1/*
2 * Copyright 2016 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23/* To compile this assembly code:
24 * PROJECT=greenland ./sp3 cwsr_trap_handler_gfx9.asm -hex tmp.hex
25 */
26
27/* HW (GFX9) source code for CWSR trap handler */
28/* Version 18 + multiple trap handler */
29
30// this performance-optimal version was originally from Seven Xu at SRDC
31
32// Revison #18	 --...
33/* Rev History
34** #1. Branch from gc dv.   //gfxip/gfx9/main/src/test/suites/block/cs/sr/cs_trap_handler.sp3#1,#50, #51, #52-53(Skip, Already Fixed by PV), #54-56(merged),#57-58(mergerd, skiped-already fixed by PV)
35** #4. SR Memory Layout:
36**			 1. VGPR-SGPR-HWREG-{LDS}
37**			 2. tba_hi.bits.26 - reconfigured as the first wave in tg bits, for defer Save LDS for a threadgroup.. performance concern..
38** #5. Update: 1. Accurate g8sr_ts_save_d timestamp
39** #6. Update: 1. Fix s_barrier usage; 2. VGPR s/r using swizzle buffer?(NoNeed, already matched the swizzle pattern, more investigation)
40** #7. Update: 1. don't barrier if noLDS
41** #8. Branch: 1. Branch to ver#0, which is very similar to gc dv version
42**	       2. Fix SQ issue by s_sleep 2
43** #9. Update: 1. Fix scc restore failed issue, restore wave_status at last
44**	       2. optimize s_buffer save by burst 16sgprs...
45** #10. Update 1. Optimize restore sgpr by busrt 16 sgprs.
46** #11. Update 1. Add 2 more timestamp for debug version
47** #12. Update 1. Add VGPR SR using DWx4, some case improve and some case drop performance
48** #13. Integ  1. Always use MUBUF for PV trap shader...
49** #14. Update 1. s_buffer_store soft clause...
50** #15. Update 1. PERF - sclar write with glc:0/mtype0 to allow L2 combine. perf improvement a lot.
51** #16. Update 1. PRRF - UNROLL LDS_DMA got 2500cycle save in IP tree
52** #17. Update 1. FUNC - LDS_DMA has issues while ATC, replace with ds_read/buffer_store for save part[TODO restore part]
53**	       2. PERF - Save LDS before save VGPR to cover LDS save long latency...
54** #18. Update 1. FUNC - Implicitly estore STATUS.VCCZ, which is not writable by s_setreg_b32
55**	       2. FUNC - Handle non-CWSR traps
56*/
57
58var G8SR_WDMEM_HWREG_OFFSET = 0
59var G8SR_WDMEM_SGPR_OFFSET  = 128  // in bytes
60
61// Keep definition same as the app shader, These 2 time stamps are part of the app shader... Should before any Save and after restore.
62
63var G8SR_DEBUG_TIMESTAMP = 0
64var G8SR_DEBUG_TS_SAVE_D_OFFSET = 40*4	// ts_save_d timestamp offset relative to SGPR_SR_memory_offset
65var s_g8sr_ts_save_s	= s[34:35]   // save start
66var s_g8sr_ts_sq_save_msg  = s[36:37]	// The save shader send SAVEWAVE msg to spi
67var s_g8sr_ts_spi_wrexec   = s[38:39]	// the SPI write the sr address to SQ
68var s_g8sr_ts_save_d	= s[40:41]   // save end
69var s_g8sr_ts_restore_s = s[42:43]   // restore start
70var s_g8sr_ts_restore_d = s[44:45]   // restore end
71
72var G8SR_VGPR_SR_IN_DWX4 = 0
73var G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4 = 0x00100000	 // DWx4 stride is 4*4Bytes
74var G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4  = G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4
75
76
77/*************************************************************************/
78/*		    control on how to run the shader			 */
79/*************************************************************************/
80//any hack that needs to be made to run this code in EMU (either because various EMU code are not ready or no compute save & restore in EMU run)
81var EMU_RUN_HACK		    =	0
82var EMU_RUN_HACK_RESTORE_NORMAL	    =	0
83var EMU_RUN_HACK_SAVE_NORMAL_EXIT   =	0
84var EMU_RUN_HACK_SAVE_SINGLE_WAVE   =	0
85var EMU_RUN_HACK_SAVE_FIRST_TIME    =	0		    //for interrupted restore in which the first save is through EMU_RUN_HACK
86var SAVE_LDS			    =	1
87var WG_BASE_ADDR_LO		    =	0x9000a000
88var WG_BASE_ADDR_HI		    =	0x0
89var WAVE_SPACE			    =	0x5000		    //memory size that each wave occupies in workgroup state mem
90var CTX_SAVE_CONTROL		    =	0x0
91var CTX_RESTORE_CONTROL		    =	CTX_SAVE_CONTROL
92var SIM_RUN_HACK		    =	0		    //any hack that needs to be made to run this code in SIM (either because various RTL code are not ready or no compute save & restore in RTL run)
93var SGPR_SAVE_USE_SQC		    =	1		    //use SQC D$ to do the write
94var USE_MTBUF_INSTEAD_OF_MUBUF	    =	0		    //because TC EMU currently asserts on 0 of // overload DFMT field to carry 4 more bits of stride for MUBUF opcodes
95var SWIZZLE_EN			    =	0		    //whether we use swizzled buffer addressing
96var ACK_SQC_STORE		    =	1		    //workaround for suspected SQC store bug causing incorrect stores under concurrency
97
98/**************************************************************************/
99/*			variables					  */
100/**************************************************************************/
101var SQ_WAVE_STATUS_INST_ATC_SHIFT  = 23
102var SQ_WAVE_STATUS_INST_ATC_MASK   = 0x00800000
103var SQ_WAVE_STATUS_SPI_PRIO_SHIFT  = 1
104var SQ_WAVE_STATUS_SPI_PRIO_MASK   = 0x00000006
105var SQ_WAVE_STATUS_HALT_MASK       = 0x2000
106
107var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT	= 12
108var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE	= 9
109var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT	= 8
110var SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE	= 6
111var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT	= 24
112var SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE	= 3			//FIXME	 sq.blk still has 4 bits at this time while SQ programming guide has 3 bits
113
114var SQ_WAVE_TRAPSTS_SAVECTX_MASK    =	0x400
115var SQ_WAVE_TRAPSTS_EXCE_MASK	    =	0x1FF			// Exception mask
116var SQ_WAVE_TRAPSTS_SAVECTX_SHIFT   =	10
117var SQ_WAVE_TRAPSTS_MEM_VIOL_MASK   =	0x100
118var SQ_WAVE_TRAPSTS_MEM_VIOL_SHIFT  =	8
119var SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK	=   0x3FF
120var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT	=   0x0
121var SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE	=   10
122var SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK	=   0xFFFFF800
123var SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT	=   11
124var SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE	=   21
125var SQ_WAVE_TRAPSTS_ILLEGAL_INST_MASK	=   0x800
126
127var SQ_WAVE_IB_STS_RCNT_SHIFT		=   16			//FIXME
128var SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT	=   15			//FIXME
129var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK	= 0x1F8000
130var SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG	= 0x00007FFF	//FIXME
131
132var SQ_BUF_RSRC_WORD1_ATC_SHIFT	    =	24
133var SQ_BUF_RSRC_WORD3_MTYPE_SHIFT   =	27
134
135var TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT	=   26			// bits [31:26] unused by SPI debug data
136var TTMP11_SAVE_RCNT_FIRST_REPLAY_MASK	=   0xFC000000
137
138/*	Save	    */
139var S_SAVE_BUF_RSRC_WORD1_STRIDE	=   0x00040000		//stride is 4 bytes
140var S_SAVE_BUF_RSRC_WORD3_MISC		=   0x00807FAC		//SQ_SEL_X/Y/Z/W, BUF_NUM_FORMAT_FLOAT, (0 for MUBUF stride[17:14] when ADD_TID_ENABLE and BUF_DATA_FORMAT_32 for MTBUF), ADD_TID_ENABLE
141
142var S_SAVE_SPI_INIT_ATC_MASK		=   0x08000000		//bit[27]: ATC bit
