mirror of https://github.com/dotnet/runtime
Merge 187f621904 into 02596ba8d9
This commit is contained in:
Xu Liangyu
GitHub
commit
9c2c6dc3c0
|
|
@ -82,14 +82,8 @@ C_FUNC(\Name\()_End):
|
|||
|
||||
////NOTE: reg1 and reg2 must be the number and GPR type !!!
|
||||
.macro PROLOG_SAVE_REG_PAIR reg1, reg2, ofs, __def_cfa_save=0
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// //NOTE:The offset of gssq/gslq must be 16-bytes aligned.
|
||||
// // here ofs must be 16-bytes aligned.
|
||||
// gssq \reg2, \reg1, \ofs(sp)
|
||||
//#else
|
||||
st.d $r\reg1, $sp, \ofs
|
||||
st.d $r\reg2, $sp, \ofs+8
|
||||
//#endif
|
||||
|
||||
.cfi_rel_offset \reg1, \ofs
|
||||
.cfi_rel_offset \reg2, \ofs + 8
|
||||
|
|
@ -123,23 +117,15 @@ C_FUNC(\Name\()_End):
|
|||
.endm
|
||||
|
||||
.macro EPILOG_RESTORE_REG_PAIR reg1, reg2, ofs
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// gslq \reg2, \reg1, \ofs(sp)
|
||||
//#else
|
||||
ld.d $r\reg2, $sp, \ofs+8
|
||||
ld.d $r\reg1, $sp, \ofs
|
||||
//#endif
|
||||
.cfi_restore \reg2
|
||||
.cfi_restore \reg1
|
||||
.endm
|
||||
|
||||
.macro EPILOG_RESTORE_REG_PAIR_INDEXED reg1, reg2, ssize
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// gslq \reg2, \reg1, 0(sp)
|
||||
//#else
|
||||
ld.d $r\reg2, $sp, 8
|
||||
ld.d $r\reg1, $sp, 0
|
||||
//#endif
|
||||
.cfi_restore \reg2
|
||||
.cfi_restore \reg1
|
||||
|
||||
|
|
@ -183,14 +169,6 @@ C_FUNC(\Name\()_End):
|
|||
// Reserve 64 bytes of memory before calling SAVE_ARGUMENT_REGISTERS
|
||||
.macro SAVE_ARGUMENT_REGISTERS reg, ofs
|
||||
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// //NOTE:The offset of gssq/gslq must be 16-bytes aligned.
|
||||
// // here ofs must be 16-bytes aligned.
|
||||
// gssq a1, a0, \ofs(\reg)
|
||||
// gssq a3, a2, \ofs+16(\reg)
|
||||
// gssq a5, a4, \ofs+32(\reg)
|
||||
// gssq a7, a6, \ofs+48(\reg)
|
||||
//#else
|
||||
st.d $a0, \reg, \ofs
|
||||
st.d $a1, \reg, \ofs+8
|
||||
st.d $a2, \reg, \ofs+16
|
||||
|
|
@ -199,21 +177,12 @@ C_FUNC(\Name\()_End):
|
|||
st.d $a5, \reg, \ofs+40
|
||||
st.d $a6, \reg, \ofs+48
|
||||
st.d $a7, \reg, \ofs+56
|
||||
//#endif
|
||||
|
||||
.endm
|
||||
|
||||
// Reserve 64 bytes of memory before calling SAVE_FLOAT_ARGUMENT_REGISTERS
|
||||
.macro SAVE_FLOAT_ARGUMENT_REGISTERS reg, ofs
|
||||
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// //NOTE:The offset of gssqc1/gslqc1 must be 16-bytes aligned.
|
||||
// // here ofs must be 16-bytes aligned.
|
||||
// gssqc1 $f13, $f12, \ofs(\reg)
|
||||
// gssqc1 $f15, $f14, \ofs+16(\reg)
|
||||
// gssqc1 $f17, $f16, \ofs+32(\reg)
|
||||
// gssqc1 $f19, $f18, \ofs+48(\reg)
|
||||
//#else
|
||||
fst.d $f0, \reg, \ofs
|
||||
fst.d $f1, \reg, \ofs+8
|
||||
fst.d $f2, \reg, \ofs+16
|
||||
|
|
@ -222,21 +191,12 @@ C_FUNC(\Name\()_End):
|
|||
fst.d $f5, \reg, \ofs+40
|
||||
fst.d $f6, \reg, \ofs+48
|
||||
fst.d $f7, \reg, \ofs+56
|
||||
//#endif
|
||||
|
||||
.endm
|
||||
|
||||
// Reserve 64 bytes of memory before calling SAVE_FLOAT_CALLEESAVED_REGISTERS
|
||||
.macro SAVE_FLOAT_CALLEESAVED_REGISTERS reg, ofs
|
||||
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// //NOTE:The offset of gssqc1/gslqc1 must be 16-bytes aligned.
