191 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GFX6-NOHSA %s3; RUN: llc -mtriple=amdgcn-amdhsa -mcpu=kaveri < %s | FileCheck -check-prefix=GFX7-HSA %s4; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=GFX8-NOHSA %s5; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s6 7; FUNC-LABEL: {{^}}constant_load_f64:8define amdgpu_kernel void @constant_load_f64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {9; GFX6-NOHSA-LABEL: constant_load_f64:10; GFX6-NOHSA: ; %bb.0:11; GFX6-NOHSA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x912; GFX6-NOHSA-NEXT: s_waitcnt lgkmcnt(0)13; GFX6-NOHSA-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x014; GFX6-NOHSA-NEXT: s_mov_b32 s3, 0xf00015; GFX6-NOHSA-NEXT: s_mov_b32 s2, -116; GFX6-NOHSA-NEXT: s_waitcnt lgkmcnt(0)17; GFX6-NOHSA-NEXT: v_mov_b32_e32 v0, s418; GFX6-NOHSA-NEXT: v_mov_b32_e32 v1, s519; GFX6-NOHSA-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 020; GFX6-NOHSA-NEXT: s_endpgm21;22; GFX7-HSA-LABEL: constant_load_f64:23; GFX7-HSA: ; %bb.0:24; GFX7-HSA-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x025; GFX7-HSA-NEXT: s_add_i32 s12, s12, s1726; GFX7-HSA-NEXT: s_mov_b32 flat_scratch_lo, s1327; GFX7-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 828; GFX7-HSA-NEXT: s_waitcnt lgkmcnt(0)29; GFX7-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x030; GFX7-HSA-NEXT: v_mov_b32_e32 v0, s031; GFX7-HSA-NEXT: v_mov_b32_e32 v1, s132; GFX7-HSA-NEXT: s_waitcnt lgkmcnt(0)33; GFX7-HSA-NEXT: v_mov_b32_e32 v2, s234; GFX7-HSA-NEXT: v_mov_b32_e32 v3, s335; GFX7-HSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3]36; GFX7-HSA-NEXT: s_endpgm37;38; GFX8-NOHSA-LABEL: constant_load_f64:39; GFX8-NOHSA: ; %bb.0:40; GFX8-NOHSA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x2441; GFX8-NOHSA-NEXT: s_waitcnt lgkmcnt(0)42; GFX8-NOHSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x043; GFX8-NOHSA-NEXT: v_mov_b32_e32 v0, s044; GFX8-NOHSA-NEXT: v_mov_b32_e32 v1, s145; GFX8-NOHSA-NEXT: s_waitcnt lgkmcnt(0)46; GFX8-NOHSA-NEXT: v_mov_b32_e32 v2, s247; GFX8-NOHSA-NEXT: v_mov_b32_e32 v3, s348; GFX8-NOHSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3]49; GFX8-NOHSA-NEXT: s_endpgm50;51; GFX12-LABEL: constant_load_f64:52; GFX12: ; %bb.0:53; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x2454; GFX12-NEXT: s_wait_kmcnt 0x055; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x056; GFX12-NEXT: v_mov_b32_e32 v2, 057; GFX12-NEXT: s_wait_kmcnt 0x058; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s359; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1]60; GFX12-NEXT: s_endpgm61 %ld = load double, ptr addrspace(4) %in62 store double %ld, ptr addrspace(1) %out63 ret void64}65 66attributes #0 = { nounwind }67 68; Tests whether a load-chain of 8 constants of 64bit each gets vectorized into a wider load.69define amdgpu_kernel void @constant_load_2v4f64(ptr addrspace(4) noalias nocapture readonly %weights, ptr addrspace(1) noalias