[mirror_edk2.git] / EdkCompatibilityPkg / Foundation / Library / Pei / PeiLib / Ipf / SetJmp.s

//++\r
// Copyright (c) 2004, Intel Corporation. All rights reserved.<BR>\r
// This program and the accompanying materials                          \r
// are licensed and made available under the terms and conditions of the BSD License         \r
// which accompanies this distribution.  The full text of the license may be found at        \r
// http://opensource.org/licenses/bsd-license.php                                            \r
//                                                                                           \r
// THE PROGRAM IS DISTRIBUTED UNDER THE BSD LICENSE ON AN "AS IS" BASIS,                     \r
// WITHOUT WARRANTIES OR REPRESENTATIONS OF ANY KIND, EITHER EXPRESS OR IMPLIED.             \r
// \r
// Module Name:\r
//\r
//  setjmp.s\r
//\r
// Abstract:\r
//\r
//  Contains an implementation of setjmp and longjmp for the\r
//  Itanium-based architecture.\r
//\r
//\r
//\r
// Revision History:\r
//\r
//--\r
\r
  .file  "setjmp.s"\r
\r
#include  <Asm.h>\r
#include  <Ia_64Gen.h>\r
\r
// int SetJump(struct jmp_buffer save)\r
//\r
//  Setup a non-local goto.\r
//\r
// Description:\r
//\r
//  SetJump stores the current register set in the area pointed to\r
//  by "save".  It returns zero.  Subsequent calls to "LongJump" will\r
//  restore the registers and return non-zero to the same location.\r
//\r
// On entry, r32 contains the pointer to the jmp_buffer\r
//\r
\r
PROCEDURE_ENTRY(SetJump)\r
  //\r
  //  Make sure buffer is aligned at 16byte boundary\r
  //\r
  mov    r32 = r33        \r
\r
    add     r10 = -0x10,r0  ;;  // mask the lower 4 bits\r
    and     r32 = r32, r10;; \r
    add     r32 = 0x10, r32;;   // move to next 16 byte boundary\r
\r
    add    r10 = J_PREDS, r32  // skip Unats & pfs save area\r
  add    r11 = J_BSP, r32\r
  //\r
  //  save immediate context\r
  //\r
  mov    r2 = ar.bsp     // save backing store pointer\r
  mov    r3 = pr         // save predicates\r
  ;;\r
  //\r
  // save user Unat register\r
  //\r
  mov    r16 = ar.lc        // save loop count register\r
  mov    r14 = ar.unat     // save user Unat register\r
\r
  st8    [r10] = r3, J_LC-J_PREDS\r
  st8    [r11] = r2, J_R4-J_BSP\r
  ;;\r
  st8    [r10] = r16, J_R5-J_LC\r
  st8    [r32] = r14, J_NATS  // Note: Unat at the \r
                    // beginning of the save area\r
  mov    r15 = ar.pfs\r
  ;;\r
  //\r
  //  save preserved general registers & NaT's\r
  //\r
  st8.spill  [r11] = r4, J_R6-J_R4\r
  ;;\r
  st8.spill  [r10] = r5, J_R7-J_R5 \r
