drm/amdkfd: Save/restore vcc on gfx10

VCC moved out of user SGPR allocation in gfx10. It's now stored
in SGPRs 106-107.

Also fixes incorrect SGPR read offsets.

Cc: Shaoyun Liu <shaoyun.liu@amd.com>
Signed-off-by: Jay Cornwall <jay.cornwall@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: shaoyunl <shaoyun.liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h
index 8089bb3..a8cf82d 100644 (file)
--- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h
+++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h
@@ -680,7 +680,7 @@ static const uint32_t cwsr_trap_gfx9_hex[] = {
 };
 
 static const uint32_t cwsr_trap_gfx10_hex[] = {
-       0xbf820001, 0xbf8201c0,
+       0xbf820001, 0xbf8201c1,
        0xb0804004, 0xb978f802,
        0x8a788678, 0xb971f803,
        0x876eff71, 0x00000400,
@@ -804,271 +804,271 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
        0x00000060, 0xbf85ffea,
        0xbe802f00, 0xbe822f02,
        0xbe842f04, 0xbe862f06,
-       0xbe882f08, 0xf469003a,
-       0xfa000000, 0xf469013a,
-       0xfa000010, 0xf465023a,
-       0xfa000020, 0x8074c074,
-       0x82758075, 0xbef40372,
-       0xbefa0380, 0xbefe03c1,
-       0x907c9973, 0x877c817c,
-       0xbf06817c, 0xbf850002,
-       0xbeff0380, 0xbf820002,
-       0xbeff03c1, 0xbf82000b,
+       0xbe882f08, 0xbe8a2f0a,
+       0xf469003a, 0xfa000000,
+       0xf469013a, 0xfa000010,
+       0xf469023a, 0xfa000020,
+       0x8074b074, 0x82758075,
+       0xbef40372, 0xbefa0380,
+       0xbefe03c1, 0x907c9973,
+       0x877c817c, 0xbf06817c,
+       0xbf850002, 0xbeff0380,
+       0xbf820002, 0xbeff03c1,
+       0xbf82000b, 0xbef603ff,
+       0x01000000, 0xe0704000,
+       0x7a5d0000, 0xe0704080,
+       0x7a5d0100, 0xe0704100,
+       0x7a5d0200, 0xe0704180,
+       0x7a5d0300, 0xbf82000a,
        0xbef603ff, 0x01000000,
        0xe0704000, 0x7a5d0000,
-       0xe0704080, 0x7a5d0100,
-       0xe0704100, 0x7a5d0200,
-       0xe0704180, 0x7a5d0300,
-       0xbf82000a, 0xbef603ff,
-       0x01000000, 0xe0704000,
-       0x7a5d0000, 0xe0704100,
-       0x7a5d0100, 0xe0704200,
-       0x7a5d0200, 0xe0704300,
-       0x7a5d0300, 0xbefe03c1,
+       0xe0704100, 0x7a5d0100,
+       0xe0704200, 0x7a5d0200,
+       0xe0704300, 0x7a5d0300,
+       0xbefe03c1, 0x907c9973,
+       0x877c817c, 0xbf06817c,
+       0xbf850002, 0xbeff0380,
+       0xbf820001, 0xbeff03c1,
+       0xb9714306, 0x8771c171,
+       0xbf840046, 0xbf8a0000,
+       0x8776ff6f, 0x04000000,
+       0xbf840042, 0x8f718671,
+       0x8f718271, 0xbef60371,
+       0xb97a2a05, 0x807a817a,
        0x907c9973, 0x877c817c,
        0xbf06817c, 0xbf850002,
-       0xbeff0380, 0xbf820001,
-       0xbeff03c1, 0xb9714306,
-       0x8771c171, 0xbf840046,
-       0xbf8a0000, 0x8776ff6f,
-       0x04000000, 0xbf840042,
-       0x8f718671, 0x8f718271,
-       0xbef60371, 0xb97a2a05,
-       0x807a817a, 0x907c9973,
+       0x8f7a897a, 0xbf820001,
+       0x8f7a8a7a, 0xb9761e06,
+       0x8f768a76, 0x807a767a,
+       0x807aff7a, 0x00000200,
+       0x807aff7a, 0x00000080,
+       0xbef603ff, 0x01000000,
+       0xd7650000, 0x000100c1,
+       0xd7660000, 0x000200c1,
+       0x16000084, 0x907c9973,
        0x877c817c, 0xbf06817c,
-       0xbf850002, 0x8f7a897a,
-       0xbf820001, 0x8f7a8a7a,
-       0xb9761e06, 0x8f768a76,
-       0x807a767a, 0x807aff7a,