nocapture %out_ptr) {70; GFX6-NOHSA-LABEL: constant_load_2v4f64:71; GFX6-NOHSA: ; %bb.0: ; %entry72; GFX6-NOHSA-NEXT: s_load_dwordx4 s[16:19], s[4:5], 0x973; GFX6-NOHSA-NEXT: s_waitcnt lgkmcnt(0)74; GFX6-NOHSA-NEXT: s_load_dwordx2 s[24:25], s[18:19], 0x075; GFX6-NOHSA-NEXT: s_load_dwordx16 s[0:15], s[16:17], 0x076; GFX6-NOHSA-NEXT: s_mov_b32 s23, 0xf00077; GFX6-NOHSA-NEXT: s_mov_b32 s22, -178; GFX6-NOHSA-NEXT: s_mov_b32 s20, s1879; GFX6-NOHSA-NEXT: s_mov_b32 s21, s1980; GFX6-NOHSA-NEXT: s_waitcnt lgkmcnt(0)81; GFX6-NOHSA-NEXT: v_mov_b32_e32 v0, s2482; GFX6-NOHSA-NEXT: v_mov_b32_e32 v1, s2583; GFX6-NOHSA-NEXT: v_add_f64 v[0:1], s[0:1], v[0:1]84; GFX6-NOHSA-NEXT: v_add_f64 v[0:1], s[2:3], v[0:1]85; GFX6-NOHSA-NEXT: v_add_f64 v[0:1], s[4:5], v[0:1]86; GFX6-NOHSA-NEXT: v_add_f64 v[0:1], s[6:7], v[0:1]87; GFX6-NOHSA-NEXT: v_add_f64 v[0:1], s[8:9], v[0:1]88; GFX6-NOHSA-NEXT: v_add_f64 v[0:1], s[10:11], v[0:1]89; GFX6-NOHSA-NEXT: v_add_f64 v[0:1], s[12:13], v[0:1]90; GFX6-NOHSA-NEXT: v_add_f64 v[0:1], s[14:15], v[0:1]91; GFX6-NOHSA-NEXT: buffer_store_dwordx2 v[0:1], off, s[20:23], 092; GFX6-NOHSA-NEXT: s_endpgm93;94; GFX7-HSA-LABEL: constant_load_2v4f64:95; GFX7-HSA: ; %bb.0: ; %entry96; GFX7-HSA-NEXT: s_add_i32 s12, s12, s1797; GFX7-HSA-NEXT: s_load_dwordx4 s[16:19], s[8:9], 0x098; GFX7-HSA-NEXT: s_mov_b32 flat_scratch_lo, s1399; GFX7-HSA-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8100; GFX7-HSA-NEXT: s_waitcnt lgkmcnt(0)101; GFX7-HSA-NEXT: s_load_dwordx2 s[20:21], s[18:19], 0x0102; GFX7-HSA-NEXT: s_load_dwordx16 s[0:15], s[16:17], 0x0103; GFX7-HSA-NEXT: v_mov_b32_e32 v2, s18104; GFX7-HSA-NEXT: v_mov_b32_e32 v3, s19105; GFX7-HSA-NEXT: s_waitcnt lgkmcnt(0)106; GFX7-HSA-NEXT: v_mov_b32_e32 v0, s20107; GFX7-HSA-NEXT: v_mov_b32_e32 v1, s21108; GFX7-HSA-NEXT: v_add_f64 v[0:1], s[0:1], v[0:1]109; GFX7-HSA-NEXT: v_add_f64 v[0:1], s[2:3], v[0:1]110; GFX7-HSA-NEXT: v_add_f64 v[0:1], s[4:5], v[0:1]111; GFX7-HSA-NEXT: v_add_f64 v[0:1], s[6:7], v[0:1]112; GFX7-HSA-NEXT: v_add_f64 v[0:1], s[8:9], v[0:1]113; GFX7-HSA-NEXT: v_add_f64 v[0:1], s[10:11], v[0:1]114; GFX7-HSA-NEXT: v_add_f64 v[0:1], s[12:13], v[0:1]115; GFX7-HSA-NEXT: v_add_f64 v[0:1], s[14:15], v[0:1]116; GFX7-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1]117; GFX7-HSA-NEXT: s_endpgm118;119; GFX8-NOHSA-LABEL: constant_load_2v4f64:120; GFX8-NOHSA: ; %bb.0: ; %entry121; GFX8-NOHSA-NEXT: s_load_dwordx4 s[16:19], s[4:5], 0x24122; GFX8-NOHSA-NEXT: s_waitcnt lgkmcnt(0)123; GFX8-NOHSA-NEXT: s_load_dwordx2 s[20:21], s[18:19], 0x0124; GFX8-NOHSA-NEXT: s_load_dwordx16 s[0:15], s[16:17], 0x0125; GFX8-NOHSA-NEXT: v_mov_b32_e32 v2, s18126; GFX8-NOHSA-NEXT: v_mov_b32_e32 v3, s19127; GFX8-NOHSA-NEXT: s_waitcnt lgkmcnt(0)128; GFX8-NOHSA-NEXT: v_mov_b32_e32 v0, s20129; GFX8-NOHSA-NEXT: v_mov_b32_e32 v1, s21130; GFX8-NOHSA-NEXT: v_add_f64 v[0:1], s[0:1], v[0:1]131; GFX8-NOHSA-NEXT: v_add_f64 v[0:1], s[2:3], v[0:1]132; GFX8-NOHSA-NEXT: v_add_f64 v[0:1], s[4:5], v[0:1]133; GFX8-NOHSA-NEXT: v_add_f64 v[0:1], s[6:7], v[0:1]134; GFX8-NOHSA-NEXT: v_add_f64 v[0:1], s[8:9], v[0:1]135; GFX8-NOHSA-NEXT: v_add_f64 v[0:1], s[10:11], v[0:1]136; GFX8-NOHSA-NEXT: v_add_f64 v[0:1], s[12:13], v[0:1]137; GFX8-NOHSA-NEXT: v_add_f64 v[0:1], s[14:15], v[0:1]138; GFX8-NOHSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1]139; GFX8-NOHSA-NEXT: s_endpgm140;141; GFX12-LABEL: constant_load_2v4f64:142; GFX12: ; %bb.0: ; %entry143; GFX12-NEXT: s_load_b128 s[16:19], s[4:5], 0x24144; GFX12-NEXT: v_mov_b32_e32 v2, 0145; GFX12-NEXT: s_wait_kmcnt 0x0146; GFX12-NEXT: s_load_b64 s[20:21], s[18:19], 0x0147; GFX12-NEXT: s_load_b512 s[0:15], s[16:17], 0x0148; GFX12-NEXT: s_wait_kmcnt 0x0149; GFX12-NEXT: v_add_f64_e64 v[0:1], s[0:1], s[20:21]150; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)151; GFX12-NEXT: v_add_f64_e32 v[0:1], s[2:3], v[0:1]152; GFX12-NEXT: v_add_f64_e32 v[0:1], s[4:5], v[0:1]153; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)154; GFX12-NEXT: v_add_f64_e32 v[0:1], s[6:7], v[0:1]155; GFX12-NEXT: v_add_f64_e32 v[0:1], s[8:9], v[0:1]156; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)157; GFX12-NEXT: v_add_f64_e32 v[0:1], s[10:11], v[0:1]158; GFX12-NEXT: v_add_f64_e32 v[0:1], s[12:13], v[0:1]159; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)160; GFX12-NEXT: v_add_f64_e32 v[0:1], s[14:15], v[0:1]161; GFX12-NEXT: global_store_b64 v2, v[0:1], s[18:19]162; GFX12-NEXT: s_endpgm163entry:164 %out_ptr.promoted = load double, ptr addrspace(1) %out_ptr, align 4165 %tmp = load double, ptr addrspace(4) %weights, align 4166 %add = fadd double %tmp, %out_ptr.promoted167 %arrayidx.1 = getelementptr inbounds double, ptr addrspace(4) %weights, i64 1168 %tmp1 = load double, ptr addrspace(4) %arrayidx.1, align 4169 %add.1 = fadd double %tmp1, %add170 %arrayidx.2 = getelementptr inbounds double, ptr addrspace(4) %weights, i64 2171 %tmp2 = load double, ptr addrspace(4) %arrayidx.2, align 4172 %add.2 = fadd double %tmp2, %add.1173 %arrayidx.3 = getelementptr inbounds double, ptr addrspace(4) %weights, i64 3174 %tmp3 = load double, ptr addrspace(4) %arrayidx.3, align 4175 %add.3 = fadd double %tmp3, %add.2176 %arrayidx.4 = getelementptr inbounds double, ptr addrspace(4) %weights, i64 4177 %tmp4 = load double, ptr addrspace(4) %arrayidx.4, align 4178 %add.4 = fadd double %tmp4, %add.3179 %arrayidx.5 = getelementptr inbounds double, ptr addrspace(4) %weights, i64 5180 %tmp5 = load double, ptr addrspace(4) %arrayidx.5, align 4181 %add.5 = fadd double %tmp5, %add.4182 %arrayidx.6 = getelementptr inbounds double, ptr addrspace(4) %weights, i64 6183 %tmp6 = load double, ptr addrspace(4) %arrayidx.6, align 4184 %add.6 = fadd double %tmp6, %add.5185 %arrayidx.7 = getelementptr inbounds double, ptr addrspace(4) %weights, i64 7186 %tmp7 = load double, ptr addrspace(4) %arrayidx.7, align 4187 %add.7 = fadd double %tmp7, %add.6188 store double %add.7, ptr addrspace(1) %out_ptr, align 4189 ret void190}191