26 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=load-store-vectorizer -mcpu=sm_35 -S < %s | FileCheck %s3 4target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"5target triple = "nvptx64-nvidia-cuda"6 7define i32 @foo(ptr %ptr) {8; CHECK-LABEL: @foo(9; CHECK-NEXT: [[PTR1:%.*]] = getelementptr i32, ptr [[PTR:%.*]], i32 110; CHECK-NEXT: [[P1:%.*]] = addrspacecast ptr [[PTR1]] to ptr addrspace(1)11; CHECK-NEXT: [[TMP2:%.*]] = load <2 x i32>, ptr [[PTR]], align 8, !invariant.load !012; CHECK-NEXT: [[V01:%.*]] = extractelement <2 x i32> [[TMP2]], i32 013; CHECK-NEXT: [[V12:%.*]] = extractelement <2 x i32> [[TMP2]], i32 114; CHECK-NEXT: [[SUM:%.*]] = add i32 [[V01]], [[V12]]15; CHECK-NEXT: ret i32 [[SUM]]16;17 %ptr1 = getelementptr i32, ptr %ptr, i32 118 %p1 = addrspacecast ptr %ptr1 to ptr addrspace(1)19 %v0 = load i32, ptr %ptr, align 8, !invariant.load !020 %v1 = load i32, ptr %ptr1, align 4, !invariant.load !021 %sum = add i32 %v0, %v122 ret i32 %sum23}24 25!0 = !{}26