143var S_SAVE_SPI_INIT_ATC_SHIFT		=   27
144var S_SAVE_SPI_INIT_MTYPE_MASK		=   0x70000000		//bit[30:28]: Mtype
145var S_SAVE_SPI_INIT_MTYPE_SHIFT		=   28
146var S_SAVE_SPI_INIT_FIRST_WAVE_MASK	=   0x04000000		//bit[26]: FirstWaveInTG
147var S_SAVE_SPI_INIT_FIRST_WAVE_SHIFT	=   26
148
149var S_SAVE_PC_HI_RCNT_SHIFT		=   28			//FIXME	 check with Brian to ensure all fields other than PC[47:0] can be used
150var S_SAVE_PC_HI_RCNT_MASK		=   0xF0000000		//FIXME
151var S_SAVE_PC_HI_FIRST_REPLAY_SHIFT	=   27			//FIXME
152var S_SAVE_PC_HI_FIRST_REPLAY_MASK	=   0x08000000		//FIXME
153
154var s_save_spi_init_lo		    =	exec_lo
155var s_save_spi_init_hi		    =	exec_hi
156
157var s_save_pc_lo	    =	ttmp0		//{TTMP1, TTMP0} = {3'h0,pc_rewind[3:0], HT[0],trapID[7:0], PC[47:0]}
158var s_save_pc_hi	    =	ttmp1
159var s_save_exec_lo	    =	ttmp2
160var s_save_exec_hi	    =	ttmp3
161var s_save_tmp		    =	ttmp4
162var s_save_trapsts	    =	ttmp5		//not really used until the end of the SAVE routine
163var s_save_xnack_mask_lo    =	ttmp6
164var s_save_xnack_mask_hi    =	ttmp7
165var s_save_buf_rsrc0	    =	ttmp8
166var s_save_buf_rsrc1	    =	ttmp9
167var s_save_buf_rsrc2	    =	ttmp10
168var s_save_buf_rsrc3	    =	ttmp11
169var s_save_status	    =	ttmp12
170var s_save_mem_offset	    =	ttmp14
171var s_save_alloc_size	    =	s_save_trapsts		//conflict
172var s_save_m0		    =	ttmp15
173var s_save_ttmps_lo	    =	s_save_tmp		//no conflict
174var s_save_ttmps_hi	    =	s_save_trapsts		//no conflict
175
176/*	Restore	    */
177var S_RESTORE_BUF_RSRC_WORD1_STRIDE	    =	S_SAVE_BUF_RSRC_WORD1_STRIDE
178var S_RESTORE_BUF_RSRC_WORD3_MISC	    =	S_SAVE_BUF_RSRC_WORD3_MISC
179
180var S_RESTORE_SPI_INIT_ATC_MASK		    =	0x08000000	    //bit[27]: ATC bit
181var S_RESTORE_SPI_INIT_ATC_SHIFT	    =	27
182var S_RESTORE_SPI_INIT_MTYPE_MASK	    =	0x70000000	    //bit[30:28]: Mtype
183var S_RESTORE_SPI_INIT_MTYPE_SHIFT	    =	28
184var S_RESTORE_SPI_INIT_FIRST_WAVE_MASK	    =	0x04000000	    //bit[26]: FirstWaveInTG
185var S_RESTORE_SPI_INIT_FIRST_WAVE_SHIFT	    =	26
186
187var S_RESTORE_PC_HI_RCNT_SHIFT		    =	S_SAVE_PC_HI_RCNT_SHIFT
188var S_RESTORE_PC_HI_RCNT_MASK		    =	S_SAVE_PC_HI_RCNT_MASK
189var S_RESTORE_PC_HI_FIRST_REPLAY_SHIFT	    =	S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
190var S_RESTORE_PC_HI_FIRST_REPLAY_MASK	    =	S_SAVE_PC_HI_FIRST_REPLAY_MASK
191
192var s_restore_spi_init_lo		    =	exec_lo
193var s_restore_spi_init_hi		    =	exec_hi
194
195var s_restore_mem_offset	=   ttmp12
196var s_restore_alloc_size	=   ttmp3
197var s_restore_tmp		=   ttmp2
198var s_restore_mem_offset_save	=   s_restore_tmp	//no conflict
199
200var s_restore_m0	    =	s_restore_alloc_size	//no conflict
201
202var s_restore_mode	    =	ttmp7
203
204var s_restore_pc_lo	    =	ttmp0
205var s_restore_pc_hi	    =	ttmp1
206var s_restore_exec_lo	    =	ttmp14
207var s_restore_exec_hi	    = 	ttmp15
208var s_restore_status	    =	ttmp4
209var s_restore_trapsts	    =	ttmp5
210var s_restore_xnack_mask_lo =	xnack_mask_lo
211var s_restore_xnack_mask_hi =	xnack_mask_hi
212var s_restore_buf_rsrc0	    =	ttmp8
213var s_restore_buf_rsrc1	    =	ttmp9
214var s_restore_buf_rsrc2	    =	ttmp10
215var s_restore_buf_rsrc3	    =	ttmp11
216var s_restore_ttmps_lo	    =	s_restore_tmp		//no conflict
217var s_restore_ttmps_hi	    =	s_restore_alloc_size	//no conflict
218
219/**************************************************************************/
220/*			trap handler entry points			  */
221/**************************************************************************/
222/* Shader Main*/
223
224shader main
225  asic(GFX9)
226  type(CS)
227
228
229    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))		    //hack to use trap_id for determining save/restore
230	//FIXME VCCZ un-init assertion s_getreg_b32	s_save_status, hwreg(HW_REG_STATUS)	    //save STATUS since we will change SCC
231	s_and_b32 s_save_tmp, s_save_pc_hi, 0xffff0000		    //change SCC
232	s_cmp_eq_u32 s_save_tmp, 0x007e0000			    //Save: trap_id = 0x7e. Restore: trap_id = 0x7f.
233	s_cbranch_scc0 L_JUMP_TO_RESTORE			    //do not need to recover STATUS here  since we are going to RESTORE
234	//FIXME	 s_setreg_b32	hwreg(HW_REG_STATUS),	s_save_status	    //need to recover STATUS since we are going to SAVE
235	s_branch L_SKIP_RESTORE					    //NOT restore, SAVE actually
236    else
237	s_branch L_SKIP_RESTORE					    //NOT restore. might be a regular trap or save
238    end
239
240L_JUMP_TO_RESTORE:
241    s_branch L_RESTORE						    //restore
242
243L_SKIP_RESTORE:
244
245    s_getreg_b32    s_save_status, hwreg(HW_REG_STATUS)				    //save STATUS since we will change SCC
246    s_andn2_b32	    s_save_status, s_save_status, SQ_WAVE_STATUS_SPI_PRIO_MASK	    //check whether this is for save
247    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
248    s_and_b32       ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_SAVECTX_MASK    //check whether this is for save
249    s_cbranch_scc1  L_SAVE					//this is the operation for save
250
251    // *********    Handle non-CWSR traps	*******************
252if (!EMU_RUN_HACK)
253    // Illegal instruction is a non-maskable exception which blocks context save.
254    // Halt the wavefront and return from the trap.
255    s_and_b32       ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_ILLEGAL_INST_MASK
256    s_cbranch_scc1  L_HALT_WAVE
257
258    // If STATUS.MEM_VIOL is asserted then we cannot fetch from the TMA.
259    // Instead, halt the wavefront and return from the trap.
260    s_and_b32       ttmp2, s_save_trapsts, SQ_WAVE_TRAPSTS_MEM_VIOL_MASK
261    s_cbranch_scc0  L_FETCH_2ND_TRAP
262
263L_HALT_WAVE:
264    // If STATUS.HALT is set then this fault must come from SQC instruction fetch.
265    // We cannot prevent further faults so just terminate the wavefront.
266    s_and_b32       ttmp2, s_save_status, SQ_WAVE_STATUS_HALT_MASK
267    s_cbranch_scc0  L_NOT_ALREADY_HALTED
268    s_endpgm
269L_NOT_ALREADY_HALTED:
270    s_or_b32        s_save_status, s_save_status, SQ_WAVE_STATUS_HALT_MASK
271
272    // If the PC points to S_ENDPGM then context save will fail if STATUS.HALT is set.
273    // Rewind the PC to prevent this from occurring. The debugger compensates for this.
274    s_sub_u32       ttmp0, ttmp0, 0x8
275    s_subb_u32      ttmp1, ttmp1, 0x0
276
277L_FETCH_2ND_TRAP:
278    // Preserve and clear scalar XNACK state before issuing scalar reads.
279    // Save IB_STS.FIRST_REPLAY[15] and IB_STS.RCNT[20:16] into unused space ttmp11[31:26].
280    s_getreg_b32    ttmp2, hwreg(HW_REG_IB_STS)
281    s_and_b32       ttmp3, ttmp2, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK
282    s_lshl_b32      ttmp3, ttmp3, (TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT - SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT)
283    s_andn2_b32     ttmp11, ttmp11, TTMP11_SAVE_RCNT_FIRST_REPLAY_MASK
284    s_or_b32        ttmp11, ttmp11, ttmp3
285
286    s_andn2_b32     ttmp2, ttmp2, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK
287    s_setreg_b32    hwreg(HW_REG_IB_STS), ttmp2
288
289    // Read second-level TBA/TMA from first-level TMA and jump if available.
