--- /dev/null
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -march=amdgcn -verify-machineinstrs | FileCheck %s -check-prefixes=GCN
+; RUN: llc < %s -march=amdgcn -mcpu=hawaii -verify-machineinstrs | FileCheck %s -check-prefixes=GFX7
+; RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck %s -check-prefixes=GFX8
+; RUN: llc < %s -march=amdgcn -mcpu=gfx900 -verify-machineinstrs | FileCheck %s -check-prefixes=GFX9
+; RUN: llc < %s -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck %s -check-prefixes=GFX10
+
+; We only have storage-only BF16 support. We can load/store those values as we treat them as u16, but
+; we don't support operations on them. As such, codegen is expected to fail for any operation other
+; than simple load/stores.
+
+define void @test_load_store(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ushort v0, v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: flat_store_short v[2:3], v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ushort v0, v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: global_store_short v[2:3], v0, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_load_ushort v0, v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_short v[2:3], v0, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load bfloat, ptr addrspace(1) %in
+ store bfloat %val, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_load_store_f32_to_bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store_f32_to_bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store_f32_to_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store_f32_to_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v0, v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: flat_store_short v[2:3], v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store_f32_to_bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_dword v0, v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: global_store_short_d16_hi v[2:3], v0, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store_f32_to_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_load_dword v0, v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_short_d16_hi v[2:3], v0, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load float, ptr addrspace(1) %in
+ %val.bf16 = fptrunc float %val to bfloat
+ store bfloat %val.bf16, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_load_store_f64_to_bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store_f64_to_bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store_f64_to_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store_f64_to_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: flat_store_short v[2:3], v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store_f64_to_bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX9-NEXT: global_store_short_d16_hi v[2:3], v0, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store_f64_to_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
+; GFX10-NEXT: global_store_short_d16_hi v[2:3], v0, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load double, ptr addrspace(1) %in
+ %val.bf16 = fptrunc double %val to bfloat
+ store bfloat %val.bf16, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_load_store_bf16_to_f32(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store_bf16_to_f32:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GCN-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store_bf16_to_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store_bf16_to_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ushort v0, v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: flat_store_dword v[2:3], v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store_bf16_to_f32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
+; GFX9-NEXT: global_load_short_d16_hi v4, v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: global_store_dword v[2:3], v4, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store_bf16_to_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v4, 0
+; GFX10-NEXT: global_load_short_d16_hi v4, v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_dword v[2:3], v4, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load bfloat, ptr addrspace(1) %in
+ %val.f32 = fpext bfloat %val to float
+ store float %val.f32, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_load_store_bf16_to_f64(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store_bf16_to_f64:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
+; GCN-NEXT: buffer_store_dwordx2 v[0:1], v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store_bf16_to_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
+; GFX7-NEXT: buffer_store_dwordx2 v[0:1], v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store_bf16_to_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ushort v0, v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
+; GFX8-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store_bf16_to_f64:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
+; GFX9-NEXT: global_load_short_d16_hi v4, v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v4
+; GFX9-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store_bf16_to_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v4, 0
+; GFX10-NEXT: global_load_short_d16_hi v4, v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v4
+; GFX10-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load bfloat, ptr addrspace(1) %in
+ %val.f64 = fpext bfloat %val to double
+ store double %val.f64, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_load_store_v2bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store_v2bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v0, v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: flat_store_dword v[2:3], v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store_v2bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_dword v0, v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: global_store_dword v[2:3], v0, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_load_dword v0, v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_dword v[2:3], v0, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load <2 x bfloat>, ptr addrspace(1) %in
+ store <2 x bfloat> %val, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_load_store_v4bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store_v4bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_dwordx2 v[0:1], v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store_v4bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_dwordx2 v[0:1], v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store_v4bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store_v4bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store_v4bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load <4 x bfloat>, ptr addrspace(1) %in
+ store <4 x bfloat> %val, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_load_store_v8bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store_v8bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_dwordx4 v[4:7], v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store_v8bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_dwordx4 v[4:7], v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store_v8bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: flat_store_dwordx4 v[2:3], v[4:7]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store_v8bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: global_store_dwordx4 v[2:3], v[4:7], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store_v8bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_dwordx4 v[2:3], v[4:7], off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load <8 x bfloat>, ptr addrspace(1) %in
+ store <8 x bfloat> %val, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_load_store_v16bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_load_store_v16bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64 offset:16
+; GCN-NEXT: buffer_load_dwordx4 v[8:11], v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(1)
+; GCN-NEXT: buffer_store_dwordx4 v[4:7], v[2:3], s[4:7], 0 addr64 offset:16
+; GCN-NEXT: s_waitcnt vmcnt(1)
+; GCN-NEXT: buffer_store_dwordx4 v[8:11], v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_load_store_v16bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64 offset:16
+; GFX7-NEXT: buffer_load_dwordx4 v[8:11], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(1)
+; GFX7-NEXT: buffer_store_dwordx4 v[4:7], v[2:3], s[4:7], 0 addr64 offset:16
+; GFX7-NEXT: s_waitcnt vmcnt(1)
+; GFX7-NEXT: buffer_store_dwordx4 v[8:11], v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_load_store_v16bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 16, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[4:5]
+; GFX8-NEXT: flat_load_dwordx4 v[8:11], v[0:1]
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v2
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v3, vcc
+; GFX8-NEXT: s_waitcnt vmcnt(1)
+; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
+; GFX8-NEXT: s_waitcnt vmcnt(1)
+; GFX8-NEXT: flat_store_dwordx4 v[2:3], v[8:11]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_load_store_v16bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_dwordx4 v[4:7], v[0:1], off offset:16
+; GFX9-NEXT: global_load_dwordx4 v[8:11], v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: global_store_dwordx4 v[2:3], v[8:11], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_load_store_v16bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: global_load_dwordx4 v[4:7], v[0:1], off offset:16
+; GFX10-NEXT: global_load_dwordx4 v[8:11], v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(1)
+; GFX10-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:16
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_dwordx4 v[2:3], v[8:11], off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load <16 x bfloat>, ptr addrspace(1) %in
