llvm/test/CodeGen/AMDGPU/llvm.amdgcn.tanh.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -global-isel=0 -mtriple=amdgpu12.50 -mattr=+real-true16 %s -o - | FileCheck -check-prefixes=GFX1250,GFX1250-SDAG-REAL16 %s
; RUN: llc -global-isel=0 -mtriple=amdgpu12.50 -mattr=-real-true16 %s -o - | FileCheck -check-prefixes=GFX1250,GFX1250-SDAG-FAKE16 %s
; xUN: llc -global-isel=1 -mtriple=amdgpu12.50 -mattr=+real-true16 %s -o - | FileCheck -check-prefixes=GFX1250,GFX1250-GISEL-REAL16 %s
; xUN: llc -global-isel=1 -mtriple=amdgpu12.50 -mattr=-real-true16 %s -o - | FileCheck -check-prefixes=GFX1250,GFX1250-GISEL-FAKE16 %s
; RUN: llc -global-isel=0 -mtriple=amdgpu13.10 -mattr=+real-true16 %s -o - | FileCheck -check-prefixes=GFX13,GFX13-SDAG-REAL16 %s
; RUN: llc -global-isel=0 -mtriple=amdgpu13.10 -mattr=-real-true16 %s -o - | FileCheck -check-prefixes=GFX13,GFX13-SDAG-FAKE16 %s
; xUN: llc -global-isel=1 -mtriple=amdgpu13.10 -mattr=+real-true16 %s -o - | FileCheck -check-prefixes=GFX13,GFX13-GISEL-REAL16 %s
; xUN: llc -global-isel=1 -mtriple=amdgpu13.10 -mattr=-real-true16 %s -o - | FileCheck -check-prefixes=GFX13,GFX13-GISEL-FAKE16 %s

; RUN: not llc -global-isel=0 -mtriple=amdgpu9.50 -filetype=null < %s 2>&1 | FileCheck -check-prefix=ERR %s
; RUN: not llc -global-isel=1 -mtriple=amdgpu9.50 -filetype=null < %s 2>&1 | FileCheck -check-prefix=ERR %s

; ERR: error: <unknown>:0:0: in function @tanh_f32 float (float): llvm.amdgcn.tanh requires target feature 'tanh-insts'

define float @tanh_f32(float %src) {
; GFX1250-LABEL: tanh_f32:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_tanh_f32_e32 v0, v0
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-LABEL: tanh_f32:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-NEXT:    s_wait_expcnt 0x0
; GFX13-NEXT:    s_wait_samplecnt 0x0
; GFX13-NEXT:    s_wait_bvhcnt 0x0
; GFX13-NEXT:    s_wait_kmcnt 0x0
; GFX13-NEXT:    v_tanh_f32_e32 v0, v0
; GFX13-NEXT:    s_set_pc_i64 s[30:31]
  %tanh = call float @llvm.amdgcn.tanh.f32(float %src)
  ret float %tanh
}

define float @tanh_f32_neg(float %src)  {
; GFX1250-LABEL: tanh_f32_neg:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_tanh_f32_e64 v0, -v0
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-LABEL: tanh_f32_neg:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-NEXT:    s_wait_expcnt 0x0
; GFX13-NEXT:    s_wait_samplecnt 0x0
; GFX13-NEXT:    s_wait_bvhcnt 0x0
; GFX13-NEXT:    s_wait_kmcnt 0x0
; GFX13-NEXT:    v_tanh_f32_e64 v0, -v0
; GFX13-NEXT:    s_set_pc_i64 s[30:31]
  %fneg = fneg float %src
  %tanh = call float @llvm.amdgcn.tanh.f32(float %fneg)
  ret float %tanh
}

define float @tanh_f32_abs(float %src)  {
; GFX1250-LABEL: tanh_f32_abs:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_tanh_f32_e64 v0, |v0|
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-LABEL: tanh_f32_abs:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-NEXT:    s_wait_expcnt 0x0
; GFX13-NEXT:    s_wait_samplecnt 0x0
; GFX13-NEXT:    s_wait_bvhcnt 0x0
; GFX13-NEXT:    s_wait_kmcnt 0x0
; GFX13-NEXT:    v_tanh_f32_e64 v0, |v0|
; GFX13-NEXT:    s_set_pc_i64 s[30:31]
  %fabs = call float @llvm.fabs.f32(float %src)
  %tanh = call float @llvm.amdgcn.tanh.f32(float %fabs)
  ret float %tanh
}