290    // ttmp[2:5] and ttmp12 can be used (others hold SPI-initialized debug data)
291    // ttmp12 holds SQ_WAVE_STATUS
292    s_getreg_b32    ttmp4, hwreg(HW_REG_SQ_SHADER_TMA_LO)
293    s_getreg_b32    ttmp5, hwreg(HW_REG_SQ_SHADER_TMA_HI)
294    s_lshl_b64      [ttmp4, ttmp5], [ttmp4, ttmp5], 0x8
295    s_load_dwordx2  [ttmp2, ttmp3], [ttmp4, ttmp5], 0x0 glc:1 // second-level TBA
296    s_waitcnt       lgkmcnt(0)
297    s_load_dwordx2  [ttmp4, ttmp5], [ttmp4, ttmp5], 0x8 glc:1 // second-level TMA
298    s_waitcnt       lgkmcnt(0)
299    s_and_b64       [ttmp2, ttmp3], [ttmp2, ttmp3], [ttmp2, ttmp3]
300    s_cbranch_scc0  L_NO_NEXT_TRAP // second-level trap handler not been set
301    s_setpc_b64     [ttmp2, ttmp3] // jump to second-level trap handler
302
303L_NO_NEXT_TRAP:
304    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
305    s_and_b32	    s_save_trapsts, s_save_trapsts, SQ_WAVE_TRAPSTS_EXCE_MASK // Check whether it is an exception
306    s_cbranch_scc1  L_EXCP_CASE	  // Exception, jump back to the shader program directly.
307    s_add_u32	    ttmp0, ttmp0, 4   // S_TRAP case, add 4 to ttmp0
308    s_addc_u32	ttmp1, ttmp1, 0
309L_EXCP_CASE:
310    s_and_b32	ttmp1, ttmp1, 0xFFFF
311
312    // Restore SQ_WAVE_IB_STS.
313    s_lshr_b32      ttmp2, ttmp11, (TTMP11_SAVE_RCNT_FIRST_REPLAY_SHIFT - SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT)
314    s_and_b32       ttmp2, ttmp2, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK
315    s_setreg_b32    hwreg(HW_REG_IB_STS), ttmp2
316
317    // Restore SQ_WAVE_STATUS.
318    s_and_b64       exec, exec, exec // Restore STATUS.EXECZ, not writable by s_setreg_b32
319    s_and_b64       vcc, vcc, vcc    // Restore STATUS.VCCZ, not writable by s_setreg_b32
320    s_setreg_b32    hwreg(HW_REG_STATUS), s_save_status
321
322    s_rfe_b64       [ttmp0, ttmp1]
323end
324    // *********	End handling of non-CWSR traps	 *******************
325
326/**************************************************************************/
327/*			save routine					  */
328/**************************************************************************/
329
330L_SAVE:
331
332if G8SR_DEBUG_TIMESTAMP
333	s_memrealtime	s_g8sr_ts_save_s
334	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
335end
336
337    s_and_b32	    s_save_pc_hi, s_save_pc_hi, 0x0000ffff    //pc[47:32]
338
339    s_mov_b32	    s_save_tmp, 0							    //clear saveCtx bit
340    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_SAVECTX_SHIFT, 1), s_save_tmp	    //clear saveCtx bit
341
342    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_RCNT_SHIFT, SQ_WAVE_IB_STS_RCNT_SIZE)		    //save RCNT
343    s_lshl_b32	    s_save_tmp, s_save_tmp, S_SAVE_PC_HI_RCNT_SHIFT
344    s_or_b32	    s_save_pc_hi, s_save_pc_hi, s_save_tmp
345    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT, SQ_WAVE_IB_STS_FIRST_REPLAY_SIZE)   //save FIRST_REPLAY
346    s_lshl_b32	    s_save_tmp, s_save_tmp, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
347    s_or_b32	    s_save_pc_hi, s_save_pc_hi, s_save_tmp
348    s_getreg_b32    s_save_tmp, hwreg(HW_REG_IB_STS)					    //clear RCNT and FIRST_REPLAY in IB_STS
349    s_and_b32	    s_save_tmp, s_save_tmp, SQ_WAVE_IB_STS_RCNT_FIRST_REPLAY_MASK_NEG
350
351    s_setreg_b32    hwreg(HW_REG_IB_STS), s_save_tmp
352
353    /*	    inform SPI the readiness and wait for SPI's go signal */
354    s_mov_b32	    s_save_exec_lo, exec_lo						    //save EXEC and use EXEC for the go signal from SPI
355    s_mov_b32	    s_save_exec_hi, exec_hi
356    s_mov_b64	    exec,   0x0								    //clear EXEC to get ready to receive
357
358if G8SR_DEBUG_TIMESTAMP
359	s_memrealtime  s_g8sr_ts_sq_save_msg
360	s_waitcnt lgkmcnt(0)
361end
362
363    if (EMU_RUN_HACK)
364
365    else
366	s_sendmsg   sendmsg(MSG_SAVEWAVE)  //send SPI a message and wait for SPI's write to EXEC
367    end
368
369    // Set SPI_PRIO=2 to avoid starving instruction fetch in the waves we're waiting for.
370    s_or_b32 s_save_tmp, s_save_status, (2 << SQ_WAVE_STATUS_SPI_PRIO_SHIFT)
371    s_setreg_b32 hwreg(HW_REG_STATUS), s_save_tmp
372
373  L_SLEEP:
374    s_sleep 0x2		       // sleep 1 (64clk) is not enough for 8 waves per SIMD, which will cause SQ hang, since the 7,8th wave could not get arbit to exec inst, while other waves are stuck into the sleep-loop and waiting for wrexec!=0
375
376    if (EMU_RUN_HACK)
377
378    else
379	s_cbranch_execz L_SLEEP
380    end
381
382if G8SR_DEBUG_TIMESTAMP
383	s_memrealtime  s_g8sr_ts_spi_wrexec
384	s_waitcnt lgkmcnt(0)
385end
386
387    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_SINGLE_WAVE))
388	//calculate wd_addr using absolute thread id
389	v_readlane_b32 s_save_tmp, v9, 0
390	s_lshr_b32 s_save_tmp, s_save_tmp, 6
391	s_mul_i32 s_save_tmp, s_save_tmp, WAVE_SPACE
392	s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO
393	s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI
394	s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL
395    else
396    end
397    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_SINGLE_WAVE))
398	s_add_i32 s_save_spi_init_lo, s_save_tmp, WG_BASE_ADDR_LO
399	s_mov_b32 s_save_spi_init_hi, WG_BASE_ADDR_HI
400	s_and_b32 s_save_spi_init_hi, s_save_spi_init_hi, CTX_SAVE_CONTROL
401    else
402    end
403
404    // Save trap temporaries 6-11, 13-15 initialized by SPI debug dispatch logic
405    // ttmp SR memory offset : size(VGPR)+size(SGPR)+0x40
406    get_vgpr_size_bytes(s_save_ttmps_lo)
407    get_sgpr_size_bytes(s_save_ttmps_hi)
408    s_add_u32	    s_save_ttmps_lo, s_save_ttmps_lo, s_save_ttmps_hi
409    s_add_u32	    s_save_ttmps_lo, s_save_ttmps_lo, s_save_spi_init_lo
410    s_addc_u32	    s_save_ttmps_hi, s_save_spi_init_hi, 0x0
411    s_and_b32	    s_save_ttmps_hi, s_save_ttmps_hi, 0xFFFF
412    s_store_dwordx2 [ttmp6, ttmp7], [s_save_ttmps_lo, s_save_ttmps_hi], 0x40 glc:1
413    ack_sqc_store_workaround()
414    s_store_dwordx4 [ttmp8, ttmp9, ttmp10, ttmp11], [s_save_ttmps_lo, s_save_ttmps_hi], 0x48 glc:1
415    ack_sqc_store_workaround()
416    s_store_dword   ttmp13, [s_save_ttmps_lo, s_save_ttmps_hi], 0x58 glc:1
417    ack_sqc_store_workaround()
418    s_store_dwordx2 [ttmp14, ttmp15], [s_save_ttmps_lo, s_save_ttmps_hi], 0x5C glc:1
419    ack_sqc_store_workaround()
420
421    /*	    setup Resource Contants    */
422    s_mov_b32	    s_save_buf_rsrc0,	s_save_spi_init_lo							//base_addr_lo
423    s_and_b32	    s_save_buf_rsrc1,	s_save_spi_init_hi, 0x0000FFFF						//base_addr_hi
424    s_or_b32	    s_save_buf_rsrc1,	s_save_buf_rsrc1,  S_SAVE_BUF_RSRC_WORD1_STRIDE
425    s_mov_b32	    s_save_buf_rsrc2,	0									//NUM_RECORDS initial value = 0 (in bytes) although not neccessarily inited
426    s_mov_b32	    s_save_buf_rsrc3,	S_SAVE_BUF_RSRC_WORD3_MISC
427    s_and_b32	    s_save_tmp,		s_save_spi_init_hi, S_SAVE_SPI_INIT_ATC_MASK
428    s_lshr_b32	    s_save_tmp,		s_save_tmp, (S_SAVE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT)	    //get ATC bit into position
429    s_or_b32	    s_save_buf_rsrc3,	s_save_buf_rsrc3,  s_save_tmp						//or ATC
430    s_and_b32	    s_save_tmp,		s_save_spi_init_hi, S_SAVE_SPI_INIT_MTYPE_MASK
431    s_lshr_b32	    s_save_tmp,		s_save_tmp, (S_SAVE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT)	    //get MTYPE bits into position
432    s_or_b32	    s_save_buf_rsrc3,	s_save_buf_rsrc3,  s_save_tmp						//or MTYPE
433
434    //FIXME  right now s_save_m0/s_save_mem_offset use tma_lo/tma_hi  (might need to save them before using them?)