-       0x00000200, 0x807aff7a,
-       0x00000080, 0xbef603ff,
-       0x01000000, 0xd7650000,
-       0x000100c1, 0xd7660000,
-       0x000200c1, 0x16000084,
-       0x907c9973, 0x877c817c,
-       0xbf06817c, 0xbefc0380,
-       0xbf850012, 0xbe8303ff,
-       0x00000080, 0xbf800000,
+       0xbefc0380, 0xbf850012,
+       0xbe8303ff, 0x00000080,
        0xbf800000, 0xbf800000,
-       0xd8d80000, 0x01000000,
-       0xbf8c0000, 0xe0704000,
-       0x7a5d0100, 0x807c037c,
-       0x807a037a, 0xd5250000,
-       0x0001ff00, 0x00000080,
-       0xbf0a717c, 0xbf85fff4,
-       0xbf820011, 0xbe8303ff,
-       0x00000100, 0xbf800000,
+       0xbf800000, 0xd8d80000,
+       0x01000000, 0xbf8c0000,
+       0xe0704000, 0x7a5d0100,
+       0x807c037c, 0x807a037a,
+       0xd5250000, 0x0001ff00,
+       0x00000080, 0xbf0a717c,
+       0xbf85fff4, 0xbf820011,
+       0xbe8303ff, 0x00000100,
        0xbf800000, 0xbf800000,
-       0xd8d80000, 0x01000000,
-       0xbf8c0000, 0xe0704000,
-       0x7a5d0100, 0x807c037c,
-       0x807a037a, 0xd5250000,
-       0x0001ff00, 0x00000100,
-       0xbf0a717c, 0xbf85fff4,
-       0xbefe03c1, 0x907c9973,
-       0x877c817c, 0xbf06817c,
-       0xbf850004, 0xbefa03ff,
-       0x00000200, 0xbeff0380,
-       0xbf820003, 0xbefa03ff,
-       0x00000400, 0xbeff03c1,
-       0xb9712a05, 0x80718171,
-       0x8f718271, 0x907c9973,
-       0x877c817c, 0xbf06817c,
-       0xbf850017, 0xbef603ff,
-       0x01000000, 0xbefc0384,
-       0xbf0a717c, 0xbf840037,
-       0x7e008700, 0x7e028701,
-       0x7e048702, 0x7e068703,
-       0xe0704000, 0x7a5d0000,
-       0xe0704080, 0x7a5d0100,
-       0xe0704100, 0x7a5d0200,
-       0xe0704180, 0x7a5d0300,
-       0x807c847c, 0x807aff7a,
-       0x00000200, 0xbf0a717c,
-       0xbf85ffef, 0xbf820025,
+       0xbf800000, 0xd8d80000,
+       0x01000000, 0xbf8c0000,
+       0xe0704000, 0x7a5d0100,
+       0x807c037c, 0x807a037a,
+       0xd5250000, 0x0001ff00,
+       0x00000100, 0xbf0a717c,
+       0xbf85fff4, 0xbefe03c1,
+       0x907c9973, 0x877c817c,
+       0xbf06817c, 0xbf850004,
+       0xbefa03ff, 0x00000200,
+       0xbeff0380, 0xbf820003,
+       0xbefa03ff, 0x00000400,
+       0xbeff03c1, 0xb9712a05,
+       0x80718171, 0x8f718271,
+       0x907c9973, 0x877c817c,
+       0xbf06817c, 0xbf850017,
        0xbef603ff, 0x01000000,
        0xbefc0384, 0xbf0a717c,
-       0xbf840020, 0x7e008700,
+       0xbf840037, 0x7e008700,
        0x7e028701, 0x7e048702,
        0x7e068703, 0xe0704000,
-       0x7a5d0000, 0xe0704100,
-       0x7a5d0100, 0xe0704200,
-       0x7a5d0200, 0xe0704300,
+       0x7a5d0000, 0xe0704080,
+       0x7a5d0100, 0xe0704100,
+       0x7a5d0200, 0xe0704180,
        0x7a5d0300, 0x807c847c,
-       0x807aff7a, 0x00000400,
+       0x807aff7a, 0x00000200,
        0xbf0a717c, 0xbf85ffef,
-       0xb9711e06, 0x8771c171,
-       0xbf84000c, 0x8f718371,
-       0x80717c71, 0xbefe03c1,
-       0xbeff0380, 0x7e008700,
+       0xbf820025, 0xbef603ff,
+       0x01000000, 0xbefc0384,
+       0xbf0a717c, 0xbf840020,
+       0x7e008700, 0x7e028701,
+       0x7e048702, 0x7e068703,
        0xe0704000, 0x7a5d0000,
-       0x807c817c, 0x807aff7a,
-       0x00000080, 0xbf0a717c,
-       0xbf85fff8, 0xbf820141,
-       0xbef4037e, 0x8775ff7f,
-       0x0000ffff, 0x8875ff75,
-       0x00040000, 0xbef60380,
-       0xbef703ff, 0x10807fac,