    ;;\r
  st8.spill  [r11] = r6, J_SP-J_R6\r
  ;;\r
  st8.spill  [r10] = r7, J_F3-J_R7 \r
    ;;\r
  st8.spill  [r11] = sp, J_F2-J_SP\r
  ;;\r
  //\r
  // save spilled Unat and pfs registers\r
  //\r
  mov    r2 = ar.unat       // save Unat register after spill\r
  ;;\r
  st8    [r32] = r2, J_PFS-J_NATS  // save unat for spilled regs\r
  ;;\r
  st8    [r32] = r15          // save pfs\r
  //\r
  //  save floating registers \r
  //\r
  stf.spill  [r11] = f2, J_F4-J_F2\r
  stf.spill  [r10] = f3, J_F5-J_F3 \r
  ;;\r
  stf.spill  [r11] = f4, J_F16-J_F4\r
  stf.spill  [r10] = f5, J_F17-J_F5 \r
  ;;\r
  stf.spill  [r11] = f16, J_F18-J_F16\r
  stf.spill  [r10] = f17, J_F19-J_F17 \r
  ;;\r
  stf.spill  [r11] = f18, J_F20-J_F18\r
  stf.spill  [r10] = f19, J_F21-J_F19 \r
  ;;\r
  stf.spill  [r11] = f20, J_F22-J_F20\r
  stf.spill  [r10] = f21, J_F23-J_F21 \r
  ;;\r
  stf.spill  [r11] = f22, J_F24-J_F22\r
  stf.spill  [r10] = f23, J_F25-J_F23 \r
  ;;\r
  stf.spill  [r11] = f24, J_F26-J_F24\r
  stf.spill  [r10] = f25, J_F27-J_F25 \r
  ;;\r
  stf.spill  [r11] = f26, J_F28-J_F26\r
  stf.spill  [r10] = f27, J_F29-J_F27 \r
  ;;\r
  stf.spill  [r11] = f28, J_F30-J_F28\r
  stf.spill  [r10] = f29, J_F31-J_F29 \r
  ;;\r
  stf.spill  [r11] = f30, J_FPSR-J_F30\r
  stf.spill  [r10] = f31, J_B0-J_F31    // size of f31 + fpsr\r
  //\r
  // save FPSR register & branch registers\r
  //\r
  mov    r2 = ar.fpsr  // save fpsr register\r
  mov    r3 = b0 \r
  ;;\r
  st8    [r11] = r2, J_B1-J_FPSR\r
  st8    [r10] = r3, J_B2-J_B0\r
  mov    r2 = b1\r
  mov    r3 = b2 \r
  ;;\r
  st8    [r11] = r2, J_B3-J_B1\r
  st8    [r10] = r3, J_B4-J_B2\r
  mov    r2 = b3\r
  mov    r3 = b4 \r
  ;;\r
  st8    [r11] = r2, J_B5-J_B3\r
  st8    [r10] = r3\r
  mov    r2 = b5 \r
  ;;\r
  st8    [r11] = r2\r
  ;;\r
  //\r
  // return\r
  //\r
  mov    r8 = r0        // return 0 from setjmp\r
  mov    ar.unat = r14  // restore unat\r
  br.ret.sptk  b0\r
\r
PROCEDURE_EXIT(SetJump)\r
\r
\r
//\r
// void LongJump(struct jmp_buffer *)\r
//\r
//  Perform a non-local goto.\r
//\r
// Description:\r
//\r
//  LongJump initializes the register set to the values saved by a\r
//  previous 'SetJump' and jumps to the return location saved by that\r
//  'SetJump'.  This has the effect of unwinding the stack and returning\r
//  for a second time to the 'SetJump'.\r
//\r
\r
PROCEDURE_ENTRY(LongJump)\r
  //\r
  //  Make sure buffer is aligned at 16byte boundary\r
  //\r
  mov    r32 = r33        \r
\r
    add     r10 = -0x10,r0  ;;  // mask the lower 4 bits\r
    and     r32 = r32, r10;; \r
