Lines Matching refs:GFX900
3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck --check-prefix=GFX900 %s
11 ; GFX900-LABEL: test_kern_empty:
12 ; GFX900: ; %bb.0: ; %entry
13 ; GFX900-NEXT: s_endpgm
34 ; GFX900-LABEL: test_kern_stack:
35 ; GFX900: ; %bb.0: ; %entry
36 ; GFX900-NEXT: s_add_u32 flat_scratch_lo, s4, s7
37 ; GFX900-NEXT: s_addc_u32 flat_scratch_hi, s5, 0
38 ; GFX900-NEXT: s_add_u32 s0, s0, s7
39 ; GFX900-NEXT: s_addc_u32 s1, s1, 0
40 ; GFX900-NEXT: v_mov_b32_e32 v0, 0
41 ; GFX900-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:4
42 ; GFX900-NEXT: s_endpgm
76 ; GFX900-LABEL: test_kern_call:
77 ; GFX900: ; %bb.0: ; %entry
78 ; GFX900-NEXT: s_add_u32 flat_scratch_lo, s4, s7
79 ; GFX900-NEXT: s_addc_u32 flat_scratch_hi, s5, 0
80 ; GFX900-NEXT: s_add_u32 s0, s0, s7
81 ; GFX900-NEXT: s_addc_u32 s1, s1, 0
82 ; GFX900-NEXT: s_getpc_b64 s[4:5]
83 ; GFX900-NEXT: s_add_u32 s4, s4, ex@rel32@lo+4
84 ; GFX900-NEXT: s_addc_u32 s5, s5, ex@rel32@hi+12
85 ; GFX900-NEXT: s_mov_b32 s32, 0
86 ; GFX900-NEXT: s_swappc_b64 s[30:31], s[4:5]
87 ; GFX900-NEXT: s_endpgm
125 ; GFX900-LABEL: test_kern_stack_and_call:
126 ; GFX900: ; %bb.0: ; %entry
127 ; GFX900-NEXT: s_add_u32 flat_scratch_lo, s4, s7
128 ; GFX900-NEXT: s_addc_u32 flat_scratch_hi, s5, 0
129 ; GFX900-NEXT: s_add_u32 s0, s0, s7
130 ; GFX900-NEXT: s_addc_u32 s1, s1, 0
131 ; GFX900-NEXT: v_mov_b32_e32 v0, 0
132 ; GFX900-NEXT: s_getpc_b64 s[4:5]
133 ; GFX900-NEXT: s_add_u32 s4, s4, ex@rel32@lo+4
134 ; GFX900-NEXT: s_addc_u32 s5, s5, ex@rel32@hi+12
135 ; GFX900-NEXT: s_movk_i32 s32, 0x400
136 ; GFX900-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:4
137 ; GFX900-NEXT: s_swappc_b64 s[30:31], s[4:5]
138 ; GFX900-NEXT: s_endpgm
169 ; GFX900-LABEL: test_force_fp_kern_empty:
170 ; GFX900: ; %bb.0: ; %entry
171 ; GFX900-NEXT: s_mov_b32 s33, 0
172 ; GFX900-NEXT: s_endpgm
195 ; GFX900-LABEL: test_force_fp_kern_stack:
196 ; GFX900: ; %bb.0: ; %entry
197 ; GFX900-NEXT: s_add_u32 flat_scratch_lo, s4, s7
198 ; GFX900-NEXT: s_addc_u32 flat_scratch_hi, s5, 0
199 ; GFX900-NEXT: s_add_u32 s0, s0, s7
200 ; GFX900-NEXT: s_mov_b32 s33, 0
201 ; GFX900-NEXT: s_addc_u32 s1, s1, 0
202 ; GFX900-NEXT: v_mov_b32_e32 v0, 0
203 ; GFX900-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:4
204 ; GFX900-NEXT: s_endpgm
240 ; GFX900-LABEL: test_force_fp_kern_call:
241 ; GFX900: ; %bb.0: ; %entry
242 ; GFX900-NEXT: s_add_u32 flat_scratch_lo, s4, s7
243 ; GFX900-NEXT: s_addc_u32 flat_scratch_hi, s5, 0
244 ; GFX900-NEXT: s_add_u32 s0, s0, s7
245 ; GFX900-NEXT: s_addc_u32 s1, s1, 0
246 ; GFX900-NEXT: s_getpc_b64 s[4:5]
247 ; GFX900-NEXT: s_add_u32 s4, s4, ex@rel32@lo+4
248 ; GFX900-NEXT: s_addc_u32 s5, s5, ex@rel32@hi+12
249 ; GFX900-NEXT: s_mov_b32 s32, 0
250 ; GFX900-NEXT: s_mov_b32 s33, 0
251 ; GFX900-NEXT: s_swappc_b64 s[30:31], s[4:5]
252 ; GFX900-NEXT: s_endpgm
292 ; GFX900-LABEL: test_force_fp_kern_stack_and_call:
293 ; GFX900: ; %bb.0: ; %entry
294 ; GFX900-NEXT: s_add_u32 flat_scratch_lo, s4, s7
295 ; GFX900-NEXT: s_addc_u32 flat_scratch_hi, s5, 0
296 ; GFX900-NEXT: s_add_u32 s0, s0, s7
297 ; GFX900-NEXT: s_addc_u32 s1, s1, 0
298 ; GFX900-NEXT: s_mov_b32 s33, 0
299 ; GFX900-NEXT: v_mov_b32_e32 v0, 0
300 ; GFX900-NEXT: s_getpc_b64 s[4:5]
301 ; GFX900-NEXT: s_add_u32 s4, s4, ex@rel32@lo+4
302 ; GFX900-NEXT: s_addc_u32 s5, s5, ex@rel32@hi+12
303 ; GFX900-NEXT: s_movk_i32 s32, 0x400
304 ; GFX900-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:4
305 ; GFX900-NEXT: s_swappc_b64 s[30:31], s[4:5]
306 ; GFX900-NEXT: s_endpgm
352 ; GFX900-LABEL: test_sgpr_offset_kernel:
353 ; GFX900: ; %bb.0: ; %entry
354 ; GFX900-NEXT: s_add_u32 flat_scratch_lo, s4, s7
355 ; GFX900-NEXT: s_addc_u32 flat_scratch_hi, s5, 0
356 ; GFX900-NEXT: s_add_u32 s0, s0, s7
357 ; GFX900-NEXT: s_addc_u32 s1, s1, 0
358 ; GFX900-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:8
359 ; GFX900-NEXT: s_mov_b32 s6, 0x40000
360 ; GFX900-NEXT: s_waitcnt vmcnt(0)
361 ; GFX900-NEXT: buffer_store_dword v0, off, s[0:3], s6 ; 4-byte Folded Spill
362 ; GFX900-NEXT: ;;#ASMSTART
363 ; GFX900-NEXT: ;;#ASMEND
364 ; GFX900-NEXT: s_mov_b32 s6, 0x40000
365 ; GFX900-NEXT: buffer_load_dword v0, off, s[0:3], s6 ; 4-byte Folded Reload
366 ; GFX900-NEXT: s_waitcnt vmcnt(0)
367 ; GFX900-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:8
368 ; GFX900-NEXT: s_endpgm