: [[P0:%.*]] = getelementptr inbounds double, double* [[P:%.*]], i64 0 46; AVX-NEXT: [[P1:%.*]] = getelementptr inbounds double, double* [[P]], i64 1 47; AVX-NEXT: [[P2:%.*]] = getelementptr inbounds double, double* [[P]], i64 2 48; AVX-NEXT: [[P3:%.*]] = getelementptr inbounds double, double* [[P]], i64 3 49; AVX-NEXT: [[P4:%.*]] = getelementptr inbounds double, double* [[P]], i64 4 50; AVX-NEXT: [[P5:%.*]] = getelementptr inbounds double, double* [[P]], i64 5 51; AVX-NEXT: [[Q0:%.*]] = getelementptr inbounds double, double* [[Q:%.*]], i64 0 52; AVX-NEXT: [[Q1:%.*]] = getelementptr inbounds double, double* [[Q]], i64 1 53; AVX-NEXT: [[Q2:%.*]] = getelementptr inbounds double, double* [[Q]], i64 2 54; AVX-NEXT: [[Q3:%.*]] = getelementptr inbounds double, double* [[Q]], i64 3 55; AVX-NEXT: [[Q4:%.*]] = getelementptr inbounds double, double* [[Q]], i64 4 56; AVX-NEXT: [[Q5:%.*]] = getelementptr inbounds double, double* [[Q]], i64 5 57; AVX-NEXT: [[TMP1:%.*]] = bitcast double* [[P0]] to <4 x double>* 58; AVX-NEXT: [[TMP2:%.*]] = load <4 x double>, <4 x double>* [[TMP1]], align 8 59; AVX-NEXT: [[TMP3:%.*]] = bitcast double* [[P4]] to <2 x double>* 60; AVX-NEXT: [[TMP4:%.*]] = load <2 x double>, <2 x double>* [[TMP3]], align 8 61; AVX-NEXT: [[TMP5:%.*]] = fadd <4 x double> [[TMP2]], <double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00> 62; AVX-NEXT: [[TMP6:%.*]] = fadd <2 x double> [[TMP4]], <double 1.000000e+00, double 1.000000e+00> 63; AVX-NEXT: [[TMP7:%.*]] = bitcast double* [[Q0]] to <4 x double>* 64; AVX-NEXT: store <4 x double> [[TMP5]], <4 x double>* [[TMP7]], align 8 65; AVX-NEXT: [[TMP8:%.*]] = bitcast double* [[Q4]] to <2 x double>* 66; AVX-NEXT: store <2 x double> [[TMP6]], <2 x double>* [[TMP8]], align 8 67; AVX-NEXT: ret void 68; 69 %p0 = getelementptr inbounds double, double* %p, i64 0 70 %p1 = getelementptr inbounds double, double* %p, i64 1 71 %p2 = getelementptr inbounds double, double* %p, i64 2 72 %p3 = getelementptr inbounds double, double* %p, i64 3 73 %p4 = getelementptr inbounds double, double* %p, i64 4 74 %p5 = getelementptr inbounds double, double* %p, i64 5 75 76 %q0 = getelementptr inbounds double, double* %q, i64 0 77 %q1 = getelementptr inbounds double, double* %q, i64 1 78 %q2 = getelementptr inbounds double, double* %q, i64 2 79 %q3 = getelementptr inbounds double, double* %q, i64 3 80 %q4 = getelementptr inbounds double, double* %q, i64 4 81 %q5 = getelementptr inbounds double, double* %q, i64 5 82 83 %d0 = load double, double* %p0 84 %d1 = load double, double* %p1 85 %d2 = load double, double* %p2 86 %d3 = load double, double* %p3 87 %d4 = load double, double* %p4 88 %d5 = load double, double* %p5 89 90 %a0 = fadd double %d0, 1.0 91 %a1 = fadd double %d1, 1.0 92 %a2 = fadd double %d2, 1.0 93 %a3 = fadd double %d3, 1.0 94 %a4 = fadd double %d4, 1.0 95 %a5 = fadd double %d5, 1.0 96 97 store double %a0, double* %q0 98 store double %a1, double* %q1 99 store double %a2, double* %q2 100 store double %a3, double* %q3 101 store double %a4, double* %q4 102 store double %a5, double* %q5 103 ret void 104} 105