define float @tanh_f32_neg_abs(float %src)  {
; GFX1250-LABEL: tanh_f32_neg_abs:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-NEXT:    s_wait_kmcnt 0x0
; GFX1250-NEXT:    v_tanh_f32_e64 v0, -|v0|
; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-LABEL: tanh_f32_neg_abs:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-NEXT:    s_wait_expcnt 0x0
; GFX13-NEXT:    s_wait_samplecnt 0x0
; GFX13-NEXT:    s_wait_bvhcnt 0x0
; GFX13-NEXT:    s_wait_kmcnt 0x0
; GFX13-NEXT:    v_tanh_f32_e64 v0, -|v0|
; GFX13-NEXT:    s_set_pc_i64 s[30:31]
  %fabs = call float @llvm.fabs.f32(float %src)
  %fneg = fneg float %fabs
  %tanh = call float @llvm.amdgcn.tanh.f32(float %fneg)
  ret float %tanh
}

define amdgpu_ps float @tanh_f32_scalar(float inreg %src) {
; GFX1250-LABEL: tanh_f32_scalar:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    global_wb
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    v_tanh_f32_e32 v0, s0
; GFX1250-NEXT:    ; return to shader part epilog
;
; GFX13-LABEL: tanh_f32_scalar:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    v_tanh_f32_e32 v0, s0
; GFX13-NEXT:    ; return to shader part epilog
  %tanh = call float @llvm.amdgcn.tanh.f32(float %src)
  ret float %tanh
}

; TODO: Really these should be constant folded
define amdgpu_ps float @tanh_f32_constant_4.0() {
; GFX1250-LABEL: tanh_f32_constant_4.0:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    global_wb
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    v_tanh_f32_e32 v0, 4.0
; GFX1250-NEXT:    ; return to shader part epilog
;
; GFX13-LABEL: tanh_f32_constant_4.0:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    v_tanh_f32_e32 v0, 4.0
; GFX13-NEXT:    ; return to shader part epilog
  %tanh = call float @llvm.amdgcn.tanh.f32(float 4.0)
  ret float %tanh
}

define amdgpu_ps float @tanh_f32_constant_100.0() {
; GFX1250-LABEL: tanh_f32_constant_100.0:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    global_wb
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    v_tanh_f32_e32 v0, 0x42c80000
; GFX1250-NEXT:    ; return to shader part epilog
;
; GFX13-LABEL: tanh_f32_constant_100.0:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    v_tanh_f32_e32 v0, 0x42c80000
; GFX13-NEXT:    ; return to shader part epilog
  %tanh = call float @llvm.amdgcn.tanh.f32(float 100.0)
  ret float %tanh
}

define amdgpu_ps float @tanh_undef_f32() {
; GFX1250-LABEL: tanh_undef_f32:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    global_wb
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    ; return to shader part epilog
;
; GFX13-LABEL: tanh_undef_f32:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    ; return to shader part epilog
  %tanh = call float @llvm.amdgcn.tanh.f32(float undef)
  ret  float %tanh
}

define half @tanh_f16(half %src) {
; GFX1250-SDAG-REAL16-LABEL: tanh_f16:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_f16_e32 v0.l, v0.l
; GFX1250-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_f16:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_f16_e32 v0, v0
; GFX1250-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-REAL16-LABEL: tanh_f16:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    v_tanh_f16_e32 v0.l, v0.l
; GFX13-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-FAKE16-LABEL: tanh_f16:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_f16_e32 v0, v0
; GFX13-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
  %tanh = call half @llvm.amdgcn.tanh.f16(half %src)
  ret half %tanh
}

