271 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 62; RUN: llc -mtriple=armv7-- < %s | FileCheck -check-prefix=NOFP16 %s3 4declare void @f16_user(half)5declare half @f16_result()6 7declare void @v2f16_user(<2 x half>)8declare <2 x half> @v2f16_result()9 10declare void @v4f16_user(<4 x half>)11declare <4 x half> @v4f16_result()12 13declare void @v8f16_user(<8 x half>)14declare <8 x half> @v8f16_result()15 16define void @f16_arg(half %arg, ptr %ptr) #0 {17; NOFP16-LABEL: f16_arg:18; NOFP16: @ %bb.0:19; NOFP16-NEXT: push {r4, lr}20; NOFP16-NEXT: uxth r0, r021; NOFP16-NEXT: mov r4, r122; NOFP16-NEXT: bl __gnu_h2f_ieee23; NOFP16-NEXT: str r0, [r4]24; NOFP16-NEXT: pop {r4, pc}25 %fpext = call float @llvm.experimental.constrained.fpext.f32.f16(half %arg, metadata !"fpexcept.strict")26 store float %fpext, ptr %ptr27 ret void28}29 30define void @v2f16_arg(<2 x half> %arg, ptr %ptr) #0 {31; NOFP16-LABEL: v2f16_arg:32; NOFP16: @ %bb.0:33; NOFP16-NEXT: push {r4, r5, r11, lr}34; NOFP16-NEXT: vpush {d8}35; NOFP16-NEXT: mov r5, r036; NOFP16-NEXT: uxth r0, r137; NOFP16-NEXT: mov r4, r238; NOFP16-NEXT: bl __gnu_h2f_ieee39; NOFP16-NEXT: uxth r1, r540; NOFP16-NEXT: vmov s17, r041; NOFP16-NEXT: mov r0, r142; NOFP16-NEXT: bl __gnu_h2f_ieee43; NOFP16-NEXT: vmov s16, r044; NOFP16-NEXT: vstr d8, [r4]45; NOFP16-NEXT: vpop {d8}46; NOFP16-NEXT: pop {r4, r5, r11, pc}47 %fpext = call <2 x float> @llvm.experimental.constrained.fpext.v2f32.v2f16(<2 x half> %arg, metadata !"fpexcept.strict")48 store <2 x float> %fpext, ptr %ptr49 ret void50}51 52define void @v3f16_arg(<3 x half> %arg, ptr %ptr) #0 {53; NOFP16-LABEL: v3f16_arg:54; NOFP16: @ %bb.0:55; NOFP16-NEXT: push {r4, r5, r6, lr}56; NOFP16-NEXT: vpush {d8}57; NOFP16-NEXT: mov r6, r058; NOFP16-NEXT: uxth r0, r159; NOFP16-NEXT: mov r4, r360; NOFP16-NEXT: mov r5, r261; NOFP16-NEXT: bl __gnu_h2f_ieee62; NOFP16-NEXT: uxth r1, r663; NOFP16-NEXT: vmov s17, r064; NOFP16-NEXT: mov r0, r165; NOFP16-NEXT: bl __gnu_h2f_ieee66; NOFP16-NEXT: vmov s16, r067; NOFP16-NEXT: uxth r0, r568; NOFP16-NEXT: vst1.32 {d8}, [r4:64]!69; NOFP16-NEXT: bl __gnu_h2f_ieee70; NOFP16-NEXT: str r0, [r4]71; NOFP16-NEXT: vpop {d8}72; NOFP16-NEXT: pop {r4, r5, r6, pc}73 %fpext = call <3 x float> @llvm.experimental.constrained.fpext.v3f32.v3f16(<3 x half> %arg, metadata !"fpexcept.strict")74 store <3 x float> %fpext, ptr %ptr75 ret void76}77 78define void @v4f16_arg(<4 x half> %arg, ptr %ptr) #0 {79; NOFP16-LABEL: v4f16_arg:80; NOFP16: @ %bb.0:81; NOFP16-NEXT: push {r4, r5, r6, r7, r11, lr}82; NOFP16-NEXT: vpush {d8, d9}83; NOFP16-NEXT: mov r6, r084; NOFP16-NEXT: uxth r0, r185; NOFP16-NEXT: mov r4, r386; NOFP16-NEXT: mov r5, r287; NOFP16-NEXT: bl __gnu_h2f_ieee88; NOFP16-NEXT: mov r7, r089; NOFP16-NEXT: uxth r0, r490; NOFP16-NEXT: bl __gnu_h2f_ieee91; NOFP16-NEXT: vmov s19, r092; NOFP16-NEXT: uxth r0, r593; NOFP16-NEXT: ldr r4, [sp, #40]94; NOFP16-NEXT: bl __gnu_h2f_ieee95; NOFP16-NEXT: vmov s18, r096; NOFP16-NEXT: uxth r0, r697; NOFP16-NEXT: vmov s17, r798; NOFP16-NEXT: bl __gnu_h2f_ieee99; NOFP16-NEXT: vmov