    add     r32 = 0x10, r32;;   // move to next 16 byte boundary\r
\r
  //\r
  // caching the return value as we do invala in the end\r
  //\r
///  mov    r8 = r33          // return value\r
  mov    r8 = 1              // For now return hard coded 1\r
\r
  //\r
  //  get immediate context\r
  //\r
  mov    r14 = ar.rsc    // get user RSC conf \r
  add    r10 = J_PFS, r32  // get address of pfs\r
  add    r11 = J_NATS, r32\r
  ;;\r
  ld8    r15 = [r10], J_BSP-J_PFS  // get pfs\r
  ld8    r2 = [r11], J_LC-J_NATS      // get unat for spilled regs\r
  ;;\r
  mov    ar.unat = r2\r
  ;;\r
  ld8    r16 = [r10], J_PREDS-J_BSP  // get backing store pointer\r
  mov    ar.rsc = r0        // put RSE in enforced lazy \r
  mov    ar.pfs = r15\r
  ;;\r
  \r
  //\r
  // while returning from longjmp the BSPSTORE and BSP needs to be\r
  // same and discard all the registers allocated after we did\r
  // setjmp. Also, we need to generate the RNAT register since we\r
  // did not flushed the RSE on setjmp.\r
  //\r
  mov    r17 = ar.bspstore  // get current BSPSTORE\r
  ;;\r
    cmp.ltu  p6,p7 = r17, r16  // is it less than BSP of \r
(p6)  br.spnt.few  .flush_rse\r
  mov    r19 = ar.rnat    // get current RNAT\r
  ;;\r
  loadrs              // invalidate dirty regs\r
  br.sptk.many  .restore_rnat    // restore RNAT\r
\r
.flush_rse:\r
  flushrs\r
  ;;\r
  mov    r19 = ar.rnat    // get current RNAT\r
  mov    r17 = r16      // current BSPSTORE\r
  ;;\r
.restore_rnat:\r
  //\r
  // check if RNAT is saved between saved BSP and curr BSPSTORE\r
  //\r
  dep    r18 = -1,r16,3,6   // get RNAT address\r
  ;;\r
  cmp.ltu  p8,p9 = r18, r17  // RNAT saved on RSE\r
  ;;\r
(p8)  ld8    r19 = [r18]    // get RNAT from RSE\r
  ;;\r
  mov    ar.bspstore = r16  // set new BSPSTORE  \r
  ;;\r
  mov    ar.rnat = r19    // restore RNAT\r
  mov    ar.rsc = r14    // restore RSC conf\r
\r
\r
  ld8    r3 = [r11], J_R4-J_LC    // get lc register\r
  ld8    r2 = [r10], J_R5-J_PREDS   // get predicates\r
  ;;\r
  mov    pr = r2, -1\r
  mov    ar.lc = r3\r
  //\r
  //  restore preserved general registers & NaT's\r
  //\r
  ld8.fill  r4 = [r11], J_R6-J_R4\r
  ;;\r
  ld8.fill  r5 = [r10], J_R7-J_R5 \r
  ld8.fill  r6 = [r11], J_SP-J_R6\r
  ;;\r
  ld8.fill  r7 = [r10], J_F2-J_R7\r
  ld8.fill  sp = [r11], J_F3-J_SP\r
  ;;\r
  //\r
  //  restore floating registers \r
  //\r
  ldf.fill  f2 = [r10], J_F4-J_F2\r
  ldf.fill  f3 = [r11], J_F5-J_F3 \r
  ;;\r
  ldf.fill  f4 = [r10], J_F16-J_F4\r
  ldf.fill  f5 = [r11], J_F17-J_F5 \r
  ;;\r
  ldf.fill  f16 = [r10], J_F18-J_F16\r
  ldf.fill  f17 = [r11], J_F19-J_F17\r
  ;;\r