|
||||
// // here ofs must be 16-bytes aligned.
|
||||
// gssqc1 $f25, $f24, \ofs(\reg)
|
||||
// gssqc1 $f27, $f26, \ofs+16(\reg)
|
||||
// gssqc1 $f29, $f28, \ofs+32(\reg)
|
||||
// gssqc1 $f31, $f30, \ofs+48(\reg)
|
||||
//#else
|
||||
fst.d $f24, \reg, \ofs
|
||||
fst.d $f25, \reg, \ofs+8
|
||||
fst.d $f26, \reg, \ofs+16
|
||||
|
|
@ -245,7 +205,6 @@ C_FUNC(\Name\()_End):
|
|||
fst.d $f29, \reg, \ofs+40
|
||||
fst.d $f30, \reg, \ofs+48
|
||||
fst.d $f31, \reg, \ofs+56
|
||||
//#endif
|
||||
|
||||
.endm
|
||||
|
||||
|
|
@ -260,14 +219,6 @@ C_FUNC(\Name\()_End):
|
|||
|
||||
.macro RESTORE_ARGUMENT_REGISTERS reg, ofs
|
||||
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// //NOTE:The offset of gssq/gslq must be 16-bytes aligned.
|
||||
// // here ofs must be 16-bytes aligned.
|
||||
// gslq a7, a6, \ofs+48(\reg)
|
||||
// gslq a5, a4, \ofs+32(\reg)
|
||||
// gslq a3, a2, \ofs+16(\reg)
|
||||
// gslq a1, a0, \ofs(\reg)
|
||||
//#else
|
||||
ld.d $a7, \reg, \ofs+56
|
||||
ld.d $a6, \reg, \ofs+48
|
||||
ld.d $a5, \reg, \ofs+40
|
||||
|
|
@ -276,18 +227,11 @@ C_FUNC(\Name\()_End):
|
|||
ld.d $a2, \reg, \ofs+16
|
||||
ld.d $a1, \reg, \ofs+8
|
||||
ld.d $a0, \reg, \ofs
|
||||
//#endif
|
||||
|
||||
.endm
|
||||
|
||||
.macro RESTORE_FLOAT_ARGUMENT_REGISTERS reg, ofs
|
||||
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// gslqc1 $f19, $f18, \ofs+48(\reg)
|
||||
// gslqc1 $f17, $f16, \ofs+32(\reg)
|
||||
// gslqc1 $f15, $f14, \ofs+16(\reg)
|
||||
// gslqc1 $f13, $f12, \ofs(\reg)
|
||||
//#else
|
||||
fld.d $f7, \reg, \ofs+56
|
||||
fld.d $f6, \reg, \ofs+48
|
||||
fld.d $f5, \reg, \ofs+40
|
||||
|
|
@ -296,20 +240,11 @@ C_FUNC(\Name\()_End):
|
|||
fld.d $f2, \reg, \ofs+16
|
||||
fld.d $f1, \reg, \ofs+8
|
||||
fld.d $f0, \reg, \ofs
|
||||
//#endif
|
||||
|
||||
.endm
|
||||
|
||||
.macro RESTORE_FLOAT_CALLEESAVED_REGISTERS reg, ofs
|
||||
|
||||
//#ifdef FEATURE_LOONGSONISA
|
||||
// //NOTE:The offset of gssqc1/gslqc1 must be 16-bytes aligned.
|
||||
// // here ofs must be 16-bytes aligned.
|
||||
// gslqc1 $f25, $f24, \ofs(\reg)
|
||||
// gslqc1 $f27, $f26, \ofs+16(\reg)
|
||||
// gslqc1 $f29, $f28, \ofs+32(\reg)
|
||||
// gslqc1 $f31, $f30, \ofs+48(\reg)
|
||||
//#else
|
||||
fld.d $f24, $r\reg, \ofs
|
||||
fld.d $f25, $r\reg, \ofs+8
|
||||
fld.d $f26, $r\reg, \ofs+16
|
||||
|
|
@ -318,7 +253,6 @@ C_FUNC(\Name\()_End):
|
|||
fld.d $f29, $r\reg, \ofs+40
|
||||
fld.d $f30, $r\reg, \ofs+48
|
||||
fld.d $f31, $r\reg, \ofs+56
|
||||
//#endif
|
||||
|
||||
.endm
|
||||
|
||||
|
|
|
|||
|
|
@ -12,6 +12,9 @@
|
|||
#define CONTEXT_FLOATING_POINT_BIT (2)
|
||||
#define CONTEXT_DEBUG_REGISTERS_BIT (3)
|
||||
|
||||
#define CONTEXT_HWCAP_LOONGARCH_LSX_BIT (1 << 4)
|
||||
#define CONTEXT_HWCAP_LOONGARCH_LASX_BIT (1 << 5)
|
||||
|
||||
#define CONTEXT_CONTROL (CONTEXT_LOONGARCH64 | (1 << CONTEXT_CONTROL_BIT))
|
||||
#define CONTEXT_INTEGER (CONTEXT_LOONGARCH64 | (1 << CONTEXT_INTEGER_BIT))
|
||||
#define CONTEXT_FLOATING_POINT (CONTEXT_LOONGARCH64 | (1 << CONTEXT_FLOATING_POINT_BIT))