+ store <16 x bfloat> %val, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_arg_store(bfloat %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_arg_store:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_store_short v0, v[1:2], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_arg_store:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_store_short v0, v[1:2], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_arg_store:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: flat_store_short v[1:2], v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_arg_store:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_store_short_d16_hi v[1:2], v0, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_arg_store:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_store_short_d16_hi v[1:2], v0, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store bfloat %in, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_arg_store_v2bf16(<2 x bfloat> %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_arg_store_v2bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_arg_store_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_arg_store_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_store_dword v[1:2], v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_arg_store_v2bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_store_dword v[1:2], v0, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_arg_store_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_store_dword v[1:2], v0, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store <2 x bfloat> %in, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_arg_store_v3bf16(<3 x bfloat> %in, <3 x bfloat> addrspace(1)* %out) {
+; GCN-LABEL: test_arg_store_v3bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_store_short v2, v[3:4], s[4:7], 0 addr64 offset:4
+; GCN-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_arg_store_v3bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_store_short v1, v[3:4], s[4:7], 0 addr64 offset:4
+; GFX7-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_arg_store_v3bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_store_dword v[2:3], v0
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v2
+; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GFX8-NEXT: flat_store_short v[2:3], v1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_arg_store_v3bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_store_short v[2:3], v1, off offset:4
+; GFX9-NEXT: global_store_dword v[2:3], v0, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_arg_store_v3bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_store_short v[2:3], v1, off offset:4
+; GFX10-NEXT: global_store_dword v[2:3], v0, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store <3 x bfloat> %in, <3 x bfloat> addrspace(1) * %out
+ ret void
+}
+
+define void @test_arg_store_v4bf16(<4 x bfloat> %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_arg_store_v4bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v1
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
+; GCN-NEXT: v_alignbit_b32 v0, v6, v0, 16
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_store_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_arg_store_v4bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_alignbit_b32 v2, v3, v2, 16
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_store_dwordx2 v[1:2], v[4:5], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_arg_store_v4bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_arg_store_v4bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_arg_store_v4bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store <4 x bfloat> %in, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_arg_store_v8bf16(<8 x bfloat> %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_arg_store_v8bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
+; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v5
+; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: v_alignbit_b32 v5, v7, v6, 16
+; GCN-NEXT: v_alignbit_b32 v4, v10, v4, 16
+; GCN-NEXT: v_alignbit_b32 v3, v3, v2, 16
+; GCN-NEXT: v_alignbit_b32 v2, v1, v0, 16
+; GCN-NEXT: buffer_store_dwordx4 v[2:5], v[8:9], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_arg_store_v8bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: v_alignbit_b32 v6, v7, v6, 16
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_alignbit_b32 v4, v3, v2, 16
+; GFX7-NEXT: v_alignbit_b32 v3, v1, v0, 16
+; GFX7-NEXT: buffer_store_dwordx4 v[3:6], v[8:9], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_arg_store_v8bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_arg_store_v8bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_store_dwordx4 v[4:5], v[0:3], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_arg_store_v8bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_store_dwordx4 v[4:5], v[0:3], off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store <8 x bfloat> %in, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_arg_store_v16bf16(<16 x bfloat> %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_arg_store_v16bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
+; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v5
+; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
+; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v13
+; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
+; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
+; GCN-NEXT: v_alignbit_b32 v5, v7, v6, 16
+; GCN-NEXT: v_alignbit_b32 v4, v18, v4, 16
+; GCN-NEXT: v_alignbit_b32 v3, v3, v2, 16
+; GCN-NEXT: v_alignbit_b32 v2, v1, v0, 16
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: v_alignbit_b32 v13, v15, v14, 16
+; GCN-NEXT: v_alignbit_b32 v12, v19, v12, 16
+; GCN-NEXT: v_alignbit_b32 v11, v11, v10, 16
+; GCN-NEXT: v_alignbit_b32 v10, v9, v8, 16
+; GCN-NEXT: buffer_store_dwordx4 v[10:13], v[16:17], s[4:7], 0 addr64 offset:16
+; GCN-NEXT: buffer_store_dwordx4 v[2:5], v[16:17], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_arg_store_v16bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_alignbit_b32 v4, v3, v2, 16
+; GFX7-NEXT: v_alignbit_b32 v3, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v15
+; GFX7-NEXT: v_alignbit_b32 v14, v0, v14, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v13
+; GFX7-NEXT: v_alignbit_b32 v13, v0, v12, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v11
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_alignbit_b32 v12, v0, v10, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v9
+; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: v_alignbit_b32 v11, v0, v8, 16
+; GFX7-NEXT: v_alignbit_b32 v6, v7, v6, 16
+; GFX7-NEXT: buffer_store_dwordx4 v[11:14], v[16:17], s[4:7], 0 addr64 offset:16
+; GFX7-NEXT: buffer_store_dwordx4 v[3:6], v[16:17], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_arg_store_v16bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[0:3]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v8
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v9, vcc
+; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_arg_store_v16bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_store_dwordx4 v[8:9], v[4:7], off offset:16
+; GFX9-NEXT: global_store_dwordx4 v[8:9], v[0:3], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_arg_store_v16bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_store_dwordx4 v[8:9], v[4:7], off offset:16
+; GFX10-NEXT: global_store_dwordx4 v[8:9], v[0:3], off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store <16 x bfloat> %in, ptr addrspace(1) %out
+ ret void
+}
+
+define amdgpu_gfx void @test_inreg_arg_store(bfloat inreg %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_inreg_arg_store:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_lshr_b32 s34, s4, 16
+; GCN-NEXT: s_mov_b32 s38, 0
+; GCN-NEXT: s_mov_b32 s39, 0xf000
+; GCN-NEXT: s_mov_b32 s36, s38
+; GCN-NEXT: s_mov_b32 s37, s38
+; GCN-NEXT: v_mov_b32_e32 v2, s34
+; GCN-NEXT: buffer_store_short v2, v[0:1], s[36:39], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_inreg_arg_store:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_lshr_b32 s34, s4, 16
+; GFX7-NEXT: s_mov_b32 s38, 0
+; GFX7-NEXT: s_mov_b32 s39, 0xf000
+; GFX7-NEXT: s_mov_b32 s36, s38
+; GFX7-NEXT: s_mov_b32 s37, s38
+; GFX7-NEXT: v_mov_b32_e32 v2, s34
+; GFX7-NEXT: buffer_store_short v2, v[0:1], s[36:39], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_inreg_arg_store:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_lshr_b32 s34, s4, 16
+; GFX8-NEXT: v_mov_b32_e32 v2, s34
+; GFX8-NEXT: flat_store_short v[0:1], v2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_inreg_arg_store:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: global_store_short_d16_hi v[0:1], v2, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_inreg_arg_store:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v2, s4
+; GFX10-NEXT: global_store_short_d16_hi v[0:1], v2, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store bfloat %in, ptr addrspace(1) %out
+ ret void
+}
+
+define bfloat @test_byval(ptr addrspace(5) byval(bfloat) %bv, bfloat %val) {
+; GCN-LABEL: test_byval:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GCN-NEXT: buffer_store_short v1, off, s[0:3], s32
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_byval:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX7-NEXT: buffer_store_short v1, off, s[0:3], s32
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_byval:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX8-NEXT: buffer_store_short v1, off, s[0:3], s32
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_byval:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v0, off, s[0:3], s32
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_byval:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v0, off, s[0:3], s32
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store bfloat %val, ptr addrspace(5) %bv
+ %retval = load bfloat, ptr addrspace(5) %bv
+ ret bfloat %retval
+}
+
+define void @test_sret(ptr addrspace(5) sret(bfloat) %sret, bfloat %val) {
+; GCN-LABEL: test_sret:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_sret:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_sret:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX8-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_sret:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v1, v0, s[0:3], 0 offen