define half @tanh_f16_neg(half %src)  {
; GFX1250-SDAG-REAL16-LABEL: tanh_f16_neg:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_f16_e64 v0.l, -v0.l
; GFX1250-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_f16_neg:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_f16_e64 v0, -v0
; GFX1250-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-REAL16-LABEL: tanh_f16_neg:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    v_tanh_f16_e64 v0.l, -v0.l
; GFX13-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-FAKE16-LABEL: tanh_f16_neg:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_f16_e64 v0, -v0
; GFX13-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
  %fneg = fneg half %src
  %tanh = call half @llvm.amdgcn.tanh.f16(half %fneg)
  ret half %tanh
}

define half @tanh_f16_abs(half %src)  {
; GFX1250-SDAG-REAL16-LABEL: tanh_f16_abs:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_f16_e64 v0.l, |v0.l|
; GFX1250-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_f16_abs:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_f16_e64 v0, |v0|
; GFX1250-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-REAL16-LABEL: tanh_f16_abs:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    v_tanh_f16_e64 v0.l, |v0.l|
; GFX13-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-FAKE16-LABEL: tanh_f16_abs:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_f16_e64 v0, |v0|
; GFX13-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
  %fabs = call half @llvm.fabs.f16(half %src)
  %tanh = call half @llvm.amdgcn.tanh.f16(half %fabs)
  ret half %tanh
}

define half @tanh_f16_neg_abs(half %src)  {
; GFX1250-SDAG-REAL16-LABEL: tanh_f16_neg_abs:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_f16_e64 v0.l, -|v0.l|
; GFX1250-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_f16_neg_abs:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_f16_e64 v0, -|v0|
; GFX1250-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-REAL16-LABEL: tanh_f16_neg_abs:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    v_tanh_f16_e64 v0.l, -|v0.l|
; GFX13-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-FAKE16-LABEL: tanh_f16_neg_abs:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_f16_e64 v0, -|v0|
; GFX13-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
  %fabs = call half @llvm.fabs.f16(half %src)
  %fneg = fneg half %fabs
  %tanh = call half @llvm.amdgcn.tanh.f16(half %fneg)
  ret half %tanh
}

define amdgpu_ps half @tanh_f16_scalar(half inreg %src) {
; GFX1250-SDAG-REAL16-LABEL: tanh_f16_scalar:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    global_wb
; GFX1250-SDAG-REAL16-NEXT:    v_nop
; GFX1250-SDAG-REAL16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_f16_e32 v0.l, s0
; GFX1250-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_f16_scalar:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    global_wb
; GFX1250-SDAG-FAKE16-NEXT:    v_nop
; GFX1250-SDAG-FAKE16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_f16_e32 v0, s0
; GFX1250-SDAG-FAKE16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-REAL16-LABEL: tanh_f16_scalar:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    v_tanh_f16_e32 v0.l, s0
; GFX13-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-FAKE16-LABEL: tanh_f16_scalar:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_f16_e32 v0, s0
; GFX13-SDAG-FAKE16-NEXT:    ; return to shader part epilog
  %tanh = call half @llvm.amdgcn.tanh.f16(half %src)
  ret half %tanh
}

define amdgpu_ps half @tanh_f16_constant_4.0() {
; GFX1250-SDAG-REAL16-LABEL: tanh_f16_constant_4.0:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    global_wb
; GFX1250-SDAG-REAL16-NEXT:    v_nop
; GFX1250-SDAG-REAL16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_f16_e32 v0.l, 4.0
; GFX1250-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_f16_constant_4.0:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    global_wb
; GFX1250-SDAG-FAKE16-NEXT:    v_nop
; GFX1250-SDAG-FAKE16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_f16_e32 v0, 4.0
; GFX1250-SDAG-FAKE16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-REAL16-LABEL: tanh_f16_constant_4.0:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    v_tanh_f16_e32 v0.l, 4.0
; GFX13-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-FAKE16-LABEL: tanh_f16_constant_4.0:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_f16_e32 v0, 4.0
; GFX13-SDAG-FAKE16-NEXT:    ; return to shader part epilog
  %tanh = call half @llvm.amdgcn.tanh.f16(half 4.0)
  ret half %tanh
}