|
||||
|
|
|
|||
|
|
@ -16,12 +16,12 @@
|
|||
LEAF_ENTRY RtlRestoreContext, _TEXT
|
||||
#ifdef HAS_ADDRESS_SANITIZER
|
||||
ld.w $r21, $a0, CONTEXT_ContextFlags
|
||||
andi $r21, $r21, (1 << CONTEXT_CONTROL_BIT)
|
||||
beq $r21, $r0, LOCAL_LABEL(Restore_CONTEXT_FLOATING_POINT)
|
||||
andi $r21, $r21, (1 << CONTEXT_CONTROL_BIT)
|
||||
beqz $r21, LOCAL_LABEL(Restore_CONTEXT_FLOATING_POINT)
|
||||
|
||||
addi.d $sp, $sp, -16
|
||||
st.d a0, $sp, 0
|
||||
st.d a1, $sp, 8
|
||||
st.d $a0, $sp, 0
|
||||
st.d $a1, $sp, 8
|
||||
|
||||
bl __asan_handle_no_return
|
||||
|
||||
|
|
@ -37,17 +37,103 @@ LOCAL_LABEL(Restore_CONTEXT_FLOATING_POINT):
|
|||
andi $t1, $r21, (1 << CONTEXT_FLOATING_POINT_BIT)
|
||||
beqz $t1, LOCAL_LABEL(No_Restore_CONTEXT_FLOATING_POINT)
|
||||
|
||||
PROLOG_SAVE_REG_PAIR_INDEXED 22, 1, 32
|
||||
st.d $a0, $fp, 16
|
||||
st.d $a1, $fp, 24
|
||||
bl C_FUNC(minipal_getcpufeatures)
|
||||
ori $t1, $a0, 0
|
||||
ld.d $a0, $fp, 16
|
||||
ld.d $a1, $fp, 24
|
||||
EPILOG_RESTORE_REG_PAIR_INDEXED 22, 1, 32
|
||||
|
||||
andi $t3, $t1, CONTEXT_HWCAP_LOONGARCH_LASX_BIT
|
||||
bnez $t3, LOCAL_LABEL(Restore_CONTEXT_LASX)
|
||||
andi $t3, $t1, CONTEXT_HWCAP_LOONGARCH_LSX_BIT
|
||||
bnez $t3, LOCAL_LABEL(Restore_CONTEXT_LSX)
|
||||
|
||||
// Neither LSX or LASX is supported.
|
||||
fld.d $f0 , $a0, CONTEXT_FPU_OFFSET
|
||||
fld.d $f1 , $a0, CONTEXT_FPU_OFFSET + 8*1
|
||||
fld.d $f2 , $a0, CONTEXT_FPU_OFFSET + 8*2
|
||||
fld.d $f3 , $a0, CONTEXT_FPU_OFFSET + 8*3
|
||||
fld.d $f4 , $a0, CONTEXT_FPU_OFFSET + 8*4
|
||||
fld.d $f5 , $a0, CONTEXT_FPU_OFFSET + 8*5
|
||||
fld.d $f6 , $a0, CONTEXT_FPU_OFFSET + 8*6
|
||||
fld.d $f7 , $a0, CONTEXT_FPU_OFFSET + 8*7
|
||||
fld.d $f8 , $a0, CONTEXT_FPU_OFFSET + 8*8
|
||||
fld.d $f9 , $a0, CONTEXT_FPU_OFFSET + 8*9
|
||||
fld.d $f10, $a0, CONTEXT_FPU_OFFSET + 8*10
|
||||
fld.d $f11, $a0, CONTEXT_FPU_OFFSET + 8*11
|
||||
fld.d $f12, $a0, CONTEXT_FPU_OFFSET + 8*12
|
||||
fld.d $f13, $a0, CONTEXT_FPU_OFFSET + 8*13
|
||||
fld.d $f14, $a0, CONTEXT_FPU_OFFSET + 8*14
|
||||
fld.d $f15, $a0, CONTEXT_FPU_OFFSET + 8*15
|
||||
fld.d $f16, $a0, CONTEXT_FPU_OFFSET + 8*16
|
||||
fld.d $f17, $a0, CONTEXT_FPU_OFFSET + 8*17
|
||||
fld.d $f18, $a0, CONTEXT_FPU_OFFSET + 8*18
|
||||
fld.d $f19, $a0, CONTEXT_FPU_OFFSET + 8*19
|
||||
fld.d $f20, $a0, CONTEXT_FPU_OFFSET + 8*20
|
||||
fld.d $f21, $a0, CONTEXT_FPU_OFFSET + 8*21
|
||||
fld.d $f22, $a0, CONTEXT_FPU_OFFSET + 8*22
|
||||
fld.d $f23, $a0, CONTEXT_FPU_OFFSET + 8*23
|
||||
fld.d $f24, $a0, CONTEXT_FPU_OFFSET + 8*24
|
||||
fld.d $f25, $a0, CONTEXT_FPU_OFFSET + 8*25
|
||||
fld.d $f26, $a0, CONTEXT_FPU_OFFSET + 8*26
|
||||
fld.d $f27, $a0, CONTEXT_FPU_OFFSET + 8*27
|
||||
fld.d $f28, $a0, CONTEXT_FPU_OFFSET + 8*28
|
||||
fld.d $f29, $a0, CONTEXT_FPU_OFFSET + 8*29
|
||||
fld.d $f30, $a0, CONTEXT_FPU_OFFSET + 8*30
|
||||
fld.d $f31, $a0, CONTEXT_FPU_OFFSET + 8*31
|
||||
b LOCAL_LABEL(Restore_CONTEXT_FLOATING_CONTROL)
|
||||
|
||||
LOCAL_LABEL(Restore_CONTEXT_LSX):
|
||||
// 128-bits SIMD:LSX.