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_sret:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v1, v0, s[0:3], 0 offen
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ store bfloat %val, ptr addrspace(5) %sret
+ ret void
+}
+
+define void @test_bitcast_from_bfloat(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GCN-LABEL: test_bitcast_from_bfloat:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_bitcast_from_bfloat:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_bitcast_from_bfloat:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ushort v0, v[0:1]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: flat_store_short v[2:3], v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_bitcast_from_bfloat:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ushort v0, v[0:1], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: global_store_short v[2:3], v0, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_bitcast_from_bfloat:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_load_ushort v0, v[0:1], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_short v[2:3], v0, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load bfloat, ptr addrspace(1) %in
+ %val_int = bitcast bfloat %val to i16
+ store i16 %val_int, ptr addrspace(1) %out
+ ret void
+}
+
+define void @test_bitcast_to_bfloat(ptr addrspace(1) %out, ptr addrspace(1) %in) {
+; GCN-LABEL: test_bitcast_to_bfloat:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_mov_b32 s6, 0
+; GCN-NEXT: s_mov_b32 s7, 0xf000
+; GCN-NEXT: s_mov_b32 s4, s6
+; GCN-NEXT: s_mov_b32 s5, s6
+; GCN-NEXT: buffer_load_ushort v2, v[2:3], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v2, v[0:1], s[4:7], 0 addr64
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_bitcast_to_bfloat:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_ushort v2, v[2:3], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_short v2, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_bitcast_to_bfloat:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ushort v2, v[2:3]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: flat_store_short v[0:1], v2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_bitcast_to_bfloat:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ushort v2, v[2:3], off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: global_store_short v[0:1], v2, off
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_bitcast_to_bfloat:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_load_ushort v2, v[2:3], off
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: global_store_short v[0:1], v2, off
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %val = load i16, ptr addrspace(1) %in
+ %val_fp = bitcast i16 %val to bfloat
+ store bfloat %val_fp, ptr addrspace(1) %out
+ ret void
+}
+
+define bfloat @test_ret(bfloat %in) {
+; GCN-LABEL: test_ret:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_ret:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_ret:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_ret:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_ret:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ ret bfloat %in
+}
+
+define <2 x bfloat> @test_ret_v2bf16(<2 x bfloat> %in) {
+; GCN-LABEL: test_ret_v2bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_ret_v2bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_ret_v2bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_ret_v2bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_ret_v2bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ ret <2 x bfloat> %in
+}
+
+define <3 x bfloat> @test_ret_v3bf16(<3 x bfloat> %in) {
+; GCN-LABEL: test_ret_v3bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_ret_v3bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_ret_v3bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_ret_v3bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX9-NEXT: s_mov_b32 s4, 0xffff
+; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v2
+; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_ret_v3bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ ret <3 x bfloat> %in
+}
+
+define <4 x bfloat> @test_ret_v4bf16(<4 x bfloat> %in) {
+; GCN-LABEL: test_ret_v4bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_ret_v4bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_ret_v4bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_ret_v4bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_ret_v4bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ ret <4 x bfloat> %in
+}
+
+define <8 x bfloat> @test_ret_v8bf16(<8 x bfloat> %in) {
+; GCN-LABEL: test_ret_v8bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_ret_v8bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_ret_v8bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_ret_v8bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_ret_v8bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ ret <8 x bfloat> %in
+}
+
+define <16 x bfloat> @test_ret_v16bf16(<16 x bfloat> %in) {
+; GCN-LABEL: test_ret_v16bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_ret_v16bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_ret_v16bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_ret_v16bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_ret_v16bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ ret <16 x bfloat> %in
+}
+
+define void @test_call(bfloat %in, ptr addrspace(5) %out) {
+; GCN-LABEL: test_call:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_writelane_b32 v2, s33, 2
+; GCN-NEXT: s_mov_b32 s33, s32
+; GCN-NEXT: s_addk_i32 s32, 0x400
+; GCN-NEXT: v_writelane_b32 v2, s30, 0
+; GCN-NEXT: v_writelane_b32 v2, s31, 1
+; GCN-NEXT: s_getpc_b64 s[4:5]
+; GCN-NEXT: s_add_u32 s4, s4, test_arg_store@gotpcrel32@lo+4
+; GCN-NEXT: s_addc_u32 s5, s5, test_arg_store@gotpcrel32@hi+12
+; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_readlane_b32 s31, v2, 1
+; GCN-NEXT: v_readlane_b32 s30, v2, 0
+; GCN-NEXT: s_addk_i32 s32, 0xfc00
+; GCN-NEXT: v_readlane_b32 s33, v2, 2
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_call:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_store_dword v2, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: v_writelane_b32 v2, s33, 2
+; GFX7-NEXT: s_mov_b32 s33, s32
+; GFX7-NEXT: s_addk_i32 s32, 0x400
+; GFX7-NEXT: s_getpc_b64 s[4:5]
+; GFX7-NEXT: s_add_u32 s4, s4, test_arg_store@gotpcrel32@lo+4
+; GFX7-NEXT: s_addc_u32 s5, s5, test_arg_store@gotpcrel32@hi+12
+; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX7-NEXT: v_writelane_b32 v2, s30, 0
+; GFX7-NEXT: v_writelane_b32 v2, s31, 1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_readlane_b32 s31, v2, 1
+; GFX7-NEXT: v_readlane_b32 s30, v2, 0
+; GFX7-NEXT: s_addk_i32 s32, 0xfc00
+; GFX7-NEXT: v_readlane_b32 s33, v2, 2
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_call:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_store_dword v2, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: v_writelane_b32 v2, s33, 2
+; GFX8-NEXT: s_mov_b32 s33, s32
+; GFX8-NEXT: s_addk_i32 s32, 0x400
+; GFX8-NEXT: s_getpc_b64 s[4:5]
+; GFX8-NEXT: s_add_u32 s4, s4, test_arg_store@gotpcrel32@lo+4
+; GFX8-NEXT: s_addc_u32 s5, s5, test_arg_store@gotpcrel32@hi+12
+; GFX8-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX8-NEXT: v_writelane_b32 v2, s30, 0
+; GFX8-NEXT: v_writelane_b32 v2, s31, 1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_readlane_b32 s31, v2, 1
+; GFX8-NEXT: v_readlane_b32 s30, v2, 0
+; GFX8-NEXT: s_addk_i32 s32, 0xfc00
+; GFX8-NEXT: v_readlane_b32 s33, v2, 2
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_call:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_writelane_b32 v2, s33, 2
+; GFX9-NEXT: s_mov_b32 s33, s32
+; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, test_arg_store@gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, test_arg_store@gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX9-NEXT: v_writelane_b32 v2, s30, 0
+; GFX9-NEXT: v_writelane_b32 v2, s31, 1
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX9-NEXT: buffer_store_short_d16_hi v0, v1, s[0:3], 0 offen
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_readlane_b32 s31, v2, 1
+; GFX9-NEXT: v_readlane_b32 s30, v2, 0
+; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: v_readlane_b32 s33, v2, 2
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_call:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_store_dword v2, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: v_writelane_b32 v2, s33, 2
+; GFX10-NEXT: s_mov_b32 s33, s32
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: s_getpc_b64 s[4:5]
+; GFX10-NEXT: s_add_u32 s4, s4, test_arg_store@gotpcrel32@lo+4
+; GFX10-NEXT: s_addc_u32 s5, s5, test_arg_store@gotpcrel32@hi+12
+; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX10-NEXT: v_writelane_b32 v2, s30, 0
+; GFX10-NEXT: v_writelane_b32 v2, s31, 1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX10-NEXT: buffer_store_short_d16_hi v0, v1, s[0:3], 0 offen
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_readlane_b32 s31, v2, 1
+; GFX10-NEXT: v_readlane_b32 s30, v2, 0
+; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: v_readlane_b32 s33, v2, 2
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %result = call bfloat @test_arg_store(bfloat %in)
+ store volatile bfloat %result, ptr addrspace(5) %out
+ ret void
+}
+
+define void @test_call_v2bf16(<2 x bfloat> %in, ptr addrspace(5) %out) {
+; GCN-LABEL: test_call_v2bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_writelane_b32 v3, s33, 2
+; GCN-NEXT: s_mov_b32 s33, s32
+; GCN-NEXT: s_addk_i32 s32, 0x400
+; GCN-NEXT: v_writelane_b32 v3, s30, 0
+; GCN-NEXT: v_writelane_b32 v3, s31, 1
+; GCN-NEXT: s_getpc_b64 s[4:5]
+; GCN-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GCN-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: v_add_i32_e32 v4, vcc, 2, v2
+; GCN-NEXT: buffer_store_short v1, v4, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_readlane_b32 s31, v3, 1
+; GCN-NEXT: v_readlane_b32 s30, v3, 0