  ldf.fill  f18 = [r10], J_F20-J_F18\r
  ldf.fill  f19 = [r11], J_F21-J_F19\r
  ;;\r
  ldf.fill  f20 = [r10], J_F22-J_F20\r
  ldf.fill  f21 = [r11], J_F23-J_F21\r
  ;;\r
  ldf.fill  f22 = [r10], J_F24-J_F22\r
  ldf.fill  f23 = [r11], J_F25-J_F23 \r
  ;;\r
  ldf.fill  f24 = [r10], J_F26-J_F24\r
  ldf.fill  f25 = [r11], J_F27-J_F25\r
  ;;\r
  ldf.fill  f26 = [r10], J_F28-J_F26\r
  ldf.fill  f27 = [r11], J_F29-J_F27\r
  ;;\r
  ldf.fill  f28 = [r10], J_F30-J_F28\r
  ldf.fill  f29 = [r11], J_F31-J_F29 \r
  ;;\r
  ldf.fill  f30 = [r10], J_FPSR-J_F30\r
  ldf.fill  f31 = [r11], J_B0-J_F31 ;;\r
\r
    //\r
  // restore branch registers and fpsr\r
  //\r
  ld8    r16 = [r10], J_B1-J_FPSR  // get fpsr\r
  ld8    r17 = [r11], J_B2-J_B0    // get return pointer\r
  ;;\r
  mov    ar.fpsr = r16\r
  mov    b0 = r17\r
  ld8    r2 = [r10], J_B3-J_B1\r
  ld8    r3 = [r11], J_B4-J_B2\r
  ;;\r
  mov    b1 = r2\r
  mov    b2 = r3\r
  ld8    r2 = [r10], J_B5-J_B3\r
  ld8    r3 = [r11]\r
  ;;\r
  mov    b3 = r2\r
  mov    b4 = r3 \r
  ld8    r2 = [r10]\r
  ld8    r21 = [r32]      // get user unat\r
  ;;\r
  mov    b5 = r2\r
  mov    ar.unat = r21\r
\r
  //\r
  // invalidate ALAT\r
  //\r
  invala ;;\r
\r
  br.ret.sptk  b0\r
PROCEDURE_EXIT(LongJump)\r
\r
\r
Commit	Line	Data
3eb9473e	1	//++\r
4ea9375a HT	2	// Copyright (c) 2004, Intel Corporation. All rights reserved.<BR>\r
4ea9375a HT	3	// This program and the accompanying materials \r
3eb9473e	4	// are licensed and made available under the terms and conditions of the BSD License \r
	5	// which accompanies this distribution. The full text of the license may be found at \r
	6	// http://opensource.org/licenses/bsd-license.php \r
	7	// \r
	8	// THE PROGRAM IS DISTRIBUTED UNDER THE BSD LICENSE ON AN "AS IS" BASIS, \r
	9	// WITHOUT WARRANTIES OR REPRESENTATIONS OF ANY KIND, EITHER EXPRESS OR IMPLIED. \r
	10	// \r
	11	// Module Name:\r
	12	//\r
	13	// setjmp.s\r
	14	//\r
	15	// Abstract:\r
	16	//\r
	17	// Contains an implementation of setjmp and longjmp for the\r
	18	// Itanium-based architecture.\r
	19	//\r
	20	//\r
	21	//\r
	22	// Revision History:\r
	23	//\r
	24	//--\r
	25	\r
	26	.file "setjmp.s"\r
	27	\r
bb4af292	28	#include <Asm.h>\r
bb4af292	29	#include <Ia_64Gen.h>\r
3eb9473e	30	\r
	31	// int SetJump(struct jmp_buffer save)\r
	32	//\r
	33	// Setup a non-local goto.\r
	34	//\r
	35	// Description:\r
	36	//\r
	37	// SetJump stores the current register set in the area pointed to\r
	38	// by "save". It returns zero. Subsequent calls to "LongJump" will\r
	39	// restore the registers and return non-zero to the same location.\r
	40	//\r