|
||||
vld $vr0 , $a0, CONTEXT_FPU_OFFSET
|
||||
vld $vr1 , $a0, CONTEXT_FPU_OFFSET + 16*1
|
||||
vld $vr2 , $a0, CONTEXT_FPU_OFFSET + 16*2
|
||||
vld $vr3 , $a0, CONTEXT_FPU_OFFSET + 16*3
|
||||
vld $vr4 , $a0, CONTEXT_FPU_OFFSET + 16*4
|
||||
vld $vr5 , $a0, CONTEXT_FPU_OFFSET + 16*5
|
||||
vld $vr6 , $a0, CONTEXT_FPU_OFFSET + 16*6
|
||||
vld $vr7 , $a0, CONTEXT_FPU_OFFSET + 16*7
|
||||
vld $vr8 , $a0, CONTEXT_FPU_OFFSET + 16*8
|
||||
vld $vr9 , $a0, CONTEXT_FPU_OFFSET + 16*9
|
||||
vld $vr10, $a0, CONTEXT_FPU_OFFSET + 16*10
|
||||
vld $vr11, $a0, CONTEXT_FPU_OFFSET + 16*11
|
||||
vld $vr12, $a0, CONTEXT_FPU_OFFSET + 16*12
|
||||
vld $vr13, $a0, CONTEXT_FPU_OFFSET + 16*13
|
||||
vld $vr14, $a0, CONTEXT_FPU_OFFSET + 16*14
|
||||
vld $vr15, $a0, CONTEXT_FPU_OFFSET + 16*15
|
||||
vld $vr16, $a0, CONTEXT_FPU_OFFSET + 16*16
|
||||
vld $vr17, $a0, CONTEXT_FPU_OFFSET + 16*17
|
||||
vld $vr18, $a0, CONTEXT_FPU_OFFSET + 16*18
|
||||
vld $vr19, $a0, CONTEXT_FPU_OFFSET + 16*19
|
||||
vld $vr20, $a0, CONTEXT_FPU_OFFSET + 16*20
|
||||
vld $vr21, $a0, CONTEXT_FPU_OFFSET + 16*21
|
||||
vld $vr22, $a0, CONTEXT_FPU_OFFSET + 16*22
|
||||
vld $vr23, $a0, CONTEXT_FPU_OFFSET + 16*23
|
||||
vld $vr24, $a0, CONTEXT_FPU_OFFSET + 16*24
|
||||
vld $vr25, $a0, CONTEXT_FPU_OFFSET + 16*25
|
||||
vld $vr26, $a0, CONTEXT_FPU_OFFSET + 16*26
|
||||
vld $vr27, $a0, CONTEXT_FPU_OFFSET + 16*27
|
||||
vld $vr28, $a0, CONTEXT_FPU_OFFSET + 16*28
|
||||
vld $vr29, $a0, CONTEXT_FPU_OFFSET + 16*29
|
||||
vld $vr30, $a0, CONTEXT_FPU_OFFSET + 16*30
|
||||
vld $vr31, $a0, CONTEXT_FPU_OFFSET + 16*31
|
||||
b LOCAL_LABEL(Restore_CONTEXT_FLOATING_CONTROL)
|
||||
|
||||
LOCAL_LABEL(Restore_CONTEXT_LASX):
|
||||
// 256-bits SIMD:LASX.
