79 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: not llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs=0 2> %t.err < %s | FileCheck %s3; RUN: FileCheck -check-prefix=ERR %s < %t.err4; FIXME: These tests cannot be tail called, and should be executed in a waterfall loop.5 6declare hidden void @void_func_i32_inreg(i32 inreg)7 8; ERR: error: <unknown>:0:0: in function tail_call_i32_inreg_divergent void (i32): illegal VGPR to SGPR copy9; ERR: error: <unknown>:0:0: in function indirect_tail_call_i32_inreg_divergent void (i32): illegal VGPR to SGPR copy10 11define void @tail_call_i32_inreg_divergent(i32 %vgpr) {12; CHECK-LABEL: tail_call_i32_inreg_divergent:13; CHECK: ; %bb.0:14; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)15; CHECK-NEXT: s_mov_b32 s16, s3316; CHECK-NEXT: s_mov_b32 s33, s3217; CHECK-NEXT: s_or_saveexec_b64 s[18:19], -118; CHECK-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill19; CHECK-NEXT: s_mov_b64 exec, s[18:19]20; CHECK-NEXT: v_writelane_b32 v40, s16, 221; CHECK-NEXT: s_addk_i32 s32, 0x40022; CHECK-NEXT: v_writelane_b32 v40, s30, 023; CHECK-NEXT: v_writelane_b32 v40, s31, 124; CHECK-NEXT: s_getpc_b64 s[16:17]25; CHECK-NEXT: s_add_u32 s16, s16, void_func_i32_inreg@rel32@lo+426; CHECK-NEXT: s_addc_u32 s17, s17, void_func_i32_inreg@rel32@hi+1227; CHECK-NEXT: ; illegal copy v0 to s028; CHECK-NEXT: s_swappc_b64 s[30:31], s[16:17]29; CHECK-NEXT: v_readlane_b32 s31, v40, 130; CHECK-NEXT: v_readlane_b32 s30, v40, 031; CHECK-NEXT: s_mov_b32 s32, s3332; CHECK-NEXT: v_readlane_b32 s4, v40, 233; CHECK-NEXT: s_or_saveexec_b64 s[6:7], -134; CHECK-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload35; CHECK-NEXT: s_mov_b64 exec, s[6:7]36; CHECK-NEXT: s_mov_b32 s33, s437; CHECK-NEXT: s_waitcnt vmcnt(0)38; CHECK-NEXT: s_setpc_b64 s[30:31]39 tail call void @void_func_i32_inreg(i32 inreg %vgpr)40 ret void41}42 43@constant = external hidden addrspace(4) constant ptr44 45define void @indirect_tail_call_i32_inreg_divergent(i32 %vgpr) {46; CHECK-LABEL: indirect_tail_call_i32_inreg_divergent:47; CHECK: ; %bb.0:48; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)49; CHECK-NEXT: s_mov_b32 s16, s3350; CHECK-NEXT: s_mov_b32 s33, s3251; CHECK-NEXT: s_or_saveexec_b64 s[18:19], -152; CHECK-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill53; CHECK-NEXT: s_mov_b64 exec, s[18:19]54; CHECK-NEXT: s_addk_i32 s32, 0x40055; CHECK-NEXT: v_writelane_b32 v40, s16, 256; CHECK-NEXT: s_getpc_b64 s[16:17]57; CHECK-NEXT: s_add_u32 s16, s16, constant@rel32@lo+458; CHECK-NEXT: s_addc_u32 s17, s17, constant@rel32@hi+1259; CHECK-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x060; CHECK-NEXT: v_writelane_b32 v40, s30, 061; CHECK-NEXT: v_writelane_b32 v40, s31, 162; CHECK-NEXT: ; illegal copy v0 to s063; CHECK-NEXT: s_waitcnt lgkmcnt(0)64; CHECK-NEXT: s_swappc_b64 s[30:31], s[16:17]65; CHECK-NEXT: v_readlane_b32 s31, v40, 166; CHECK-NEXT: v_readlane_b32 s30, v40, 067; CHECK-NEXT: s_mov_b32 s32, s3368; CHECK-NEXT: v_readlane_b32 s4, v40, 269; CHECK-NEXT: s_or_saveexec_b64 s[6:7], -170; CHECK-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload71; CHECK-NEXT: s_mov_b64 exec, s[6:7]72; CHECK-NEXT: s_mov_b32 s33, s473; CHECK-NEXT: s_waitcnt vmcnt(0)74; CHECK-NEXT: s_setpc_b64 s[30:31]75 %fptr = load ptr, ptr addrspace(4) @constant, align 876 tail call void %fptr(i32 inreg %vgpr)77 ret void78}79