@@ -357,30 +357,29 @@ define amdgpu_kernel void @infinite_loop_nest_ret_callbr(ptr addrspace(1) %out)
357357; SI-NEXT: ;;#ASMSTART
358358; SI-NEXT: ;;#ASMEND
359359; SI-NEXT: ; %bb.1: ; %outer_loop.preheader
360- ; SI-NEXT: s_load_dwordx2 s[0:1 ], s[4:5], 0x9
361- ; SI-NEXT: s_mov_b32 s3 , 0xf000
362- ; SI-NEXT: s_mov_b32 s2 , -1
360+ ; SI-NEXT: s_load_dwordx2 s[4:5 ], s[4:5], 0x9
361+ ; SI-NEXT: s_mov_b32 s7 , 0xf000
362+ ; SI-NEXT: s_mov_b32 s6 , -1
363363; SI-NEXT: v_mov_b32_e32 v0, 0x3e7
364- ; SI-NEXT: ; implicit-def: $sgpr4_sgpr5
364+ ; SI-NEXT: s_and_b64 s[0:1], exec, 0
365365; SI-NEXT: s_branch .LBB7_3
366366; SI-NEXT: .LBB7_2: ; %loop.exit.guard
367367; SI-NEXT: ; in Loop: Header=BB7_3 Depth=1
368- ; SI-NEXT: s_and_b64 vcc, exec, s[4:5]
369- ; SI-NEXT: s_mov_b64 s[4:5], 0
368+ ; SI-NEXT: s_and_b64 vcc, exec, s[2:3]
370369; SI-NEXT: s_cbranch_vccnz .LBB7_5
371370; SI-NEXT: .LBB7_3: ; %outer_loop
372371; SI-NEXT: ; =>This Inner Loop Header: Depth=1
373372; SI-NEXT: ;;#ASMSTART
374373; SI-NEXT: ;;#ASMEND
375374; SI-NEXT: s_waitcnt lgkmcnt(0)
376- ; SI-NEXT: buffer_store_dword v0, off, s[0:3 ], 0
375+ ; SI-NEXT: buffer_store_dword v0, off, s[4:7 ], 0
377376; SI-NEXT: s_waitcnt vmcnt(0)
378- ; SI-NEXT: s_and_b64 vcc, exec, s[4:5]
379- ; SI-NEXT: s_mov_b64 s[4:5], -1
377+ ; SI-NEXT: s_mov_b64 s[2:3], -1
378+ ; SI-NEXT: s_mov_b64 vcc, s[0:1]
380379; SI-NEXT: s_cbranch_vccz .LBB7_2
381380; SI-NEXT: ; %bb.4: ; %TransitionBlock.target.outer_loop
382381; SI-NEXT: ; in Loop: Header=BB7_3 Depth=1
383- ; SI-NEXT: s_mov_b64 s[4:5 ], 0
382+ ; SI-NEXT: s_mov_b64 s[2:3 ], 0
384383; SI-NEXT: s_branch .LBB7_2
385384; SI-NEXT: .LBB7_5: ; Inline asm indirect target
386385; SI-NEXT: ; %UnifiedReturnBlock
0 commit comments