|
||||
xvld $xr0, $a0, CONTEXT_FPU_OFFSET + 0
|
||||
xvld $xr1, $a0, CONTEXT_FPU_OFFSET + 32*1
|
||||
xvld $xr2, $a0, CONTEXT_FPU_OFFSET + 32*2
|
||||
xvld $xr3, $a0, CONTEXT_FPU_OFFSET + 32*3
|
||||
xvld $xr4, $a0, CONTEXT_FPU_OFFSET + 32*4
|
||||
xvld $xr5, $a0, CONTEXT_FPU_OFFSET + 32*5
|
||||
xvld $xr6, $a0, CONTEXT_FPU_OFFSET + 32*6
|
||||
xvld $xr7, $a0, CONTEXT_FPU_OFFSET + 32*7
|
||||
xvld $xr8, $a0, CONTEXT_FPU_OFFSET + 32*8
|
||||
xvld $xr9, $a0, CONTEXT_FPU_OFFSET + 32*9
|
||||
xvld $xr0 , $a0, CONTEXT_FPU_OFFSET
|
||||
xvld $xr1 , $a0, CONTEXT_FPU_OFFSET + 32*1
|
||||
xvld $xr2 , $a0, CONTEXT_FPU_OFFSET + 32*2
|
||||
xvld $xr3 , $a0, CONTEXT_FPU_OFFSET + 32*3
|
||||
xvld $xr4 , $a0, CONTEXT_FPU_OFFSET + 32*4
|
||||
xvld $xr5 , $a0, CONTEXT_FPU_OFFSET + 32*5
|
||||
xvld $xr6 , $a0, CONTEXT_FPU_OFFSET + 32*6
|
||||
xvld $xr7 , $a0, CONTEXT_FPU_OFFSET + 32*7
|
||||
xvld $xr8 , $a0, CONTEXT_FPU_OFFSET + 32*8
|
||||
xvld $xr9 , $a0, CONTEXT_FPU_OFFSET + 32*9
|
||||
xvld $xr10, $a0, CONTEXT_FPU_OFFSET + 32*10
|
||||
xvld $xr11, $a0, CONTEXT_FPU_OFFSET + 32*11
|
||||
xvld $xr12, $a0, CONTEXT_FPU_OFFSET + 32*12
|
||||
|
|
@ -71,6 +157,7 @@ LOCAL_LABEL(Restore_CONTEXT_FLOATING_POINT):
|
|||
xvld $xr30, $a0, CONTEXT_FPU_OFFSET + 32*30
|
||||
xvld $xr31, $a0, CONTEXT_FPU_OFFSET + 32*31
|
||||
|
||||
LOCAL_LABEL(Restore_CONTEXT_FLOATING_CONTROL):
|
||||
ld.d $t1, $a0, CONTEXT_FLOAT_CONTROL_OFFSET
|
||||
movgr2cf $fcc0, $t1
|
||||
srli.d $t1, $t1, 8
|
||||
|
|
@ -127,7 +214,7 @@ LOCAL_LABEL(No_Restore_CONTEXT_FLOATING_POINT):
|
|||
LOCAL_LABEL(No_Restore_CONTEXT_INTEGER):
|
||||
|
||||
andi $r21, $r21, (1 << CONTEXT_CONTROL_BIT)
|
||||
beq $r21, $r0, LOCAL_LABEL(No_Restore_CONTEXT_CONTROL)
|
||||
beqz $r21, LOCAL_LABEL(No_Restore_CONTEXT_CONTROL)
|
||||
|
||||
ld.d $ra, $t4, CONTEXT_Ra
|
||||
ld.d $fp, $t4, CONTEXT_Fp
|
||||
|
|
@ -218,8 +305,92 @@ LOCAL_LABEL(Done_CONTEXT_INTEGER):
|
|||
andi $t3, $t1, (1 << CONTEXT_FLOATING_POINT_BIT)
|
||||
beqz $t3, LOCAL_LABEL(Done_CONTEXT_FLOATING_POINT)
|
||||
|
||||
PROLOG_SAVE_REG_PAIR_INDEXED 22, 1, 32
|
||||
st.d $a0, $fp, 16
|
||||
bl C_FUNC(minipal_getcpufeatures)
|
||||
ori $t1, $a0, 0
|
||||
ld.d $a0, $fp, 16
|
||||
EPILOG_RESTORE_REG_PAIR_INDEXED 22, 1, 32
|
||||
|
||||
andi $t3, $t1, CONTEXT_HWCAP_LOONGARCH_LASX_BIT
|
||||
bnez $t3, LOCAL_LABEL(Store_CONTEXT_LASX)
|
||||
andi $t3, $t1, CONTEXT_HWCAP_LOONGARCH_LSX_BIT
|
||||
bnez $t3, LOCAL_LABEL(Store_CONTEXT_LSX)
|
||||
|
||||
// Neither LSX or LASX is supported.