-       0x8772ff7f, 0x08000000,
-       0x90728372, 0x88777277,
-       0x8772ff7f, 0x70000000,
-       0x90728172, 0x88777277,
-       0xb97302dc, 0x8f739973,
-       0x8873737f, 0x8772ff7f,
-       0x04000000, 0xbf840036,
-       0xbefe03c1, 0x907c9973,
-       0x877c817c, 0xbf06817c,
-       0xbf850002, 0xbeff0380,
-       0xbf820001, 0xbeff03c1,
-       0xb96f4306, 0x876fc16f,
-       0xbf84002b, 0x8f6f866f,
-       0x8f6f826f, 0xbef6036f,
-       0xb9782a05, 0x80788178,
+       0xe0704100, 0x7a5d0100,
+       0xe0704200, 0x7a5d0200,
+       0xe0704300, 0x7a5d0300,
+       0x807c847c, 0x807aff7a,
+       0x00000400, 0xbf0a717c,
+       0xbf85ffef, 0xb9711e06,
+       0x8771c171, 0xbf84000c,
+       0x8f718371, 0x80717c71,
+       0xbefe03c1, 0xbeff0380,
+       0x7e008700, 0xe0704000,
+       0x7a5d0000, 0x807c817c,
+       0x807aff7a, 0x00000080,
+       0xbf0a717c, 0xbf85fff8,
+       0xbf820141, 0xbef4037e,
+       0x8775ff7f, 0x0000ffff,
+       0x8875ff75, 0x00040000,
+       0xbef60380, 0xbef703ff,
+       0x10807fac, 0x8772ff7f,
+       0x08000000, 0x90728372,
+       0x88777277, 0x8772ff7f,
+       0x70000000, 0x90728172,
+       0x88777277, 0xb97302dc,
+       0x8f739973, 0x8873737f,
+       0x8772ff7f, 0x04000000,
+       0xbf840036, 0xbefe03c1,
        0x907c9973, 0x877c817c,
        0xbf06817c, 0xbf850002,
-       0x8f788978, 0xbf820001,
-       0x8f788a78, 0xb9721e06,
-       0x8f728a72, 0x80787278,
-       0x8078ff78, 0x00000200,
-       0x8078ff78, 0x00000080,
-       0xbef603ff, 0x01000000,
-       0x907c9973, 0x877c817c,
-       0xbf06817c, 0xbefc0380,
-       0xbf850009, 0xe0310000,
-       0x781d0000, 0x807cff7c,
-       0x00000080, 0x8078ff78,
-       0x00000080, 0xbf0a6f7c,
-       0xbf85fff8, 0xbf820008,
+       0xbeff0380, 0xbf820001,
+       0xbeff03c1, 0xb96f4306,
+       0x876fc16f, 0xbf84002b,
+       0x8f6f866f, 0x8f6f826f,
+       0xbef6036f, 0xb9782a05,
+       0x80788178, 0x907c9973,
+       0x877c817c, 0xbf06817c,
+       0xbf850002, 0x8f788978,
+       0xbf820001, 0x8f788a78,
+       0xb9721e06, 0x8f728a72,
+       0x80787278, 0x8078ff78,
+       0x00000200, 0x8078ff78,
+       0x00000080, 0xbef603ff,
+       0x01000000, 0x907c9973,
+       0x877c817c, 0xbf06817c,
+       0xbefc0380, 0xbf850009,
        0xe0310000, 0x781d0000,
-       0x807cff7c, 0x00000100,
-       0x8078ff78, 0x00000100,
+       0x807cff7c, 0x00000080,
+       0x8078ff78, 0x00000080,
        0xbf0a6f7c, 0xbf85fff8,
-       0xbef80380, 0xbefe03c1,
-       0x907c9973, 0x877c817c,
-       0xbf06817c, 0xbf850002,
-       0xbeff0380, 0xbf820001,
-       0xbeff03c1, 0xb96f2a05,
-       0x806f816f, 0x8f6f826f,
-       0x907c9973, 0x877c817c,
-       0xbf06817c, 0xbf850021,
-       0xbef603ff, 0x01000000,
-       0xbef20378, 0x8078ff78,
-       0x00000200, 0xbefc0384,
-       0xe0304000, 0x785d0000,
-       0xe0304080, 0x785d0100,
-       0xe0304100, 0x785d0200,
-       0xe0304180, 0x785d0300,
-       0xbf8c3f70, 0x7e008500,
-       0x7e028501, 0x7e048502,
-       0x7e068503, 0x807c847c,
-       0x8078ff78, 0x00000200,
-       0xbf0a6f7c, 0xbf85ffee,
-       0xe0304000, 0x725d0000,
-       0xe0304080, 0x725d0100,
-       0xe0304100, 0x725d0200,
-       0xe0304180, 0x725d0300,
-       0xbf820031, 0xbef603ff,
+       0xbf820008, 0xe0310000,
+       0x781d0000, 0x807cff7c,