	41	// On entry, r32 contains the pointer to the jmp_buffer\r
	42	//\r
	43	\r
	44	PROCEDURE_ENTRY(SetJump)\r
	45	//\r
	46	// Make sure buffer is aligned at 16byte boundary\r
	47	//\r
	48	mov r32 = r33 \r
	49	\r
	50	add r10 = -0x10,r0 ;; // mask the lower 4 bits\r
	51	and r32 = r32, r10;; \r
	52	add r32 = 0x10, r32;; // move to next 16 byte boundary\r
	53	\r
	54	add r10 = J_PREDS, r32 // skip Unats & pfs save area\r
	55	add r11 = J_BSP, r32\r
	56	//\r
	57	// save immediate context\r
	58	//\r
	59	mov r2 = ar.bsp // save backing store pointer\r
	60	mov r3 = pr // save predicates\r
	61	;;\r
	62	//\r
	63	// save user Unat register\r
	64	//\r
	65	mov r16 = ar.lc // save loop count register\r
	66	mov r14 = ar.unat // save user Unat register\r
	67	\r
	68	st8 [r10] = r3, J_LC-J_PREDS\r
	69	st8 [r11] = r2, J_R4-J_BSP\r
	70	;;\r
	71	st8 [r10] = r16, J_R5-J_LC\r
	72	st8 [r32] = r14, J_NATS // Note: Unat at the \r
	73	// beginning of the save area\r
	74	mov r15 = ar.pfs\r
	75	;;\r
	76	//\r
	77	// save preserved general registers & NaT's\r
	78	//\r
	79	st8.spill [r11] = r4, J_R6-J_R4\r
	80	;;\r
	81	st8.spill [r10] = r5, J_R7-J_R5 \r
	82	;;\r
	83	st8.spill [r11] = r6, J_SP-J_R6\r
	84	;;\r
	85	st8.spill [r10] = r7, J_F3-J_R7 \r
	86	;;\r
	87	st8.spill [r11] = sp, J_F2-J_SP\r
	88	;;\r
	89	//\r
	90	// save spilled Unat and pfs registers\r
	91	//\r
	92	mov r2 = ar.unat // save Unat register after spill\r
	93	;;\r
94	st8 [r32] = r2, J_PFS-J_NATS // save unat for spilled regs\r
95	;;\r
96	st8 [r32] = r15 // save pfs\r
97	//\r
98	// save floating registers \r
99	//\r
100	stf.spill [r11] = f2, J_F4-J_F2\r
101	stf.spill [r10] = f3, J_F5-J_F3 \r
102	;;\r
103	stf.spill [r11] = f4, J_F16-J_F4\r
104	stf.spill [r10] = f5, J_F17-J_F5 \r
105	;;\r
106	stf.spill [r11] = f16, J_F18-J_F16\r
107	stf.spill [r10] = f17, J_F19-J_F17 \r
108	;;\r
109	stf.spill [r11] = f18, J_F20-J_F18\r
110	stf.spill [r10] = f19, J_F21-J_F19 \r
111	;;\r
112	stf.spill [r11] = f20, J_F22-J_F20\r
113	stf.spill [r10] = f21, J_F23-J_F21 \r
114	;;\r
115	stf.spill [r11] = f22, J_F24-J_F22\r
116	stf.spill [r10] = f23, J_F25-J_F23 \r
117	;;\r
118	stf.spill [r11] = f24, J_F26-J_F24\r
119	stf.spill [r10] = f25, J_F27-J_F25 \r
120	;;\r
121	stf.spill [r11] = f26, J_F28-J_F26\r
122	stf.spill [r10] = f27, J_F29-J_F27 \r
123	;;\r
124	stf.spill [r11] = f28, J_F30-J_F28\r
125	stf.spill [r10] = f29, J_F31-J_F29 \r
126	;;\r
127	stf.spill [r11] = f30, J_FPSR-J_F30\r
128	stf.spill [r10] = f31, J_B0-J_F31 // size of f31 + fpsr\r
129	//\r
130	// save FPSR register & branch registers\r
131	//\r
132	mov r2 = ar.fpsr // save fpsr register\r