define amdgpu_ps half @tanh_f16_constant_100.0() {
; GFX1250-SDAG-REAL16-LABEL: tanh_f16_constant_100.0:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    global_wb
; GFX1250-SDAG-REAL16-NEXT:    v_nop
; GFX1250-SDAG-REAL16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_f16_e32 v0.l, 0x5640
; GFX1250-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_f16_constant_100.0:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    global_wb
; GFX1250-SDAG-FAKE16-NEXT:    v_nop
; GFX1250-SDAG-FAKE16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_f16_e32 v0, 0x5640
; GFX1250-SDAG-FAKE16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-REAL16-LABEL: tanh_f16_constant_100.0:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    v_tanh_f16_e32 v0.l, 0x5640
; GFX13-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-FAKE16-LABEL: tanh_f16_constant_100.0:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_f16_e32 v0, 0x5640
; GFX13-SDAG-FAKE16-NEXT:    ; return to shader part epilog
  %tanh = call half @llvm.amdgcn.tanh.f16(half 100.0)
  ret half %tanh
}

define amdgpu_ps half @tanh_undef_f16() {
; GFX1250-LABEL: tanh_undef_f16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    global_wb
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    ; return to shader part epilog
;
; GFX13-LABEL: tanh_undef_f16:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    ; return to shader part epilog
  %tanh = call half @llvm.amdgcn.tanh.f16(half undef)
  ret  half %tanh
}

define bfloat @tanh_bf16(bfloat %src) {
; GFX1250-SDAG-REAL16-LABEL: tanh_bf16:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_bf16_e32 v0.l, v0.l
; GFX1250-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_bf16:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_bf16_e32 v0, v0
; GFX1250-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-REAL16-LABEL: tanh_bf16:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    v_tanh_bf16_e32 v0.l, v0.l
; GFX13-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-FAKE16-LABEL: tanh_bf16:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_bf16_e32 v0, v0
; GFX13-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
  %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat %src)
  ret bfloat %tanh
}

define bfloat @tanh_bf16_neg(bfloat %src)  {
; GFX1250-SDAG-REAL16-LABEL: tanh_bf16_neg:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_bf16_e64 v0.l, -v0.l
; GFX1250-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_bf16_neg:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_bf16_e64 v0, -v0
; GFX1250-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-REAL16-LABEL: tanh_bf16_neg:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    v_tanh_bf16_e64 v0.l, -v0.l
; GFX13-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-FAKE16-LABEL: tanh_bf16_neg:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_bf16_e64 v0, -v0
; GFX13-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
  %fneg = fneg bfloat %src
  %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat %fneg)
  ret bfloat %tanh
}

define bfloat @tanh_bf16_abs(bfloat %src)  {
; GFX1250-SDAG-REAL16-LABEL: tanh_bf16_abs:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_bf16_e64 v0.l, |v0.l|
; GFX1250-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_bf16_abs:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_bf16_e64 v0, |v0|
; GFX1250-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-REAL16-LABEL: tanh_bf16_abs:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    v_tanh_bf16_e64 v0.l, |v0.l|
; GFX13-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-FAKE16-LABEL: tanh_bf16_abs:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_bf16_e64 v0, |v0|
; GFX13-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
  %fabs = call bfloat @llvm.fabs.bf16(bfloat %src)
  %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat %fabs)
  ret bfloat %tanh
}

define bfloat @tanh_bf16_neg_abs(bfloat %src)  {
; GFX1250-SDAG-REAL16-LABEL: tanh_bf16_neg_abs:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_bf16_e64 v0.l, -|v0.l|
; GFX1250-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_bf16_neg_abs:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_bf16_e64 v0, -|v0|
; GFX1250-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-REAL16-LABEL: tanh_bf16_neg_abs:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-REAL16-NEXT:    v_tanh_bf16_e64 v0.l, -|v0.l|
; GFX13-SDAG-REAL16-NEXT:    s_set_pc_i64 s[30:31]
;
; GFX13-SDAG-FAKE16-LABEL: tanh_bf16_neg_abs:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_expcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_samplecnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    s_wait_kmcnt 0x0
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_bf16_e64 v0, -|v0|
; GFX13-SDAG-FAKE16-NEXT:    s_set_pc_i64 s[30:31]
  %fabs = call bfloat @llvm.fabs.bf16(bfloat %src)
  %fneg = fneg bfloat %fabs
  %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat %fneg)
  ret bfloat %tanh
}

