1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -global-isel -march=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX906 %s 3; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1011 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX10 %s 4; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1012 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX10 %s 5 6define float @v_fdot2(<2 x half> %a, <2 x half> %b, float %c) { 7; GFX906-LABEL: v_fdot2: 8; GFX906: ; %bb.0: 9; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 10; GFX906-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 11; GFX906-NEXT: s_setpc_b64 s[30:31] 12; 13; GFX10-LABEL: v_fdot2: 14; GFX10: ; %bb.0: 15; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 16; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 17; GFX10-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 18; GFX10-NEXT: s_setpc_b64 s[30:31] 19 %r = call float @llvm.amdgcn.fdot2(<2 x half> %a, <2 x half> %b, float %c, i1 false) 20 ret float %r 21} 22 23define float @v_fdot2_clamp(<2 x half> %a, <2 x half> %b, float %c) { 24; GFX906-LABEL: v_fdot2_clamp: 25; GFX906: ; %bb.0: 26; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 27; GFX906-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 clamp 28; GFX906-NEXT: s_setpc_b64 s[30:31] 29; 30; GFX10-LABEL: v_fdot2_clamp: 31; GFX10: ; %bb.0: 32; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 33; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 34; GFX10-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 clamp 35; GFX10-NEXT: s_setpc_b64 s[30:31] 36 %r = call float @llvm.amdgcn.fdot2(<2 x half> %a, <2 x half> %b, float %c, i1 true) 37 ret float %r 38} 39 40define float @v_fdot2_neg_a(<2 x half> %a, <2 x half> %b, float %c) { 41; GFX906-LABEL: v_fdot2_neg_a: 42; GFX906: ; %bb.0: 43; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 44; GFX906-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 neg_lo:[1,0,0] neg_hi:[1,0,0] 45; GFX906-NEXT: s_setpc_b64 s[30:31] 46; 47; GFX10-LABEL: v_fdot2_neg_a: 48; GFX10: ; %bb.0: 49; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 50; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 51; GFX10-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 neg_lo:[1,0,0] neg_hi:[1,0,0] 52; GFX10-NEXT: s_setpc_b64 s[30:31] 53 %neg.a = fneg <2 x half> %a 54 %r = call float @llvm.amdgcn.fdot2(<2 x half> %neg.a, <2 x half> %b, float %c, i1 false) 55 ret float %r 56} 57 58define float @v_fdot2_neg_b(<2 x half> %a, <2 x half> %b, float %c) { 59; GFX906-LABEL: v_fdot2_neg_b: 60; GFX906: ; %bb.0: 61; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 62; GFX906-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 neg_lo:[0,1,0] neg_hi:[0,1,0] 63; GFX906-NEXT: s_setpc_b64 s[30:31] 64; 65; GFX10-LABEL: v_fdot2_neg_b: 66; GFX10: ; %bb.0: 67; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 68; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 69; GFX10-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 neg_lo:[0,1,0] neg_hi:[0,1,0] 70; GFX10-NEXT: s_setpc_b64 s[30:31] 71 %neg.b = fneg <2 x half> %b 72 %r = call float @llvm.amdgcn.fdot2(<2 x half> %a, <2 x half> %neg.b, float %c, i1 false) 73 ret float %r 74} 75 76define float @v_fdot2_neg_a_neg_b(<2 x half> %a, <2 x half> %b, float %c) { 77; GFX906-LABEL: v_fdot2_neg_a_neg_b: 78; GFX906: ; %bb.0: 79; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 80; GFX906-NEXT: v_dot2_f32_f16 v0, v1, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0] 81; GFX906-NEXT: s_setpc_b64 s[30:31] 82; 83; GFX10-LABEL: v_fdot2_neg_a_neg_b: 84; GFX10: ; %bb.0: 85; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 86; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 87; GFX10-NEXT: v_dot2_f32_f16 v0, v1, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0] 88; GFX10-NEXT: s_setpc_b64 s[30:31] 89 %neg.a = fneg <2 x half> %b 90 %neg.b = fneg <2 x half> %b 91 %r = call float @llvm.amdgcn.fdot2(<2 x half> %neg.a, <2 x half> %neg.b, float %c, i1 false) 92 ret float %r 93} 94 95define float @v_fdot2_neg_c(<2 x half> %a, <2 x half> %b, float %c) { 96; GFX906-LABEL: v_fdot2_neg_c: 97; GFX906: ; %bb.0: 98; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 99; GFX906-NEXT: v_xor_b32_e32 v2, 0x80000000, v2 100; GFX906-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 101; GFX906-NEXT: s_setpc_b64 s[30:31] 102; 103; GFX10-LABEL: v_fdot2_neg_c: 104; GFX10: ; %bb.0: 105; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 106; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 107; GFX10-NEXT: v_xor_b32_e32 v2, 0x80000000, v2 108; GFX10-NEXT: v_dot2_f32_f16 v0, v0, v1, v2 109; GFX10-NEXT: s_setpc_b64 s[30:31] 110 %neg.c = fneg float %c 111 %r = call float @llvm.amdgcn.fdot2(<2 x half> %a, <2 x half> %b, float %neg.c, i1 false) 112 ret float %r 113} 114 115define float @v_fdot2_inline_literal_a(<2 x half> %b, float %c) { 116; GFX906-LABEL: v_fdot2_inline_literal_a: 117; GFX906: ; %bb.0: 118; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 119; GFX906-NEXT: s_movk_i32 s4, 0x4000 120; GFX906-NEXT: s_pack_ll_b32_b16 s4, s4, s4 121; GFX906-NEXT: v_dot2_f32_f16 v0, s4, v0, v1 122; GFX906-NEXT: s_setpc_b64 s[30:31] 123; 124; GFX10-LABEL: v_fdot2_inline_literal_a: 125; GFX10: ; %bb.0: 126; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 127; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 128; GFX10-NEXT: s_movk_i32 s4, 0x4000 129; GFX10-NEXT: s_pack_ll_b32_b16 s4, s4, s4 130; GFX10-NEXT: v_dot2_f32_f16 v0, s4, v0, v1 131; GFX10-NEXT: s_setpc_b64 s[30:31] 132 %ret = tail call float @llvm.amdgcn.fdot2(<2 x half> <half 2.0, half 2.0>, <2 x half> %b, float %c, i1 false) 133 ret float %ret 134} 135 136define float @v_fdot2_inline_literal_b(<2 x half> %a, float %c) { 137; GFX906-LABEL: v_fdot2_inline_literal_b: 138; GFX906: ; %bb.0: 139; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 140; GFX906-NEXT: s_movk_i32 s4, 0x4000 141; GFX906-NEXT: s_pack_ll_b32_b16 s4, s4, s4 142; GFX906-NEXT: v_dot2_f32_f16 v0, v0, s4, v1 143; GFX906-NEXT: s_setpc_b64 s[30:31] 144; 145; GFX10-LABEL: v_fdot2_inline_literal_b: 146; GFX10: ; %bb.0: 147; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 148; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 149; GFX10-NEXT: s_movk_i32 s4, 0x4000 150; GFX10-NEXT: s_pack_ll_b32_b16 s4, s4, s4 151; GFX10-NEXT: v_dot2_f32_f16 v0, v0, s4, v1 152; GFX10-NEXT: s_setpc_b64 s[30:31] 153 %ret = tail call float @llvm.amdgcn.fdot2(<2 x half> %a, <2 x half> <half 2.0, half 2.0>, float %c, i1 false) 154 ret float %ret 155} 156 157define float @v_fdot2_inline_literal_c(<2 x half> %a, <2 x half> %b) { 158; GFX906-LABEL: v_fdot2_inline_literal_c: 159; GFX906: ; %bb.0: 160; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 161; GFX906-NEXT: v_dot2_f32_f16 v0, v0, v1, 1.0 162; GFX906-NEXT: s_setpc_b64 s[30:31] 163; 164; GFX10-LABEL: v_fdot2_inline_literal_c: 165; GFX10: ; %bb.0: 166; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) 167; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 168; GFX10-NEXT: v_dot2_f32_f16 v0, v0, v1, 1.0 169; GFX10-NEXT: s_setpc_b64 s[30:31] 170 %ret = tail call float @llvm.amdgcn.fdot2(<2 x half> %a, <2 x half> %b, float 1.0, i1 false) 171 ret float %ret 172} 173 174declare float @llvm.amdgcn.fdot2(<2 x half>, <2 x half>, float, i1 immarg) #0 175 176attributes #0 = { nounwind readnone speculatable } 177