133	mov r3 = b0 \r
134	;;\r
135	st8 [r11] = r2, J_B1-J_FPSR\r
136	st8 [r10] = r3, J_B2-J_B0\r
137	mov r2 = b1\r
138	mov r3 = b2 \r
139	;;\r
140	st8 [r11] = r2, J_B3-J_B1\r
141	st8 [r10] = r3, J_B4-J_B2\r
142	mov r2 = b3\r
143	mov r3 = b4 \r
144	;;\r
145	st8 [r11] = r2, J_B5-J_B3\r
146	st8 [r10] = r3\r
147	mov r2 = b5 \r
148	;;\r
149	st8 [r11] = r2\r
150	;;\r
151	//\r
152	// return\r
153	//\r
154	mov r8 = r0 // return 0 from setjmp\r
155	mov ar.unat = r14 // restore unat\r
156	br.ret.sptk b0\r
157	\r
158	PROCEDURE_EXIT(SetJump)\r
159	\r
160	\r
161	//\r
162	// void LongJump(struct jmp_buffer *)\r
163	//\r
164	// Perform a non-local goto.\r
165	//\r
166	// Description:\r
167	//\r
168	// LongJump initializes the register set to the values saved by a\r
169	// previous 'SetJump' and jumps to the return location saved by that\r
170	// 'SetJump'. This has the effect of unwinding the stack and returning\r
171	// for a second time to the 'SetJump'.\r
172	//\r
173	\r
174	PROCEDURE_ENTRY(LongJump)\r
175	//\r
176	// Make sure buffer is aligned at 16byte boundary\r
177	//\r
178	mov r32 = r33 \r
179	\r
180	add r10 = -0x10,r0 ;; // mask the lower 4 bits\r
181	and r32 = r32, r10;; \r
182	add r32 = 0x10, r32;; // move to next 16 byte boundary\r
183	\r
184	//\r
185	// caching the return value as we do invala in the end\r
186	//\r
187	/// mov r8 = r33 // return value\r
188	mov r8 = 1 // For now return hard coded 1\r
189	\r
190	//\r
191	// get immediate context\r
192	//\r
193	mov r14 = ar.rsc // get user RSC conf \r
194	add r10 = J_PFS, r32 // get address of pfs\r
195	add r11 = J_NATS, r32\r
196	;;\r
197	ld8 r15 = [r10], J_BSP-J_PFS // get pfs\r
198	ld8 r2 = [r11], J_LC-J_NATS // get unat for spilled regs\r
199	;;\r
200	mov ar.unat = r2\r
201	;;\r
202	ld8 r16 = [r10], J_PREDS-J_BSP // get backing store pointer\r
203	mov ar.rsc = r0 // put RSE in enforced lazy \r
204	mov ar.pfs = r15\r
205	;;\r
206	\r
207	//\r
208	// while returning from longjmp the BSPSTORE and BSP needs to be\r
209	// same and discard all the registers allocated after we did\r
210	// setjmp. Also, we need to generate the RNAT register since we\r
211	// did not flushed the RSE on setjmp.\r
212	//\r
213	mov r17 = ar.bspstore // get current BSPSTORE\r
214	;;\r
215	cmp.ltu p6,p7 = r17, r16 // is it less than BSP of \r
216	(p6) br.spnt.few .flush_rse\r
217	mov r19 = ar.rnat // get current RNAT\r
218	;;\r
219	loadrs // invalidate dirty regs\r
220	br.sptk.many .restore_rnat // restore RNAT\r
221	\r
222	.flush_rse:\r
223	flushrs\r
224	;;\r
225	mov r19 = ar.rnat // get current RNAT\r
226	mov r17 = r16 // current BSPSTORE\r
