diff options
Diffstat (limited to 'llvm/test/CodeGen/ARM/llvm.sincos.ll')
| -rw-r--r-- | llvm/test/CodeGen/ARM/llvm.sincos.ll | 1131 |
1 files changed, 956 insertions, 175 deletions
diff --git a/llvm/test/CodeGen/ARM/llvm.sincos.ll b/llvm/test/CodeGen/ARM/llvm.sincos.ll index 9628405..1448fac 100644 --- a/llvm/test/CodeGen/ARM/llvm.sincos.ll +++ b/llvm/test/CodeGen/ARM/llvm.sincos.ll @@ -1,223 +1,1004 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2 -; RUN: llc -mtriple=thumbv7-gnu-linux < %s | FileCheck -check-prefixes=CHECK %s +; RUN: llc -mtriple=thumbv7-gnu-linux < %s | FileCheck -check-prefix=GNU %s +; RUN: llc -mtriple=armv7-linux-gnueabi -mcpu=cortex-a8 < %s | FileCheck -check-prefix=GNUEABI %s +; RUN: llc -mtriple=armv7-apple-ios6 -mcpu=cortex-a8 < %s | FileCheck -check-prefixes=IOS,IOS-NO-STRET %s +; RUN: llc -mtriple=armv7-apple-ios7 -mcpu=cortex-a8 < %s | FileCheck -check-prefixes=IOS,IOS-WITH-STRET %s +; RUN: llc -mtriple=thumbv7k-apple-watchos2.0 < %s | FileCheck -check-prefix=WATCHABI %s define { half, half } @test_sincos_f16(half %a) { -; CHECK-LABEL: test_sincos_f16: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r4, lr} -; CHECK-NEXT: sub sp, #8 -; CHECK-NEXT: bl __gnu_h2f_ieee -; CHECK-NEXT: add r1, sp, #4 -; CHECK-NEXT: mov r2, sp -; CHECK-NEXT: bl sincosf -; CHECK-NEXT: ldr r0, [sp, #4] -; CHECK-NEXT: bl __gnu_f2h_ieee -; CHECK-NEXT: mov r4, r0 -; CHECK-NEXT: ldr r0, [sp] -; CHECK-NEXT: bl __gnu_f2h_ieee -; CHECK-NEXT: mov r1, r0 -; CHECK-NEXT: mov r0, r4 -; CHECK-NEXT: add sp, #8 -; CHECK-NEXT: pop {r4, pc} +; GNU-LABEL: test_sincos_f16: +; GNU: @ %bb.0: +; GNU-NEXT: push {r4, lr} +; GNU-NEXT: sub sp, #8 +; GNU-NEXT: bl __gnu_h2f_ieee +; GNU-NEXT: add r1, sp, #4 +; GNU-NEXT: mov r2, sp +; GNU-NEXT: bl sincosf +; GNU-NEXT: ldr r0, [sp, #4] +; GNU-NEXT: bl __gnu_f2h_ieee +; GNU-NEXT: mov r4, r0 +; GNU-NEXT: ldr r0, [sp] +; GNU-NEXT: bl __gnu_f2h_ieee +; GNU-NEXT: mov r1, r0 +; GNU-NEXT: mov r0, r4 +; GNU-NEXT: add sp, #8 +; GNU-NEXT: pop {r4, pc} +; +; GNUEABI-LABEL: test_sincos_f16: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r4, lr} +; GNUEABI-NEXT: push {r4, lr} +; GNUEABI-NEXT: .pad #8 +; GNUEABI-NEXT: sub sp, sp, #8 +; GNUEABI-NEXT: bl __gnu_h2f_ieee +; GNUEABI-NEXT: add r1, sp, #4 +; GNUEABI-NEXT: mov r2, sp +; GNUEABI-NEXT: bl sincosf +; GNUEABI-NEXT: ldr r0, [sp, #4] +; GNUEABI-NEXT: bl __gnu_f2h_ieee +; GNUEABI-NEXT: mov r4, r0 +; GNUEABI-NEXT: ldr r0, [sp] +; GNUEABI-NEXT: bl __gnu_f2h_ieee +; GNUEABI-NEXT: mov r1, r0 +; GNUEABI-NEXT: mov r0, r4 +; GNUEABI-NEXT: add sp, sp, #8 +; GNUEABI-NEXT: pop {r4, pc} +; +; IOS-NO-STRET-LABEL: test_sincos_f16: +; IOS-NO-STRET: @ %bb.0: +; IOS-NO-STRET-NEXT: push {r4, r5, lr} +; IOS-NO-STRET-NEXT: bl ___extendhfsf2 +; IOS-NO-STRET-NEXT: mov r4, r0 +; IOS-NO-STRET-NEXT: bl _sinf +; IOS-NO-STRET-NEXT: bl ___truncsfhf2 +; IOS-NO-STRET-NEXT: mov r5, r0 +; IOS-NO-STRET-NEXT: mov r0, r4 +; IOS-NO-STRET-NEXT: bl _cosf +; IOS-NO-STRET-NEXT: bl ___truncsfhf2 +; IOS-NO-STRET-NEXT: mov r1, r0 +; IOS-NO-STRET-NEXT: mov r0, r5 +; IOS-NO-STRET-NEXT: pop {r4, r5, pc} +; +; IOS-WITH-STRET-LABEL: test_sincos_f16: +; IOS-WITH-STRET: @ %bb.0: +; IOS-WITH-STRET-NEXT: push {r4, r5, lr} +; IOS-WITH-STRET-NEXT: sub sp, sp, #8 +; IOS-WITH-STRET-NEXT: bl ___extendhfsf2 +; IOS-WITH-STRET-NEXT: mov r1, r0 +; IOS-WITH-STRET-NEXT: mov r0, sp +; IOS-WITH-STRET-NEXT: bl ___sincosf_stret +; IOS-WITH-STRET-NEXT: ldm sp, {r0, r4} +; IOS-WITH-STRET-NEXT: bl ___truncsfhf2 +; IOS-WITH-STRET-NEXT: mov r5, r0 +; IOS-WITH-STRET-NEXT: mov r0, r4 +; IOS-WITH-STRET-NEXT: bl ___truncsfhf2 +; IOS-WITH-STRET-NEXT: mov r1, r0 +; IOS-WITH-STRET-NEXT: mov r0, r5 +; IOS-WITH-STRET-NEXT: add sp, sp, #8 +; IOS-WITH-STRET-NEXT: pop {r4, r5, pc} +; +; WATCHABI-LABEL: test_sincos_f16: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push {r7, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 8 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: sub sp, #8 +; WATCHABI-NEXT: .cfi_def_cfa_offset 16 +; WATCHABI-NEXT: vcvtb.f32.f16 s0, s0 +; WATCHABI-NEXT: bl ___sincosf_stret +; WATCHABI-NEXT: vcvtb.f16.f32 s0, s0 +; WATCHABI-NEXT: vcvtb.f16.f32 s1, s1 +; WATCHABI-NEXT: add sp, #8 +; WATCHABI-NEXT: pop {r7, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { half, half } @llvm.sincos.f16(half %a) ret { half, half } %result } define half @test_sincos_f16_only_use_sin(half %a) { -; CHECK-LABEL: test_sincos_f16_only_use_sin: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r7, lr} -; CHECK-NEXT: sub sp, #8 -; CHECK-NEXT: bl __gnu_h2f_ieee -; CHECK-NEXT: add r1, sp, #4 -; CHECK-NEXT: mov r2, sp -; CHECK-NEXT: bl sincosf -; CHECK-NEXT: ldr r0, [sp, #4] -; CHECK-NEXT: bl __gnu_f2h_ieee -; CHECK-NEXT: add sp, #8 -; CHECK-NEXT: pop {r7, pc} +; GNU-LABEL: test_sincos_f16_only_use_sin: +; GNU: @ %bb.0: +; GNU-NEXT: push {r7, lr} +; GNU-NEXT: sub sp, #8 +; GNU-NEXT: bl __gnu_h2f_ieee +; GNU-NEXT: add r1, sp, #4 +; GNU-NEXT: mov r2, sp +; GNU-NEXT: bl sincosf +; GNU-NEXT: ldr r0, [sp, #4] +; GNU-NEXT: bl __gnu_f2h_ieee +; GNU-NEXT: add sp, #8 +; GNU-NEXT: pop {r7, pc} +; +; GNUEABI-LABEL: test_sincos_f16_only_use_sin: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r11, lr} +; GNUEABI-NEXT: push {r11, lr} +; GNUEABI-NEXT: .pad #8 +; GNUEABI-NEXT: sub sp, sp, #8 +; GNUEABI-NEXT: bl __gnu_h2f_ieee +; GNUEABI-NEXT: add r1, sp, #4 +; GNUEABI-NEXT: mov r2, sp +; GNUEABI-NEXT: bl sincosf +; GNUEABI-NEXT: ldr r0, [sp, #4] +; GNUEABI-NEXT: bl __gnu_f2h_ieee +; GNUEABI-NEXT: add sp, sp, #8 +; GNUEABI-NEXT: pop {r11, pc} +; +; IOS-NO-STRET-LABEL: test_sincos_f16_only_use_sin: +; IOS-NO-STRET: @ %bb.0: +; IOS-NO-STRET-NEXT: push {lr} +; IOS-NO-STRET-NEXT: bl ___extendhfsf2 +; IOS-NO-STRET-NEXT: bl _sinf +; IOS-NO-STRET-NEXT: bl ___truncsfhf2 +; IOS-NO-STRET-NEXT: pop {lr} +; IOS-NO-STRET-NEXT: bx lr +; +; IOS-WITH-STRET-LABEL: test_sincos_f16_only_use_sin: +; IOS-WITH-STRET: @ %bb.0: +; IOS-WITH-STRET-NEXT: push {lr} +; IOS-WITH-STRET-NEXT: sub sp, sp, #8 +; IOS-WITH-STRET-NEXT: bl ___extendhfsf2 +; IOS-WITH-STRET-NEXT: mov r1, r0 +; IOS-WITH-STRET-NEXT: mov r0, sp +; IOS-WITH-STRET-NEXT: bl ___sincosf_stret +; IOS-WITH-STRET-NEXT: ldr r0, [sp] +; IOS-WITH-STRET-NEXT: bl ___truncsfhf2 +; IOS-WITH-STRET-NEXT: add sp, sp, #8 +; IOS-WITH-STRET-NEXT: pop {lr} +; IOS-WITH-STRET-NEXT: bx lr +; +; WATCHABI-LABEL: test_sincos_f16_only_use_sin: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push {r7, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 8 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: sub sp, #8 +; WATCHABI-NEXT: .cfi_def_cfa_offset 16 +; WATCHABI-NEXT: vcvtb.f32.f16 s0, s0 +; WATCHABI-NEXT: bl ___sincosf_stret +; WATCHABI-NEXT: vcvtb.f16.f32 s0, s0 +; WATCHABI-NEXT: add sp, #8 +; WATCHABI-NEXT: pop {r7, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { half, half } @llvm.sincos.f16(half %a) %result.0 = extractvalue { half, half } %result, 0 ret half %result.0 } define half @test_sincos_f16_only_use_cos(half %a) { -; CHECK-LABEL: test_sincos_f16_only_use_cos: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r7, lr} -; CHECK-NEXT: sub sp, #8 -; CHECK-NEXT: bl __gnu_h2f_ieee -; CHECK-NEXT: add r1, sp, #4 -; CHECK-NEXT: mov r2, sp -; CHECK-NEXT: bl sincosf -; CHECK-NEXT: ldr r0, [sp] -; CHECK-NEXT: bl __gnu_f2h_ieee -; CHECK-NEXT: add sp, #8 -; CHECK-NEXT: pop {r7, pc} +; GNU-LABEL: test_sincos_f16_only_use_cos: +; GNU: @ %bb.0: +; GNU-NEXT: push {r7, lr} +; GNU-NEXT: sub sp, #8 +; GNU-NEXT: bl __gnu_h2f_ieee +; GNU-NEXT: add r1, sp, #4 +; GNU-NEXT: mov r2, sp +; GNU-NEXT: bl sincosf +; GNU-NEXT: ldr r0, [sp] +; GNU-NEXT: bl __gnu_f2h_ieee +; GNU-NEXT: add sp, #8 +; GNU-NEXT: pop {r7, pc} +; +; GNUEABI-LABEL: test_sincos_f16_only_use_cos: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r11, lr} +; GNUEABI-NEXT: push {r11, lr} +; GNUEABI-NEXT: .pad #8 +; GNUEABI-NEXT: sub sp, sp, #8 +; GNUEABI-NEXT: bl __gnu_h2f_ieee +; GNUEABI-NEXT: add r1, sp, #4 +; GNUEABI-NEXT: mov r2, sp +; GNUEABI-NEXT: bl sincosf +; GNUEABI-NEXT: ldr r0, [sp] +; GNUEABI-NEXT: bl __gnu_f2h_ieee +; GNUEABI-NEXT: add sp, sp, #8 +; GNUEABI-NEXT: pop {r11, pc} +; +; IOS-NO-STRET-LABEL: test_sincos_f16_only_use_cos: +; IOS-NO-STRET: @ %bb.0: +; IOS-NO-STRET-NEXT: push {lr} +; IOS-NO-STRET-NEXT: bl ___extendhfsf2 +; IOS-NO-STRET-NEXT: bl _cosf +; IOS-NO-STRET-NEXT: bl ___truncsfhf2 +; IOS-NO-STRET-NEXT: pop {lr} +; IOS-NO-STRET-NEXT: bx lr +; +; IOS-WITH-STRET-LABEL: test_sincos_f16_only_use_cos: +; IOS-WITH-STRET: @ %bb.0: +; IOS-WITH-STRET-NEXT: push {lr} +; IOS-WITH-STRET-NEXT: sub sp, sp, #8 +; IOS-WITH-STRET-NEXT: bl ___extendhfsf2 +; IOS-WITH-STRET-NEXT: mov r1, r0 +; IOS-WITH-STRET-NEXT: mov r0, sp +; IOS-WITH-STRET-NEXT: bl ___sincosf_stret +; IOS-WITH-STRET-NEXT: ldr r0, [sp, #4] +; IOS-WITH-STRET-NEXT: bl ___truncsfhf2 +; IOS-WITH-STRET-NEXT: add sp, sp, #8 +; IOS-WITH-STRET-NEXT: pop {lr} +; IOS-WITH-STRET-NEXT: bx lr +; +; WATCHABI-LABEL: test_sincos_f16_only_use_cos: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push {r7, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 8 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: sub sp, #8 +; WATCHABI-NEXT: .cfi_def_cfa_offset 16 +; WATCHABI-NEXT: vcvtb.f32.f16 s0, s0 +; WATCHABI-NEXT: bl ___sincosf_stret +; WATCHABI-NEXT: vcvtb.f16.f32 s0, s1 +; WATCHABI-NEXT: add sp, #8 +; WATCHABI-NEXT: pop {r7, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { half, half } @llvm.sincos.f16(half %a) %result.1 = extractvalue { half, half } %result, 1 ret half %result.1 } define { <2 x half>, <2 x half> } @test_sincos_v2f16(<2 