|
||||
fst.d $f0 , $a0, CONTEXT_FPU_OFFSET
|
||||
fst.d $f1 , $a0, CONTEXT_FPU_OFFSET + 8*1
|
||||
fst.d $f2 , $a0, CONTEXT_FPU_OFFSET + 8*2
|
||||
fst.d $f3 , $a0, CONTEXT_FPU_OFFSET + 8*3
|
||||
fst.d $f4 , $a0, CONTEXT_FPU_OFFSET + 8*4
|
||||
fst.d $f5 , $a0, CONTEXT_FPU_OFFSET + 8*5
|
||||
fst.d $f6 , $a0, CONTEXT_FPU_OFFSET + 8*6
|
||||
fst.d $f7 , $a0, CONTEXT_FPU_OFFSET + 8*7
|
||||
fst.d $f8 , $a0, CONTEXT_FPU_OFFSET + 8*8
|
||||
fst.d $f9 , $a0, CONTEXT_FPU_OFFSET + 8*9
|
||||
fst.d $f10, $a0, CONTEXT_FPU_OFFSET + 8*10
|
||||
fst.d $f11, $a0, CONTEXT_FPU_OFFSET + 8*11
|
||||
fst.d $f12, $a0, CONTEXT_FPU_OFFSET + 8*12
|
||||
fst.d $f13, $a0, CONTEXT_FPU_OFFSET + 8*13
|
||||
fst.d $f14, $a0, CONTEXT_FPU_OFFSET + 8*14
|
||||
fst.d $f15, $a0, CONTEXT_FPU_OFFSET + 8*15
|
||||
fst.d $f16, $a0, CONTEXT_FPU_OFFSET + 8*16
|
||||
fst.d $f17, $a0, CONTEXT_FPU_OFFSET + 8*17
|
||||
fst.d $f18, $a0, CONTEXT_FPU_OFFSET + 8*18
|
||||
fst.d $f19, $a0, CONTEXT_FPU_OFFSET + 8*19
|
||||
fst.d $f20, $a0, CONTEXT_FPU_OFFSET + 8*20
|
||||
fst.d $f21, $a0, CONTEXT_FPU_OFFSET + 8*21
|
||||
fst.d $f22, $a0, CONTEXT_FPU_OFFSET + 8*22
|
||||
fst.d $f23, $a0, CONTEXT_FPU_OFFSET + 8*23
|
||||
fst.d $f24, $a0, CONTEXT_FPU_OFFSET + 8*24
|
||||
fst.d $f25, $a0, CONTEXT_FPU_OFFSET + 8*25
|
||||
fst.d $f26, $a0, CONTEXT_FPU_OFFSET + 8*26
|
||||
fst.d $f27, $a0, CONTEXT_FPU_OFFSET + 8*27
|
||||
fst.d $f28, $a0, CONTEXT_FPU_OFFSET + 8*28
|
||||
fst.d $f29, $a0, CONTEXT_FPU_OFFSET + 8*29
|
||||
fst.d $f30, $a0, CONTEXT_FPU_OFFSET + 8*30
|
||||
fst.d $f31, $a0, CONTEXT_FPU_OFFSET + 8*31
|
||||
b LOCAL_LABEL(Store_CONTEXT_FLOAT_CONTROL)
|
||||
|
||||
LOCAL_LABEL(Store_CONTEXT_LSX):
|
||||
// 128-bits SIMD:LSX.
|
||||
vst $vr0 , $a0, CONTEXT_FPU_OFFSET
|
||||
vst $vr1 , $a0, CONTEXT_FPU_OFFSET + 16*1
|
||||
vst $vr2 , $a0, CONTEXT_FPU_OFFSET + 16*2
|
||||
vst $vr3 , $a0, CONTEXT_FPU_OFFSET + 16*3
|
||||
vst $vr4 , $a0, CONTEXT_FPU_OFFSET + 16*4
|
||||
vst $vr5 , $a0, CONTEXT_FPU_OFFSET + 16*5
|
||||
vst $vr6 , $a0, CONTEXT_FPU_OFFSET + 16*6
|
||||
vst $vr7 , $a0, CONTEXT_FPU_OFFSET + 16*7
|
||||
vst $vr8 , $a0, CONTEXT_FPU_OFFSET + 16*8
|
||||
vst $vr9 , $a0, CONTEXT_FPU_OFFSET + 16*9
|
||||
vst $vr10, $a0, CONTEXT_FPU_OFFSET + 16*10
|
||||
vst $vr11, $a0, CONTEXT_FPU_OFFSET + 16*11
|
||||
vst $vr12, $a0, CONTEXT_FPU_OFFSET + 16*12
|
||||
vst $vr13, $a0, CONTEXT_FPU_OFFSET + 16*13
|
||||
vst $vr14, $a0, CONTEXT_FPU_OFFSET + 16*14
|
||||
vst $vr15, $a0, CONTEXT_FPU_OFFSET + 16*15
|
||||
vst $vr16, $a0, CONTEXT_FPU_OFFSET + 16*16
|
||||
vst $vr17, $a0, CONTEXT_FPU_OFFSET + 16*17
|
||||
vst $vr18, $a0, CONTEXT_FPU_OFFSET + 16*18
|
||||
vst $vr19, $a0, CONTEXT_FPU_OFFSET + 16*19
|
||||
vst $vr20, $a0, CONTEXT_FPU_OFFSET + 16*20
|
||||
vst $vr21, $a0, CONTEXT_FPU_OFFSET + 16*21
