@@ -19,22 +19,24 @@ define amdgpu_kernel void @foo(ptr addrspace(5) %ptr5, ptr %p0, double %v0, <4 x
1919; CHECK-NEXT: s_mov_b64 s[4:5], src_private_base
2020; CHECK-NEXT: s_waitcnt lgkmcnt(0)
2121; CHECK-NEXT: s_cmp_lg_u32 s68, -1
22- ; CHECK-NEXT: s_cselect_b32 s4, s5, 0
23- ; CHECK-NEXT: s_cselect_b32 s5, s68, 0
22+ ; CHECK-NEXT: s_mov_b64 s[38:39], s[6:7]
23+ ; CHECK-NEXT: s_mov_b32 s4, 0
24+ ; CHECK-NEXT: s_cselect_b32 s5, s5, 0
25+ ; CHECK-NEXT: s_cselect_b32 s6, s68, 0
2426; CHECK-NEXT: s_add_u32 s50, s34, 48
27+ ; CHECK-NEXT: v_mov_b32_e32 v47, s5
28+ ; CHECK-NEXT: s_mov_b32 s5, s4
2529; CHECK-NEXT: s_addc_u32 s51, s35, 0
26- ; CHECK-NEXT: v_mov_b32_e32 v46, s5
27- ; CHECK-NEXT: v_mov_b32_e32 v47, s4
30+ ; CHECK-NEXT: v_pk_mov_b32 v[62:63], s[4:5], s[4:5] op_sel:[0,1]
2831; CHECK-NEXT: s_getpc_b64 s[4:5]
2932; CHECK-NEXT: s_add_u32 s4, s4, G@gotpcrel32@lo+4
3033; CHECK-NEXT: s_addc_u32 s5, s5, G@gotpcrel32@hi+12
31- ; CHECK-NEXT: v_pk_mov_b32 v[56:57], s[64:65], s[64:65] op_sel:[0,1]
32- ; CHECK-NEXT: s_load_dwordx2 s[64:65], s[4:5], 0x0
33- ; CHECK-NEXT: s_mov_b32 s54, 0
34- ; CHECK-NEXT: s_mov_b32 s55, s54
34+ ; CHECK-NEXT: s_load_dwordx2 s[54:55], s[4:5], 0x0
3535; CHECK-NEXT: s_mov_b32 s53, s14
36- ; CHECK-NEXT: v_pk_mov_b32 v[62:63], s[54:55], s[54:55] op_sel:[0,1]
36+ ; CHECK-NEXT: v_mov_b32_e32 v46, s6
37+ ; CHECK-NEXT: v_pk_mov_b32 v[56:57], s[64:65], s[64:65] op_sel:[0,1]
3738; CHECK-NEXT: s_mov_b64 s[4:5], s[48:49]
39+ ; CHECK-NEXT: s_mov_b64 s[6:7], s[38:39]
3840; CHECK-NEXT: s_mov_b64 s[8:9], s[50:51]
3941; CHECK-NEXT: s_mov_b32 s12, s14
4042; CHECK-NEXT: s_mov_b32 s13, s15
@@ -44,14 +46,13 @@ define amdgpu_kernel void @foo(ptr addrspace(5) %ptr5, ptr %p0, double %v0, <4 x
4446; CHECK-NEXT: s_mov_b32 s33, s16
4547; CHECK-NEXT: s_mov_b32 s52, s15
4648; CHECK-NEXT: s_mov_b64 s[36:37], s[10:11]
47- ; CHECK-NEXT: s_mov_b64 s[38:39], s[6:7]
4849; CHECK-NEXT: v_mov_b32_e32 v40, v0
4950; CHECK-NEXT: v_mov_b32_e32 v60, s66
5051; CHECK-NEXT: v_mov_b32_e32 v61, s67
5152; CHECK-NEXT: flat_store_dwordx2 v[56:57], v[62:63]
5253; CHECK-NEXT: ; kill: def $sgpr15 killed $sgpr15
5354; CHECK-NEXT: s_waitcnt lgkmcnt(0)
54- ; CHECK-NEXT: s_swappc_b64 s[30:31], s[64:65 ]
55+ ; CHECK-NEXT: s_swappc_b64 s[30:31], s[54:55 ]
5556; CHECK-NEXT: flat_load_dwordx2 v[58:59], v[56:57]
5657; CHECK-NEXT: v_mov_b32_e32 v0, 0
5758; CHECK-NEXT: v_mov_b32_e32 v1, 0x3ff00000
@@ -67,7 +68,7 @@ define amdgpu_kernel void @foo(ptr addrspace(5) %ptr5, ptr %p0, double %v0, <4 x
6768; CHECK-NEXT: flat_store_dwordx2 v[56:57], v[62:63]
6869; CHECK-NEXT: s_waitcnt vmcnt(0)
6970; CHECK-NEXT: ; kill: def $sgpr15 killed $sgpr15
70- ; CHECK-NEXT: s_swappc_b64 s[30:31], s[64:65 ]
71+ ; CHECK-NEXT: s_swappc_b64 s[30:31], s[54:55 ]
7172; CHECK-NEXT: flat_load_dwordx2 v[0:1], v[46:47] glc
7273; CHECK-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
7374; CHECK-NEXT: v_mov_b32_e32 v1, s67
@@ -76,7 +77,7 @@ define amdgpu_kernel void @foo(ptr addrspace(5) %ptr5, ptr %p0, double %v0, <4 x
7677; CHECK-NEXT: s_waitcnt vmcnt(0)
7778; CHECK-NEXT: flat_store_dwordx2 v[56:57], v[60:61]
7879; CHECK-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
79- ; CHECK-NEXT: v_mov_b32_e32 v1, s54
80+ ; CHECK-NEXT: v_mov_b32_e32 v1, 0
8081; CHECK-NEXT: v_cmp_lt_i32_e32 vcc, 0, v42
8182; CHECK-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
8283; CHECK-NEXT: ; implicit-def: $vgpr4
0 commit comments