x half> %a) { -; CHECK-LABEL: test_sincos_v2f16: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r4, lr} -; CHECK-NEXT: vpush {d8} -; CHECK-NEXT: sub sp, #24 -; CHECK-NEXT: mov r4, r0 -; CHECK-NEXT: mov r0, r1 -; CHECK-NEXT: bl __gnu_h2f_ieee -; CHECK-NEXT: add r1, sp, #12 -; CHECK-NEXT: add r2, sp, #8 -; CHECK-NEXT: bl sincosf -; CHECK-NEXT: mov r0, r4 -; CHECK-NEXT: bl __gnu_h2f_ieee -; CHECK-NEXT: add r1, sp, #4 -; CHECK-NEXT: mov r2, sp -; CHECK-NEXT: bl sincosf -; CHECK-NEXT: ldr r0, [sp, #12] -; CHECK-NEXT: bl __gnu_f2h_ieee -; CHECK-NEXT: ldr r1, [sp, #4] -; CHECK-NEXT: strh.w r0, [sp, #22] -; CHECK-NEXT: mov r0, r1 -; CHECK-NEXT: bl __gnu_f2h_ieee -; CHECK-NEXT: strh.w r0, [sp, #20] -; CHECK-NEXT: add r0, sp, #20 -; CHECK-NEXT: vld1.32 {d8[0]}, [r0:32] -; CHECK-NEXT: ldr r0, [sp, #8] -; CHECK-NEXT: bl __gnu_f2h_ieee -; CHECK-NEXT: ldr r1, [sp] -; CHECK-NEXT: strh.w r0, [sp, #18] -; CHECK-NEXT: mov r0, r1 -; CHECK-NEXT: bl __gnu_f2h_ieee -; CHECK-NEXT: strh.w r0, [sp, #16] -; CHECK-NEXT: add r0, sp, #16 -; CHECK-NEXT: vmovl.u16 q9, d8 -; CHECK-NEXT: vld1.32 {d16[0]}, [r0:32] -; CHECK-NEXT: vmovl.u16 q8, d16 -; CHECK-NEXT: vmov.32 r0, d18[0] -; CHECK-NEXT: vmov.32 r1, d18[1] -; CHECK-NEXT: vmov.32 r2, d16[0] -; CHECK-NEXT: vmov.32 r3, d16[1] -; CHECK-NEXT: add sp, #24 -; CHECK-NEXT: vpop {d8} -; CHECK-NEXT: pop {r4, pc} +; GNU-LABEL: test_sincos_v2f16: +; GNU: @ %bb.0: +; GNU-NEXT: push {r4, lr} +; GNU-NEXT: vpush {d8} +; GNU-NEXT: sub sp, #24 +; GNU-NEXT: mov r4, r0 +; GNU-NEXT: mov r0, r1 +; GNU-NEXT: bl __gnu_h2f_ieee +; GNU-NEXT: add r1, sp, #12 +; GNU-NEXT: add r2, sp, #8 +; GNU-NEXT: bl sincosf +; GNU-NEXT: mov r0, r4 +; GNU-NEXT: bl __gnu_h2f_ieee +; GNU-NEXT: add r1, sp, #4 +; GNU-NEXT: mov r2, sp +; GNU-NEXT: bl sincosf +; GNU-NEXT: ldr r0, [sp, #12] +; GNU-NEXT: bl __gnu_f2h_ieee +; GNU-NEXT: ldr r1, [sp, #4] +; GNU-NEXT: strh.w r0, [sp, #22] +; GNU-NEXT: mov r0, r1 +; GNU-NEXT: bl __gnu_f2h_ieee +; GNU-NEXT: strh.w r0, [sp, #20] +; GNU-NEXT: add r0, sp, #20 +; GNU-NEXT: vld1.32 {d8[0]}, [r0:32] +; GNU-NEXT: ldr r0, [sp, #8] +; GNU-NEXT: bl __gnu_f2h_ieee +; GNU-NEXT: ldr r1, [sp] +; GNU-NEXT: strh.w r0, [sp, #18] +; GNU-NEXT: mov r0, r1 +; GNU-NEXT: bl __gnu_f2h_ieee +; GNU-NEXT: strh.w r0, [sp, #16] +; GNU-NEXT: add r0, sp, #16 +; GNU-NEXT: vmovl.u16 q9, d8 +; GNU-NEXT: vld1.32 {d16[0]}, [r0:32] +; GNU-NEXT: vmovl.u16 q8, d16 +; GNU-NEXT: vmov.32 r0, d18[0] +; GNU-NEXT: vmov.32 r1, d18[1] +; GNU-NEXT: vmov.32 r2, d16[0] +; GNU-NEXT: vmov.32 r3, d16[1] +; GNU-NEXT: add sp, #24 +; GNU-NEXT: vpop {d8} +; GNU-NEXT: pop {r4, pc} +; +; GNUEABI-LABEL: test_sincos_v2f16: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r4, lr} +; GNUEABI-NEXT: push {r4, lr} +; GNUEABI-NEXT: .vsave {d8} +; GNUEABI-NEXT: vpush {d8} +; GNUEABI-NEXT: .pad #24 +; GNUEABI-NEXT: sub sp, sp, #24 +; GNUEABI-NEXT: mov r4, r0 +; GNUEABI-NEXT: mov r0, r1 +; GNUEABI-NEXT: bl __gnu_h2f_ieee +; GNUEABI-NEXT: add r1, sp, #12 +; GNUEABI-NEXT: add r2, sp, #8 +; GNUEABI-NEXT: bl sincosf +; GNUEABI-NEXT: mov r0, r4 +; GNUEABI-NEXT: bl __gnu_h2f_ieee +; GNUEABI-NEXT: add r1, sp, #4 +; GNUEABI-NEXT: mov r2, sp +; GNUEABI-NEXT: bl sincosf +; GNUEABI-NEXT: ldr r0, [sp, #12] +; GNUEABI-NEXT: bl __gnu_f2h_ieee +; GNUEABI-NEXT: ldr r1, [sp, #4] +; GNUEABI-NEXT: strh r0, [sp, #22] +; GNUEABI-NEXT: mov r0, r1 +; GNUEABI-NEXT: bl __gnu_f2h_ieee +; GNUEABI-NEXT: strh r0, [sp, #20] +; GNUEABI-NEXT: add r0, sp, #20 +; GNUEABI-NEXT: vld1.32 {d8[0]}, [r0:32] +; GNUEABI-NEXT: ldr r0, [sp, #8] +; GNUEABI-NEXT: bl __gnu_f2h_ieee +; GNUEABI-NEXT: ldr r1, [sp] +; GNUEABI-NEXT: strh r0, [sp, #18] +; GNUEABI-NEXT: mov r0, r1 +; GNUEABI-NEXT: bl __gnu_f2h_ieee +; GNUEABI-NEXT: strh r0, [sp, #16] +; GNUEABI-NEXT: add r0, sp, #16 +; GNUEABI-NEXT: vmovl.u16 q9, d8 +; GNUEABI-NEXT: vld1.32 {d16[0]}, [r0:32] +; GNUEABI-NEXT: vmovl.u16 q8, d16 +; GNUEABI-NEXT: vmov.32 r0, d18[0] +; GNUEABI-NEXT: vmov.32 r1, d18[1] +; GNUEABI-NEXT: vmov.32 r2, d16[0] +; GNUEABI-NEXT: vmov.32 r3, d16[1] +; GNUEABI-NEXT: add sp, sp, #24 +; GNUEABI-NEXT: vpop {d8} +; GNUEABI-NEXT: pop {r4, pc} +; +; IOS-NO-STRET-LABEL: test_sincos_v2f16: +; IOS-NO-STRET: @ %bb.0: +; IOS-NO-STRET-NEXT: push {r4, r5, lr} +; IOS-NO-STRET-NEXT: vpush {d8} +; IOS-NO-STRET-NEXT: sub sp, sp, #8 +; IOS-NO-STRET-NEXT: mov r5, r0 +; IOS-NO-STRET-NEXT: mov r0, r1 +; IOS-NO-STRET-NEXT: bl ___extendhfsf2 +; IOS-NO-STRET-NEXT: mov r4, r0 +; IOS-NO-STRET-NEXT: bl _sinf +; IOS-NO-STRET-NEXT: bl ___truncsfhf2 +; IOS-NO-STRET-NEXT: strh r0, [sp, #6] +; IOS-NO-STRET-NEXT: mov r0, r5 +; IOS-NO-STRET-NEXT: bl ___extendhfsf2 +; IOS-NO-STRET-NEXT: mov r5, r0 +; IOS-NO-STRET-NEXT: bl _sinf +; IOS-NO-STRET-NEXT: bl ___truncsfhf2 +; IOS-NO-STRET-NEXT: strh r0, [sp, #4] +; IOS-NO-STRET-NEXT: add r0, sp, #4 +; IOS-NO-STRET-NEXT: vld1.32 {d8[0]}, [r0:32] +; IOS-NO-STRET-NEXT: mov r0, r4 +; IOS-NO-STRET-NEXT: bl _cosf +; IOS-NO-STRET-NEXT: bl ___truncsfhf2 +; IOS-NO-STRET-NEXT: strh r0, [sp, #2] +; IOS-NO-STRET-NEXT: mov r0, r5 +; IOS-NO-STRET-NEXT: bl _cosf +; IOS-NO-STRET-NEXT: bl ___truncsfhf2 +; IOS-NO-STRET-NEXT: strh r0, [sp] +; IOS-NO-STRET-NEXT: mov r0, sp +; IOS-NO-STRET-NEXT: vld1.32 {d16[0]}, [r0:32] +; IOS-NO-STRET-NEXT: vmovl.u16 q9, d8 +; IOS-NO-STRET-NEXT: vmovl.u16 q8, d16 +; IOS-NO-STRET-NEXT: vmov.32 r0, d18[0] +; IOS-NO-STRET-NEXT: vmov.32 r1, d18[1] +; IOS-NO-STRET-NEXT: vmov.32 r2, d16[0] +; IOS-NO-STRET-NEXT: vmov.32 r3, d16[1] +; IOS-NO-STRET-NEXT: add sp, sp, #8 +; IOS-NO-STRET-NEXT: vpop {d8} +; IOS-NO-STRET-NEXT: pop {r4, r5, pc} +; +; IOS-WITH-STRET-LABEL: test_sincos_v2f16: +; IOS-WITH-STRET: @ %bb.0: +; IOS-WITH-STRET-NEXT: push {r4, r5, lr} +; IOS-WITH-STRET-NEXT: vpush {d8} +; IOS-WITH-STRET-NEXT: sub sp, sp, #24 +; IOS-WITH-STRET-NEXT: mov r4, r0 +; IOS-WITH-STRET-NEXT: mov r0, r1 +; IOS-WITH-STRET-NEXT: bl ___extendhfsf2 +; IOS-WITH-STRET-NEXT: mov r1, r0 +; IOS-WITH-STRET-NEXT: add r0, sp, #8 +; IOS-WITH-STRET-NEXT: bl ___sincosf_stret +; IOS-WITH-STRET-NEXT: mov r0, r4 +; IOS-WITH-STRET-NEXT: bl ___extendhfsf2 +; IOS-WITH-STRET-NEXT: mov r1, r0 +; IOS-WITH-STRET-NEXT: mov r0, sp +; IOS-WITH-STRET-NEXT: bl ___sincosf_stret +; IOS-WITH-STRET-NEXT: ldr r0, [sp, #8] +; IOS-WITH-STRET-NEXT: ldr r4, [sp, #12] +; IOS-WITH-STRET-NEXT: bl ___truncsfhf2 +; IOS-WITH-STRET-NEXT: ldm sp, {r1, r5} +; IOS-WITH-STRET-NEXT: strh r0, [sp, #22] +; IOS-WITH-STRET-NEXT: mov r0, r1 +; IOS-WITH-STRET-NEXT: bl ___truncsfhf2 +; IOS-WITH-STRET-NEXT: strh r0, [sp, #20] +; IOS-WITH-STRET-NEXT: add r0, sp, #20 +; IOS-WITH-STRET-NEXT: vld1.32 {d8[0]}, [r0:32] +; IOS-WITH-STRET-NEXT: mov r0, r4 +; IOS-WITH-STRET-NEXT: bl ___truncsfhf2 +; IOS-WITH-STRET-NEXT: strh r0, [sp, #18] +; IOS-WITH-STRET-NEXT: mov r0, r5 +; IOS-WITH-STRET-NEXT: bl ___truncsfhf2 +; IOS-WITH-STRET-NEXT: strh r0, [sp, #16] +; IOS-WITH-STRET-NEXT: add r0, sp, #16 +; IOS-WITH-STRET-NEXT: vmovl.u16 q9, d8 +; IOS-WITH-STRET-NEXT: vld1.32 {d16[0]}, [r0:32] +; IOS-WITH-STRET-NEXT: vmovl.u16 q8, d16 +; IOS-WITH-STRET-NEXT: vmov.32 r0, d18[0] +; IOS-WITH-STRET-NEXT: vmov.32 r1, d18[1] +; IOS-WITH-STRET-NEXT: vmov.32 r2, d16[0] +; IOS-WITH-STRET-NEXT: vmov.32 r3, d16[1] +; IOS-WITH-STRET-NEXT: add sp, sp, #24 +; IOS-WITH-STRET-NEXT: vpop {d8} +; IOS-WITH-STRET-NEXT: pop {r4, r5, pc} +; +; WATCHABI-LABEL: test_sincos_v2f16: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push {r7, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 8 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: vpush {d10} +; WATCHABI-NEXT: .cfi_def_cfa_offset 16 +; WATCHABI-NEXT: vpush {d8} +; WATCHABI-NEXT: .cfi_def_cfa_offset 24 +; WATCHABI-NEXT: .cfi_offset d10, -16 +; WATCHABI-NEXT: .cfi_offset d8, -24 +; WATCHABI-NEXT: sub sp, #8 +; WATCHABI-NEXT: .cfi_def_cfa_offset 32 +; WATCHABI-NEXT: vmov.f32 s16, s0 +; WATCHABI-NEXT: vcvtb.f32.f16 s0, s1 +; WATCHABI-NEXT: bl ___sincosf_stret +; WATCHABI-NEXT: vcvtb.f16.f32 s0, s0 +; WATCHABI-NEXT: vcvtb.f32.f16 s4, s16 +; WATCHABI-NEXT: vmov r0, s0 +; WATCHABI-NEXT: vmov.f32 s0, s4 +; WATCHABI-NEXT: vmov.f32 s20, s1 +; WATCHABI-NEXT: strh.w r0, [sp, #6] +; WATCHABI-NEXT: bl ___sincosf_stret +; WATCHABI-NEXT: vcvtb.f16.f32 s0, s0 +; WATCHABI-NEXT: vmov r0, s0 +; WATCHABI-NEXT: vcvtb.f16.f32 s0, s20 +; WATCHABI-NEXT: strh.w r0, [sp, #4] +; WATCHABI-NEXT: add r0, sp, #4 +; WATCHABI-NEXT: vld1.32 {d16[0]}, [r0:32] +; WATCHABI-NEXT: vmov r0, s0 +; WATCHABI-NEXT: vcvtb.f16.f32 s0, s1 +; WATCHABI-NEXT: strh.w r0, [sp, #2] +; WATCHABI-NEXT: vmov r0, s0 +; WATCHABI-NEXT: vmovl.u16 q0, d16 +; WATCHABI-NEXT: strh.w r0, [sp] +; WATCHABI-NEXT: mov r0, sp +; WATCHABI-NEXT: vld1.32 {d18[0]}, [r0:32] +; WATCHABI-NEXT: vmovl.u16 q1, d18 +; WATCHABI-NEXT: vmov.f32 s2, s4 +; WATCHABI-NEXT: vmov.f32 s3, s5 +; WATCHABI-NEXT: add sp, #8 +; WATCHABI-NEXT: vpop {d8} +; WATCHABI-NEXT: vpop {d10} +; WATCHABI-NEXT: pop {r7, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { <2 x half>, <2 x half> } @llvm.sincos.v2f16(<2 x half> %a) ret { <2 x half>, <2 x half> } %result } define { float, float } @test_sincos_f32(float %a) { -; CHECK-LABEL: test_sincos_f32: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r7, lr} -; CHECK-NEXT: sub sp, #8 -; CHECK-NEXT: add r1, sp, #4 -; CHECK-NEXT: mov r2, sp -; CHECK-NEXT: bl sincosf -; CHECK-NEXT: ldrd r1, r0, [sp], #8 -; CHECK-NEXT: pop {r7, pc} +; GNU-LABEL: test_sincos_f32: +; GNU: @ %bb.0: +; GNU-NEXT: push {r7, lr} +; GNU-NEXT: sub sp, #8 +; GNU-NEXT: add r1, sp, #4 +; GNU-NEXT: mov r2, sp +; GNU-NEXT: bl sincosf +; GNU-NEXT: ldrd r1, r0, [sp], #8 +; GNU-NEXT: pop {r7, pc} +; +; GNUEABI-LABEL: test_sincos_f32: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r11, lr} +; GNUEABI-NEXT: push {r11, lr} +; GNUEABI-NEXT: .pad #8 +; GNUEABI-NEXT: sub sp, sp, #8 +; GNUEABI-NEXT: add r1, sp, #4 +; GNUEABI-NEXT: mov r2, sp +; GNUEABI-NEXT: bl sincosf +; GNUEABI-NEXT: ldr r0, [sp, #4] +; GNUEABI-NEXT: ldr r1, [sp], #8 +; GNUEABI-NEXT: pop {r11, pc} +; +; IOS-NO-STRET-LABEL: test_sincos_f32: +; IOS-NO-STRET: @ %bb.0: +; IOS-NO-STRET-NEXT: push {r4, r5, lr} +; IOS-NO-STRET-NEXT: mov r4, r0 +; IOS-NO-STRET-NEXT: bl _sinf +; IOS-NO-STRET-NEXT: mov r5, r0 +; IOS-NO-STRET-NEXT: mov r0, r4 +; IOS-NO-STRET-NEXT: bl _cosf +; IOS-NO-STRET-NEXT: mov r1, r0 +; IOS-NO-STRET-NEXT: mov r0, r5 +; IOS-NO-STRET-NEXT: pop {r4, r5, pc} +; +; IOS-WITH-STRET-LABEL: test_sincos_f32: +; IOS-WITH-STRET: @ %bb.0: +; IOS-WITH-STRET-NEXT: push {lr} +; IOS-WITH-STRET-NEXT: sub sp, sp, #8 +; IOS-WITH-STRET-NEXT: mov r1, r0 +; IOS-WITH-STRET-NEXT: mov r0, sp +; IOS-WITH-STRET-NEXT: bl ___sincosf_stret +; IOS-WITH-STRET-NEXT: pop {r0, r1} +; IOS-WITH-STRET-NEXT: pop {lr} +; IOS-WITH-STRET-NEXT: bx lr +; +; WATCHABI-LABEL: test_sincos_f32: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push {r7, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 8 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: sub sp, #8 +; WATCHABI-NEXT: .cfi_def_cfa_offset 16 +; WATCHABI-NEXT: bl ___sincosf_stret +; WATCHABI-NEXT: add sp, #8 +; WATCHABI-NEXT: pop {r7, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { float, float } @llvm.sincos.f32(float %a) ret { float, float } %result } define { <2 x float>, <2 x float> } @test_sincos_v2f32(<2 x float> %a) { -; CHECK-LABEL: test_sincos_v2f32: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r7, lr} -; CHECK-NEXT: vpush {d8} -; CHECK-NEXT: sub sp, #16 -; CHECK-NEXT: vmov d8, r0, r1 -; CHECK-NEXT: add r1, sp, #4 -; CHECK-NEXT: mov r2, sp -; CHECK-NEXT: vmov r0, s17 -; CHECK-NEXT: bl sincosf -; CHECK-NEXT: vmov r0, s16 -; CHECK-NEXT: add r1, sp, #12 -; CHECK-NEXT: add r2, sp, #8 -; CHECK-NEXT: bl sincosf -; CHECK-NEXT: vldr s1, [sp, #4] -; CHECK-NEXT: vldr s3, [sp] -; CHECK-NEXT: vldr s0, [sp, #12] -; CHECK-NEXT: vldr s2, [sp, #8] -; CHECK-NEXT: vmov r0, r1, d0 -; CHECK-NEXT: vmov r2, r3, d1 -; CHECK-NEXT: add sp, #16 -; CHECK-NEXT: vpop {d8} -; CHECK-NEXT: pop {r7, pc} +; GNU-LABEL: test_sincos_v2f32: +; GNU: @ %bb.0: +; GNU-NEXT: push {r7, lr} +; GNU-NEXT: vpush {d8} +; GNU-NEXT: sub sp, #16 +; GNU-NEXT: vmov d8, r0, r1 +; GNU-NEXT: add r1, sp, #4 +; GNU-NEXT: mov r2, sp +; GNU-NEXT: vmov r0, s17 +; GNU-NEXT: bl sincosf +; GNU-NEXT: vmov r0, s16 +; GNU-NEXT: add r1, sp, #12 +; GNU-NEXT: add r2, sp, #8 +; GNU-NEXT: bl sincosf +; GNU-NEXT: vldr s1, [sp, #4] +; GNU-NEXT: vldr s3, [sp] +; GNU-NEXT: vldr s0, [sp, #12] +; GNU-NEXT: vldr s2, [sp, #8] +; GNU-NEXT: vmov r0, r1, d0 +; GNU-NEXT: vmov r2, r3, d1 +; GNU-NEXT: add sp, #16 +; GNU-NEXT: vpop {d8} +; GNU-NEXT: pop {r7, pc} +; +; GNUEABI-LABEL: test_sincos_v2f32: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r11, lr} +; GNUEABI-NEXT: push {r11, lr} +; GNUEABI-NEXT: .vsave {d8} +; GNUEABI-NEXT: vpush {d8} +; GNUEABI-NEXT: .pad #16 +; GNUEABI-NEXT: sub sp, sp, #16 +; GNUEABI-NEXT: vmov d8, r0, r1 +; GNUEABI-NEXT: add r1, sp, #4 +; GNUEABI-NEXT: mov r2, sp +; GNUEABI-NEXT: vmov r0, s17 +; GNUEABI-NEXT: bl sincosf +; GNUEABI-NEXT: vmov r0, s16 +; GNUEABI-NEXT: add r1, sp, #12 +; GNUEABI-NEXT: add r2, sp, #8 +; GNUEABI-NEXT: bl sincosf +; GNUEABI-NEXT: vldr s1, [sp, #4] +; GNUEABI-NEXT: vldr s3, [sp] +; GNUEABI-NEXT: vldr s0, [sp, #12] +; GNUEABI-NEXT: vldr s2, [sp, #8] +; GNUEABI-NEXT: vmov r0, r1, d0 +; GNUEABI-NEXT: vmov r2, r3, d1 +; GNUEABI-NEXT: add sp, sp, #16 +; GNUEABI-NEXT: vpop {d8} +; GNUEABI-NEXT: pop {r11, pc} +; +; IOS-NO-STRET-LABEL: test_sincos_v2f32: +; IOS-NO-STRET: @ %bb.0: +; IOS-NO-STRET-NEXT: push {r4, r5, r6, r7, lr} +; IOS-NO-STRET-NEXT: vpush {d8} +; IOS-NO-STRET-NEXT: vmov d8, r0, r1 +; IOS-NO-STRET-NEXT: vmov r4, s17 +; IOS-NO-STRET-NEXT: mov r0, r4 +; IOS-NO-STRET-NEXT: bl _sinf +; IOS-NO-STRET-NEXT: mov r5, r0 +; IOS-NO-STRET-NEXT: mov r0, r4 +; IOS-NO-STRET-NEXT: bl _cosf +; IOS-NO-STRET-NEXT: vmov r6, s16 +; IOS-NO-STRET-NEXT: mov r4, r0 +; IOS-NO-STRET-NEXT: mov r0, r6 +; IOS-NO-STRET-NEXT: bl _sinf +; IOS-NO-STRET-NEXT: mov r7, r0 +; IOS-NO-STRET-NEXT: mov r0, r6 +; IOS-NO-STRET-NEXT: bl _cosf +; IOS-NO-STRET-NEXT: mov r2, r0 +; IOS-NO-STRET-NEXT: mov r0, r7 +; IOS-NO-STRET-NEXT: mov r1, r5 +; IOS-NO-STRET-NEXT: mov r3, r4 +; IOS-NO-STRET-NEXT: vpop {d8} +; IOS-NO-STRET-NEXT: pop {r4, r5, r6, r7, pc} +; +; IOS-WITH-STRET-LABEL: test_sincos_v2f32: +; IOS-WITH-STRET: @ %bb.0: +; IOS-WITH-STRET-NEXT: push {lr} +; IOS-WITH-STRET-NEXT: vpush {d8} +; IOS-WITH-STRET-NEXT: sub sp, sp, #16 +; IOS-WITH-STRET-NEXT: vmov d8, r0, r1 +; IOS-WITH-STRET-NEXT: mov r0, sp +; IOS-WITH-STRET-NEXT: vmov r1, s17 +; IOS-WITH-STRET-NEXT: bl ___sincosf_stret +; IOS-WITH-STRET-NEXT: vmov r1, s16 +; IOS-WITH-STRET-NEXT: add r0, sp, #8 +; IOS-WITH-STRET-NEXT: bl ___sincosf_stret +; IOS-WITH-STRET-NEXT: vldr s1, [sp] +; IOS-WITH-STRET-NEXT: vldr s3, [sp, #4] +; IOS-WITH-STRET-NEXT: vldr s0, [sp, #8] +; IOS-WITH-STRET-NEXT: vldr s2, [sp, #12] +; IOS-WITH-STRET-NEXT: vmov r0, r1, d0 +; IOS-WITH-STRET-NEXT: vmov r2, r3, d1 +; IOS-WITH-STRET-NEXT: add sp, sp, #16 +; IOS-WITH-STRET-NEXT: vpop {d8} +; IOS-WITH-STRET-NEXT: pop {lr} +; IOS-WITH-STRET-NEXT: bx lr +; +; WATCHABI-LABEL: test_sincos_v2f32: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push {r7, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 8 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: vpush {d8, d9, d10} +; WATCHABI-NEXT: .cfi_def_cfa_offset 32 +; WATCHABI-NEXT: .cfi_offset d10, -16 +; WATCHABI-NEXT: .cfi_offset d9, -24 +; WATCHABI-NEXT: .cfi_offset d8, -32 +; WATCHABI-NEXT: vmov.f64 d8, d0 +; WATCHABI-NEXT: vmov.f32 s0, s17 +; WATCHABI-NEXT: bl ___sincosf_stret +; WATCHABI-NEXT: vmov.f32 s19, s0 +; WATCHABI-NEXT: vmov.f32 s0, s16 +; WATCHABI-NEXT: vmov.f32 s21, s1 +; WATCHABI-NEXT: bl ___sincosf_stret +; WATCHABI-NEXT: vmov.f32 s20, s1 +; WATCHABI-NEXT: vmov.f32 s18, s0 +; WATCHABI-NEXT: vmov.f64 d1, d10 +; WATCHABI-NEXT: vmov.f64 d0, d9 +; WATCHABI-NEXT: vpop {d8, d9, d10} +; WATCHABI-NEXT: pop {r7, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { <2 x float>, <2 x float> } @llvm.sincos.v2f32(<2 x float> %a) ret { <2 x float>, <2 x float> } %result } define { double, double } @test_sincos_f64(double %a) { -; CHECK-LABEL: test_sincos_f64: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r7, lr} -; CHECK-NEXT: sub sp, #16 -; CHECK-NEXT: add r2, sp, #8 -; CHECK-NEXT: mov r3, sp -; CHECK-NEXT: bl sincos -; CHECK-NEXT: ldrd r0, r1, [sp, #8] -; CHECK-NEXT: ldrd r2, r3, [sp], #16 -; CHECK-NEXT: pop {r7, pc} +; GNU-LABEL: test_sincos_f64: +; GNU: @ %bb.0: +; GNU-NEXT: push {r7, lr} +; GNU-NEXT: sub sp, #16 +; GNU-NEXT: add r2, sp, #8 +; GNU-NEXT: mov r3, sp +; GNU-NEXT: bl sincos +; GNU-NEXT: ldrd r0, r1, [sp, #8] +; GNU-NEXT: ldrd r2, r3, [sp], #16 +; GNU-NEXT: pop {r7, pc} +; +; GNUEABI-LABEL: test_sincos_f64: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r11, lr} +; GNUEABI-NEXT: push {r11, lr} +; GNUEABI-NEXT: .pad #16 +; GNUEABI-NEXT: sub sp, sp, #16 +; GNUEABI-NEXT: add r2, sp, #8 +; GNUEABI-NEXT: mov r3, sp +; GNUEABI-NEXT: bl sincos +; GNUEABI-NEXT: ldm sp, {r2, r3} +; GNUEABI-NEXT: ldr r0, [sp, #8] +; GNUEABI-NEXT: ldr r1, [sp, #12] +; GNUEABI-NEXT: add sp, sp, #16 +; GNUEABI-NEXT: pop {r11, pc} +; +; IOS-NO-STRET-LABEL: test_sincos_f64: +; IOS-NO-STRET: @ %bb.0: +; IOS-NO-STRET-NEXT: push {r4, r5, r6, r7, lr} +; IOS-NO-STRET-NEXT: mov r4, r1 +; IOS-NO-STRET-NEXT: mov r5, r0 +; IOS-NO-STRET-NEXT: bl _sin +; IOS-NO-STRET-NEXT: mov r6, r0 +; IOS-NO-STRET-NEXT: mov r7, r1 +; IOS-NO-STRET-NEXT: mov r0, r5 +; IOS-NO-STRET-NEXT: mov r1, r4 +; IOS-NO-STRET-NEXT: bl _cos +; IOS-NO-STRET-NEXT: mov r2, r0 +; IOS-NO-STRET-NEXT: mov r3, r1 +; IOS-NO-STRET-NEXT: mov r0, r6 +; IOS-NO-STRET-NEXT: mov r1, r7 +; IOS-NO-STRET-NEXT: pop {r4, r5, r6, r7, pc} +; +; IOS-WITH-STRET-LABEL: test_sincos_f64: +; IOS-WITH-STRET: @ %bb.0: +; IOS-WITH-STRET-NEXT: push {lr} +; IOS-WITH-STRET-NEXT: sub sp, sp, #16 +; IOS-WITH-STRET-NEXT: mov r2, r1 +; IOS-WITH-STRET-NEXT: mov r1, r0 +; IOS-WITH-STRET-NEXT: mov r0, sp +; IOS-WITH-STRET-NEXT: bl ___sincos_stret +; IOS-WITH-STRET-NEXT: vldr d16, [sp, #8] +; IOS-WITH-STRET-NEXT: ldm sp, {r0, r1} +; IOS-WITH-STRET-NEXT: vmov r2, r3, d16 +; IOS-WITH-STRET-NEXT: add sp, sp, #16 +; IOS-WITH-STRET-NEXT: pop {lr} +; IOS-WITH-STRET-NEXT: bx lr +; +; WATCHABI-LABEL: test_sincos_f64: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push {r7, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 8 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: sub sp, #8 +; WATCHABI-NEXT: .cfi_def_cfa_offset 16 +; WATCHABI-NEXT: bl ___sincos_stret +; WATCHABI-NEXT: add sp, #8 +; WATCHABI-NEXT: pop {r7, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { double, double } @llvm.sincos.f64(double %a) ret { double, double } %result } define { <2 x double>, <2 x double> } @test_sincos_v2f64(<2 x double> %a) { -; CHECK-LABEL: test_sincos_v2f64: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r4, lr} -; CHECK-NEXT: sub sp, #32 -; CHECK-NEXT: mov r1, r3 -; CHECK-NEXT: mov r12, r2 -; CHECK-NEXT: add r2, sp, #24 -; CHECK-NEXT: add r3, sp, #16 -; CHECK-NEXT: mov r4, r0 -; CHECK-NEXT: mov r0, r12 -; CHECK-NEXT: bl sincos -; CHECK-NEXT: ldrd r0, r1, [sp, #40] -; CHECK-NEXT: add r2, sp, #8 -; CHECK-NEXT: mov r3, sp -; CHECK-NEXT: bl sincos -; CHECK-NEXT: vldr