|
||||
vst $vr22, $a0, CONTEXT_FPU_OFFSET + 16*22
|
||||
vst $vr23, $a0, CONTEXT_FPU_OFFSET + 16*23
|
||||
vst $vr24, $a0, CONTEXT_FPU_OFFSET + 16*24
|
||||
vst $vr25, $a0, CONTEXT_FPU_OFFSET + 16*25
|
||||
vst $vr26, $a0, CONTEXT_FPU_OFFSET + 16*26
|
||||
vst $vr27, $a0, CONTEXT_FPU_OFFSET + 16*27
|
||||
vst $vr28, $a0, CONTEXT_FPU_OFFSET + 16*28
|
||||
vst $vr29, $a0, CONTEXT_FPU_OFFSET + 16*29
|
||||
vst $vr30, $a0, CONTEXT_FPU_OFFSET + 16*30
|
||||
vst $vr31, $a0, CONTEXT_FPU_OFFSET + 16*31
|
||||
b LOCAL_LABEL(Store_CONTEXT_FLOAT_CONTROL)
|
||||
|
||||
LOCAL_LABEL(Store_CONTEXT_LASX):
|
||||
// 256-bits SIMD:LASX.
|
||||
xvst $xr0 , $a0, CONTEXT_FPU_OFFSET + 32*0
|
||||
xvst $xr0 , $a0, CONTEXT_FPU_OFFSET
|
||||
xvst $xr1 , $a0, CONTEXT_FPU_OFFSET + 32*1
|
||||
xvst $xr2 , $a0, CONTEXT_FPU_OFFSET + 32*2
|
||||
xvst $xr3 , $a0, CONTEXT_FPU_OFFSET + 32*3
|
||||
|
|
@ -252,6 +423,7 @@ LOCAL_LABEL(Done_CONTEXT_INTEGER):
|
|||
xvst $xr30, $a0, CONTEXT_FPU_OFFSET + 32*30
|
||||
xvst $xr31, $a0, CONTEXT_FPU_OFFSET + 32*31
|
||||
|
||||
LOCAL_LABEL(Store_CONTEXT_FLOAT_CONTROL):
|
||||
ori $t0, $r0, 0
|
||||
movcf2gr $t0, $fcc0
|
||||
st.b $t0, $a0, CONTEXT_FLOAT_CONTROL_OFFSET
|
||||
|
|
|
|||
|
|
@ -364,7 +364,6 @@ NESTED_ENTRY NDirectImportThunk, _TEXT, NoHandler
|
|||
|
||||
// $fp,$ra
|
||||
PROLOG_SAVE_REG_PAIR_INDEXED 22, 1, 0xa0
|
||||
//PROLOG_SAVE_REG gp, 16
|
||||
SAVE_ARGUMENT_REGISTERS $sp, 0x20
|
||||
SAVE_FLOAT_ARGUMENT_REGISTERS $sp, 0x60
|
||||
|
||||
|
|
@ -375,7 +374,6 @@ NESTED_ENTRY NDirectImportThunk, _TEXT, NoHandler
|
|||
// pop the stack and restore original register state
|
||||
RESTORE_FLOAT_ARGUMENT_REGISTERS $sp, 0x60
|
||||
RESTORE_ARGUMENT_REGISTERS $sp, 0x20
|
||||
//EPILOG_RESTORE_REG gp, 16
|
||||
// $fp,$ra
|
||||
EPILOG_RESTORE_REG_PAIR_INDEXED 22, 1, 0xa0
|
||||
|
||||
|
|
@ -406,7 +404,6 @@ NESTED_ENTRY VirtualMethodFixupStub, _TEXT, NoHandler
|
|||
// Save arguments and return address
|
||||
// $fp,$ra
|
||||
PROLOG_SAVE_REG_PAIR_INDEXED 22, 1, 0xa0
|
||||
//PROLOG_SAVE_REG gp, 16
|
||||
SAVE_ARGUMENT_REGISTERS $sp, 32
|
||||
SAVE_FLOAT_ARGUMENT_REGISTERS $sp, 96
|
||||
|
||||
|
|
@ -425,7 +422,6 @@ NESTED_ENTRY VirtualMethodFixupStub, _TEXT, NoHandler
|
|||
// pop the stack and restore original register state
|
||||
RESTORE_FLOAT_ARGUMENT_REGISTERS $sp, 96
|
||||
RESTORE_ARGUMENT_REGISTERS $sp, 32
|
||||
//EPILOG_RESTORE_REG gp, 16
|
||||
// $fp,$ra
|
||||
EPILOG_RESTORE_REG_PAIR_INDEXED 22, 1, 0xa0
|
||||
|
||||
|
|
@ -476,7 +472,6 @@ NESTED_ENTRY TheUMEntryPrestub, _TEXT, UnhandledExceptionHandlerUnix
|
|||
// Save arguments and return address
|
||||
// $fp,$ra
|
||||
PROLOG_SAVE_REG_PAIR_INDEXED 22, 1, 0xa0
|
||||
//PROLOG_SAVE_REG gp, 16
|
||||
SAVE_ARGUMENT_REGISTERS $sp, 32
