45 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 %s -o - | FileCheck -check-prefix=GCN %s3 4define float @test_canonicalize_amdgcn_tanh_f32(float %a) {5; GCN-LABEL: test_canonicalize_amdgcn_tanh_f32:6; GCN: ; %bb.0:7; GCN-NEXT: s_wait_loadcnt_dscnt 0x08; GCN-NEXT: s_wait_kmcnt 0x09; GCN-NEXT: v_tanh_f32_e32 v0, v010; GCN-NEXT: s_set_pc_i64 s[30:31]11 %tanh = call float @llvm.amdgcn.tanh.f32(float %a)12 %canonicalized = call float @llvm.canonicalize.f32(float %tanh)13 ret float %canonicalized14}15 16define bfloat @test_canonicalize_amdgcn_tanh_bf16(bfloat %a) {17; GCN-LABEL: test_canonicalize_amdgcn_tanh_bf16:18; GCN: ; %bb.0:19; GCN-NEXT: s_wait_loadcnt_dscnt 0x020; GCN-NEXT: s_wait_kmcnt 0x021; GCN-NEXT: v_tanh_bf16_e32 v0, v022; GCN-NEXT: v_nop23; GCN-NEXT: s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)24; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v025; GCN-NEXT: v_max_num_f32_e32 v0, v0, v026; GCN-NEXT: s_delay_alu instid0(VALU_DEP_1)27; GCN-NEXT: v_cvt_pk_bf16_f32 v0, v0, s028; GCN-NEXT: s_set_pc_i64 s[30:31]29 %tanh = call bfloat @llvm.amdgcn.tanh.bf16(bfloat %a)30 %canonicalized = call bfloat @llvm.canonicalize.bf16(bfloat %tanh)31 ret bfloat %canonicalized32}33 34define half @test_canonicalize_amdgcn_tanh_f16(half %a) {35; GCN-LABEL: test_canonicalize_amdgcn_tanh_f16:36; GCN: ; %bb.0:37; GCN-NEXT: s_wait_loadcnt_dscnt 0x038; GCN-NEXT: s_wait_kmcnt 0x039; GCN-NEXT: v_tanh_f16_e32 v0, v040; GCN-NEXT: s_set_pc_i64 s[30:31]41 %tanh = call half @llvm.amdgcn.tanh.f16(half %a)42 %canonicalized = call half @llvm.canonicalize.f16(half %tanh)43 ret half %canonicalized44}45