d19, [sp, #8] -; CHECK-NEXT: vldr d18, [sp, #24] -; CHECK-NEXT: vldr d17, [sp] -; CHECK-NEXT: vldr d16, [sp, #16] -; CHECK-NEXT: vst1.64 {d18, d19}, [r4]! -; CHECK-NEXT: vst1.64 {d16, d17}, [r4] -; CHECK-NEXT: add sp, #32 -; CHECK-NEXT: pop {r4, pc} +; GNU-LABEL: test_sincos_v2f64: +; GNU: @ %bb.0: +; GNU-NEXT: push {r4, lr} +; GNU-NEXT: sub sp, #32 +; GNU-NEXT: mov r1, r3 +; GNU-NEXT: mov r12, r2 +; GNU-NEXT: add r2, sp, #24 +; GNU-NEXT: add r3, sp, #16 +; GNU-NEXT: mov r4, r0 +; GNU-NEXT: mov r0, r12 +; GNU-NEXT: bl sincos +; GNU-NEXT: ldrd r0, r1, [sp, #40] +; GNU-NEXT: add r2, sp, #8 +; GNU-NEXT: mov r3, sp +; GNU-NEXT: bl sincos +; GNU-NEXT: vldr d19, [sp, #8] +; GNU-NEXT: vldr d18, [sp, #24] +; GNU-NEXT: vldr d17, [sp] +; GNU-NEXT: vldr d16, [sp, #16] +; GNU-NEXT: vst1.64 {d18, d19}, [r4]! +; GNU-NEXT: vst1.64 {d16, d17}, [r4] +; GNU-NEXT: add sp, #32 +; GNU-NEXT: pop {r4, pc} +; +; GNUEABI-LABEL: test_sincos_v2f64: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r4, lr} +; GNUEABI-NEXT: push {r4, lr} +; GNUEABI-NEXT: .pad #32 +; GNUEABI-NEXT: sub sp, sp, #32 +; GNUEABI-NEXT: mov r1, r3 +; GNUEABI-NEXT: mov r12, r2 +; GNUEABI-NEXT: add r2, sp, #24 +; GNUEABI-NEXT: add r3, sp, #16 +; GNUEABI-NEXT: mov r4, r0 +; GNUEABI-NEXT: mov r0, r12 +; GNUEABI-NEXT: bl sincos +; GNUEABI-NEXT: ldr r0, [sp, #40] +; GNUEABI-NEXT: add r2, sp, #8 +; GNUEABI-NEXT: ldr r1, [sp, #44] +; GNUEABI-NEXT: mov r3, sp +; GNUEABI-NEXT: bl sincos +; GNUEABI-NEXT: vldr d19, [sp, #8] +; GNUEABI-NEXT: vldr d18, [sp, #24] +; GNUEABI-NEXT: vldr d17, [sp] +; GNUEABI-NEXT: vldr d16, [sp, #16] +; GNUEABI-NEXT: vst1.64 {d18, d19}, [r4]! +; GNUEABI-NEXT: vst1.64 {d16, d17}, [r4] +; GNUEABI-NEXT: add sp, sp, #32 +; GNUEABI-NEXT: pop {r4, pc} +; +; IOS-NO-STRET-LABEL: test_sincos_v2f64: +; IOS-NO-STRET: @ %bb.0: +; IOS-NO-STRET-NEXT: push {r4, r5, r6, r7, r8, r10, r11, lr} +; IOS-NO-STRET-NEXT: vpush {d8, d9, d10, d11} +; IOS-NO-STRET-NEXT: ldr r8, [sp, #64] +; IOS-NO-STRET-NEXT: mov r7, r1 +; IOS-NO-STRET-NEXT: mov r4, r0 +; IOS-NO-STRET-NEXT: mov r0, r3 +; IOS-NO-STRET-NEXT: mov r6, r3 +; IOS-NO-STRET-NEXT: mov r10, r2 +; IOS-NO-STRET-NEXT: mov r1, r8 +; IOS-NO-STRET-NEXT: bl _sin +; IOS-NO-STRET-NEXT: mov r11, r0 +; IOS-NO-STRET-NEXT: mov r5, r1 +; IOS-NO-STRET-NEXT: mov r0, r6 +; IOS-NO-STRET-NEXT: mov r1, r8 +; IOS-NO-STRET-NEXT: bl _cos +; IOS-NO-STRET-NEXT: vmov d9, r0, r1 +; IOS-NO-STRET-NEXT: mov r0, r7 +; IOS-NO-STRET-NEXT: mov r1, r10 +; IOS-NO-STRET-NEXT: vmov d11, r11, r5 +; IOS-NO-STRET-NEXT: bl _sin +; IOS-NO-STRET-NEXT: vmov d10, r0, r1 +; IOS-NO-STRET-NEXT: mov r0, r7 +; IOS-NO-STRET-NEXT: mov r1, r10 +; IOS-NO-STRET-NEXT: bl _cos +; IOS-NO-STRET-NEXT: vmov d8, r0, r1 +; IOS-NO-STRET-NEXT: vst1.32 {d10, d11}, [r4]! +; IOS-NO-STRET-NEXT: vst1.32 {d8, d9}, [r4] +; IOS-NO-STRET-NEXT: vpop {d8, d9, d10, d11} +; IOS-NO-STRET-NEXT: pop {r4, r5, r6, r7, r8, r10, r11, pc} +; +; IOS-WITH-STRET-LABEL: test_sincos_v2f64: +; IOS-WITH-STRET: @ %bb.0: +; IOS-WITH-STRET-NEXT: push {r4, r5, r6, lr} +; IOS-WITH-STRET-NEXT: sub sp, sp, #32 +; IOS-WITH-STRET-NEXT: mov r4, r2 +; IOS-WITH-STRET-NEXT: ldr r2, [sp, #48] +; IOS-WITH-STRET-NEXT: mov r6, r0 +; IOS-WITH-STRET-NEXT: add r0, sp, #16 +; IOS-WITH-STRET-NEXT: mov r5, r1 +; IOS-WITH-STRET-NEXT: mov r1, r3 +; IOS-WITH-STRET-NEXT: bl ___sincos_stret +; IOS-WITH-STRET-NEXT: mov r0, sp +; IOS-WITH-STRET-NEXT: mov r1, r5 +; IOS-WITH-STRET-NEXT: mov r2, r4 +; IOS-WITH-STRET-NEXT: bl ___sincos_stret +; IOS-WITH-STRET-NEXT: vldr d17, [sp, #16] +; IOS-WITH-STRET-NEXT: vldr d16, [sp] +; IOS-WITH-STRET-NEXT: vldr d19, [sp, #24] +; IOS-WITH-STRET-NEXT: vldr d18, [sp, #8] +; IOS-WITH-STRET-NEXT: vst1.32 {d16, d17}, [r6]! +; IOS-WITH-STRET-NEXT: vst1.32 {d18, d19}, [r6] +; IOS-WITH-STRET-NEXT: add sp, sp, #32 +; IOS-WITH-STRET-NEXT: pop {r4, r5, r6, pc} +; +; WATCHABI-LABEL: test_sincos_v2f64: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push {r7, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 8 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: vpush {d8, d9, d10, d11, d12, d13} +; WATCHABI-NEXT: .cfi_def_cfa_offset 56 +; WATCHABI-NEXT: .cfi_offset d13, -16 +; WATCHABI-NEXT: .cfi_offset d12, -24 +; WATCHABI-NEXT: .cfi_offset d11, -32 +; WATCHABI-NEXT: .cfi_offset d10, -40 +; WATCHABI-NEXT: .cfi_offset d9, -48 +; WATCHABI-NEXT: .cfi_offset d8, -56 +; WATCHABI-NEXT: sub sp, #8 +; WATCHABI-NEXT: .cfi_def_cfa_offset 64 +; WATCHABI-NEXT: vorr q4, q0, q0 +; WATCHABI-NEXT: vorr d0, d9, d9 +; WATCHABI-NEXT: bl ___sincos_stret +; WATCHABI-NEXT: vorr d11, d0, d0 +; WATCHABI-NEXT: vorr d0, d8, d8 +; WATCHABI-NEXT: vorr d13, d1, d1 +; WATCHABI-NEXT: bl ___sincos_stret +; WATCHABI-NEXT: vorr d12, d1, d1 +; WATCHABI-NEXT: vorr d10, d0, d0 +; WATCHABI-NEXT: vorr q1, q6, q6 +; WATCHABI-NEXT: vorr