435    s_mov_b32	    s_save_m0,		m0								    //save M0
436
437    /*	    global mem offset		*/
438    s_mov_b32	    s_save_mem_offset,	0x0									//mem offset initial value = 0
439
440
441
442
443    /*	    save HW registers	*/
444    //////////////////////////////
445
446  L_SAVE_HWREG:
447	// HWREG SR memory offset : size(VGPR)+size(SGPR)
448       get_vgpr_size_bytes(s_save_mem_offset)
449       get_sgpr_size_bytes(s_save_tmp)
450       s_add_u32 s_save_mem_offset, s_save_mem_offset, s_save_tmp
451
452
453    s_mov_b32	    s_save_buf_rsrc2, 0x4				//NUM_RECORDS	in bytes
454    if (SWIZZLE_EN)
455	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
456    else
457	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
458    end
459
460
461    write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset)			//M0
462
463    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_SAVE_FIRST_TIME))
464	s_add_u32 s_save_pc_lo, s_save_pc_lo, 4		    //pc[31:0]+4
465	s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0	    //carry bit over
466    end
467
468    write_hwreg_to_mem(s_save_pc_lo, s_save_buf_rsrc0, s_save_mem_offset)		    //PC
469    write_hwreg_to_mem(s_save_pc_hi, s_save_buf_rsrc0, s_save_mem_offset)
470    write_hwreg_to_mem(s_save_exec_lo, s_save_buf_rsrc0, s_save_mem_offset)		//EXEC
471    write_hwreg_to_mem(s_save_exec_hi, s_save_buf_rsrc0, s_save_mem_offset)
472    write_hwreg_to_mem(s_save_status, s_save_buf_rsrc0, s_save_mem_offset)		//STATUS
473
474    //s_save_trapsts conflicts with s_save_alloc_size
475    s_getreg_b32    s_save_trapsts, hwreg(HW_REG_TRAPSTS)
476    write_hwreg_to_mem(s_save_trapsts, s_save_buf_rsrc0, s_save_mem_offset)		//TRAPSTS
477
478    write_hwreg_to_mem(xnack_mask_lo, s_save_buf_rsrc0, s_save_mem_offset)	    //XNACK_MASK_LO
479    write_hwreg_to_mem(xnack_mask_hi, s_save_buf_rsrc0, s_save_mem_offset)	    //XNACK_MASK_HI
480
481    //use s_save_tmp would introduce conflict here between s_save_tmp and s_save_buf_rsrc2
482    s_getreg_b32    s_save_m0, hwreg(HW_REG_MODE)						    //MODE
483    write_hwreg_to_mem(s_save_m0, s_save_buf_rsrc0, s_save_mem_offset)
484
485
486
487    /*	    the first wave in the threadgroup	 */
488    s_and_b32	    s_save_tmp, s_save_spi_init_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK	// extract fisrt wave bit
489    s_mov_b32	     s_save_exec_hi, 0x0
490    s_or_b32	     s_save_exec_hi, s_save_tmp, s_save_exec_hi				 // save first wave bit to s_save_exec_hi.bits[26]
491
492
493    /*		save SGPRs	*/
494	// Save SGPR before LDS save, then the s0 to s4 can be used during LDS save...
495    //////////////////////////////
496
497    // SGPR SR memory offset : size(VGPR)
498    get_vgpr_size_bytes(s_save_mem_offset)
499    // TODO, change RSRC word to rearrange memory layout for SGPRS
500
501    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)		//spgr_size
502    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 1
503    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 4			    //Number of SGPRs = (sgpr_size + 1) * 16   (non-zero value)
504
505    if (SGPR_SAVE_USE_SQC)
506	s_lshl_b32	s_save_buf_rsrc2,   s_save_alloc_size, 2		    //NUM_RECORDS in bytes
507    else
508	s_lshl_b32	s_save_buf_rsrc2,   s_save_alloc_size, 8		    //NUM_RECORDS in bytes (64 threads)
509    end
510
511    if (SWIZZLE_EN)
512	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
513    else
514	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
515    end
516
517
518    // backup s_save_buf_rsrc0,1 to s_save_pc_lo/hi, since write_16sgpr_to_mem function will change the rsrc0
519    //s_mov_b64 s_save_pc_lo, s_save_buf_rsrc0
520    s_mov_b64 s_save_xnack_mask_lo, s_save_buf_rsrc0
521    s_add_u32 s_save_buf_rsrc0, s_save_buf_rsrc0, s_save_mem_offset
522    s_addc_u32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0
523
524    s_mov_b32	    m0, 0x0			    //SGPR initial index value =0
525    s_nop	    0x0				    //Manually inserted wait states
526  L_SAVE_SGPR_LOOP:
527    // SGPR is allocated in 16 SGPR granularity
528    s_movrels_b64   s0, s0     //s0 = s[0+m0], s1 = s[1+m0]
529    s_movrels_b64   s2, s2     //s2 = s[2+m0], s3 = s[3+m0]
530    s_movrels_b64   s4, s4     //s4 = s[4+m0], s5 = s[5+m0]
531    s_movrels_b64   s6, s6     //s6 = s[6+m0], s7 = s[7+m0]
532    s_movrels_b64   s8, s8     //s8 = s[8+m0], s9 = s[9+m0]
533    s_movrels_b64   s10, s10   //s10 = s[10+m0], s11 = s[11+m0]
534    s_movrels_b64   s12, s12   //s12 = s[12+m0], s13 = s[13+m0]
535    s_movrels_b64   s14, s14   //s14 = s[14+m0], s15 = s[15+m0]
536
537    write_16sgpr_to_mem(s0, s_save_buf_rsrc0, s_save_mem_offset) //PV: the best performance should be using s_buffer_store_dwordx4
538    s_add_u32	    m0, m0, 16							    //next sgpr index
539    s_cmp_lt_u32    m0, s_save_alloc_size					    //scc = (m0 < s_save_alloc_size) ? 1 : 0
540    s_cbranch_scc1  L_SAVE_SGPR_LOOP					//SGPR save is complete?
541    // restore s_save_buf_rsrc0,1
542    //s_mov_b64 s_save_buf_rsrc0, s_save_pc_lo
543    s_mov_b64 s_save_buf_rsrc0, s_save_xnack_mask_lo
544
545
546
547
548    /*		save first 4 VGPR, then LDS save could use   */
549	// each wave will alloc 4 vgprs at least...
550    /////////////////////////////////////////////////////////////////////////////////////
551
552    s_mov_b32	    s_save_mem_offset, 0
553    s_mov_b32	    exec_lo, 0xFFFFFFFF						    //need every thread from now on
554    s_mov_b32	    exec_hi, 0xFFFFFFFF
555    s_mov_b32	    xnack_mask_lo, 0x0
556    s_mov_b32	    xnack_mask_hi, 0x0
557
558    if (SWIZZLE_EN)
559	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
560    else
561	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
562    end
563
564
565    // VGPR Allocated in 4-GPR granularity
566
567if G8SR_VGPR_SR_IN_DWX4
568	// the const stride for DWx4 is 4*4 bytes
569	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
570	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
571
572	buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
573
574	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
575	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE  // reset const stride to 4 bytes
576else
577	buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
578	buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256
579	buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*2
580	buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*3
581end
582
583
584
585    /*		save LDS	*/
586    //////////////////////////////
587
588  L_SAVE_LDS:
589
590	// Change EXEC to all threads...
591    s_mov_b32	    exec_lo, 0xFFFFFFFF	  //need every thread from now on
592    s_mov_b32	    exec_hi, 0xFFFFFFFF
593
594    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		    //lds_size
595    s_and_b32	    s_save_alloc_size, s_save_alloc_size, 0xFFFFFFFF		    //lds_size is zero?
596    s_cbranch_scc0  L_SAVE_LDS_DONE									       //no lds used? jump to L_SAVE_DONE
597
598    s_barrier		    //LDS is used? wait for other waves in the same TG
599    s_and_b32	    s_save_tmp, s_save_exec_hi, S_SAVE_SPI_INIT_FIRST_WAVE_MASK		       //exec is still used here
600    s_cbranch_scc0  L_SAVE_LDS_DONE
601
602	// first wave do LDS save;
603
604    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 6			    //LDS size in dwords = lds_size * 64dw
605    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 2			    //LDS size in bytes
606    s_mov_b32	    s_save_buf_rsrc2,  s_save_alloc_size			    //NUM_RECORDS in bytes
607
608    // LDS at offset: size(VGPR)+SIZE(SGPR)+SIZE(HWREG)
609    //
610    get_vgpr_size_bytes(s_save_mem_offset)
611    get_sgpr_size_bytes(s_save_tmp)
612    s_add_u32  s_save_mem_offset, s_save_mem_offset, s_save_tmp
613    s_add_u32 s_save_mem_offset, s_save_mem_offset, get_hwreg_size_bytes()
614
615
616    if (SWIZZLE_EN)
617	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0	      //FIXME need to use swizzle to enable bounds checking?
