1 ; Some shufflevector optimized lowering. This list is by no means exhaustive. It 2 ; is only a **basic** smoke test. the vector_ops crosstest has a broader range 3 ; of test cases. 4 5 ; RUN: %p2i -i %s --target=x8632 --filetype=obj --disassemble -a -O2 \ 6 ; RUN: --allow-externally-defined-symbols | FileCheck %s --check-prefix=X86 7 8 ; RUN: %if --need=target_MIPS32 --need=allow_dump \ 9 ; RUN: --command %p2i --filetype=asm --assemble --disassemble --target \ 10 ; RUN: mips32 -i %s --args -O2 -allow-externally-defined-symbols \ 11 ; RUN: | %if --need=target_MIPS32 --need=allow_dump \ 12 ; RUN: --command FileCheck --check-prefix MIPS32 %s 13 14 declare void @useV4I32(<4 x i32> %t); 15 16 define internal void @shuffleV4I32(<4 x i32> %a, <4 x i32> %b) { 17 ; X86-LABEL: shuffleV4I32 18 %a_0 = extractelement <4 x i32> %a, i32 0 19 %a_1 = extractelement <4 x i32> %a, i32 1 20 %a_2 = extractelement <4 x i32> %a, i32 2 21 %a_3 = extractelement <4 x i32> %a, i32 3 22 23 %b_0 = extractelement <4 x i32> %b, i32 0 24 %b_1 = extractelement <4 x i32> %b, i32 1 25 %b_2 = extractelement <4 x i32> %b, i32 2 26 %b_3 = extractelement <4 x i32> %b, i32 3 27 28 %t0_0 = insertelement <4 x i32> undef, i32 %a_0, i32 0 29 %t0_1 = insertelement <4 x i32> %t0_0, i32 %b_0, i32 1 30 %t0_2 = insertelement <4 x i32> %t0_1, i32 %a_1, i32 2 31 %t0 = insertelement <4 x i32> %t0_2, i32 %b_1, i32 3 32 ; X86: punpckldq {{.*}} 33 34 call void @useV4I32(<4 x i32> %t0) 35 ; X86: call 36 37 %t1_0 = insertelement <4 x i32> undef, i32 %a_0, i32 0 38 %t1_1 = insertelement <4 x i32> %t1_0, i32 %b_1, i32 1 39 %t1_2 = insertelement <4 x i32> %t1_1, i32 %b_1, i32 2 40 %t1 = insertelement <4 x i32> %t1_2, i32 %a_0, i32 3 41 ; X86: shufps [[T:xmm[0-9]+]],{{.*}},0x10 42 ; X86: pshufd {{.*}},[[T]],0x28 43 44 call void @useV4I32(<4 x i32> %t1) 45 ; X86: call 46 47 %t2_0 = insertelement <4 x i32> undef, i32 %a_0, i32 0 48 %t2_1 = insertelement <4 x i32> %t2_0, i32 %b_3, i32 1 49 %t2_2 = insertelement <4 x i32> %t2_1, i32 %a_2, i32 2 50 %t2 = insertelement <4 x i32> %t2_2, i32 %b_2, i32 3 51 ; X86: shufps {{.*}},0x30 52 ; X86: shufps {{.*}},0x22 53 ; X86: shufps {{.*}},0x88 54 55 call void @useV4I32(<4 x i32> %t2) 56 ; X86: call 57 58 ret void 59 } 60 ; MIPS32-LABEL: shuffleV4I32 61 ; MIPS32: move 62 ; MIPS32: move 63 ; MIPS32: move 64 ; MIPS32: move 65 ; MIPS32: jal 66 ; MIPS32: move 67 ; MIPS32: move 68 ; MIPS32: move 69 ; MIPS32: move 70 ; MIPS32: jal 71 ; MIPS32: move 72 ; MIPS32: move 73 ; MIPS32: move 74 ; MIPS32: move 75 ; MIPS32: jal 76
