@@ -23,9 +23,9 @@ define amdgpu_kernel void @raw_buffer_atomic_add_noret_f64(<4 x i32> %rsrc, doub
2323; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
2424; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
2525; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
26- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
27- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
28- ; GFX90A-NEXT: buffer_atomic_add_f64 v[0:1 ], v2 , s[0:3], 0 offen
26+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
27+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
28+ ; GFX90A-NEXT: buffer_atomic_add_f64 v[2:3 ], v0 , s[0:3], 0 offen
2929; GFX90A-NEXT: s_endpgm
3030;
3131; GFX942-LABEL: raw_buffer_atomic_add_noret_f64:
@@ -34,9 +34,9 @@ define amdgpu_kernel void @raw_buffer_atomic_add_noret_f64(<4 x i32> %rsrc, doub
3434; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
3535; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
3636; GFX942-NEXT: s_waitcnt lgkmcnt(0)
37- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
38- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
39- ; GFX942-NEXT: buffer_atomic_add_f64 v[0:1 ], v2 , s[0:3], 0 offen
37+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
38+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
39+ ; GFX942-NEXT: buffer_atomic_add_f64 v[2:3 ], v0 , s[0:3], 0 offen
4040; GFX942-NEXT: s_endpgm
4141;
4242; GFX1250-LABEL: raw_buffer_atomic_add_noret_f64:
@@ -142,9 +142,9 @@ define amdgpu_kernel void @raw_ptr_buffer_atomic_add_noret_f64(ptr addrspace(8)
142142; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
143143; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
144144; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
145- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
146- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
147- ; GFX90A-NEXT: buffer_atomic_add_f64 v[0:1 ], v2 , s[0:3], 0 offen
145+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
146+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
147+ ; GFX90A-NEXT: buffer_atomic_add_f64 v[2:3 ], v0 , s[0:3], 0 offen
148148; GFX90A-NEXT: s_endpgm
149149;
150150; GFX942-LABEL: raw_ptr_buffer_atomic_add_noret_f64:
@@ -153,9 +153,9 @@ define amdgpu_kernel void @raw_ptr_buffer_atomic_add_noret_f64(ptr addrspace(8)
153153; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
154154; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
155155; GFX942-NEXT: s_waitcnt lgkmcnt(0)
156- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
157- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
158- ; GFX942-NEXT: buffer_atomic_add_f64 v[0:1 ], v2 , s[0:3], 0 offen
156+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
157+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
158+ ; GFX942-NEXT: buffer_atomic_add_f64 v[2:3 ], v0 , s[0:3], 0 offen
159159; GFX942-NEXT: s_endpgm
160160;
161161; GFX1250-LABEL: raw_ptr_buffer_atomic_add_noret_f64:
@@ -261,9 +261,9 @@ define amdgpu_kernel void @struct_buffer_atomic_add_noret_f64(<4 x i32> %rsrc, d
261261; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
262262; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
263263; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
264- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
265- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
266- ; GFX90A-NEXT: buffer_atomic_add_f64 v[0:1 ], v2 , s[0:3], 0 idxen
264+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
265+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
266+ ; GFX90A-NEXT: buffer_atomic_add_f64 v[2:3 ], v0 , s[0:3], 0 idxen
267267; GFX90A-NEXT: s_endpgm
268268;
269269; GFX942-LABEL: struct_buffer_atomic_add_noret_f64:
@@ -272,9 +272,9 @@ define amdgpu_kernel void @struct_buffer_atomic_add_noret_f64(<4 x i32> %rsrc, d
272272; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
273273; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
274274; GFX942-NEXT: s_waitcnt lgkmcnt(0)
275- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
276- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
277- ; GFX942-NEXT: buffer_atomic_add_f64 v[0:1 ], v2 , s[0:3], 0 idxen
275+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
276+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
277+ ; GFX942-NEXT: buffer_atomic_add_f64 v[2:3 ], v0 , s[0:3], 0 idxen
278278; GFX942-NEXT: s_endpgm
279279;
280280; GFX1250-LABEL: struct_buffer_atomic_add_noret_f64:
@@ -379,9 +379,9 @@ define amdgpu_kernel void @struct_ptr_buffer_atomic_add_noret_f64(ptr addrspace(