618    else
619	s_mov_b32	s_save_buf_rsrc2,  0x1000000		      //NUM_RECORDS in bytes
620    end
621
622    s_mov_b32	    m0, 0x0						  //lds_offset initial value = 0
623
624
625var LDS_DMA_ENABLE = 0
626var UNROLL = 0
627if UNROLL==0 && LDS_DMA_ENABLE==1
628	s_mov_b32  s3, 256*2
629	s_nop 0
630	s_nop 0
631	s_nop 0
632  L_SAVE_LDS_LOOP:
633	//TODO: looks the 2 buffer_store/load clause for s/r will hurt performance.???
634    if (SAVE_LDS)     //SPI always alloc LDS space in 128DW granularity
635	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1		// first 64DW
636	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256 // second 64DW
637    end
638
639    s_add_u32	    m0, m0, s3						//every buffer_store_lds does 256 bytes
640    s_add_u32	    s_save_mem_offset, s_save_mem_offset, s3				//mem offset increased by 256 bytes
641    s_cmp_lt_u32    m0, s_save_alloc_size						//scc=(m0 < s_save_alloc_size) ? 1 : 0
642    s_cbranch_scc1  L_SAVE_LDS_LOOP							//LDS save is complete?
643
644elsif LDS_DMA_ENABLE==1 && UNROLL==1 // UNROOL	, has ichace miss
645      // store from higest LDS address to lowest
646      s_mov_b32	 s3, 256*2
647      s_sub_u32	 m0, s_save_alloc_size, s3
648      s_add_u32 s_save_mem_offset, s_save_mem_offset, m0
649      s_lshr_b32 s_save_alloc_size, s_save_alloc_size, 9   // how many 128 trunks...
650      s_sub_u32 s_save_alloc_size, 128, s_save_alloc_size   // store from higheset addr to lowest
651      s_mul_i32 s_save_alloc_size, s_save_alloc_size, 6*4   // PC offset increment,  each LDS save block cost 6*4 Bytes instruction
652      s_add_u32 s_save_alloc_size, s_save_alloc_size, 3*4   //2is the below 2 inst...//s_addc and s_setpc
653      s_nop 0
654      s_nop 0
655      s_nop 0	//pad 3 dw to let LDS_DMA align with 64Bytes
656      s_getpc_b64 s[0:1]			      // reuse s[0:1], since s[0:1] already saved
657      s_add_u32	  s0, s0,s_save_alloc_size
658      s_addc_u32  s1, s1, 0
659      s_setpc_b64 s[0:1]
660
661
662       for var i =0; i< 128; i++
663	    // be careful to make here a 64Byte aligned address, which could improve performance...
664	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:0		// first 64DW
665	    buffer_store_lds_dword s_save_buf_rsrc0, s_save_mem_offset lds:1 offset:256		  // second 64DW
666
667	if i!=127
668	s_sub_u32  m0, m0, s3	   // use a sgpr to shrink 2DW-inst to 1DW inst to improve performance , i.e.  pack more LDS_DMA inst to one Cacheline
669	    s_sub_u32  s_save_mem_offset, s_save_mem_offset,  s3
670	    end
671       end
672
673else   // BUFFER_STORE
674      v_mbcnt_lo_u32_b32 v2, 0xffffffff, 0x0
675      v_mbcnt_hi_u32_b32 v3, 0xffffffff, v2	// tid
676      v_mul_i32_i24 v2, v3, 8	// tid*8
677      v_mov_b32 v3, 256*2
678      s_mov_b32 m0, 0x10000
679      s_mov_b32 s0, s_save_buf_rsrc3
680      s_and_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, 0xFF7FFFFF	  // disable add_tid
681      s_or_b32 s_save_buf_rsrc3, s_save_buf_rsrc3, 0x58000   //DFMT
682
683L_SAVE_LDS_LOOP_VECTOR:
684      ds_read_b64 v[0:1], v2	//x =LDS[a], byte address
685      s_waitcnt lgkmcnt(0)
686      buffer_store_dwordx2  v[0:1], v2, s_save_buf_rsrc0, s_save_mem_offset offen:1  glc:1  slc:1
687//	s_waitcnt vmcnt(0)
688//	v_add_u32 v2, vcc[0:1], v2, v3
689      v_add_u32 v2, v2, v3
690      v_cmp_lt_u32 vcc[0:1], v2, s_save_alloc_size
691      s_cbranch_vccnz L_SAVE_LDS_LOOP_VECTOR
692
693      // restore rsrc3
694      s_mov_b32 s_save_buf_rsrc3, s0
695
696end
697
698L_SAVE_LDS_DONE:
699
700
701    /*		save VGPRs  - set the Rest VGPRs	*/
702    //////////////////////////////////////////////////////////////////////////////////////
703  L_SAVE_VGPR:
704    // VGPR SR memory offset: 0
705    // TODO rearrange the RSRC words to use swizzle for VGPR save...
706
707    s_mov_b32	    s_save_mem_offset, (0+256*4)				    // for the rest VGPRs
708    s_mov_b32	    exec_lo, 0xFFFFFFFF						    //need every thread from now on
709    s_mov_b32	    exec_hi, 0xFFFFFFFF
710
711    s_getreg_b32    s_save_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)		    //vpgr_size
712    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 1
713    s_lshl_b32	    s_save_alloc_size, s_save_alloc_size, 2			    //Number of VGPRs = (vgpr_size + 1) * 4    (non-zero value)	  //FIXME for GFX, zero is possible
714    s_lshl_b32	    s_save_buf_rsrc2,  s_save_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads*4)
715    if (SWIZZLE_EN)
716	s_add_u32	s_save_buf_rsrc2, s_save_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
717    else
718	s_mov_b32	s_save_buf_rsrc2,  0x1000000				    //NUM_RECORDS in bytes
719    end
720
721
722    // VGPR Allocated in 4-GPR granularity
723
724if G8SR_VGPR_SR_IN_DWX4
725	// the const stride for DWx4 is 4*4 bytes
726	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
727	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, G8SR_SAVE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
728
729	s_mov_b32	  m0, 4	    // skip first 4 VGPRs
730	s_cmp_lt_u32	  m0, s_save_alloc_size
731	s_cbranch_scc0	  L_SAVE_VGPR_LOOP_END	    // no more vgprs
732
733	s_set_gpr_idx_on  m0, 0x1   // This will change M0
734	s_add_u32	  s_save_alloc_size, s_save_alloc_size, 0x1000	// because above inst change m0
735L_SAVE_VGPR_LOOP:
736	v_mov_b32	  v0, v0   // v0 = v[0+m0]
737	v_mov_b32	  v1, v1
738	v_mov_b32	  v2, v2
739	v_mov_b32	  v3, v3
740
741
742	buffer_store_dwordx4 v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
743	s_add_u32	  m0, m0, 4
744	s_add_u32	  s_save_mem_offset, s_save_mem_offset, 256*4
745	s_cmp_lt_u32	  m0, s_save_alloc_size
746    s_cbranch_scc1  L_SAVE_VGPR_LOOP						    //VGPR save is complete?
747    s_set_gpr_idx_off
748L_SAVE_VGPR_LOOP_END:
749
750	s_and_b32 s_save_buf_rsrc1, s_save_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
751	s_or_b32  s_save_buf_rsrc1, s_save_buf_rsrc1, S_SAVE_BUF_RSRC_WORD1_STRIDE  // reset const stride to 4 bytes
752else
753    // VGPR store using dw burst
754    s_mov_b32	      m0, 0x4	//VGPR initial index value =0
755    s_cmp_lt_u32      m0, s_save_alloc_size
756    s_cbranch_scc0    L_SAVE_VGPR_END
757
758
759    s_set_gpr_idx_on	m0, 0x1 //M0[7:0] = M0[7:0] and M0[15:12] = 0x1
760    s_add_u32	    s_save_alloc_size, s_save_alloc_size, 0x1000		    //add 0x1000 since we compare m0 against it later
761
762  L_SAVE_VGPR_LOOP:
763    v_mov_b32	    v0, v0		//v0 = v[0+m0]
764    v_mov_b32	    v1, v1		//v0 = v[0+m0]
765    v_mov_b32	    v2, v2		//v0 = v[0+m0]
766    v_mov_b32	    v3, v3		//v0 = v[0+m0]
767
768    if(USE_MTBUF_INSTEAD_OF_MUBUF)
769	tbuffer_store_format_x v0, v0, s_save_buf_rsrc0, s_save_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
770    else
771	buffer_store_dword v0, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1
772	buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256
773	buffer_store_dword v2, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*2
774	buffer_store_dword v3, v0, s_save_buf_rsrc0, s_save_mem_offset slc:1 glc:1  offset:256*3
775    end
776
777    s_add_u32	    m0, m0, 4							    //next vgpr index
778    s_add_u32	    s_save_mem_offset, s_save_mem_offset, 256*4			    //every buffer_store_dword does 256 bytes
779    s_cmp_lt_u32    m0, s_save_alloc_size					    //scc = (m0 < s_save_alloc_size) ? 1 : 0
780    s_cbranch_scc1  L_SAVE_VGPR_LOOP						    //VGPR save is complete?