+       0x00000100, 0x8078ff78,
+       0x00000100, 0xbf0a6f7c,
+       0xbf85fff8, 0xbef80380,
+       0xbefe03c1, 0x907c9973,
+       0x877c817c, 0xbf06817c,
+       0xbf850002, 0xbeff0380,
+       0xbf820001, 0xbeff03c1,
+       0xb96f2a05, 0x806f816f,
+       0x8f6f826f, 0x907c9973,
+       0x877c817c, 0xbf06817c,
+       0xbf850021, 0xbef603ff,
        0x01000000, 0xbef20378,
-       0x8078ff78, 0x00000400,
+       0x8078ff78, 0x00000200,
        0xbefc0384, 0xe0304000,
-       0x785d0000, 0xe0304100,
-       0x785d0100, 0xe0304200,
-       0x785d0200, 0xe0304300,
+       0x785d0000, 0xe0304080,
+       0x785d0100, 0xe0304100,
+       0x785d0200, 0xe0304180,
        0x785d0300, 0xbf8c3f70,
        0x7e008500, 0x7e028501,
        0x7e048502, 0x7e068503,
        0x807c847c, 0x8078ff78,
-       0x00000400, 0xbf0a6f7c,
-       0xbf85ffee, 0xb96f1e06,
-       0x876fc16f, 0xbf84000e,
-       0x8f6f836f, 0x806f7c6f,
-       0xbefe03c1, 0xbeff0380,
+       0x00000200, 0xbf0a6f7c,
+       0xbf85ffee, 0xe0304000,
+       0x725d0000, 0xe0304080,
+       0x725d0100, 0xe0304100,
+       0x725d0200, 0xe0304180,
+       0x725d0300, 0xbf820031,
+       0xbef603ff, 0x01000000,
+       0xbef20378, 0x8078ff78,
+       0x00000400, 0xbefc0384,
        0xe0304000, 0x785d0000,
+       0xe0304100, 0x785d0100,
+       0xe0304200, 0x785d0200,
+       0xe0304300, 0x785d0300,
        0xbf8c3f70, 0x7e008500,
-       0x807c817c, 0x8078ff78,
-       0x00000080, 0xbf0a6f7c,
-       0xbf85fff7, 0xbeff03c1,
-       0xe0304000, 0x725d0000,
-       0xe0304100, 0x725d0100,
-       0xe0304200, 0x725d0200,
-       0xe0304300, 0x725d0300,
-       0xb9782a05, 0x80788178,
-       0x907c9973, 0x877c817c,
-       0xbf06817c, 0xbf850002,
-       0x8f788978, 0xbf820001,
-       0x8f788a78, 0xb9721e06,
-       0x8f728a72, 0x80787278,
-       0x8078ff78, 0x00000200,
-       0x80f8ff78, 0x00000058,
-       0x80f88878, 0xbef603ff,
+       0x7e028501, 0x7e048502,
+       0x7e068503, 0x807c847c,
+       0x8078ff78, 0x00000400,
+       0xbf0a6f7c, 0xbf85ffee,
+       0xb96f1e06, 0x876fc16f,
+       0xbf84000e, 0x8f6f836f,
+       0x806f7c6f, 0xbefe03c1,
+       0xbeff0380, 0xe0304000,
+       0x785d0000, 0xbf8c3f70,
+       0x7e008500, 0x807c817c,
+       0x8078ff78, 0x00000080,
+       0xbf0a6f7c, 0xbf85fff7,
+       0xbeff03c1, 0xe0304000,
+       0x725d0000, 0xe0304100,
+       0x725d0100, 0xe0304200,
+       0x725d0200, 0xe0304300,
+       0x725d0300, 0xb9782a05,
+       0x80788178, 0x907c9973,
+       0x877c817c, 0xbf06817c,
+       0xbf850002, 0x8f788978,
+       0xbf820001, 0x8f788a78,
+       0xb9721e06, 0x8f728a72,
+       0x80787278, 0x8078ff78,
+       0x00000200, 0x80f8ff78,
+       0x00000050, 0xbef603ff,
        0x01000000, 0xbefc03ff,
-       0x0000006a, 0xf425003a,
-       0xf0000000, 0x80f8a078,
-       0xbf8cc07f, 0x80fc827c,
+       0x0000006c, 0x80f89078,
+       0xf429003a, 0xf0000000,
+       0xbf8cc07f, 0x80fc847c,
        0xbf800000, 0xbe803100,
+       0xbe823102, 0x80f8a078,
        0xf42d003a, 0xf0000000,
-       0x80f8c078, 0xbf8cc07f,
-       0x80fc887c, 0xbf800000,
-       0xbe803100, 0xbe823102,
-       0xbe843104, 0xbe863106,
+       0xbf8cc07f, 0x80fc887c,
+       0xbf800000, 0xbe803100,
+       0xbe823102, 0xbe843104,
+       0xbe863106, 0x80f8c078,
        0xf431003a, 0xf0000000,
-       0x80f8c078, 0xbf8cc07f,