+; GCN-NEXT: s_addk_i32 s32, 0xfc00
+; GCN-NEXT: v_readlane_b32 s33, v3, 2
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_call_v2bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: v_writelane_b32 v3, s33, 2
+; GFX7-NEXT: s_mov_b32 s33, s32
+; GFX7-NEXT: s_addk_i32 s32, 0x400
+; GFX7-NEXT: s_getpc_b64 s[4:5]
+; GFX7-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX7-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX7-NEXT: v_writelane_b32 v3, s30, 0
+; GFX7-NEXT: v_writelane_b32 v3, s31, 1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 2, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_short v1, v4, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_readlane_b32 s31, v3, 1
+; GFX7-NEXT: v_readlane_b32 s30, v3, 0
+; GFX7-NEXT: s_addk_i32 s32, 0xfc00
+; GFX7-NEXT: v_readlane_b32 s33, v3, 2
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_call_v2bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_store_dword v2, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: v_writelane_b32 v2, s33, 2
+; GFX8-NEXT: s_mov_b32 s33, s32
+; GFX8-NEXT: s_addk_i32 s32, 0x400
+; GFX8-NEXT: s_getpc_b64 s[4:5]
+; GFX8-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX8-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX8-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX8-NEXT: v_writelane_b32 v2, s30, 0
+; GFX8-NEXT: v_writelane_b32 v2, s31, 1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX8-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_readlane_b32 s31, v2, 1
+; GFX8-NEXT: v_readlane_b32 s30, v2, 0
+; GFX8-NEXT: s_addk_i32 s32, 0xfc00
+; GFX8-NEXT: v_readlane_b32 s33, v2, 2
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_call_v2bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_writelane_b32 v2, s33, 2
+; GFX9-NEXT: s_mov_b32 s33, s32
+; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX9-NEXT: v_writelane_b32 v2, s30, 0
+; GFX9-NEXT: v_writelane_b32 v2, s31, 1
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX9-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_readlane_b32 s31, v2, 1
+; GFX9-NEXT: v_readlane_b32 s30, v2, 0
+; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: v_readlane_b32 s33, v2, 2
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_call_v2bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_store_dword v2, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: v_writelane_b32 v2, s33, 2
+; GFX10-NEXT: s_mov_b32 s33, s32
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: s_getpc_b64 s[4:5]
+; GFX10-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX10-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX10-NEXT: v_writelane_b32 v2, s30, 0
+; GFX10-NEXT: v_writelane_b32 v2, s31, 1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX10-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_readlane_b32 s31, v2, 1
+; GFX10-NEXT: v_readlane_b32 s30, v2, 0
+; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: v_readlane_b32 s33, v2, 2
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %result = call <2 x bfloat> @test_arg_store_v2bf16(<2 x bfloat> %in)
+ store volatile <2 x bfloat> %result, ptr addrspace(5) %out
+ ret void
+}
+
+define void @test_call_v3bf16(<3 x bfloat> %in, ptr addrspace(5) %out) {
+; GCN-LABEL: test_call_v3bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_writelane_b32 v4, s33, 2
+; GCN-NEXT: s_mov_b32 s33, s32
+; GCN-NEXT: s_addk_i32 s32, 0x400
+; GCN-NEXT: v_writelane_b32 v4, s30, 0
+; GCN-NEXT: v_writelane_b32 v4, s31, 1
+; GCN-NEXT: s_getpc_b64 s[4:5]
+; GCN-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GCN-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GCN-NEXT: v_add_i32_e32 v5, vcc, 4, v3
+; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GCN-NEXT: buffer_store_short v2, v5, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_dword v0, v3, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_readlane_b32 s31, v4, 1
+; GCN-NEXT: v_readlane_b32 s30, v4, 0
+; GCN-NEXT: s_addk_i32 s32, 0xfc00
+; GCN-NEXT: v_readlane_b32 s33, v4, 2
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_call_v3bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_store_dword v4, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: v_writelane_b32 v4, s33, 2
+; GFX7-NEXT: s_mov_b32 s33, s32
+; GFX7-NEXT: s_addk_i32 s32, 0x400
+; GFX7-NEXT: s_getpc_b64 s[4:5]
+; GFX7-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX7-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX7-NEXT: v_writelane_b32 v4, s30, 0
+; GFX7-NEXT: v_writelane_b32 v4, s31, 1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 4, v3
+; GFX7-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_dword v0, v3, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_readlane_b32 s31, v4, 1
+; GFX7-NEXT: v_readlane_b32 s30, v4, 0
+; GFX7-NEXT: s_addk_i32 s32, 0xfc00
+; GFX7-NEXT: v_readlane_b32 s33, v4, 2
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_load_dword v4, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_call_v3bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: v_writelane_b32 v3, s33, 2
+; GFX8-NEXT: s_mov_b32 s33, s32
+; GFX8-NEXT: s_addk_i32 s32, 0x400
+; GFX8-NEXT: s_getpc_b64 s[4:5]
+; GFX8-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX8-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX8-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX8-NEXT: v_writelane_b32 v3, s30, 0
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX8-NEXT: v_writelane_b32 v3, s31, 1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 4, v2
+; GFX8-NEXT: buffer_store_short v1, v4, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_readlane_b32 s31, v3, 1
+; GFX8-NEXT: v_readlane_b32 s30, v3, 0
+; GFX8-NEXT: s_addk_i32 s32, 0xfc00
+; GFX8-NEXT: v_readlane_b32 s33, v3, 2
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_call_v3bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_writelane_b32 v3, s33, 2
+; GFX9-NEXT: s_mov_b32 s33, s32
+; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
+; GFX9-NEXT: s_mov_b32 s4, 0xffff
+; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v4
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX9-NEXT: v_writelane_b32 v3, s30, 0
+; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX9-NEXT: v_writelane_b32 v3, s31, 1
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX9-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen offset:4
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_readlane_b32 s31, v3, 1
+; GFX9-NEXT: v_readlane_b32 s30, v3, 0
+; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: v_readlane_b32 s33, v3, 2
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_call_v3bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: v_writelane_b32 v3, s33, 2
+; GFX10-NEXT: s_mov_b32 s33, s32
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: s_getpc_b64 s[4:5]
+; GFX10-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX10-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
+; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX10-NEXT: v_writelane_b32 v3, s30, 0
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v0, v4
+; GFX10-NEXT: v_writelane_b32 v3, s31, 1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX10-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen offset:4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_readlane_b32 s31, v3, 1
+; GFX10-NEXT: v_readlane_b32 s30, v3, 0
+; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: v_readlane_b32 s33, v3, 2
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %result = call <3 x bfloat> @test_arg_store_v2bf16(<3 x bfloat> %in)
+ store volatile <3 x bfloat> %result, ptr addrspace(5) %out
+ ret void
+}
+
+define void @test_call_v4bf16(<4 x bfloat> %in, ptr addrspace(5) %out) {
+; GCN-LABEL: test_call_v4bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_writelane_b32 v5, s33, 2
+; GCN-NEXT: s_mov_b32 s33, s32
+; GCN-NEXT: s_addk_i32 s32, 0x400
+; GCN-NEXT: v_writelane_b32 v5, s30, 0
+; GCN-NEXT: v_writelane_b32 v5, s31, 1
+; GCN-NEXT: s_getpc_b64 s[4:5]
+; GCN-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GCN-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GCN-NEXT: v_add_i32_e32 v6, vcc, 6, v4
+; GCN-NEXT: v_add_i32_e32 v7, vcc, 4, v4
+; GCN-NEXT: v_add_i32_e32 v8, vcc, 2, v4
+; GCN-NEXT: buffer_store_short v3, v6, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v2, v7, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v1, v8, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v0, v4, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_readlane_b32 s31, v5, 1
+; GCN-NEXT: v_readlane_b32 s30, v5, 0
+; GCN-NEXT: s_addk_i32 s32, 0xfc00
+; GCN-NEXT: v_readlane_b32 s33, v5, 2
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_call_v4bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: v_writelane_b32 v5, s33, 2
+; GFX7-NEXT: s_mov_b32 s33, s32
+; GFX7-NEXT: s_addk_i32 s32, 0x400
+; GFX7-NEXT: s_getpc_b64 s[4:5]
+; GFX7-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX7-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX7-NEXT: v_writelane_b32 v5, s30, 0
+; GFX7-NEXT: v_writelane_b32 v5, s31, 1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 6, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: buffer_store_short v3, v6, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 4, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: buffer_store_short v2, v3, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 2, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_short v0, v4, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_readlane_b32 s31, v5, 1
+; GFX7-NEXT: v_readlane_b32 s30, v5, 0
+; GFX7-NEXT: s_addk_i32 s32, 0xfc00
+; GFX7-NEXT: v_readlane_b32 s33, v5, 2
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_call_v4bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: v_writelane_b32 v3, s33, 2
+; GFX8-NEXT: s_mov_b32 s33, s32
+; GFX8-NEXT: s_addk_i32 s32, 0x400
+; GFX8-NEXT: s_getpc_b64 s[4:5]
+; GFX8-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX8-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX8-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX8-NEXT: v_writelane_b32 v3, s30, 0