379379; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
380380; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
381381; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
382- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
383- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
384- ; GFX90A-NEXT: buffer_atomic_add_f64 v[0:1 ], v2 , s[0:3], 0 idxen
382+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
383+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
384+ ; GFX90A-NEXT: buffer_atomic_add_f64 v[2:3 ], v0 , s[0:3], 0 idxen
385385; GFX90A-NEXT: s_endpgm
386386;
387387; GFX942-LABEL: struct_ptr_buffer_atomic_add_noret_f64:
@@ -390,9 +390,9 @@ define amdgpu_kernel void @struct_ptr_buffer_atomic_add_noret_f64(ptr addrspace(
390390; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
391391; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
392392; GFX942-NEXT: s_waitcnt lgkmcnt(0)
393- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
394- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
395- ; GFX942-NEXT: buffer_atomic_add_f64 v[0:1 ], v2 , s[0:3], 0 idxen
393+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
394+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
395+ ; GFX942-NEXT: buffer_atomic_add_f64 v[2:3 ], v0 , s[0:3], 0 idxen
396396; GFX942-NEXT: s_endpgm
397397;
398398; GFX1250-LABEL: struct_ptr_buffer_atomic_add_noret_f64:
@@ -497,9 +497,9 @@ define amdgpu_kernel void @raw_buffer_atomic_min_noret_f64(<4 x i32> %rsrc, doub
497497; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
498498; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
499499; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
500- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
501- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
502- ; GFX90A-NEXT: buffer_atomic_min_f64 v[0:1 ], v2 , s[0:3], 0 offen
500+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
501+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
502+ ; GFX90A-NEXT: buffer_atomic_min_f64 v[2:3 ], v0 , s[0:3], 0 offen
503503; GFX90A-NEXT: s_endpgm
504504;
505505; GFX942-LABEL: raw_buffer_atomic_min_noret_f64:
@@ -508,9 +508,9 @@ define amdgpu_kernel void @raw_buffer_atomic_min_noret_f64(<4 x i32> %rsrc, doub
508508; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
509509; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
510510; GFX942-NEXT: s_waitcnt lgkmcnt(0)
511- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
512- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
513- ; GFX942-NEXT: buffer_atomic_min_f64 v[0:1 ], v2 , s[0:3], 0 offen
511+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
512+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
513+ ; GFX942-NEXT: buffer_atomic_min_f64 v[2:3 ], v0 , s[0:3], 0 offen
514514; GFX942-NEXT: s_endpgm
515515;
516516; GFX1250-LABEL: raw_buffer_atomic_min_noret_f64:
@@ -616,9 +616,9 @@ define amdgpu_kernel void @raw_ptr_buffer_atomic_min_noret_f64(ptr addrspace(8)
616616; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
617617; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
618618; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
619- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
620- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
621- ; GFX90A-NEXT: buffer_atomic_min_f64 v[0:1 ], v2 , s[0:3], 0 offen
619+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
620+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
621+ ; GFX90A-NEXT: buffer_atomic_min_f64 v[2:3 ], v0 , s[0:3], 0 offen
622622; GFX90A-NEXT: s_endpgm
623623;
624624; GFX942-LABEL: raw_ptr_buffer_atomic_min_noret_f64:
@@ -627,9 +627,9 @@ define amdgpu_kernel void @raw_ptr_buffer_atomic_min_noret_f64(ptr addrspace(8)
627627; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
628628; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
629629; GFX942-NEXT: s_waitcnt lgkmcnt(0)
630- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
631- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
632- ; GFX942-NEXT: buffer_atomic_min_f64 v[0:1 ], v2 , s[0:3], 0 offen
630+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
631+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
632+ ; GFX942-NEXT: buffer_atomic_min_f64 v[2:3 ], v0 , s[0:3], 0 offen
633633; GFX942-NEXT: s_endpgm
634634;
635635; GFX1250-LABEL: raw_ptr_buffer_atomic_min_noret_f64:
@@ -735,9 +735,9 @@ define amdgpu_kernel void @struct_buffer_atomic_min_noret_f64(<4 x i32> %rsrc, d