781    s_set_gpr_idx_off
782end
783
784L_SAVE_VGPR_END:
785
786
787
788
789
790
791    /*	   S_PGM_END_SAVED  */				    //FIXME  graphics ONLY
792    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_SAVE_NORMAL_EXIT))
793	s_and_b32 s_save_pc_hi, s_save_pc_hi, 0x0000ffff    //pc[47:32]
794	s_add_u32 s_save_pc_lo, s_save_pc_lo, 4		    //pc[31:0]+4
795	s_addc_u32 s_save_pc_hi, s_save_pc_hi, 0x0	    //carry bit over
796	s_rfe_b64 s_save_pc_lo				    //Return to the main shader program
797    else
798    end
799
800// Save Done timestamp
801if G8SR_DEBUG_TIMESTAMP
802	s_memrealtime	s_g8sr_ts_save_d
803	// SGPR SR memory offset : size(VGPR)
804	get_vgpr_size_bytes(s_save_mem_offset)
805	s_add_u32 s_save_mem_offset, s_save_mem_offset, G8SR_DEBUG_TS_SAVE_D_OFFSET
806	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
807	// Need reset rsrc2??
808	s_mov_b32 m0, s_save_mem_offset
809	s_mov_b32 s_save_buf_rsrc2,  0x1000000					//NUM_RECORDS in bytes
810	s_buffer_store_dwordx2 s_g8sr_ts_save_d, s_save_buf_rsrc0, m0	    glc:1
811end
812
813
814    s_branch	L_END_PGM
815
816
817
818/**************************************************************************/
819/*			restore routine					  */
820/**************************************************************************/
821
822L_RESTORE:
823    /*	    Setup Resource Contants    */
824    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))
825	//calculate wd_addr using absolute thread id
826	v_readlane_b32 s_restore_tmp, v9, 0
827	s_lshr_b32 s_restore_tmp, s_restore_tmp, 6
828	s_mul_i32 s_restore_tmp, s_restore_tmp, WAVE_SPACE
829	s_add_i32 s_restore_spi_init_lo, s_restore_tmp, WG_BASE_ADDR_LO
830	s_mov_b32 s_restore_spi_init_hi, WG_BASE_ADDR_HI
831	s_and_b32 s_restore_spi_init_hi, s_restore_spi_init_hi, CTX_RESTORE_CONTROL
832    else
833    end
834
835if G8SR_DEBUG_TIMESTAMP
836	s_memrealtime	s_g8sr_ts_restore_s
837	s_waitcnt lgkmcnt(0)	     //FIXME, will cause xnack??
838	// tma_lo/hi are sgpr 110, 111, which will not used for 112 SGPR allocated case...
839	s_mov_b32 s_restore_pc_lo, s_g8sr_ts_restore_s[0]
840	s_mov_b32 s_restore_pc_hi, s_g8sr_ts_restore_s[1]   //backup ts to ttmp0/1, sicne exec will be finally restored..
841end
842
843
844
845    s_mov_b32	    s_restore_buf_rsrc0,    s_restore_spi_init_lo							    //base_addr_lo
846    s_and_b32	    s_restore_buf_rsrc1,    s_restore_spi_init_hi, 0x0000FFFF						    //base_addr_hi
847    s_or_b32	    s_restore_buf_rsrc1,    s_restore_buf_rsrc1,  S_RESTORE_BUF_RSRC_WORD1_STRIDE
848    s_mov_b32	    s_restore_buf_rsrc2,    0										    //NUM_RECORDS initial value = 0 (in bytes)
849    s_mov_b32	    s_restore_buf_rsrc3,    S_RESTORE_BUF_RSRC_WORD3_MISC
850    s_and_b32	    s_restore_tmp,	    s_restore_spi_init_hi, S_RESTORE_SPI_INIT_ATC_MASK
851    s_lshr_b32	    s_restore_tmp,	    s_restore_tmp, (S_RESTORE_SPI_INIT_ATC_SHIFT-SQ_BUF_RSRC_WORD1_ATC_SHIFT)	    //get ATC bit into position
852    s_or_b32	    s_restore_buf_rsrc3,    s_restore_buf_rsrc3,  s_restore_tmp						    //or ATC
853    s_and_b32	    s_restore_tmp,	    s_restore_spi_init_hi, S_RESTORE_SPI_INIT_MTYPE_MASK
854    s_lshr_b32	    s_restore_tmp,	    s_restore_tmp, (S_RESTORE_SPI_INIT_MTYPE_SHIFT-SQ_BUF_RSRC_WORD3_MTYPE_SHIFT)   //get MTYPE bits into position
855    s_or_b32	    s_restore_buf_rsrc3,    s_restore_buf_rsrc3,  s_restore_tmp						    //or MTYPE
856
857    /*	    global mem offset		*/
858//  s_mov_b32	    s_restore_mem_offset, 0x0				    //mem offset initial value = 0
859
860    /*	    the first wave in the threadgroup	 */
861    s_and_b32	    s_restore_tmp, s_restore_spi_init_hi, S_RESTORE_SPI_INIT_FIRST_WAVE_MASK
862    s_cbranch_scc0  L_RESTORE_VGPR
863
864    /*		restore LDS	*/
865    //////////////////////////////
866  L_RESTORE_LDS:
867
868    s_mov_b32	    exec_lo, 0xFFFFFFFF							    //need every thread from now on   //be consistent with SAVE although can be moved ahead
869    s_mov_b32	    exec_hi, 0xFFFFFFFF
870
871    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_LDS_ALLOC,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT,SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		//lds_size
872    s_and_b32	    s_restore_alloc_size, s_restore_alloc_size, 0xFFFFFFFF		    //lds_size is zero?
873    s_cbranch_scc0  L_RESTORE_VGPR							    //no lds used? jump to L_RESTORE_VGPR
874    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 6			    //LDS size in dwords = lds_size * 64dw
875    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 2			    //LDS size in bytes
876    s_mov_b32	    s_restore_buf_rsrc2,    s_restore_alloc_size			    //NUM_RECORDS in bytes
877
878    // LDS at offset: size(VGPR)+SIZE(SGPR)+SIZE(HWREG)
879    //
880    get_vgpr_size_bytes(s_restore_mem_offset)
881    get_sgpr_size_bytes(s_restore_tmp)
882    s_add_u32  s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
883    s_add_u32  s_restore_mem_offset, s_restore_mem_offset, get_hwreg_size_bytes()	     //FIXME, Check if offset overflow???
884
885
886    if (SWIZZLE_EN)
887	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
888    else
889	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
890    end
891    s_mov_b32	    m0, 0x0								    //lds_offset initial value = 0
892
893  L_RESTORE_LDS_LOOP:
894    if (SAVE_LDS)
895	buffer_load_dword   v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1		       // first 64DW
896	buffer_load_dword   v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 offset:256	       // second 64DW
897    end
898    s_add_u32	    m0, m0, 256*2						// 128 DW
899    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*2		//mem offset increased by 128DW
900    s_cmp_lt_u32    m0, s_restore_alloc_size					//scc=(m0 < s_restore_alloc_size) ? 1 : 0
901    s_cbranch_scc1  L_RESTORE_LDS_LOOP							    //LDS restore is complete?