+; GFX8-NEXT: v_writelane_b32 v3, s31, 1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 4, v2
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v0
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v1
+; GFX8-NEXT: buffer_store_short v1, v6, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 6, v2
+; GFX8-NEXT: buffer_store_short v5, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 2, v2
+; GFX8-NEXT: buffer_store_short v4, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_readlane_b32 s31, v3, 1
+; GFX8-NEXT: v_readlane_b32 s30, v3, 0
+; GFX8-NEXT: s_addk_i32 s32, 0xfc00
+; GFX8-NEXT: v_readlane_b32 s33, v3, 2
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_call_v4bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_writelane_b32 v3, s33, 2
+; GFX9-NEXT: s_mov_b32 s33, s32
+; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX9-NEXT: v_writelane_b32 v3, s30, 0
+; GFX9-NEXT: v_writelane_b32 v3, s31, 1
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX9-NEXT: buffer_store_short_d16_hi v1, v2, s[0:3], 0 offen offset:6
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen offset:4
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v0, v2, s[0:3], 0 offen offset:2
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_readlane_b32 s31, v3, 1
+; GFX9-NEXT: v_readlane_b32 s30, v3, 0
+; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: v_readlane_b32 s33, v3, 2
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_call_v4bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_store_dword v3, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: v_writelane_b32 v3, s33, 2
+; GFX10-NEXT: s_mov_b32 s33, s32
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: s_getpc_b64 s[4:5]
+; GFX10-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX10-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX10-NEXT: v_writelane_b32 v3, s30, 0
+; GFX10-NEXT: v_writelane_b32 v3, s31, 1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX10-NEXT: buffer_store_short_d16_hi v1, v2, s[0:3], 0 offen offset:6
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen offset:4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v0, v2, s[0:3], 0 offen offset:2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_readlane_b32 s31, v3, 1
+; GFX10-NEXT: v_readlane_b32 s30, v3, 0
+; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: v_readlane_b32 s33, v3, 2
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %result = call <4 x bfloat> @test_arg_store_v2bf16(<4 x bfloat> %in)
+ store volatile <4 x bfloat> %result, ptr addrspace(5) %out
+ ret void
+}
+
+define void @test_call_v8bf16(<8 x bfloat> %in, ptr addrspace(5) %out) {
+; GCN-LABEL: test_call_v8bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_writelane_b32 v9, s33, 2
+; GCN-NEXT: s_mov_b32 s33, s32
+; GCN-NEXT: s_addk_i32 s32, 0x400
+; GCN-NEXT: v_writelane_b32 v9, s30, 0
+; GCN-NEXT: v_writelane_b32 v9, s31, 1
+; GCN-NEXT: s_getpc_b64 s[4:5]
+; GCN-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GCN-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
+; GCN-NEXT: v_add_i32_e32 v10, vcc, 14, v8
+; GCN-NEXT: v_add_i32_e32 v11, vcc, 12, v8
+; GCN-NEXT: v_add_i32_e32 v12, vcc, 10, v8
+; GCN-NEXT: v_add_i32_e32 v13, vcc, 8, v8
+; GCN-NEXT: v_add_i32_e32 v14, vcc, 6, v8
+; GCN-NEXT: v_add_i32_e32 v15, vcc, 4, v8
+; GCN-NEXT: v_add_i32_e32 v16, vcc, 2, v8
+; GCN-NEXT: buffer_store_short v7, v10, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v6, v11, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v5, v12, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v4, v13, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v3, v14, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v2, v15, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v1, v16, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v0, v8, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_readlane_b32 s31, v9, 1
+; GCN-NEXT: v_readlane_b32 s30, v9, 0
+; GCN-NEXT: s_addk_i32 s32, 0xfc00
+; GCN-NEXT: v_readlane_b32 s33, v9, 2
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_call_v8bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: v_writelane_b32 v9, s33, 2
+; GFX7-NEXT: s_mov_b32 s33, s32
+; GFX7-NEXT: s_addk_i32 s32, 0x400
+; GFX7-NEXT: s_getpc_b64 s[4:5]
+; GFX7-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX7-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX7-NEXT: v_writelane_b32 v9, s30, 0
+; GFX7-NEXT: v_writelane_b32 v9, s31, 1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
+; GFX7-NEXT: v_add_i32_e32 v10, vcc, 14, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX7-NEXT: buffer_store_short v7, v10, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v7, vcc, 12, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: buffer_store_short v6, v7, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 10, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: buffer_store_short v5, v6, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v5, vcc, 8, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: buffer_store_short v4, v5, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 6, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: buffer_store_short v3, v4, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 4, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: buffer_store_short v2, v3, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_short v0, v8, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_readlane_b32 s31, v9, 1
+; GFX7-NEXT: v_readlane_b32 s30, v9, 0
+; GFX7-NEXT: s_addk_i32 s32, 0xfc00
+; GFX7-NEXT: v_readlane_b32 s33, v9, 2
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_call_v8bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: v_writelane_b32 v5, s33, 2
+; GFX8-NEXT: s_mov_b32 s33, s32
+; GFX8-NEXT: s_addk_i32 s32, 0x400
+; GFX8-NEXT: s_getpc_b64 s[4:5]
+; GFX8-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX8-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX8-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX8-NEXT: v_writelane_b32 v5, s30, 0
+; GFX8-NEXT: v_writelane_b32 v5, s31, 1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX8-NEXT: v_add_u32_e32 v10, vcc, 12, v4
+; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v3
+; GFX8-NEXT: buffer_store_short v3, v10, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 8, v4
+; GFX8-NEXT: v_lshrrev_b32_e32 v8, 16, v2
+; GFX8-NEXT: buffer_store_short v2, v3, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v4
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v0
+; GFX8-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_store_short v0, v4, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 14, v4
+; GFX8-NEXT: buffer_store_short v9, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 10, v4
+; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v1
+; GFX8-NEXT: buffer_store_short v8, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 6, v4
+; GFX8-NEXT: buffer_store_short v7, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 2, v4
+; GFX8-NEXT: buffer_store_short v6, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_readlane_b32 s31, v5, 1
+; GFX8-NEXT: v_readlane_b32 s30, v5, 0
+; GFX8-NEXT: s_addk_i32 s32, 0xfc00
+; GFX8-NEXT: v_readlane_b32 s33, v5, 2
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_call_v8bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_writelane_b32 v5, s33, 2
+; GFX9-NEXT: s_mov_b32 s33, s32
+; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX9-NEXT: v_writelane_b32 v5, s30, 0
+; GFX9-NEXT: v_writelane_b32 v5, s31, 1
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX9-NEXT: buffer_store_short_d16_hi v3, v4, s[0:3], 0 offen offset:14
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v3, v4, s[0:3], 0 offen offset:12
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v2, v4, s[0:3], 0 offen offset:10
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v2, v4, s[0:3], 0 offen offset:8
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v1, v4, s[0:3], 0 offen offset:6
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v1, v4, s[0:3], 0 offen offset:4
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v0, v4, s[0:3], 0 offen offset:2
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v0, v4, s[0:3], 0 offen
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_readlane_b32 s31, v5, 1
+; GFX9-NEXT: v_readlane_b32 s30, v5, 0
+; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: v_readlane_b32 s33, v5, 2
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_call_v8bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: v_writelane_b32 v5, s33, 2
+; GFX10-NEXT: s_mov_b32 s33, s32
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: s_getpc_b64 s[4:5]
+; GFX10-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX10-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX10-NEXT: v_writelane_b32 v5, s30, 0
+; GFX10-NEXT: v_writelane_b32 v5, s31, 1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX10-NEXT: buffer_store_short_d16_hi v3, v4, s[0:3], 0 offen offset:14
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v3, v4, s[0:3], 0 offen offset:12
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v2, v4, s[0:3], 0 offen offset:10
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v2, v4, s[0:3], 0 offen offset:8
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v1, v4, s[0:3], 0 offen offset:6
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v1, v4, s[0:3], 0 offen offset:4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v0, v4, s[0:3], 0 offen offset:2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v0, v4, s[0:3], 0 offen
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_readlane_b32 s31, v5, 1
+; GFX10-NEXT: v_readlane_b32 s30, v5, 0
+; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: v_readlane_b32 s33, v5, 2
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %result = call <8 x bfloat> @test_arg_store_v2bf16(<8 x bfloat> %in)
+ store volatile <8 x bfloat> %result, ptr addrspace(5) %out
+ ret void
+}
+
+define void @test_call_v16bf16(<16 x bfloat> %in, ptr addrspace(5) %out) {