735735; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
736736; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
737737; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
738- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
739- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
740- ; GFX90A-NEXT: buffer_atomic_min_f64 v[0:1 ], v2 , s[0:3], 0 idxen
738+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
739+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
740+ ; GFX90A-NEXT: buffer_atomic_min_f64 v[2:3 ], v0 , s[0:3], 0 idxen
741741; GFX90A-NEXT: s_endpgm
742742;
743743; GFX942-LABEL: struct_buffer_atomic_min_noret_f64:
@@ -746,9 +746,9 @@ define amdgpu_kernel void @struct_buffer_atomic_min_noret_f64(<4 x i32> %rsrc, d
746746; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
747747; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
748748; GFX942-NEXT: s_waitcnt lgkmcnt(0)
749- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
750- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
751- ; GFX942-NEXT: buffer_atomic_min_f64 v[0:1 ], v2 , s[0:3], 0 idxen
749+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
750+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
751+ ; GFX942-NEXT: buffer_atomic_min_f64 v[2:3 ], v0 , s[0:3], 0 idxen
752752; GFX942-NEXT: s_endpgm
753753;
754754; GFX1250-LABEL: struct_buffer_atomic_min_noret_f64:
@@ -853,9 +853,9 @@ define amdgpu_kernel void @struct_ptr_buffer_atomic_min_noret_f64(ptr addrspace(
853853; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
854854; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
855855; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
856- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
857- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
858- ; GFX90A-NEXT: buffer_atomic_min_f64 v[0:1 ], v2 , s[0:3], 0 idxen
856+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
857+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
858+ ; GFX90A-NEXT: buffer_atomic_min_f64 v[2:3 ], v0 , s[0:3], 0 idxen
859859; GFX90A-NEXT: s_endpgm
860860;
861861; GFX942-LABEL: struct_ptr_buffer_atomic_min_noret_f64:
@@ -864,9 +864,9 @@ define amdgpu_kernel void @struct_ptr_buffer_atomic_min_noret_f64(ptr addrspace(
864864; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
865865; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
866866; GFX942-NEXT: s_waitcnt lgkmcnt(0)
867- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
868- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
869- ; GFX942-NEXT: buffer_atomic_min_f64 v[0:1 ], v2 , s[0:3], 0 idxen
867+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
868+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
869+ ; GFX942-NEXT: buffer_atomic_min_f64 v[2:3 ], v0 , s[0:3], 0 idxen
870870; GFX942-NEXT: s_endpgm
871871;
872872; GFX1250-LABEL: struct_ptr_buffer_atomic_min_noret_f64:
@@ -971,9 +971,9 @@ define amdgpu_kernel void @raw_buffer_atomic_max_noret_f64(<4 x i32> %rsrc, doub
971971; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
972972; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
973973; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
974- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
975- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
976- ; GFX90A-NEXT: buffer_atomic_max_f64 v[0:1 ], v2 , s[0:3], 0 offen
974+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
975+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
976+ ; GFX90A-NEXT: buffer_atomic_max_f64 v[2:3 ], v0 , s[0:3], 0 offen
977977; GFX90A-NEXT: s_endpgm
978978;
979979; GFX942-LABEL: raw_buffer_atomic_max_noret_f64:
@@ -982,9 +982,9 @@ define amdgpu_kernel void @raw_buffer_atomic_max_noret_f64(<4 x i32> %rsrc, doub
982982; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
983983; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
984984; GFX942-NEXT: s_waitcnt lgkmcnt(0)
985- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
986- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
987- ; GFX942-NEXT: buffer_atomic_max_f64 v[0:1 ], v2 , s[0:3], 0 offen
985+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
986+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
987+ ; GFX942-NEXT: buffer_atomic_max_f64 v[2:3 ], v0 , s[0:3], 0 offen
988988; GFX942-NEXT: s_endpgm
989989;
990990; GFX1250-LABEL: raw_buffer_atomic_max_noret_f64:
@@ -1090,9 +1090,9 @@ define amdgpu_kernel void @raw_ptr_buffer_atomic_max_noret_f64(ptr addrspace(8)