q0, q5, q5 +; WATCHABI-NEXT: add sp, #8 +; WATCHABI-NEXT: vpop {d8, d9, d10, d11, d12, d13} +; WATCHABI-NEXT: pop {r7, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { <2 x double>, <2 x double> } @llvm.sincos.v2f64(<2 x double> %a) ret { <2 x double>, <2 x double> } %result } define { fp128, fp128 } @test_sincos_f128(fp128 %a) { -; CHECK-LABEL: test_sincos_f128: -; CHECK: @ %bb.0: -; CHECK-NEXT: push {r4, r5, r7, lr} -; CHECK-NEXT: sub sp, #40 -; CHECK-NEXT: mov r12, r3 -; CHECK-NEXT: ldr r3, [sp, #56] -; CHECK-NEXT: add.w lr, sp, #8 -; CHECK-NEXT: mov r4, r0 -; CHECK-NEXT: add r0, sp, #24 -; CHECK-NEXT: strd r0, lr, [sp] -; CHECK-NEXT: mov r0, r1 -; CHECK-NEXT: mov r1, r2 -; CHECK-NEXT: mov r2, r12 -; CHECK-NEXT: bl sincosl -; CHECK-NEXT: ldrd r2, r3, [sp, #16] -; CHECK-NEXT: ldrd r12, r1, [sp, #8] -; CHECK-NEXT: str r3, [r4, #28] -; CHECK-NEXT: ldrd r3, r5, [sp, #32] -; CHECK-NEXT: ldrd lr, r0, [sp, #24] -; CHECK-NEXT: strd r1, r2, [r4, #20] -; CHECK-NEXT: add.w r1, r4, #8 -; CHECK-NEXT: stm.w r1, {r3, r5, r12} -; CHECK-NEXT: strd lr, r0, [r4] -; CHECK-NEXT: add sp, #40 -; CHECK-NEXT: pop {r4, r5, r7, pc} +; GNU-LABEL: test_sincos_f128: +; GNU: @ %bb.0: +; GNU-NEXT: push {r4, r5, r7, lr} +; GNU-NEXT: sub sp, #40 +; GNU-NEXT: mov r12, r3 +; GNU-NEXT: ldr r3, [sp, #56] +; GNU-NEXT: add.w lr, sp, #8 +; GNU-NEXT: mov r4, r0 +; GNU-NEXT: add r0, sp, #24 +; GNU-NEXT: strd r0, lr, [sp] +; GNU-NEXT: mov r0, r1 +; GNU-NEXT: mov r1, r2 +; GNU-NEXT: mov r2, r12 +; GNU-NEXT: bl sincosl +; GNU-NEXT: ldrd r2, r3, [sp, #16] +; GNU-NEXT: ldrd r12, r1, [sp, #8] +; GNU-NEXT: str r3, [r4, #28] +; GNU-NEXT: ldrd r3, r5, [sp, #32] +; GNU-NEXT: ldrd lr, r0, [sp, #24] +; GNU-NEXT: strd r1, r2, [r4, #20] +; GNU-NEXT: add.w r1, r4, #8 +; GNU-NEXT: stm.w r1, {r3, r5, r12} +; GNU-NEXT: strd lr, r0, [r4] +; GNU-NEXT: add sp, #40 +; GNU-NEXT: pop {r4, r5, r7, pc} +; +; GNUEABI-LABEL: test_sincos_f128: +; GNUEABI: @ %bb.0: +; GNUEABI-NEXT: .save {r4, r5, r11, lr} +; GNUEABI-NEXT: push {r4, r5, r11, lr} +; GNUEABI-NEXT: .pad #40 +; GNUEABI-NEXT: sub sp, sp, #40 +; GNUEABI-NEXT: mov r12, r3 +; GNUEABI-NEXT: ldr r3, [sp, #56] +; GNUEABI-NEXT: mov r4, r0 +; GNUEABI-NEXT: add r0, sp, #24 +; GNUEABI-NEXT: add r5, sp, #8 +; GNUEABI-NEXT: stm sp, {r0, r5} +; GNUEABI-NEXT: mov r0, r1 +; GNUEABI-NEXT: mov r1, r2 +; GNUEABI-NEXT: mov r2, r12 +; GNUEABI-NEXT: bl sincosl +; GNUEABI-NEXT: add r3, sp, #12 +; GNUEABI-NEXT: ldr r12, [sp, #8] +; GNUEABI-NEXT: ldm r3, {r1, r2, r3} +; GNUEABI-NEXT: str r3, [r4, #28] +; GNUEABI-NEXT: ldr r0, [sp, #32] +; GNUEABI-NEXT: ldr lr, [sp, #24] +; GNUEABI-NEXT: ldr r5, [sp, #28] +; GNUEABI-NEXT: ldr r3, [sp, #36] +; GNUEABI-NEXT: str r2, [r4, #24] +; GNUEABI-NEXT: str r1, [r4, #20] +; GNUEABI-NEXT: add r1, r4, #8 +; GNUEABI-NEXT: stm r1, {r0, r3, r12} +; GNUEABI-NEXT: str r5, [r4, #4] +; GNUEABI-NEXT: str lr, [r4] +; GNUEABI-NEXT: add sp, sp, #40 +; GNUEABI-NEXT: pop {r4, r5, r11, pc} +; +; IOS-LABEL: test_sincos_f128: +; IOS: @ %bb.0: +; IOS-NEXT: push {r4, r5, r6, r7, r8, lr} +; IOS-NEXT: ldr r8, [sp, #24] +; IOS-NEXT: mov r4, r0 +; IOS-NEXT: mov r5, r3 +; IOS-NEXT: mov r6, r2 +; IOS-NEXT: mov r7, r1 +; IOS-NEXT: mov r0, r1 +; IOS-NEXT: mov r1, r2 +; IOS-NEXT: mov r2, r3 +; IOS-NEXT: mov r3, r8 +; IOS-NEXT: bl _cosl +; IOS-NEXT: add r9, r4, #16 +; IOS-NEXT: stm r9, {r0, r1, r2, r3} +; IOS-NEXT: mov r0, r7 +; IOS-NEXT: mov r1, r6 +; IOS-NEXT: mov r2, r5 +; IOS-NEXT: mov r3, r8 +; IOS-NEXT: bl _sinl +; IOS-NEXT: stm r4, {r0, r1, r2, r3} +; IOS-NEXT: pop {r4, r5, r6, r7, r8, pc} +; +; WATCHABI-LABEL: test_sincos_f128: +; WATCHABI: .cfi_startproc +; WATCHABI-NEXT: @ %bb.0: +; WATCHABI-NEXT: push.w {r4, r5, r6, r7, r8, lr} +; WATCHABI-NEXT: .cfi_def_cfa_offset 24 +; WATCHABI-NEXT: .cfi_offset lr, -4 +; WATCHABI-NEXT: .cfi_offset r7, -8 +; WATCHABI-NEXT: .cfi_offset r6, -12 +; WATCHABI-NEXT: .cfi_offset r5, -16 +; WATCHABI-NEXT: .cfi_offset r4, -20 +; WATCHABI-NEXT: .cfi_offset r8, -24 +; WATCHABI-NEXT: sub sp, #8 +; WATCHABI-NEXT: .cfi_def_cfa_offset 32 +; WATCHABI-NEXT: ldr.w r8, [sp, #32] +; WATCHABI-NEXT: mov r4, r0 +; WATCHABI-NEXT: mov r5, r3 +; WATCHABI-NEXT: mov r6, r2 +; WATCHABI-NEXT: mov r7, r1 +; WATCHABI-NEXT: mov r0, r1 +; WATCHABI-NEXT: mov r1, r2 +; WATCHABI-NEXT: mov r2, r3 +; WATCHABI-NEXT: mov r3, r8 +; WATCHABI-NEXT: bl _cosl +; WATCHABI-NEXT: add.w r9, r4, #16 +; WATCHABI-NEXT: stm.w r9, {r0, r1, r2, r3} +; WATCHABI-NEXT: mov r0, r7 +; WATCHABI-NEXT: mov r1, r6 +; WATCHABI-NEXT: mov r2, r5 +; WATCHABI-NEXT: mov r3, r8 +; WATCHABI-NEXT: bl _sinl +; WATCHABI-NEXT: stm r4!, {r0, r1, r2, r3} +; WATCHABI-NEXT: add sp, #8 +; WATCHABI-NEXT: pop.w {r4, r5, r6, r7, r8, pc} +; WATCHABI-NEXT: .cfi_endproc %result = call { fp128, fp128 } @llvm.sincos.f16(fp128 %a) ret { fp128, fp128 } %result } |