+; GCN-LABEL: test_call_v16bf16:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s32 ; 4-byte Folded Spill
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_writelane_b32 v17, s33, 2
+; GCN-NEXT: s_mov_b32 s33, s32
+; GCN-NEXT: s_addk_i32 s32, 0x400
+; GCN-NEXT: v_writelane_b32 v17, s30, 0
+; GCN-NEXT: v_writelane_b32 v17, s31, 1
+; GCN-NEXT: s_getpc_b64 s[4:5]
+; GCN-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GCN-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GCN-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
+; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
+; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
+; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
+; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
+; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
+; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
+; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
+; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
+; GCN-NEXT: v_add_i32_e32 v18, vcc, 30, v16
+; GCN-NEXT: v_add_i32_e32 v19, vcc, 28, v16
+; GCN-NEXT: v_add_i32_e32 v20, vcc, 26, v16
+; GCN-NEXT: buffer_store_short v15, v18, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v15, vcc, 24, v16
+; GCN-NEXT: v_add_i32_e32 v18, vcc, 22, v16
+; GCN-NEXT: buffer_store_short v14, v19, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v14, vcc, 20, v16
+; GCN-NEXT: v_add_i32_e32 v19, vcc, 18, v16
+; GCN-NEXT: buffer_store_short v13, v20, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v13, vcc, 16, v16
+; GCN-NEXT: v_add_i32_e32 v20, vcc, 14, v16
+; GCN-NEXT: buffer_store_short v12, v15, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v12, vcc, 12, v16
+; GCN-NEXT: v_add_i32_e32 v15, vcc, 10, v16
+; GCN-NEXT: buffer_store_short v11, v18, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v11, vcc, 8, v16
+; GCN-NEXT: v_add_i32_e32 v18, vcc, 6, v16
+; GCN-NEXT: buffer_store_short v10, v14, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v10, vcc, 4, v16
+; GCN-NEXT: v_add_i32_e32 v14, vcc, 2, v16
+; GCN-NEXT: buffer_store_short v9, v19, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v8, v13, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v7, v20, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v6, v12, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v5, v15, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v4, v11, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v3, v18, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v2, v10, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v1, v14, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_store_short v0, v16, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_readlane_b32 s31, v17, 1
+; GCN-NEXT: v_readlane_b32 s30, v17, 0
+; GCN-NEXT: s_addk_i32 s32, 0xfc00
+; GCN-NEXT: v_readlane_b32 s33, v17, 2
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 ; 4-byte Folded Reload
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_call_v16bf16:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_store_dword v17, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: v_writelane_b32 v17, s33, 2
+; GFX7-NEXT: s_mov_b32 s33, s32
+; GFX7-NEXT: s_addk_i32 s32, 0x400
+; GFX7-NEXT: s_getpc_b64 s[4:5]
+; GFX7-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX7-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX7-NEXT: v_writelane_b32 v17, s30, 0
+; GFX7-NEXT: v_writelane_b32 v17, s31, 1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v15, 16, v15
+; GFX7-NEXT: v_add_i32_e32 v18, vcc, 30, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
+; GFX7-NEXT: buffer_store_short v15, v18, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v15, vcc, 28, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v13, 16, v13
+; GFX7-NEXT: buffer_store_short v14, v15, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v14, vcc, 26, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v12, 16, v12
+; GFX7-NEXT: buffer_store_short v13, v14, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v13, vcc, 24, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v11, 16, v11
+; GFX7-NEXT: buffer_store_short v12, v13, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v12, vcc, 22, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v10, 16, v10
+; GFX7-NEXT: buffer_store_short v11, v12, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v11, vcc, 20, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v9, 16, v9
+; GFX7-NEXT: buffer_store_short v10, v11, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v10, vcc, 18, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v8, 16, v8
+; GFX7-NEXT: buffer_store_short v9, v10, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v9, vcc, 16, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
+; GFX7-NEXT: buffer_store_short v8, v9, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v8, vcc, 14, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX7-NEXT: buffer_store_short v7, v8, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v7, vcc, 12, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: buffer_store_short v6, v7, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 10, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: buffer_store_short v5, v6, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v5, vcc, 8, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: buffer_store_short v4, v5, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 6, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: buffer_store_short v3, v4, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 4, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: buffer_store_short v2, v3, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 2, v16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_short v0, v16, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_readlane_b32 s31, v17, 1
+; GFX7-NEXT: v_readlane_b32 s30, v17, 0
+; GFX7-NEXT: s_addk_i32 s32, 0xfc00
+; GFX7-NEXT: v_readlane_b32 s33, v17, 2
+; GFX7-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX7-NEXT: buffer_load_dword v17, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX7-NEXT: s_mov_b64 exec, s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_call_v16bf16:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: v_writelane_b32 v9, s33, 2
+; GFX8-NEXT: s_mov_b32 s33, s32
+; GFX8-NEXT: s_addk_i32 s32, 0x400
+; GFX8-NEXT: s_getpc_b64 s[4:5]
+; GFX8-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX8-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX8-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX8-NEXT: v_writelane_b32 v9, s30, 0
+; GFX8-NEXT: v_writelane_b32 v9, s31, 1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX8-NEXT: v_add_u32_e32 v18, vcc, 28, v8
+; GFX8-NEXT: v_lshrrev_b32_e32 v17, 16, v7
+; GFX8-NEXT: buffer_store_short v7, v18, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 24, v8
+; GFX8-NEXT: v_lshrrev_b32_e32 v16, 16, v6
+; GFX8-NEXT: buffer_store_short v6, v7, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 20, v8
+; GFX8-NEXT: v_lshrrev_b32_e32 v15, 16, v5
+; GFX8-NEXT: buffer_store_short v5, v6, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 16, v8
+; GFX8-NEXT: v_lshrrev_b32_e32 v14, 16, v4
+; GFX8-NEXT: buffer_store_short v4, v5, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 12, v8
+; GFX8-NEXT: v_lshrrev_b32_e32 v13, 16, v3
+; GFX8-NEXT: buffer_store_short v3, v4, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 8, v8
+; GFX8-NEXT: v_lshrrev_b32_e32 v12, 16, v2
+; GFX8-NEXT: buffer_store_short v2, v3, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v8
+; GFX8-NEXT: v_lshrrev_b32_e32 v10, 16, v0
+; GFX8-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_store_short v0, v8, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 30, v8
+; GFX8-NEXT: buffer_store_short v17, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 26, v8
+; GFX8-NEXT: buffer_store_short v16, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 22, v8
+; GFX8-NEXT: buffer_store_short v15, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 18, v8
+; GFX8-NEXT: buffer_store_short v14, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 14, v8
+; GFX8-NEXT: buffer_store_short v13, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 10, v8
+; GFX8-NEXT: v_lshrrev_b32_e32 v11, 16, v1
+; GFX8-NEXT: buffer_store_short v12, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 6, v8
+; GFX8-NEXT: buffer_store_short v11, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 2, v8
+; GFX8-NEXT: buffer_store_short v10, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_readlane_b32 s31, v9, 1
+; GFX8-NEXT: v_readlane_b32 s30, v9, 0
+; GFX8-NEXT: s_addk_i32 s32, 0xfc00
+; GFX8-NEXT: v_readlane_b32 s33, v9, 2
+; GFX8-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX8-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX8-NEXT: s_mov_b64 exec, s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_call_v16bf16:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_writelane_b32 v9, s33, 2
+; GFX9-NEXT: s_mov_b32 s33, s32
+; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX9-NEXT: v_writelane_b32 v9, s30, 0
+; GFX9-NEXT: v_writelane_b32 v9, s31, 1
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX9-NEXT: buffer_store_short_d16_hi v7, v8, s[0:3], 0 offen offset:30
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v7, v8, s[0:3], 0 offen offset:28
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v6, v8, s[0:3], 0 offen offset:26
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v6, v8, s[0:3], 0 offen offset:24
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v5, v8, s[0:3], 0 offen offset:22
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v5, v8, s[0:3], 0 offen offset:20
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v4, v8, s[0:3], 0 offen offset:18
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v4, v8, s[0:3], 0 offen offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v3, v8, s[0:3], 0 offen offset:14
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v3, v8, s[0:3], 0 offen offset:12
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v2, v8, s[0:3], 0 offen offset:10