|
||||
SAVE_FLOAT_ARGUMENT_REGISTERS $sp, 96
|
||||
|
||||
|
|
@ -488,7 +483,6 @@ NESTED_ENTRY TheUMEntryPrestub, _TEXT, UnhandledExceptionHandlerUnix
|
|||
// pop the stack and restore original register state
|
||||
RESTORE_FLOAT_ARGUMENT_REGISTERS $sp, 96
|
||||
RESTORE_ARGUMENT_REGISTERS $sp, 32
|
||||
//EPILOG_RESTORE_REG gp, 16
|
||||
// $fp,$ra
|
||||
EPILOG_RESTORE_REG_PAIR_INDEXED 22, 1, 0xa0
|
||||
|
||||
|
|
@ -604,34 +598,6 @@ LOCAL_LABEL(Fail):
|
|||
b C_FUNC(ResolveWorkerAsmStub) // call the ResolveWorkerAsmStub method to transition into the VM
|
||||
NESTED_END ResolveWorkerChainLookupAsmStub, _TEXT
|
||||
|
||||
|
||||
//NOTE: Frame_Size = SIZEOF__ArgumentRegisters + SIZEOF__FloatArgumentRegisters + extra.
|
||||
//
|
||||
// |gp |
|
||||
// |s0 |
|
||||
// |$t2 |
|
||||
// |t9 |
|
||||
// |$a7 |
|
||||
// |$a6 |
|
||||
// |$a5 |
|
||||
// |$a4 |
|
||||
// |$a3 |
|
||||
// |$a2 |
|
||||
// |$a1 |
|
||||
// |$a0 |
|
||||
// |$ra | $sp+8
|
||||
// |fp | $sp
|
||||
//
|
||||
// |f19 | if needed.
|
||||
// |f18 |
|
||||
// |f17 |
|
||||
// |f16 |
|
||||
// |f15 |
|
||||
// |f14 |
|
||||
// |f13 |
|
||||
// |f12 |
|
||||
//
|
||||
|
||||
// ------------------------------------------------------------------
|
||||
// void ResolveWorkerAsmStub(args in regs $a0-$a7 & stack, t8:IndirectionCellAndFlags, $t2:DispatchToken)
|
||||
//
|
||||
|
|
|
|||
|
|
@ -32,7 +32,6 @@ EXTERN_C VOID STDCALL PrecodeRemotingThunk();
|
|||
#elif defined(TARGET_LOONGARCH64)
|
||||
|
||||
#define SIZEOF_PRECODE_BASE CODE_SIZE_ALIGN
|
||||
#define SHIFTOF_PRECODE_TYPE 5
|
||||
|
||||
#elif defined(TARGET_RISCV64)
|
||||
|
||||
|
|
|
|||
|
|
@ -614,6 +614,26 @@ int minipal_getcpufeatures(void)
|
|||
|
||||
#endif // HOST_RISCV64
|
||||
|
||||
#if defined(HOST_LOONGARCH64)
|
||||
|
||||
#if defined(HOST_UNIX)
|
||||
|
||||
#if HAVE_AUXV_HWCAP_H
|
||||
|
||||
unsigned long hwCap = getauxval(AT_HWCAP);
|
||||
|
||||
if (hwCap & HWCAP_LOONGARCH_LSX)
|
||||
result |= LoongArch64IntrinsicConstants_LSX;
|
||||
|
||||
if (hwCap & HWCAP_LOONGARCH_LASX)
|
||||
result |= LoongArch64IntrinsicConstants_LASX;
|
||||
|
||||
#endif // HAVE_GETAUXVAL
|
||||
|
||||
#endif // HOST_UNIX
|
||||
|
||||
#endif // HOST_LOONGARCH64
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -58,6 +58,11 @@ static_assert((1 << ARM64_ATOMICS_FEATURE_FLAG_BIT) == ARM64IntrinsicConstants_A
|
|||
#define RiscV64IntrinsicConstants_Zbb (1 << 1)
|
||||
#endif // HOST_RISCV64
|
||||
|
||||
#if defined(HOST_LOONGARCH64)
|
||||
#define LoongArch64IntrinsicConstants_LSX (1 << 4)
|
||||
#define LoongArch64IntrinsicConstants_LASX (1 << 5)
|
||||
#endif // HOST_LOONGARCH64
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C"
|
||||
{
|
||||
|
|
|
|||
Loading…
Reference in New Issue