90 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 62; RUN: opt -passes=debugify < %s | llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 | FileCheck %s3 4@lds = addrspace(3) global [512 x float] poison, align 45 6define amdgpu_kernel void @simple_write2_one_val_f32(ptr addrspace(1) %C, ptr addrspace(1) %in) #0 {7; CHECK-LABEL: simple_write2_one_val_f32:8; CHECK: .Lfunc_begin0:9; CHECK-NEXT: .cfi_sections .debug_frame10; CHECK-NEXT: .cfi_startproc11; CHECK-NEXT: ; %bb.0:12; CHECK-NEXT: .file 1 "/" "<stdin>"13; CHECK-NEXT: .loc 1 1 1 prologue_end ; <stdin>:1:114; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x815; CHECK-NEXT: .Ltmp0:16; CHECK-NEXT: ;DEBUG_VALUE: simple_write2_one_val_f32:1 <- $vgpr017; CHECK-NEXT: ;DEBUG_VALUE: simple_write2_one_val_f32:5 <- [DW_OP_plus_uconst 8, DW_OP_stack_value] $vgpr018; CHECK-NEXT: ;DEBUG_VALUE: simple_write2_one_val_f32:3 <- undef19; CHECK-NEXT: .loc 1 2 1 ; <stdin>:2:120; CHECK-NEXT: v_lshlrev_b32_e32 v0, 2, v021; CHECK-NEXT: .Ltmp1:22; CHECK-NEXT: ;DEBUG_VALUE: simple_write2_one_val_f32:4 <- $vgpr023; CHECK-NEXT: ;DEBUG_VALUE: simple_write2_one_val_f32:2 <- undef24; CHECK-NEXT: .loc 1 3 1 ; <stdin>:3:125; CHECK-NEXT: s_waitcnt lgkmcnt(0)26; CHECK-NEXT: global_load_dword v1, v0, s[0:1]27; CHECK-NEXT: .Ltmp2:28; CHECK-NEXT: ;DEBUG_VALUE: simple_write2_one_val_f32:6 <- [DW_OP_plus_uconst 32, DW_OP_stack_value] $vgpr029; CHECK-NEXT: .loc 1 0 0 is_stmt 0 ; <stdin>:030; CHECK-NEXT: s_waitcnt vmcnt(0)31; CHECK-NEXT: ds_write2_b32 v0, v1, v1 offset1:832; CHECK-NEXT: .loc 1 9 1 is_stmt 1 ; <stdin>:9:133; CHECK-NEXT: s_endpgm34; CHECK-NEXT: .Ltmp3:35 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #136 %in.gep = getelementptr float, ptr addrspace(1) %in, i32 %x.i37 %val = load float, ptr addrspace(1) %in.gep, align 438 %arrayidx0 = getelementptr inbounds [512 x float], ptr addrspace(3) @lds, i32 0, i32 %x.i39 store float %val, ptr addrspace(3) %arrayidx0, align 440 %add.x = add nsw i32 %x.i, 841 %arrayidx1 = getelementptr inbounds [512 x float], ptr addrspace(3) @lds, i32 0, i32 %add.x42 store float %val, ptr addrspace(3) %arrayidx1, align 443 ret void44}45 46define amdgpu_kernel void @simple_read2_f32(ptr addrspace(1) %out) #0 {47; CHECK-LABEL: simple_read2_f32:48; CHECK: .Lfunc_begin1:49; CHECK-NEXT: .cfi_startproc50; CHECK-NEXT: ; %bb.0:51; CHECK-NEXT: .loc 1 11 1 prologue_end ; <stdin>:11:152; CHECK-NEXT: v_lshlrev_b32_e32 v2, 2, v053; CHECK-NEXT: .Ltmp4:54; CHECK-NEXT: ;DEBUG_VALUE: simple_read2_f32:8 <- $vgpr255; CHECK-NEXT: .loc 1 0 0 is_stmt 0 ; <stdin>:056; CHECK-NEXT: ds_read2_b32 v[0:1], v2 offset1:857; CHECK-NEXT: .Ltmp5:58; CHECK-NEXT: ;DEBUG_VALUE: simple_read2_f32:9 <- undef59; CHECK-NEXT: ;DEBUG_VALUE: simple_read2_f32:11 <- [DW_OP_plus_uconst 32, DW_OP_stack_value] $vgpr260; CHECK-NEXT: ;DEBUG_VALUE: simple_read2_f32:12 <- undef61; CHECK-NEXT: .loc 1 10 1 is_stmt 1 ; <stdin>:10:162; CHECK-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x063; CHECK-NEXT: .Ltmp6:64; CHECK-NEXT: ;DEBUG_VALUE: simple_read2_f32:7 <- undef65; CHECK-NEXT: ;DEBUG_VALUE: simple_read2_f32:10 <- [DW_OP_plus_uconst 8, DW_OP_stack_value] undef66; CHECK-NEXT: .loc 1 16 1 ; <stdin>:16:167; CHECK-NEXT: s_waitcnt lgkmcnt(0)68; CHECK-NEXT: v_add_f32_e32 v0, v0, v169; CHECK-NEXT: .Ltmp7:70; CHECK-NEXT: ;DEBUG_VALUE: simple_read2_f32:13 <- $vgpr071; CHECK-NEXT: ;DEBUG_VALUE: simple_read2_f32:14 <- undef72; CHECK-NEXT: .loc 1 18 1 ; <stdin>:18:173; CHECK-NEXT: global_store_dword v2, v0, s[0:1]74; CHECK-NEXT: .loc 1 19 1 ; <stdin>:19:175; CHECK-NEXT: s_endpgm76; CHECK-NEXT: .Ltmp8:77 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #178 %arrayidx0 = getelementptr inbounds [512 x float], ptr addrspace(3) @lds, i32 0, i32 %x.i79 %val0 = load float, ptr addrspace(3) %arrayidx0, align 480 %add.x = add nsw i32 %x.i, 881 %arrayidx1 = getelementptr inbounds [512 x float], ptr addrspace(3) @lds, i32 0, i32 %add.x82 %val1 = load float, ptr addrspace(3) %arrayidx1, align 483 %sum = fadd float %val0, %val184 %out.gep = getelementptr inbounds float, ptr addrspace(1) %out, i32 %x.i85 store float %sum, ptr addrspace(1) %out.gep, align 486 ret void87}88 89attributes #0 = { nounwind }90