-       0x80fc907c, 0xbf800000,
-       0xbe803100, 0xbe823102,
-       0xbe843104, 0xbe863106,
-       0xbe883108, 0xbe8a310a,
-       0xbe8c310c, 0xbe8e310e,
-       0xbf06807c, 0xbf84fff0,
-       0xb9782a05, 0x80788178,
-       0x907c9973, 0x877c817c,
-       0xbf06817c, 0xbf850002,
-       0x8f788978, 0xbf820001,
-       0x8f788a78, 0xb9721e06,
-       0x8f728a72, 0x80787278,
-       0x8078ff78, 0x00000200,
-       0xbef603ff, 0x01000000,
-       0xf4211bfa, 0xf0000000,
-       0x80788478, 0xf4211b3a,
+       0xbf8cc07f, 0x80fc907c,
+       0xbf800000, 0xbe803100,
+       0xbe823102, 0xbe843104,
+       0xbe863106, 0xbe883108,
+       0xbe8a310a, 0xbe8c310c,
+       0xbe8e310e, 0xbf06807c,
+       0xbf84fff0, 0xb9782a05,
+       0x80788178, 0x907c9973,
+       0x877c817c, 0xbf06817c,
+       0xbf850002, 0x8f788978,
+       0xbf820001, 0x8f788a78,
+       0xb9721e06, 0x8f728a72,
+       0x80787278, 0x8078ff78,
+       0x00000200, 0xbef603ff,
+       0x01000000, 0xf4211bfa,
        0xf0000000, 0x80788478,
-       0xf4211b7a, 0xf0000000,
-       0x80788478, 0xf4211eba,
+       0xf4211b3a, 0xf0000000,
+       0x80788478, 0xf4211b7a,
        0xf0000000, 0x80788478,
-       0xf4211efa, 0xf0000000,
-       0x80788478, 0xf4211c3a,
+       0xf4211eba, 0xf0000000,
+       0x80788478, 0xf4211efa,
        0xf0000000, 0x80788478,
-       0xf4211c7a, 0xf0000000,
-       0x80788478, 0xf4211e7a,
+       0xf4211c3a, 0xf0000000,
+       0x80788478, 0xf4211c7a,
        0xf0000000, 0x80788478,
-       0xf4211cfa, 0xf0000000,
-       0x80788478, 0xf4211bba,
+       0xf4211e7a, 0xf0000000,
+       0x80788478, 0xf4211cfa,
        0xf0000000, 0x80788478,
-       0xbf8cc07f, 0xb9eef814,
        0xf4211bba, 0xf0000000,
        0x80788478, 0xbf8cc07f,
-       0xb9eef815, 0xbef2036d,
-       0x876dff72, 0x0000ffff,
-       0xbefc036f, 0xbefe037a,
-       0xbeff037b, 0x876f71ff,
-       0x000003ff, 0xb9ef4803,
-       0xb9f9f816, 0x876f71ff,
-       0xfffff800, 0x906f8b6f,
-       0xb9efa2c3, 0xb9f3f801,
-       0x876fff72, 0xfc000000,
-       0x906f9a6f, 0x8f6f906f,
-       0xbef30380, 0x88736f73,
-       0x876fff72, 0x02000000,
-       0x906f996f, 0x8f6f8f6f,
+       0xb9eef814, 0xf4211bba,
+       0xf0000000, 0x80788478,
+       0xbf8cc07f, 0xb9eef815,
+       0xbef2036d, 0x876dff72,
+       0x0000ffff, 0xbefc036f,
+       0xbefe037a, 0xbeff037b,
+       0x876f71ff, 0x000003ff,
+       0xb9ef4803, 0xb9f9f816,
+       0x876f71ff, 0xfffff800,
+       0x906f8b6f, 0xb9efa2c3,
+       0xb9f3f801, 0x876fff72,
+       0xfc000000, 0x906f9a6f,
+       0x8f6f906f, 0xbef30380,
        0x88736f73, 0x876fff72,
-       0x01000000, 0x906f986f,
-       0x8f6f996f, 0x88736f73,
-       0x876fff70, 0x00800000,
-       0x906f976f, 0xb9f3f807,
-       0x87fe7e7e, 0x87ea6a6a,
-       0xb9f0f802, 0xbf8a0000,
-       0xbe80226c, 0xbf810000,
+       0x02000000, 0x906f996f,
+       0x8f6f8f6f, 0x88736f73,
+       0x876fff72, 0x01000000,
+       0x906f986f, 0x8f6f996f,
+       0x88736f73, 0x876fff70,
+       0x00800000, 0x906f976f,
+       0xb9f3f807, 0x87fe7e7e,
+       0x87ea6a6a, 0xb9f0f802,
+       0xbf8a0000, 0xbe80226c,
+       0xbf810000, 0xbf9f0000,
        0xbf9f0000, 0xbf9f0000,
        0xbf9f0000, 0xbf9f0000,
-       0xbf9f0000, 0x00000000,
 };
 static const uint32_t cwsr_trap_arcturus_hex[] = {
        0xbf820001, 0xbf8202c4,

diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm
index fafdfd2..3598621 100644 (file)
--- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm
+++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm
@@ -87,7 +87,7 @@ var S_SAVE_PC_HI_FIRST_REPLAY_MASK            = 0x02000000
 var S_SAVE_PC_HI_REPLAY_W64H_SHIFT             = 24
 var S_SAVE_PC_HI_REPLAY_W64H_MASK              = 0x01000000
 
-var s_sgpr_save_num                            = 106
+var s_sgpr_save_num                            = 108
 
 var s_save_spi_init_lo                         = exec_lo
 var s_save_spi_init_hi                         = exec_hi
@@ -357,13 +357,14 @@ L_SAVE_SGPR_LOOP:
        s_cmp_lt_u32    m0, 96                                                  //scc = (m0 < first 96 SGPR) ? 1 : 0
        s_cbranch_scc1  L_SAVE_SGPR_LOOP                                        //first 96 SGPR save is complete?
 
-       //save the rest 10 SGPR
+       //save the rest 12 SGPR
        s_movrels_b64   s0, s0                                                  //s0 = s[0+m0], s1 = s[1+m0]
        s_movrels_b64   s2, s2                                                  //s2 = s[2+m0], s3 = s[3+m0]
        s_movrels_b64   s4, s4                                                  //s4 = s[4+m0], s5 = s[5+m0]
        s_movrels_b64   s6, s6                                                  //s6 = s[6+m0], s7 = s[7+m0]
        s_movrels_b64   s8, s8                                                  //s8 = s[8+m0], s9 = s[9+m0]
-       write_10sgpr_to_mem(s0, s_save_buf_rsrc0, s_save_mem_offset)
+       s_movrels_b64   s10, s10                                                //s10 = s[10+m0], s11 = s[11+m0]
+       write_12sgpr_to_mem(s0, s_save_buf_rsrc0, s_save_mem_offset)
 
