diff options
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/flat-scratch-i8-i16.ll')
| -rw-r--r-- | llvm/test/CodeGen/AMDGPU/flat-scratch-i8-i16.ll | 30 |
1 files changed, 18 insertions, 12 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/flat-scratch-i8-i16.ll b/llvm/test/CodeGen/AMDGPU/flat-scratch-i8-i16.ll index 47910f5280bf..04b036cafd81 100644 --- a/llvm/test/CodeGen/AMDGPU/flat-scratch-i8-i16.ll +++ b/llvm/test/CodeGen/AMDGPU/flat-scratch-i8-i16.ll @@ -263,7 +263,8 @@ define amdgpu_ps void @test_scratch_load_i8_zext_to_d16_hi_v(ptr addrspace(5) %i ; ; GFX11-LABEL: test_scratch_load_i8_zext_to_d16_hi_v: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_dual_mov_b32 v3, -1 :: v_dual_add_nc_u32 v0, 1, v0 +; GFX11-NEXT: v_add_nc_u32_e32 v0, 1, v0 +; GFX11-NEXT: v_mov_b16_e32 v3.l, -1 ; GFX11-NEXT: scratch_load_d16_hi_u8 v3, v0, off ; GFX11-NEXT: s_waitcnt vmcnt(0) ; GFX11-NEXT: flat_store_b32 v[1:2], v3 @@ -301,7 +302,8 @@ define amdgpu_ps void @test_scratch_load_i8_sext_to_d16_hi_v(ptr addrspace(5) %i ; ; GFX11-LABEL: test_scratch_load_i8_sext_to_d16_hi_v: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_dual_mov_b32 v3, -1 :: v_dual_add_nc_u32 v0, 1, v0 +; GFX11-NEXT: v_add_nc_u32_e32 v0, 1, v0 +; GFX11-NEXT: v_mov_b16_e32 v3.l, -1 ; GFX11-NEXT: scratch_load_d16_hi_i8 v3, v0, off ; GFX11-NEXT: s_waitcnt vmcnt(0) ; GFX11-NEXT: flat_store_b32 v[1:2], v3 @@ -339,7 +341,8 @@ define amdgpu_ps void @test_scratch_load_i16_to_d16_hi_v(ptr addrspace(5) %in, p ; ; GFX11-LABEL: test_scratch_load_i16_to_d16_hi_v: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_dual_mov_b32 v3, -1 :: v_dual_add_nc_u32 v0, 2, v0 +; GFX11-NEXT: v_add_nc_u32_e32 v0, 2, v0 +; GFX11-NEXT: v_mov_b16_e32 v3.l, -1 ; GFX11-NEXT: scratch_load_d16_hi_b16 v3, v0, off ; GFX11-NEXT: s_waitcnt vmcnt(0) ; GFX11-NEXT: flat_store_b32 v[1:2], v3 @@ -696,7 +699,7 @@ define amdgpu_ps void @test_scratch_load_i8_zext_to_d16_hi_s(ptr addrspace(5) in ; ; GFX11-LABEL: test_scratch_load_i8_zext_to_d16_hi_s: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_mov_b32_e32 v2, -1 +; GFX11-NEXT: v_mov_b16_e32 v2.l, -1 ; GFX11-NEXT: s_add_i32 s0, s0, 1 ; GFX11-NEXT: scratch_load_d16_hi_u8 v2, off, s0 ; GFX11-NEXT: s_waitcnt vmcnt(0) @@ -735,7 +738,7 @@ define amdgpu_ps void @test_scratch_load_i8_sext_to_d16_hi_s(ptr addrspace(5) in ; ; GFX11-LABEL: test_scratch_load_i8_sext_to_d16_hi_s: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_mov_b32_e32 v2, -1 +; GFX11-NEXT: v_mov_b16_e32 v2.l, -1 ; GFX11-NEXT: s_add_i32 s0, s0, 1 ; GFX11-NEXT: scratch_load_d16_hi_i8 v2, off, s0 ; GFX11-NEXT: s_waitcnt vmcnt(0) @@ -774,7 +777,7 @@ define amdgpu_ps void @test_scratch_load_i16_to_d16_hi_s(ptr addrspace(5) inreg ; ; GFX11-LABEL: test_scratch_load_i16_to_d16_hi_s: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_mov_b32_e32 v2, -1 +; GFX11-NEXT: v_mov_b16_e32 v2.l, -1 ; GFX11-NEXT: s_add_i32 s0, s0, 2 ; GFX11-NEXT: scratch_load_d16_hi_b16 v2, off, s0 ; GFX11-NEXT: s_waitcnt vmcnt(0) @@ -1165,8 +1168,9 @@ define amdgpu_ps void @test_scratch_load_i8_zext_to_d16_hi_svs(ptr addrspace(5) ; ; GFX11-LABEL: test_scratch_load_i8_zext_to_d16_hi_svs: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_dual_mov_b32 v3, -1 :: v_dual_lshlrev_b32 v0, 2, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_lshlrev_b32_e32 v0, 2, v0 +; GFX11-NEXT: v_mov_b16_e32 v3.l, -1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_add3_u32 v0, s0, v0, 1 ; GFX11-NEXT: scratch_load_d16_hi_u8 v3, v0, off ; GFX11-NEXT: s_waitcnt vmcnt(0) @@ -1208,8 +1212,9 @@ define amdgpu_ps void @test_scratch_load_i8_sext_to_d16_hi_svs(ptr addrspace(5) ; ; GFX11-LABEL: test_scratch_load_i8_sext_to_d16_hi_svs: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_dual_mov_b32 v3, -1 :: v_dual_lshlrev_b32 v0, 2, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_lshlrev_b32_e32 v0, 2, v0 +; GFX11-NEXT: v_mov_b16_e32 v3.l, -1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_add3_u32 v0, s0, v0, 1 ; GFX11-NEXT: scratch_load_d16_hi_i8 v3, v0, off ; GFX11-NEXT: s_waitcnt vmcnt(0) @@ -1251,8 +1256,9 @@ define amdgpu_ps void @test_scratch_load_i16_to_d16_hi_svs(ptr addrspace(5) inre ; ; GFX11-LABEL: test_scratch_load_i16_to_d16_hi_svs: ; GFX11: ; %bb.0: ; %bb -; GFX11-NEXT: v_dual_mov_b32 v3, -1 :: v_dual_lshlrev_b32 v0, 2, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_lshlrev_b32_e32 v0, 2, v0 +; GFX11-NEXT: v_mov_b16_e32 v3.l, -1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_add3_u32 v0, s0, v0, 2 ; GFX11-NEXT: scratch_load_d16_hi_b16 v3, v0, off ; GFX11-NEXT: s_waitcnt vmcnt(0) |