s16, r0100; NOFP16-NEXT: vst1.64 {d8, d9}, [r4]101; NOFP16-NEXT: vpop {d8, d9}102; NOFP16-NEXT: pop {r4, r5, r6, r7, r11, pc}103 %fpext = call <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half> %arg, metadata !"fpexcept.strict")104 store <4 x float> %fpext, ptr %ptr105 ret void106}107 108 define half @f16_return(float %arg) #0 {109; NOFP16-LABEL: f16_return:110; NOFP16: @ %bb.0:111; NOFP16-NEXT: push {r11, lr}112; NOFP16-NEXT: bl __gnu_f2h_ieee113; NOFP16-NEXT: pop {r11, pc}114 %fptrunc = call half @llvm.experimental.constrained.fptrunc.f16.f32(float %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")115 ret half %fptrunc116 }117 118 define <2 x half> @v2f16_return(<2 x float> %arg) #0 {119; NOFP16-LABEL: v2f16_return:120; NOFP16: @ %bb.0:121; NOFP16-NEXT: push {r11, lr}122; NOFP16-NEXT: vpush {d8}123; NOFP16-NEXT: sub sp, sp, #8124; NOFP16-NEXT: vmov d8, r0, r1125; NOFP16-NEXT: vmov r0, s17126; NOFP16-NEXT: bl __gnu_f2h_ieee127; NOFP16-NEXT: vmov r1, s16128; NOFP16-NEXT: strh r0, [sp, #6]129; NOFP16-NEXT: mov r0, r1130; NOFP16-NEXT: bl __gnu_f2h_ieee131; NOFP16-NEXT: strh r0, [sp, #4]132; NOFP16-NEXT: add r0, sp, #4133; NOFP16-NEXT: vld1.32 {d16[0]}, [r0:32]134; NOFP16-NEXT: vmovl.u16 q8, d16135; NOFP16-NEXT: vmov.32 r0, d16[0]136; NOFP16-NEXT: vmov.32 r1, d16[1]137; NOFP16-NEXT: add sp, sp, #8138; NOFP16-NEXT: vpop {d8}139; NOFP16-NEXT: pop {r11, pc}140 %fptrunc = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")141 ret <2 x half> %fptrunc142 }143 144 define <3 x half> @v3f16_return(<3 x float> %arg) #0 {145; NOFP16-LABEL: v3f16_return:146; NOFP16: @ %bb.0:147; NOFP16-NEXT: push {r4, r5, r6, lr}148; NOFP16-NEXT: vmov d1, r2, r3149; NOFP16-NEXT: mov r5, r0150; NOFP16-NEXT: vmov d0, r0, r1151; NOFP16-NEXT: mov r4, r1152; NOFP16-NEXT: vmov r0, s2153; NOFP16-NEXT: bl __gnu_f2h_ieee154; NOFP16-NEXT: uxth r6, r0155; NOFP16-NEXT: mov r0, r4156; NOFP16-NEXT: bl __gnu_f2h_ieee157; NOFP16-NEXT: mov r4, r0158; NOFP16-NEXT: mov r0, r5159; NOFP16-NEXT: bl __gnu_f2h_ieee160; NOFP16-NEXT: pkhbt r0, r0, r4, lsl #16161; NOFP16-NEXT: vmov d16, r0, r6162; NOFP16-NEXT: vmov.u16 r0, d16[0]163; NOFP16-NEXT: vmov.u16 r1, d16[1]164; NOFP16-NEXT: vmov.u16 r2, d16[2]165; NOFP16-NEXT: pop {r4, r5, r6, pc}166 %fptrunc = call <3 x half> @llvm.experimental.constrained.fptrunc.v3f16.v3f32(<3 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")167 ret <3 x half> %fptrunc168 }169 170 define <4 x half> @v4f16_return(<4 x float> %arg) #0 {171; NOFP16-LABEL: v4f16_return:172; NOFP16: @ %bb.0:173; NOFP16-NEXT: push {r4, r5, r11, lr}174; NOFP16-NEXT: vpush {d8, d9}175; NOFP16-NEXT: vmov d8, r2, r3176; NOFP16-NEXT: vmov d9, r0, r1177; NOFP16-NEXT: vmov r2, s17178; NOFP16-NEXT: mov r0, r2179; NOFP16-NEXT: bl __gnu_f2h_ieee180; NOFP16-NEXT: mov r4, r0181; NOFP16-NEXT: vmov r0, s16182; NOFP16-NEXT: bl __gnu_f2h_ieee183; NOFP16-NEXT: vmov r1, s19184; NOFP16-NEXT: pkhbt r5, r0, r4, lsl #16185; NOFP16-NEXT: mov r0, r1186; NOFP16-NEXT: bl __gnu_f2h_ieee187; NOFP16-NEXT: mov r4, r0188; NOFP16-NEXT: vmov r0, s18189; NOFP16-NEXT: bl __gnu_f2h_ieee190; NOFP16-NEXT: pkhbt r0, r0, r4, lsl #16191; NOFP16-NEXT: vmov d16, r0, r5192; NOFP16-NEXT: vmov.u16 r0, d16[0]193; NOFP16-NEXT: vmov.u16 r1, d16[1]194; NOFP16-NEXT: vmov.u16 r2, d16[2]195; NOFP16-NEXT: vmov.u16 r3, d16[3]196; NOFP16-NEXT: vpop {d8, d9}197; NOFP16-NEXT: pop {r4, r5, r11, pc}198 %fptrunc = call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(<4 x float> %arg, metadata !"round.tonearest", metadata !"fpexcept.strict")199 ret <4 x half> %fptrunc200 }201 202define void @outgoing_v4f16_return(ptr %ptr) #0 {203; NOFP16-LABEL: outgoing_v4f16_return:204; NOFP16: @ %bb.0:205; NOFP16-NEXT: push {r4, lr}206; NOFP16-NEXT: mov r4, r0207; NOFP16-NEXT: bl v4f16_result208; NOFP16-NEXT: strh r3, [r4, #6]209; NOFP16-NEXT: strh r2, [r4, #4]210; NOFP16-NEXT: strh r1, [r4, #2]211; NOFP16-NEXT: strh r0, [r4]212; NOFP16-NEXT: pop {r4, pc}213 %val = call <4 x half> @v4f16_result() #0214 store <4 x half> %val, ptr %ptr215 ret void216}217 218define void @outgoing_v8f16_return(ptr %ptr) #0 {219; NOFP16-LABEL: outgoing_v8f16_return:220; NOFP16: @ %bb.0:221; NOFP16-NEXT: push {r4, r10, r11, lr}222; NOFP16-NEXT: add r11, sp, #8223; NOFP16-NEXT: sub sp, sp, #16224; NOFP16-NEXT: bfc sp, #0, #4225; NOFP16-NEXT: mov r4, r0226; NOFP16-NEXT: mov r0, sp227; NOFP16-NEXT: bl v8f16_result228; NOFP16-NEXT: ldm sp, {r0, r1, r2, r3}229; NOFP16-NEXT: stm r4, {r0, r1, r2, r3}230; NOFP16-NEXT: sub sp, r11, #8231; NOFP16-NEXT: pop {r4, r10, r11, pc}232 %val = call <8 x half> @v8f16_result() #0233 store <8 x half> %val, ptr %ptr234 ret void235}236 237define half @call_split_type_used_outside_block_v8f16() #0 {238; NOFP16-LABEL: call_split_type_used_outside_block_v8f16:239; NOFP16: @ %bb.0: @ %bb0240; NOFP16-NEXT: push {r4, r10, r11, lr}241; NOFP16-NEXT: add r11, sp, #8242; NOFP16-NEXT: sub sp, sp, #16243; NOFP16-NEXT: bfc sp, #0, #4244; NOFP16-NEXT: mov r4, sp245; NOFP16-NEXT: mov r0, r4246; NOFP16-NEXT: bl v8f16_result247; NOFP16-NEXT: vld1.32 {d16[0]}, [r4:32]248; NOFP16-NEXT: vmov.u16 r0, d16[0]249; NOFP16-NEXT: sub sp, r11, #8250; NOFP16-NEXT: pop {r4, r10, r11, pc}251bb0:252 %split.ret.type = call <8 x half> @v8f16_result() #0253 br label %bb1254 255bb1:256 %extract = extractelement <8 x half> %split.ret.type, i32 0257 ret half %extract258}259 260declare float @llvm.experimental.constrained.fpext.f32.f16(half, metadata) #0261declare <2 x float> @llvm.experimental.constrained.fpext.v2f32.v2f16(<2 x half>, metadata) #0262declare <3 x float> @llvm.experimental.constrained.fpext.v3f32.v3f16(<3 x half>, metadata) #0263declare <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half>, metadata) #0264 265declare half @llvm.experimental.constrained.fptrunc.f16.f32(float, metadata, metadata) #0266declare <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float>, metadata, metadata) #0267declare <3 x half> @llvm.experimental.constrained.fptrunc.v3f16.v3f32(<3 x float>, metadata, metadata) #0268declare <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(<4 x float>, metadata, metadata) #0269 270attributes #0 = { strictfp }271