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v2, v8, s[0:3], 0 offen offset:8
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v1, v8, s[0:3], 0 offen offset:6
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v1, v8, s[0:3], 0 offen offset:4
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v0, v8, s[0:3], 0 offen offset:2
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: buffer_store_short v0, v8, s[0:3], 0 offen
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_readlane_b32 s31, v9, 1
+; GFX9-NEXT: v_readlane_b32 s30, v9, 0
+; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: v_readlane_b32 s33, v9, 2
+; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_call_v16bf16:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_store_dword v9, off, s[0:3], s32 ; 4-byte Folded Spill
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: v_writelane_b32 v9, s33, 2
+; GFX10-NEXT: s_mov_b32 s33, s32
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: s_getpc_b64 s[4:5]
+; GFX10-NEXT: s_add_u32 s4, s4, test_arg_store_v2bf16@gotpcrel32@lo+4
+; GFX10-NEXT: s_addc_u32 s5, s5, test_arg_store_v2bf16@gotpcrel32@hi+12
+; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX10-NEXT: v_writelane_b32 v9, s30, 0
+; GFX10-NEXT: v_writelane_b32 v9, s31, 1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_swappc_b64 s[30:31], s[4:5]
+; GFX10-NEXT: buffer_store_short_d16_hi v7, v8, s[0:3], 0 offen offset:30
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v7, v8, s[0:3], 0 offen offset:28
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v6, v8, s[0:3], 0 offen offset:26
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v6, v8, s[0:3], 0 offen offset:24
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v5, v8, s[0:3], 0 offen offset:22
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v5, v8, s[0:3], 0 offen offset:20
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v4, v8, s[0:3], 0 offen offset:18
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v4, v8, s[0:3], 0 offen offset:16
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v3, v8, s[0:3], 0 offen offset:14
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v3, v8, s[0:3], 0 offen offset:12
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v2, v8, s[0:3], 0 offen offset:10
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v2, v8, s[0:3], 0 offen offset:8
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v1, v8, s[0:3], 0 offen offset:6
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v1, v8, s[0:3], 0 offen offset:4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short_d16_hi v0, v8, s[0:3], 0 offen offset:2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_store_short v0, v8, s[0:3], 0 offen
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_readlane_b32 s31, v9, 1
+; GFX10-NEXT: v_readlane_b32 s30, v9, 0
+; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: v_readlane_b32 s33, v9, 2
+; GFX10-NEXT: s_or_saveexec_b32 s4, -1
+; GFX10-NEXT: buffer_load_dword v9, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %result = call <16 x bfloat> @test_arg_store_v2bf16(<16 x bfloat> %in)
+ store volatile <16 x bfloat> %result, ptr addrspace(5) %out
+ ret void
+}
+
+define bfloat @test_alloca_load_store_ret(bfloat %in) {
+; GCN-LABEL: test_alloca_load_store_ret:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: buffer_store_short v0, off, s[0:3], s32
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: buffer_load_ushort v0, off, s[0:3], s32 glc
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_alloca_load_store_ret:
+; GFX7: ; %bb.0: ; %entry
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: buffer_store_short v0, off, s[0:3], s32
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_load_ushort v0, off, s[0:3], s32 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_alloca_load_store_ret:
+; GFX8: ; %bb.0: ; %entry
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: buffer_store_short v0, off, s[0:3], s32
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_load_ushort v0, off, s[0:3], s32 glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_alloca_load_store_ret:
+; GFX9: ; %bb.0: ; %entry
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: buffer_store_short_d16_hi v0, off, s[0:3], s32
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v0, 0
+; GFX9-NEXT: buffer_load_short_d16_hi v0, off, s[0:3], s32 glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_alloca_load_store_ret:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v1, 0
+; GFX10-NEXT: buffer_store_short_d16_hi v0, off, s[0:3], s32
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: buffer_load_short_d16_hi v1, off, s[0:3], s32 glc dlc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %in.addr = alloca bfloat, align 2, addrspace(5)
+ store volatile bfloat %in, ptr addrspace(5) %in.addr, align 2
+ %loaded = load volatile bfloat, ptr addrspace(5) %in.addr, align 2
+ ret bfloat %loaded
+}
+
+define { <32 x i32>, bfloat } @test_overflow_stack(bfloat %a, <32 x i32> %b) {
+; GCN-LABEL: test_overflow_stack:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8
+; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x7c, v0
+; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32
+; GCN-NEXT: s_waitcnt vmcnt(2)
+; GCN-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
+; GCN-NEXT: s_waitcnt vmcnt(2)
+; GCN-NEXT: buffer_store_dword v32, v2, s[0:3], 0 offen
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x74, v0
+; GCN-NEXT: s_waitcnt vmcnt(2)
+; GCN-NEXT: buffer_store_dword v33, v2, s[0:3], 0 offen
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x70, v0
+; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x6c, v0
+; GCN-NEXT: buffer_store_dword v30, v2, s[0:3], 0 offen
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x68, v0
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x64, v0
+; GCN-NEXT: buffer_store_dword v29, v31, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x60, v0
+; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x5c, v0
+; GCN-NEXT: buffer_store_dword v28, v2, s[0:3], 0 offen
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x54, v0
+; GCN-NEXT: buffer_store_dword v27, v30, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x50, v0
+; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x4c, v0
+; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v26, vcc, 0x48, v0
+; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x44, v0
+; GCN-NEXT: buffer_store_dword v25, v31, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v25, vcc, 64, v0
+; GCN-NEXT: v_add_i32_e32 v31, vcc, 60, v0
+; GCN-NEXT: buffer_store_dword v24, v2, s[0:3], 0 offen
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 56, v0
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v24, vcc, 52, v0
+; GCN-NEXT: buffer_store_dword v23, v28, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v23, vcc, 48, v0
+; GCN-NEXT: v_add_i32_e32 v28, vcc, 44, v0
+; GCN-NEXT: buffer_store_dword v22, v27, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v22, vcc, 40, v0
+; GCN-NEXT: v_add_i32_e32 v27, vcc, 36, v0
+; GCN-NEXT: buffer_store_dword v21, v30, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v21, vcc, 32, v0
+; GCN-NEXT: v_add_i32_e32 v30, vcc, 28, v0
+; GCN-NEXT: buffer_store_dword v20, v26, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v20, vcc, 24, v0
+; GCN-NEXT: v_add_i32_e32 v26, vcc, 20, v0
+; GCN-NEXT: buffer_store_dword v19, v29, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v19, vcc, 16, v0
+; GCN-NEXT: v_add_i32_e32 v29, vcc, 12, v0
+; GCN-NEXT: buffer_store_dword v18, v25, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt expcnt(0)
+; GCN-NEXT: v_add_i32_e32 v18, vcc, 8, v0
+; GCN-NEXT: v_add_i32_e32 v25, vcc, 4, v0
+; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x80, v0
+; GCN-NEXT: buffer_store_dword v17, v31, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v16, v2, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v15, v24, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v14, v23, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v13, v28, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v12, v22, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v11, v27, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v10, v21, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v9, v30, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v8, v20, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v7, v26, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v6, v19, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v5, v29, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v4, v18, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_dword v3, v25, s[0:3], 0 offen
+; GCN-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: test_overflow_stack:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
+; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8
+; GFX7-NEXT: v_add_i32_e32 v31, vcc, 0x7c, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
+; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
+; GFX7-NEXT: v_add_i32_e32 v31, vcc, 0x78, v0
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
+; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32
+; GFX7-NEXT: v_add_i32_e32 v31, vcc, 0x74, v0
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x70, v0
+; GFX7-NEXT: buffer_store_dword v30, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x6c, v0
+; GFX7-NEXT: buffer_store_dword v29, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x68, v0
+; GFX7-NEXT: buffer_store_dword v28, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x64, v0
+; GFX7-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x60, v0
+; GFX7-NEXT: buffer_store_dword v26, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x5c, v0
+; GFX7-NEXT: buffer_store_dword v25, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
+; GFX7-NEXT: buffer_store_dword v24, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x54, v0
+; GFX7-NEXT: buffer_store_dword v23, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x50, v0
