1 ; RUN: llc < %s -march=x86 -mattr=+sse42 | FileCheck %s 2 ; CHECK: psraw 3 ; CHECK: psraw 4 5 define void @update(i64* %dst_i, i64* %src_i, i32 %n) nounwind { 6 entry: 7 %dst_i.addr = alloca i64* ; <i64**> [#uses=2] 8 %src_i.addr = alloca i64* ; <i64**> [#uses=2] 9 %n.addr = alloca i32 ; <i32*> [#uses=2] 10 %i = alloca i32, align 4 ; <i32*> [#uses=8] 11 %dst = alloca <8 x i8>*, align 4 ; <<8 x i8>**> [#uses=2] 12 %src = alloca <8 x i8>*, align 4 ; <<8 x i8>**> [#uses=2] 13 store i64* %dst_i, i64** %dst_i.addr 14 store i64* %src_i, i64** %src_i.addr 15 store i32 %n, i32* %n.addr 16 store i32 0, i32* %i 17 br label %forcond 18 19 forcond: ; preds = %forinc, %entry 20 %tmp = load i32* %i ; <i32> [#uses=1] 21 %tmp1 = load i32* %n.addr ; <i32> [#uses=1] 22 %cmp = icmp slt i32 %tmp, %tmp1 ; <i1> [#uses=1] 23 br i1 %cmp, label %forbody, label %afterfor 24 25 forbody: ; preds = %forcond 26 %tmp2 = load i32* %i ; <i32> [#uses=1] 27 %tmp3 = load i64** %dst_i.addr ; <i64*> [#uses=1] 28 %arrayidx = getelementptr i64* %tmp3, i32 %tmp2 ; <i64*> [#uses=1] 29 %conv = bitcast i64* %arrayidx to <8 x i8>* ; <<8 x i8>*> [#uses=1] 30 store <8 x i8>* %conv, <8 x i8>** %dst 31 %tmp4 = load i32* %i ; <i32> [#uses=1] 32 %tmp5 = load i64** %src_i.addr ; <i64*> [#uses=1] 33 %arrayidx6 = getelementptr i64* %tmp5, i32 %tmp4 ; <i64*> [#uses=1] 34 %conv7 = bitcast i64* %arrayidx6 to <8 x i8>* ; <<8 x i8>*> [#uses=1] 35 store <8 x i8>* %conv7, <8 x i8>** %src 36 %tmp8 = load i32* %i ; <i32> [#uses=1] 37 %tmp9 = load <8 x i8>** %dst ; <<8 x i8>*> [#uses=1] 38 %arrayidx10 = getelementptr <8 x i8>* %tmp9, i32 %tmp8 ; <<8 x i8>*> [#uses=1] 39 %tmp11 = load i32* %i ; <i32> [#uses=1] 40 %tmp12 = load <8 x i8>** %src ; <<8 x i8>*> [#uses=1] 41 %arrayidx13 = getelementptr <8 x i8>* %tmp12, i32 %tmp11 ; <<8 x i8>*> [#uses=1] 42 %tmp14 = load <8 x i8>* %arrayidx13 ; <<8 x i8>> [#uses=1] 43 %add = add <8 x i8> %tmp14, < i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1 > ; <<8 x i8>> [#uses=1] 44 %shr = ashr <8 x i8> %add, < i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2 > ; <<8 x i8>> [#uses=1] 45 store <8 x i8> %shr, <8 x i8>* %arrayidx10 46 br label %forinc 47 48 forinc: ; preds = %forbody 49 %tmp15 = load i32* %i ; <i32> [#uses=1] 50 %inc = add i32 %tmp15, 1 ; <i32> [#uses=1] 51 store i32 %inc, i32* %i 52 br label %forcond 53 54 afterfor: ; preds = %forcond 55 ret void 56 } 57 58