227	;;\r
228	.restore_rnat:\r
229	//\r
230	// check if RNAT is saved between saved BSP and curr BSPSTORE\r
231	//\r
88b6bcec	232	dep r18 = -1,r16,3,6 // get RNAT address\r
3eb9473e	233	;;\r
	234	cmp.ltu p8,p9 = r18, r17 // RNAT saved on RSE\r
	235	;;\r
	236	(p8) ld8 r19 = [r18] // get RNAT from RSE\r
	237	;;\r
	238	mov ar.bspstore = r16 // set new BSPSTORE \r
	239	;;\r
	240	mov ar.rnat = r19 // restore RNAT\r
	241	mov ar.rsc = r14 // restore RSC conf\r
	242	\r
	243	\r
	244	ld8 r3 = [r11], J_R4-J_LC // get lc register\r
	245	ld8 r2 = [r10], J_R5-J_PREDS // get predicates\r
	246	;;\r
	247	mov pr = r2, -1\r
	248	mov ar.lc = r3\r
	249	//\r
	250	// restore preserved general registers & NaT's\r
	251	//\r
	252	ld8.fill r4 = [r11], J_R6-J_R4\r
	253	;;\r
	254	ld8.fill r5 = [r10], J_R7-J_R5 \r
	255	ld8.fill r6 = [r11], J_SP-J_R6\r
	256	;;\r
	257	ld8.fill r7 = [r10], J_F2-J_R7\r
	258	ld8.fill sp = [r11], J_F3-J_SP\r
	259	;;\r
	260	//\r
	261	// restore floating registers \r
	262	//\r
	263	ldf.fill f2 = [r10], J_F4-J_F2\r
	264	ldf.fill f3 = [r11], J_F5-J_F3 \r
	265	;;\r
	266	ldf.fill f4 = [r10], J_F16-J_F4\r
	267	ldf.fill f5 = [r11], J_F17-J_F5 \r
	268	;;\r
	269	ldf.fill f16 = [r10], J_F18-J_F16\r
	270	ldf.fill f17 = [r11], J_F19-J_F17\r
	271	;;\r
	272	ldf.fill f18 = [r10], J_F20-J_F18\r
	273	ldf.fill f19 = [r11], J_F21-J_F19\r
	274	;;\r
	275	ldf.fill f20 = [r10], J_F22-J_F20\r
	276	ldf.fill f21 = [r11], J_F23-J_F21\r
	277	;;\r
	278	ldf.fill f22 = [r10], J_F24-J_F22\r
	279	ldf.fill f23 = [r11], J_F25-J_F23 \r
	280	;;\r
	281	ldf.fill f24 = [r10], J_F26-J_F24\r
	282	ldf.fill f25 = [r11], J_F27-J_F25\r
	283	;;\r
	284	ldf.fill f26 = [r10], J_F28-J_F26\r
	285	ldf.fill f27 = [r11], J_F29-J_F27\r
	286	;;\r
	287	ldf.fill f28 = [r10], J_F30-J_F28\r
	288	ldf.fill f29 = [r11], J_F31-J_F29 \r
	289	;;\r
	290	ldf.fill f30 = [r10], J_FPSR-J_F30\r
	291	ldf.fill f31 = [r11], J_B0-J_F31 ;;\r
	292	\r
	293	//\r
	294	// restore branch registers and fpsr\r
	295	//\r
	296	ld8 r16 = [r10], J_B1-J_FPSR // get fpsr\r
297	ld8 r17 = [r11], J_B2-J_B0 // get return pointer\r
298	;;\r
299	mov ar.fpsr = r16\r
300	mov b0 = r17\r
301	ld8 r2 = [r10], J_B3-J_B1\r
302	ld8 r3 = [r11], J_B4-J_B2\r
303	;;\r
304	mov b1 = r2\r
305	mov b2 = r3\r
306	ld8 r2 = [r10], J_B5-J_B3\r
307	ld8 r3 = [r11]\r
308	;;\r
309	mov b3 = r2\r
310	mov b4 = r3 \r
311	ld8 r2 = [r10]\r
312	ld8 r21 = [r32] // get user unat\r
313	;;\r
314	mov b5 = r2\r
315	mov ar.unat = r21\r
316	\r
317	//\r
318	// invalidate ALAT\r
319	//\r
320	invala ;;\r
321	\r
322	br.ret.sptk b0\r
323	PROCEDURE_EXIT(LongJump)\r
324	\r
325	\r