902
903
904    /*		restore VGPRs	    */
905    //////////////////////////////
906  L_RESTORE_VGPR:
907	// VGPR SR memory offset : 0
908    s_mov_b32	    s_restore_mem_offset, 0x0
909    s_mov_b32	    exec_lo, 0xFFFFFFFF							    //need every thread from now on   //be consistent with SAVE although can be moved ahead
910    s_mov_b32	    exec_hi, 0xFFFFFFFF
911
912    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)	//vpgr_size
913    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 1
914    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 2			    //Number of VGPRs = (vgpr_size + 1) * 4    (non-zero value)
915    s_lshl_b32	    s_restore_buf_rsrc2,  s_restore_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads*4)
916    if (SWIZZLE_EN)
917	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
918    else
919	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
920    end
921
922if G8SR_VGPR_SR_IN_DWX4
923     get_vgpr_size_bytes(s_restore_mem_offset)
924     s_sub_u32	       s_restore_mem_offset, s_restore_mem_offset, 256*4
925
926     // the const stride for DWx4 is 4*4 bytes
927     s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
928     s_or_b32  s_restore_buf_rsrc1, s_restore_buf_rsrc1, G8SR_RESTORE_BUF_RSRC_WORD1_STRIDE_DWx4  // const stride to 4*4 bytes
929
930     s_mov_b32	       m0, s_restore_alloc_size
931     s_set_gpr_idx_on  m0, 0x8	  // Note.. This will change m0
932
933L_RESTORE_VGPR_LOOP:
934     buffer_load_dwordx4 v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1
935     s_waitcnt vmcnt(0)
936     s_sub_u32	       m0, m0, 4
937     v_mov_b32	       v0, v0	// v[0+m0] = v0
938     v_mov_b32	       v1, v1
939     v_mov_b32	       v2, v2
940     v_mov_b32	       v3, v3
941     s_sub_u32	       s_restore_mem_offset, s_restore_mem_offset, 256*4
942     s_cmp_eq_u32      m0, 0x8000
943     s_cbranch_scc0    L_RESTORE_VGPR_LOOP
944     s_set_gpr_idx_off
945
946     s_and_b32 s_restore_buf_rsrc1, s_restore_buf_rsrc1, 0x0000FFFF   // reset const stride to 0
947     s_or_b32  s_restore_buf_rsrc1, s_restore_buf_rsrc1, S_RESTORE_BUF_RSRC_WORD1_STRIDE  // const stride to 4*4 bytes
948
949else
950    // VGPR load using dw burst
951    s_mov_b32	    s_restore_mem_offset_save, s_restore_mem_offset	// restore start with v1, v0 will be the last
952    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*4
953    s_mov_b32	    m0, 4				//VGPR initial index value = 1
954    s_set_gpr_idx_on  m0, 0x8			    //M0[7:0] = M0[7:0] and M0[15:12] = 0x8
955    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 0x8000			    //add 0x8000 since we compare m0 against it later
956
957  L_RESTORE_VGPR_LOOP:
958    if(USE_MTBUF_INSTEAD_OF_MUBUF)
959	tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
960    else
961	buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1
962	buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256
963	buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*2
964	buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset slc:1 glc:1 offset:256*3
965    end
966    s_waitcnt	    vmcnt(0)								    //ensure data ready
967    v_mov_b32	    v0, v0								    //v[0+m0] = v0
968    v_mov_b32	    v1, v1
969    v_mov_b32	    v2, v2
970    v_mov_b32	    v3, v3
971    s_add_u32	    m0, m0, 4								    //next vgpr index
972    s_add_u32	    s_restore_mem_offset, s_restore_mem_offset, 256*4				//every buffer_load_dword does 256 bytes
973    s_cmp_lt_u32    m0, s_restore_alloc_size						    //scc = (m0 < s_restore_alloc_size) ? 1 : 0
974    s_cbranch_scc1  L_RESTORE_VGPR_LOOP							    //VGPR restore (except v0) is complete?
975    s_set_gpr_idx_off
976											    /* VGPR restore on v0 */
977    if(USE_MTBUF_INSTEAD_OF_MUBUF)
978	tbuffer_load_format_x v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save format:BUF_NUM_FORMAT_FLOAT format: BUF_DATA_FORMAT_32 slc:1 glc:1
979    else
980	buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1
981	buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256
982	buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256*2
983	buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save    slc:1 glc:1 offset:256*3
984    end
985
986end
987
988    /*		restore SGPRs	    */
989    //////////////////////////////
990
991    // SGPR SR memory offset : size(VGPR)
992    get_vgpr_size_bytes(s_restore_mem_offset)
993    get_sgpr_size_bytes(s_restore_tmp)
994    s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
995    s_sub_u32 s_restore_mem_offset, s_restore_mem_offset, 16*4	   // restore SGPR from S[n] to S[0], by 16 sgprs group
996    // TODO, change RSRC word to rearrange memory layout for SGPRS
997
998    s_getreg_b32    s_restore_alloc_size, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)		    //spgr_size
999    s_add_u32	    s_restore_alloc_size, s_restore_alloc_size, 1
1000    s_lshl_b32	    s_restore_alloc_size, s_restore_alloc_size, 4			    //Number of SGPRs = (sgpr_size + 1) * 16   (non-zero value)
1001
1002    if (SGPR_SAVE_USE_SQC)
1003	s_lshl_b32	s_restore_buf_rsrc2,	s_restore_alloc_size, 2			    //NUM_RECORDS in bytes
1004    else
1005	s_lshl_b32	s_restore_buf_rsrc2,	s_restore_alloc_size, 8			    //NUM_RECORDS in bytes (64 threads)
1006    end
1007    if (SWIZZLE_EN)
1008	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
1009    else
1010	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
1011    end
1012
1013    s_mov_b32 m0, s_restore_alloc_size
1014
1015 L_RESTORE_SGPR_LOOP:
1016    read_16sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)	 //PV: further performance improvement can be made
1017    s_waitcnt	    lgkmcnt(0)								    //ensure data ready
1018
1019    s_sub_u32 m0, m0, 16    // Restore from S[n] to S[0]
1020    s_nop 0 // hazard SALU M0=> S_MOVREL
1021
1022    s_movreld_b64   s0, s0	//s[0+m0] = s0
1023    s_movreld_b64   s2, s2
1024    s_movreld_b64   s4, s4
1025    s_movreld_b64   s6, s6
1026    s_movreld_b64   s8, s8
1027    s_movreld_b64   s10, s10
1028    s_movreld_b64   s12, s12
1029    s_movreld_b64   s14, s14
1030
1031    s_cmp_eq_u32    m0, 0		//scc = (m0 < s_restore_alloc_size) ? 1 : 0
1032    s_cbranch_scc0  L_RESTORE_SGPR_LOOP		    //SGPR restore (except s0) is complete?
1033
1034    /*	    restore HW registers    */
1035    //////////////////////////////
1036  L_RESTORE_HWREG:
1037
1038
1039if G8SR_DEBUG_TIMESTAMP
1040      s_mov_b32 s_g8sr_ts_restore_s[0], s_restore_pc_lo
1041      s_mov_b32 s_g8sr_ts_restore_s[1], s_restore_pc_hi
1042end
1043
1044    // HWREG SR memory offset : size(VGPR)+size(SGPR)
1045    get_vgpr_size_bytes(s_restore_mem_offset)
1046    get_sgpr_size_bytes(s_restore_tmp)
1047    s_add_u32 s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
1048
1049
1050    s_mov_b32	    s_restore_buf_rsrc2, 0x4						    //NUM_RECORDS   in bytes
1051    if (SWIZZLE_EN)
1052	s_add_u32	s_restore_buf_rsrc2, s_restore_buf_rsrc2, 0x0			    //FIXME need to use swizzle to enable bounds checking?
