75 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -mattr=sse2 -passes=slp-vectorizer -S | FileCheck %s --check-prefix=SSE3; RUN: opt < %s -mattr=avx2 -passes=slp-vectorizer -S | FileCheck %s --check-prefix=AVX4 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 using8; 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(ptr noalias nocapture align 32 %q, ptr noalias nocapture readonly align 32 %p) {14; SSE-LABEL: @PR28457(15; SSE-NEXT: [[P2:%.*]] = getelementptr inbounds double, ptr [[P:%.*]], i64 216; SSE-NEXT: [[P4:%.*]] = getelementptr inbounds double, ptr [[P]], i64 417; SSE-NEXT: [[Q2:%.*]] = getelementptr inbounds double, ptr [[Q:%.*]], i64 218; SSE-NEXT: [[Q4:%.*]] = getelementptr inbounds double, ptr [[Q]], i64 419; SSE-NEXT: [[TMP2:%.*]] = load <2 x double>, ptr [[P]], align 820; SSE-NEXT: [[TMP3:%.*]] = fadd <2 x double> [[TMP2]], splat (double 1.000000e+00)21; SSE-NEXT: store <2 x double> [[TMP3]], ptr [[Q]], align 822; SSE-NEXT: [[TMP6:%.*]] = load <2 x double>, ptr [[P2]], align 823; SSE-NEXT: [[TMP7:%.*]] = fadd <2 x double> [[TMP6]], splat (double 1.000000e+00)24; SSE-NEXT: store <2 x double> [[TMP7]], ptr [[Q2]], align 825; SSE-NEXT: [[TMP10:%.*]] = load <2 x double>, ptr [[P4]], align 826; SSE-NEXT: [[TMP11:%.*]] = fadd <2 x double> [[TMP10]], splat (double 1.000000e+00)27; SSE-NEXT: store <2 x double> [[TMP11]], ptr [[Q4]], align 828; SSE-NEXT: ret void29;30; AVX-LABEL: @PR28457(31; AVX-NEXT: [[P4:%.*]] = getelementptr inbounds double, ptr [[P:%.*]], i64 432; AVX-NEXT: [[Q4:%.*]] = getelementptr inbounds double, ptr [[Q:%.*]], i64 433; AVX-NEXT: [[TMP2:%.*]] = load <4 x double>, ptr [[P]], align 834; AVX-NEXT: [[TMP3:%.*]] = fadd <4 x double> [[TMP2]], splat (double 1.000000e+00)35; AVX-NEXT: store <4 x double> [[TMP3]], ptr [[Q]], align 836; AVX-NEXT: [[TMP6:%.*]] = load <2 x double>, ptr [[P4]], align 837; AVX-NEXT: [[TMP7:%.*]] = fadd <2 x double> [[TMP6]], splat (double 1.000000e+00)38; AVX-NEXT: store <2 x double> [[TMP7]], ptr [[Q4]], align 839; AVX-NEXT: ret void40;41 %p1 = getelementptr inbounds double, ptr %p, i64 142 %p2 = getelementptr inbounds double, ptr %p, i64 243 %p3 = getelementptr inbounds double, ptr %p, i64 344 %p4 = getelementptr inbounds double, ptr %p, i64 445 %p5 = getelementptr inbounds double, ptr %p, i64 546 47 %q1 = getelementptr inbounds double, ptr %q, i64 148 %q2 = getelementptr inbounds double, ptr %q, i64 249 %q3 = getelementptr inbounds double, ptr %q, i64 350 %q4 = getelementptr inbounds double, ptr %q, i64 451 %q5 = getelementptr inbounds double, ptr %q, i64 552 53 %d0 = load double, ptr %p54 %d1 = load double, ptr %p155 %d2 = load double, ptr %p256 %d3 = load double, ptr %p357 %d4 = load double, ptr %p458 %d5 = load double, ptr %p559 60 %a0 = fadd double %d0, 1.061 %a1 = fadd double %d1, 1.062 %a2 = fadd double %d2, 1.063 %a3 = fadd double %d3, 1.064 %a4 = fadd double %d4, 1.065 %a5 = fadd double %d5, 1.066 67 store double %a0, ptr %q68 store double %a1, ptr %q169 store double %a2, ptr %q270 store double %a3, ptr %q371 store double %a4, ptr %q472 store double %a5, ptr %q573 ret void74}75