+; GFX7-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x4c, v0
+; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x48, v0
+; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x44, v0
+; GFX7-NEXT: buffer_store_dword v19, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 64, v0
+; GFX7-NEXT: buffer_store_dword v18, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 60, v0
+; GFX7-NEXT: buffer_store_dword v17, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 56, v0
+; GFX7-NEXT: buffer_store_dword v16, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 52, v0
+; GFX7-NEXT: buffer_store_dword v15, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 48, v0
+; GFX7-NEXT: buffer_store_dword v14, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 44, v0
+; GFX7-NEXT: buffer_store_dword v13, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 40, v0
+; GFX7-NEXT: buffer_store_dword v12, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 36, v0
+; GFX7-NEXT: buffer_store_dword v11, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 32, v0
+; GFX7-NEXT: buffer_store_dword v10, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 28, v0
+; GFX7-NEXT: buffer_store_dword v9, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 24, v0
+; GFX7-NEXT: buffer_store_dword v8, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 20, v0
+; GFX7-NEXT: buffer_store_dword v7, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 16, v0
+; GFX7-NEXT: buffer_store_dword v6, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 12, v0
+; GFX7-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 8, v0
+; GFX7-NEXT: buffer_store_dword v4, v2, s[0:3], 0 offen
+; GFX7-NEXT: v_add_i32_e32 v2, vcc, 4, v0
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x80, v0
+; GFX7-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
+; GFX7-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: test_overflow_stack:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
+; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8
+; GFX8-NEXT: v_add_u32_e32 v31, vcc, 0x7c, v0
+; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
+; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
+; GFX8-NEXT: v_add_u32_e32 v31, vcc, 0x78, v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
+; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32
+; GFX8-NEXT: v_add_u32_e32 v31, vcc, 0x74, v0
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
+; GFX8-NEXT: buffer_store_dword v30, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
+; GFX8-NEXT: buffer_store_dword v29, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
+; GFX8-NEXT: buffer_store_dword v28, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
+; GFX8-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
+; GFX8-NEXT: buffer_store_dword v26, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
+; GFX8-NEXT: buffer_store_dword v25, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
+; GFX8-NEXT: buffer_store_dword v24, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
+; GFX8-NEXT: buffer_store_dword v23, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
+; GFX8-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
+; GFX8-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
+; GFX8-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
+; GFX8-NEXT: buffer_store_dword v19, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 64, v0
+; GFX8-NEXT: buffer_store_dword v18, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 60, v0
+; GFX8-NEXT: buffer_store_dword v17, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 56, v0
+; GFX8-NEXT: buffer_store_dword v16, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 52, v0
+; GFX8-NEXT: buffer_store_dword v15, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 48, v0
+; GFX8-NEXT: buffer_store_dword v14, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 44, v0
+; GFX8-NEXT: buffer_store_dword v13, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 40, v0
+; GFX8-NEXT: buffer_store_dword v12, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 36, v0
+; GFX8-NEXT: buffer_store_dword v11, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 32, v0
+; GFX8-NEXT: buffer_store_dword v10, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 28, v0
+; GFX8-NEXT: buffer_store_dword v9, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 24, v0
+; GFX8-NEXT: buffer_store_dword v8, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 20, v0
+; GFX8-NEXT: buffer_store_dword v7, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 16, v0
+; GFX8-NEXT: buffer_store_dword v6, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 12, v0
+; GFX8-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 8, v0
+; GFX8-NEXT: buffer_store_dword v4, v2, s[0:3], 0 offen
+; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x80, v0
+; GFX8-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
+; GFX8-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: test_overflow_stack:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:112
+; GFX9-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:108
+; GFX9-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:104
+; GFX9-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:100
+; GFX9-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:96
+; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:4
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:8
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:92
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32
+; GFX9-NEXT: s_nop 0
+; GFX9-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:88
+; GFX9-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:84
+; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:80
+; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:76
+; GFX9-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:72
+; GFX9-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:68
+; GFX9-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:64
+; GFX9-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:60
+; GFX9-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:56
+; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:52
+; GFX9-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:48
+; GFX9-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:44
+; GFX9-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:40
+; GFX9-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:36
+; GFX9-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:32
+; GFX9-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:28
+; GFX9-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:24
+; GFX9-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:20
+; GFX9-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:12
+; GFX9-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:8
+; GFX9-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:4
+; GFX9-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
+; GFX9-NEXT: s_waitcnt vmcnt(25)
+; GFX9-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:124
+; GFX9-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:120
+; GFX9-NEXT: s_waitcnt vmcnt(25)
+; GFX9-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:116
+; GFX9-NEXT: buffer_store_short_d16_hi v1, v0, s[0:3], 0 offen offset:128
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: test_overflow_stack:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_clause 0x2
+; GFX10-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
+; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
+; GFX10-NEXT: buffer_load_dword v33, off, s[0:3], s32
+; GFX10-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:112
+; GFX10-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:108
+; GFX10-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:104
+; GFX10-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:100
+; GFX10-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:96
+; GFX10-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:92
+; GFX10-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:88
+; GFX10-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:84
+; GFX10-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:80
+; GFX10-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:76
+; GFX10-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:72
+; GFX10-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:68
+; GFX10-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:64
+; GFX10-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:60
+; GFX10-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:56
+; GFX10-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:52
+; GFX10-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:48
+; GFX10-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:44
+; GFX10-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:40
+; GFX10-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:36
+; GFX10-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:32
+; GFX10-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:28
+; GFX10-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:24
+; GFX10-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:20
+; GFX10-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:16
+; GFX10-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:12
+; GFX10-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:8
+; GFX10-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:4
+; GFX10-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
+; GFX10-NEXT: s_waitcnt vmcnt(2)
+; GFX10-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:124
+; GFX10-NEXT: s_waitcnt vmcnt(1)
+; GFX10-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:120
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_store_dword v33, v0, s[0:3], 0 offen offset:116
+; GFX10-NEXT: buffer_store_short_d16_hi v1, v0, s[0:3], 0 offen offset:128
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+ %ins.0 = insertvalue { <32 x i32>, bfloat } poison, <32 x i32> %b, 0
+ %ins.1 = insertvalue { <32 x i32>, bfloat } %ins.0 ,bfloat %a, 1
+ ret { <32 x i32>, bfloat } %ins.1
+}