10901090; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
10911091; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
10921092; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
1093- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
1094- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
1095- ; GFX90A-NEXT: buffer_atomic_max_f64 v[0:1 ], v2 , s[0:3], 0 offen
1093+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
1094+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
1095+ ; GFX90A-NEXT: buffer_atomic_max_f64 v[2:3 ], v0 , s[0:3], 0 offen
10961096; GFX90A-NEXT: s_endpgm
10971097;
10981098; GFX942-LABEL: raw_ptr_buffer_atomic_max_noret_f64:
@@ -1101,9 +1101,9 @@ define amdgpu_kernel void @raw_ptr_buffer_atomic_max_noret_f64(ptr addrspace(8)
11011101; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
11021102; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
11031103; GFX942-NEXT: s_waitcnt lgkmcnt(0)
1104- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
1105- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
1106- ; GFX942-NEXT: buffer_atomic_max_f64 v[0:1 ], v2 , s[0:3], 0 offen
1104+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
1105+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
1106+ ; GFX942-NEXT: buffer_atomic_max_f64 v[2:3 ], v0 , s[0:3], 0 offen
11071107; GFX942-NEXT: s_endpgm
11081108;
11091109; GFX1250-LABEL: raw_ptr_buffer_atomic_max_noret_f64:
@@ -1209,9 +1209,9 @@ define amdgpu_kernel void @struct_buffer_atomic_max_noret_f64(<4 x i32> %rsrc, d
12091209; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
12101210; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
12111211; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
1212- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
1213- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
1214- ; GFX90A-NEXT: buffer_atomic_max_f64 v[0:1 ], v2 , s[0:3], 0 idxen
1212+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
1213+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
1214+ ; GFX90A-NEXT: buffer_atomic_max_f64 v[2:3 ], v0 , s[0:3], 0 idxen
12151215; GFX90A-NEXT: s_endpgm
12161216;
12171217; GFX942-LABEL: struct_buffer_atomic_max_noret_f64:
@@ -1220,9 +1220,9 @@ define amdgpu_kernel void @struct_buffer_atomic_max_noret_f64(<4 x i32> %rsrc, d
12201220; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
12211221; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
12221222; GFX942-NEXT: s_waitcnt lgkmcnt(0)
1223- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
1224- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
1225- ; GFX942-NEXT: buffer_atomic_max_f64 v[0:1 ], v2 , s[0:3], 0 idxen
1223+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
1224+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
1225+ ; GFX942-NEXT: buffer_atomic_max_f64 v[2:3 ], v0 , s[0:3], 0 idxen
12261226; GFX942-NEXT: s_endpgm
12271227;
12281228; GFX1250-LABEL: struct_buffer_atomic_max_noret_f64:
@@ -1327,9 +1327,9 @@ define amdgpu_kernel void @struct_ptr_buffer_atomic_max_noret_f64(ptr addrspace(
13271327; GFX90A-NEXT: s_load_dword s8, s[4:5], 0x3c
13281328; GFX90A-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
13291329; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
1330- ; GFX90A-NEXT: v_pk_mov_b32 v[0:1 ], s[6:7], s[6:7] op_sel:[0,1]
1331- ; GFX90A-NEXT: v_mov_b32_e32 v2 , s8
1332- ; GFX90A-NEXT: buffer_atomic_max_f64 v[0:1 ], v2 , s[0:3], 0 idxen
1330+ ; GFX90A-NEXT: v_pk_mov_b32 v[2:3 ], s[6:7], s[6:7] op_sel:[0,1]
1331+ ; GFX90A-NEXT: v_mov_b32_e32 v0 , s8
1332+ ; GFX90A-NEXT: buffer_atomic_max_f64 v[2:3 ], v0 , s[0:3], 0 idxen
13331333; GFX90A-NEXT: s_endpgm
13341334;
13351335; GFX942-LABEL: struct_ptr_buffer_atomic_max_noret_f64:
@@ -1338,9 +1338,9 @@ define amdgpu_kernel void @struct_ptr_buffer_atomic_max_noret_f64(ptr addrspace(
13381338; GFX942-NEXT: s_load_dword s8, s[4:5], 0x3c
13391339; GFX942-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
13401340; GFX942-NEXT: s_waitcnt lgkmcnt(0)
1341- ; GFX942-NEXT: v_mov_b64_e32 v[0:1 ], s[6:7]
1342- ; GFX942-NEXT: v_mov_b32_e32 v2 , s8
1343- ; GFX942-NEXT: buffer_atomic_max_f64 v[0:1 ], v2 , s[0:3], 0 idxen
1341+ ; GFX942-NEXT: v_mov_b64_e32 v[2:3 ], s[6:7]
1342+ ; GFX942-NEXT: v_mov_b32_e32 v0 , s8
1343+ ; GFX942-NEXT: buffer_atomic_max_f64 v[2:3 ], v0 , s[0:3], 0 idxen
13441344; GFX942-NEXT: s_endpgm
13451345;
13461346; GFX1250-LABEL: struct_ptr_buffer_atomic_max_noret_f64:
0 commit comments