1053    else
1054	s_mov_b32	s_restore_buf_rsrc2,  0x1000000					    //NUM_RECORDS in bytes
1055    end
1056
1057    read_hwreg_from_mem(s_restore_m0, s_restore_buf_rsrc0, s_restore_mem_offset)		    //M0
1058    read_hwreg_from_mem(s_restore_pc_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		//PC
1059    read_hwreg_from_mem(s_restore_pc_hi, s_restore_buf_rsrc0, s_restore_mem_offset)
1060    read_hwreg_from_mem(s_restore_exec_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		    //EXEC
1061    read_hwreg_from_mem(s_restore_exec_hi, s_restore_buf_rsrc0, s_restore_mem_offset)
1062    read_hwreg_from_mem(s_restore_status, s_restore_buf_rsrc0, s_restore_mem_offset)		    //STATUS
1063    read_hwreg_from_mem(s_restore_trapsts, s_restore_buf_rsrc0, s_restore_mem_offset)		    //TRAPSTS
1064    read_hwreg_from_mem(xnack_mask_lo, s_restore_buf_rsrc0, s_restore_mem_offset)		    //XNACK_MASK_LO
1065    read_hwreg_from_mem(xnack_mask_hi, s_restore_buf_rsrc0, s_restore_mem_offset)		    //XNACK_MASK_HI
1066    read_hwreg_from_mem(s_restore_mode, s_restore_buf_rsrc0, s_restore_mem_offset)		//MODE
1067
1068    s_waitcnt	    lgkmcnt(0)											    //from now on, it is safe to restore STATUS and IB_STS
1069
1070    //for normal save & restore, the saved PC points to the next inst to execute, no adjustment needs to be made, otherwise:
1071    if ((EMU_RUN_HACK) && (!EMU_RUN_HACK_RESTORE_NORMAL))
1072	s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 8		 //pc[31:0]+8	  //two back-to-back s_trap are used (first for save and second for restore)
1073	s_addc_u32  s_restore_pc_hi, s_restore_pc_hi, 0x0	 //carry bit over
1074    end
1075    if ((EMU_RUN_HACK) && (EMU_RUN_HACK_RESTORE_NORMAL))
1076	s_add_u32 s_restore_pc_lo, s_restore_pc_lo, 4		 //pc[31:0]+4	  // save is hack through s_trap but restore is normal
1077	s_addc_u32  s_restore_pc_hi, s_restore_pc_hi, 0x0	 //carry bit over
1078    end
1079
1080    s_mov_b32	    m0,		s_restore_m0
1081    s_mov_b32	    exec_lo,	s_restore_exec_lo
1082    s_mov_b32	    exec_hi,	s_restore_exec_hi
1083
1084    s_and_b32	    s_restore_m0, SQ_WAVE_TRAPSTS_PRE_SAVECTX_MASK, s_restore_trapsts
1085    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_PRE_SAVECTX_SHIFT, SQ_WAVE_TRAPSTS_PRE_SAVECTX_SIZE), s_restore_m0
1086    s_and_b32	    s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_MASK, s_restore_trapsts
1087    s_lshr_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT
1088    s_setreg_b32    hwreg(HW_REG_TRAPSTS, SQ_WAVE_TRAPSTS_POST_SAVECTX_SHIFT, SQ_WAVE_TRAPSTS_POST_SAVECTX_SIZE), s_restore_m0
1089    //s_setreg_b32  hwreg(HW_REG_TRAPSTS),  s_restore_trapsts	   //don't overwrite SAVECTX bit as it may be set through external SAVECTX during restore
1090    s_setreg_b32    hwreg(HW_REG_MODE),	    s_restore_mode
1091
1092    // Restore trap temporaries 6-11, 13-15 initialized by SPI debug dispatch logic
1093    // ttmp SR memory offset : size(VGPR)+size(SGPR)+0x40
1094    get_vgpr_size_bytes(s_restore_ttmps_lo)
1095    get_sgpr_size_bytes(s_restore_ttmps_hi)
1096    s_add_u32	    s_restore_ttmps_lo, s_restore_ttmps_lo, s_restore_ttmps_hi
1097    s_add_u32	    s_restore_ttmps_lo, s_restore_ttmps_lo, s_restore_buf_rsrc0
1098    s_addc_u32	    s_restore_ttmps_hi, s_restore_buf_rsrc1, 0x0
1099    s_and_b32	    s_restore_ttmps_hi, s_restore_ttmps_hi, 0xFFFF
1100    s_load_dwordx2  [ttmp6, ttmp7], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x40 glc:1
1101    s_load_dwordx4  [ttmp8, ttmp9, ttmp10, ttmp11], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x48 glc:1
1102    s_load_dword    ttmp13, [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x58 glc:1
1103    s_load_dwordx2  [ttmp14, ttmp15], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x5C glc:1
1104    s_waitcnt	    lgkmcnt(0)
1105
1106    //reuse s_restore_m0 as a temp register
1107    s_and_b32	    s_restore_m0, s_restore_pc_hi, S_SAVE_PC_HI_RCNT_MASK
1108    s_lshr_b32	    s_restore_m0, s_restore_m0, S_SAVE_PC_HI_RCNT_SHIFT
1109    s_lshl_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_RCNT_SHIFT
1110    s_mov_b32	    s_restore_tmp, 0x0										    //IB_STS is zero
1111    s_or_b32	    s_restore_tmp, s_restore_tmp, s_restore_m0
1112    s_and_b32	    s_restore_m0, s_restore_pc_hi, S_SAVE_PC_HI_FIRST_REPLAY_MASK
1113    s_lshr_b32	    s_restore_m0, s_restore_m0, S_SAVE_PC_HI_FIRST_REPLAY_SHIFT
1114    s_lshl_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_IB_STS_FIRST_REPLAY_SHIFT
1115    s_or_b32	    s_restore_tmp, s_restore_tmp, s_restore_m0
1116    s_and_b32	    s_restore_m0, s_restore_status, SQ_WAVE_STATUS_INST_ATC_MASK
1117    s_lshr_b32	    s_restore_m0, s_restore_m0, SQ_WAVE_STATUS_INST_ATC_SHIFT
1118    s_setreg_b32    hwreg(HW_REG_IB_STS),   s_restore_tmp
1119
1120    s_and_b32 s_restore_pc_hi, s_restore_pc_hi, 0x0000ffff	//pc[47:32]	   //Do it here in order not to affect STATUS
1121    s_and_b64	 exec, exec, exec  // Restore STATUS.EXECZ, not writable by s_setreg_b32
1122    s_and_b64	 vcc, vcc, vcc	// Restore STATUS.VCCZ, not writable by s_setreg_b32
1123    s_setreg_b32    hwreg(HW_REG_STATUS),   s_restore_status	 // SCC is included, which is changed by previous salu
1124
1125    s_barrier							//barrier to ensure the readiness of LDS before access attempts from any other wave in the same TG //FIXME not performance-optimal at this time
1126
1127if G8SR_DEBUG_TIMESTAMP
1128    s_memrealtime s_g8sr_ts_restore_d
1129    s_waitcnt lgkmcnt(0)
1130end
1131
1132//  s_rfe_b64 s_restore_pc_lo					//Return to the main shader program and resume execution
1133    s_rfe_restore_b64  s_restore_pc_lo, s_restore_m0		// s_restore_m0[0] is used to set STATUS.inst_atc
1134
1135
1136/**************************************************************************/
1137/*			the END						  */
1138/**************************************************************************/
1139L_END_PGM:
1140    s_endpgm
1141
1142end
1143
1144
1145/**************************************************************************/
1146/*			the helper functions				  */
1147/**************************************************************************/
1148
1149//Only for save hwreg to mem
1150function write_hwreg_to_mem(s, s_rsrc, s_mem_offset)
1151	s_mov_b32 exec_lo, m0			//assuming exec_lo is not needed anymore from this point on
1152	s_mov_b32 m0, s_mem_offset
1153	s_buffer_store_dword s, s_rsrc, m0	glc:1
1154	ack_sqc_store_workaround()
1155	s_add_u32	s_mem_offset, s_mem_offset, 4
1156	s_mov_b32   m0, exec_lo
1157end
1158
1159
1160// HWREG are saved before SGPRs, so all HWREG could be use.
1161function write_16sgpr_to_mem(s, s_rsrc, s_mem_offset)
1162
1163	s_buffer_store_dwordx4 s[0], s_rsrc, 0	glc:1
1164	ack_sqc_store_workaround()
1165	s_buffer_store_dwordx4 s[4], s_rsrc, 16	 glc:1
1166	ack_sqc_store_workaround()
1167	s_buffer_store_dwordx4 s[8], s_rsrc, 32	 glc:1
1168	ack_sqc_store_workaround()
1169	s_buffer_store_dwordx4 s[12], s_rsrc, 48 glc:1
1170	ack_sqc_store_workaround()
1171	s_add_u32	s_rsrc[0], s_rsrc[0], 4*16
1172	s_addc_u32	s_rsrc[1], s_rsrc[1], 0x0	      // +scc
1173end
1174
1175
1176function read_hwreg_from_mem(s, s_rsrc, s_mem_offset)
1177    s_buffer_load_dword s, s_rsrc, s_mem_offset	    glc:1
1178    s_add_u32	    s_mem_offset, s_mem_offset, 4
1179end
1180
1181function read_16sgpr_from_mem(s, s_rsrc, s_mem_offset)
1182    s_buffer_load_dwordx16 s, s_rsrc, s_mem_offset	glc:1
1183    s_sub_u32	    s_mem_offset, s_mem_offset, 4*16
1184end
1185
1186
1187
1188function get_lds_size_bytes(s_lds_size_byte)
1189    // SQ LDS granularity is 64DW, while PGM_RSRC2.lds_size is in granularity 128DW
1190    s_getreg_b32   s_lds_size_byte, hwreg(HW_REG_LDS_ALLOC, SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT, SQ_WAVE_LDS_ALLOC_LDS_SIZE_SIZE)		// lds_size
1191    s_lshl_b32	   s_lds_size_byte, s_lds_size_byte, 8			    //LDS size in dwords = lds_size * 64 *4Bytes    // granularity 64DW
1192end
1193
1194function get_vgpr_size_bytes(s_vgpr_size_byte)
1195    s_getreg_b32   s_vgpr_size_byte, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_VGPR_SIZE_SIZE)	 //vpgr_size
1196    s_add_u32	   s_vgpr_size_byte, s_vgpr_size_byte, 1
1197    s_lshl_b32	   s_vgpr_size_byte, s_vgpr_size_byte, (2+8) //Number of VGPRs = (vgpr_size + 1) * 4 * 64 * 4	(non-zero value)   //FIXME for GFX, zero is possible
1198end
1199
1200function get_sgpr_size_bytes(s_sgpr_size_byte)
1201    s_getreg_b32   s_sgpr_size_byte, hwreg(HW_REG_GPR_ALLOC,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SHIFT,SQ_WAVE_GPR_ALLOC_SGPR_SIZE_SIZE)	 //spgr_size
1202    s_add_u32	   s_sgpr_size_byte, s_sgpr_size_byte, 1
1203    s_lshl_b32	   s_sgpr_size_byte, s_sgpr_size_byte, 6 //Number of SGPRs = (sgpr_size + 1) * 16 *4   (non-zero value)
1204end
1205
1206function get_hwreg_size_bytes
1207    return 128 //HWREG size 128 bytes
1208end
1209
1210function ack_sqc_store_workaround
1211    if ACK_SQC_STORE
1212        s_waitcnt lgkmcnt(0)
1213    end
1214end
1215