        // restore s_save_buf_rsrc0,1
        s_mov_b32       s_save_buf_rsrc0, s_save_xnack_mask
@@ -766,26 +767,25 @@ L_RESTORE_SGPR:
        get_svgpr_size_bytes(s_restore_tmp)
        s_add_u32       s_restore_mem_offset, s_restore_mem_offset, s_restore_tmp
        s_add_u32       s_restore_mem_offset, s_restore_mem_offset, get_sgpr_size_bytes()
-       s_sub_u32       s_restore_mem_offset, s_restore_mem_offset, 22*4        //s106~s127 is not saved
-       s_sub_u32       s_restore_mem_offset, s_restore_mem_offset, 2*4         // restore SGPR from S[n] to S[0], by 2 sgprs group
+       s_sub_u32       s_restore_mem_offset, s_restore_mem_offset, 20*4        //s108~s127 is not saved
 
        s_mov_b32       s_restore_buf_rsrc2, 0x1000000                          //NUM_RECORDS in bytes
 
        s_mov_b32       m0, s_sgpr_save_num
 
-       read_2sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)
-
+       read_4sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)
        s_waitcnt       lgkmcnt(0)
 
-       s_sub_u32       m0, m0, 2                                               // Restore from S[n] to S[0]
+       s_sub_u32       m0, m0, 4                                               // Restore from S[0] to S[104]
        s_nop           0                                                       // hazard SALU M0=> S_MOVREL
 
        s_movreld_b64   s0, s0                                                  //s[0+m0] = s0
+       s_movreld_b64   s2, s2
 
        read_8sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)
        s_waitcnt       lgkmcnt(0)
 
-       s_sub_u32       m0, m0, 8                                               // Restore from S[n] to S[0]
+       s_sub_u32       m0, m0, 8                                               // Restore from S[0] to S[96]
        s_nop           0                                                       // hazard SALU M0=> S_MOVREL
 
        s_movreld_b64   s0, s0                                                  //s[0+m0] = s0
@@ -903,11 +903,11 @@ function write_16sgpr_to_mem(s, s_rsrc, s_mem_offset)
        s_addc_u32      s_rsrc[1], s_rsrc[1], 0x0
 end
 
-function write_10sgpr_to_mem(s, s_rsrc, s_mem_offset)
+function write_12sgpr_to_mem(s, s_rsrc, s_mem_offset)
        s_buffer_store_dwordx4  s[0], s_rsrc, 0 glc:1
        s_buffer_store_dwordx4  s[4], s_rsrc, 16 glc:1
-       s_buffer_store_dwordx2  s[8], s_rsrc, 32 glc:1
-       s_add_u32       s_rsrc[0], s_rsrc[0], 4*16
+       s_buffer_store_dwordx4  s[8], s_rsrc, 32 glc:1
+       s_add_u32       s_rsrc[0], s_rsrc[0], 4*12
        s_addc_u32      s_rsrc[1], s_rsrc[1], 0x0
 end
 
@@ -918,18 +918,18 @@ function read_hwreg_from_mem(s, s_rsrc, s_mem_offset)
 end
 
 function read_16sgpr_from_mem(s, s_rsrc, s_mem_offset)
-       s_buffer_load_dwordx16  s, s_rsrc, s_mem_offset glc:1
        s_sub_u32       s_mem_offset, s_mem_offset, 4*16
+       s_buffer_load_dwordx16  s, s_rsrc, s_mem_offset glc:1
 end
 
 function read_8sgpr_from_mem(s, s_rsrc, s_mem_offset)
+       s_sub_u32       s_mem_offset, s_mem_offset, 4*8
        s_buffer_load_dwordx8   s, s_rsrc, s_mem_offset glc:1
-       s_sub_u32       s_mem_offset, s_mem_offset, 4*16
 end
 
-function read_2sgpr_from_mem(s, s_rsrc, s_mem_offset)
-       s_buffer_load_dwordx2 s, s_rsrc, s_mem_offset glc:1
-       s_sub_u32       s_mem_offset, s_mem_offset, 4*8
+function read_4sgpr_from_mem(s, s_rsrc, s_mem_offset)
+       s_sub_u32       s_mem_offset, s_mem_offset, 4*4
+       s_buffer_load_dwordx4   s, s_rsrc, s_mem_offset glc:1
 end