define amdgpu_ps bfloat @tanh_bf16_scalar(bfloat inreg %src) {
; GFX1250-SDAG-REAL16-LABEL: tanh_bf16_scalar:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    global_wb
; GFX1250-SDAG-REAL16-NEXT:    v_nop
; GFX1250-SDAG-REAL16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_bf16_e32 v0.l, s0
; GFX1250-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_bf16_scalar:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    global_wb
; GFX1250-SDAG-FAKE16-NEXT:    v_nop
; GFX1250-SDAG-FAKE16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_bf16_e32 v0, s0
; GFX1250-SDAG-FAKE16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-REAL16-LABEL: tanh_bf16_scalar:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    v_tanh_bf16_e32 v0.l, s0
; GFX13-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-FAKE16-LABEL: tanh_bf16_scalar:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_bf16_e32 v0, s0
; GFX13-SDAG-FAKE16-NEXT:    ; return to shader part epilog
  %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat %src)
  ret bfloat %tanh
}

define amdgpu_ps bfloat @tanh_bf16_constant_4.0() {
; GFX1250-SDAG-REAL16-LABEL: tanh_bf16_constant_4.0:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    global_wb
; GFX1250-SDAG-REAL16-NEXT:    v_nop
; GFX1250-SDAG-REAL16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_bf16_e32 v0.l, 4.0
; GFX1250-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_bf16_constant_4.0:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    global_wb
; GFX1250-SDAG-FAKE16-NEXT:    v_nop
; GFX1250-SDAG-FAKE16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_bf16_e32 v0, 4.0
; GFX1250-SDAG-FAKE16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-REAL16-LABEL: tanh_bf16_constant_4.0:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    v_tanh_bf16_e32 v0.l, 4.0
; GFX13-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-FAKE16-LABEL: tanh_bf16_constant_4.0:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_bf16_e32 v0, 4.0
; GFX13-SDAG-FAKE16-NEXT:    ; return to shader part epilog
  %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat 4.0)
  ret bfloat %tanh
}

define amdgpu_ps bfloat @tanh_bf16_constant_100.0() {
; GFX1250-SDAG-REAL16-LABEL: tanh_bf16_constant_100.0:
; GFX1250-SDAG-REAL16:       ; %bb.0:
; GFX1250-SDAG-REAL16-NEXT:    global_wb
; GFX1250-SDAG-REAL16-NEXT:    v_nop
; GFX1250-SDAG-REAL16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-REAL16-NEXT:    v_tanh_bf16_e32 v0.l, 0x42c8
; GFX1250-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX1250-SDAG-FAKE16-LABEL: tanh_bf16_constant_100.0:
; GFX1250-SDAG-FAKE16:       ; %bb.0:
; GFX1250-SDAG-FAKE16-NEXT:    global_wb
; GFX1250-SDAG-FAKE16-NEXT:    v_nop
; GFX1250-SDAG-FAKE16-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-SDAG-FAKE16-NEXT:    v_tanh_bf16_e32 v0, 0x42c8
; GFX1250-SDAG-FAKE16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-REAL16-LABEL: tanh_bf16_constant_100.0:
; GFX13-SDAG-REAL16:       ; %bb.0:
; GFX13-SDAG-REAL16-NEXT:    v_tanh_bf16_e32 v0.l, 0x42c8
; GFX13-SDAG-REAL16-NEXT:    ; return to shader part epilog
;
; GFX13-SDAG-FAKE16-LABEL: tanh_bf16_constant_100.0:
; GFX13-SDAG-FAKE16:       ; %bb.0:
; GFX13-SDAG-FAKE16-NEXT:    v_tanh_bf16_e32 v0, 0x42c8
; GFX13-SDAG-FAKE16-NEXT:    ; return to shader part epilog
  %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat 100.0)
  ret bfloat %tanh
}

define amdgpu_ps bfloat @tanh_undef_bf16() {
; GFX1250-LABEL: tanh_undef_bf16:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    global_wb
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    ; return to shader part epilog
;
; GFX13-LABEL: tanh_undef_bf16:
; GFX13:       ; %bb.0:
; GFX13-NEXT:    ; return to shader part epilog
  %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat undef)
  ret  bfloat %tanh
}