mirror of
https://github.com/ptitSeb/box64.git
synced 2025-05-09 00:21:32 +08:00
Improved x87 FIST(T)(P) opcode ([ARM64_DYNAREC] too)
This commit is contained in:
ptitSeb
parent
dbaee7c49c
commit
2ebde976db
@ -178,6 +178,7 @@ uintptr_t dynarec64_DF(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
|
||||
#else
|
||||
MRS_fpsr(x5);
|
||||
BFCw(x5, FPSR_IOC, 1); // reset IOC bit
|
||||
BFCw(x5, FPSR_QC, 1); // reset QC bit
|
||||
MSR_fpsr(x5);
|
||||
if(ST_IS_F(0)) {
|
||||
VFCVTZSs(s0, v1);
|
||||
@ -185,13 +186,16 @@ uintptr_t dynarec64_DF(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
|
||||
VFCVTZSd(s0, v1);
|
||||
SQXTN_S_D(s0, s0);
|
||||
}
|
||||
SQXTN_H_S(s0, s0);
|
||||
VST16(s0, wback, fixedaddress);
|
||||
VMOVSto(x3, s0, 0);
|
||||
MRS_fpsr(x5); // get back FPSR to check the IOC bit
|
||||
TBZ_MARK3(x5, FPSR_IOC);
|
||||
MOV32w(x5, 0x8000);
|
||||
STH(x5, wback, fixedaddress);
|
||||
TBNZ_MARK2(x5, FPSR_IOC);
|
||||
SXTHw(x5, x3); // check if 16bits value is fine
|
||||
SUBw_REG(x5, x5, x3);
|
||||
CBZw_MARK3(x5);
|
||||
MARK2;
|
||||
MOV32w(x3, 0x8000);
|
||||
MARK3;
|
||||
STH(x3, wback, fixedaddress);
|
||||
#endif
|
||||
x87_do_pop(dyn, ninst, x3);
|
||||
break;
|
||||
@ -212,6 +216,7 @@ uintptr_t dynarec64_DF(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
|
||||
#else
|
||||
MRS_fpsr(x5);
|
||||
BFCw(x5, FPSR_IOC, 1); // reset IOC bit
|
||||
BFCw(x5, FPSR_QC, 1); // reset QC bit
|
||||
MSR_fpsr(x5);
|
||||
if(ST_IS_F(0)) {
|
||||
FRINTXS(s0, v1);
|
||||
@ -221,13 +226,16 @@ uintptr_t dynarec64_DF(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
|
||||
VFCVTZSd(s0, s0);
|
||||
SQXTN_S_D(s0, s0);
|
||||
}
|
||||
SQXTN_H_S(s0, s0);
|
||||
VST16(s0, wback, fixedaddress);
|
||||
VMOVSto(x3, s0, 0);
|
||||
MRS_fpsr(x5); // get back FPSR to check the IOC bit
|
||||
TBZ_MARK3(x5, FPSR_IOC);
|
||||
MOV32w(x5, 0x8000);
|
||||
STH(x5, wback, fixedaddress);
|
||||
TBNZ_MARK2(x5, FPSR_IOC);
|
||||
SXTHw(x5, x3); // check if 16bits value is fine
|
||||
SUBw_REG(x5, x5, x3);
|
||||
CBZw_MARK3(x5);
|
||||
MARK2;
|
||||
MOV32w(x3, 0x8000);
|
||||
MARK3;
|
||||
STH(x3, wback, fixedaddress);
|
||||
#endif
|
||||
x87_restoreround(dyn, ninst, u8);
|
||||
break;
|
||||
@ -257,13 +265,16 @@ uintptr_t dynarec64_DF(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
|
||||
VFCVTZSd(s0, s0);
|
||||
SQXTN_S_D(s0, s0);
|
||||
}
|
||||
SQXTN_H_S(s0, s0);
|
||||
VST16(s0, wback, fixedaddress);
|
||||
VMOVSto(x3, s0, 0);
|
||||
MRS_fpsr(x5); // get back FPSR to check the IOC bit
|
||||
TBZ_MARK3(x5, FPSR_IOC);
|
||||
MOV32w(x5, 0x8000);
|
||||
STH(x5, wback, fixedaddress);
|
||||
TBNZ_MARK2(x5, FPSR_IOC);
|
||||
SXTHw(x5, x3); // check if 16bits value is fine
|
||||
SUBw_REG(x5, x5, x3);
|
||||
CBZw_MARK3(x5);
|
||||
MARK2;
|
||||
MOV32w(x3, 0x8000);
|
||||
MARK3;
|
||||
STH(x3, wback, fixedaddress);
|
||||
#endif
|
||||
x87_do_pop(dyn, ninst, x3);
|
||||
x87_restoreround(dyn, ninst, u8);
|
||||
|
||||
@ -125,7 +125,14 @@ uintptr_t RunDD(x64emu_t *emu, rex_t rex, uintptr_t addr)
|
||||
break;
|
||||
case 1: /* FISTTP ED qword */
|
||||
GETE8(0);
|
||||
*(int64_t*)ED = ST0.d;
|
||||
if(STll(0).sref==ST(0).sq)
|
||||
ED->sq[0] = STll(0).sq;
|
||||
else {
|
||||
if(isgreater(ST0.d, (double)0x7fffffffffffffffLL) || isless(ST0.d, -(double)0x8000000000000000LL) || !isfinite(ST0.d))
|
||||
*(uint64_t*)ED = 0x8000000000000000LL;
|
||||
else
|
||||
*(int64_t*)ED = ST0.d;
|
||||
}
|
||||
fpu_do_pop(emu);
|
||||
break;
|
||||
case 2: /* FST double */
|
||||
|
||||
@ -175,7 +175,7 @@ uintptr_t RunDF(x64emu_t *emu, rex_t rex, uintptr_t addr)
|
||||
if(STll(0).sref==ST(0).sq)
|
||||
ED->sq[0] = STll(0).sq;
|
||||
else {
|
||||
if(isgreater(ST0.d, (double)(int64_t)0x7fffffffffffffffLL) || isless(ST0.d, (double)(int64_t)0x8000000000000000LL) || !isfinite(ST0.d))
|
||||
if(isgreater(ST0.d, (double)0x7fffffffffffffffLL) || isless(ST0.d, -(double)0x8000000000000000LL) || !isfinite(ST0.d))
|
||||
ED->sq[0] = 0x8000000000000000LL;
|
||||
else
|
||||
ED->sq[0] = fpu_round(emu, ST0.d);
|
||||
|
||||
1354
tests/ref22.txt
1354
tests/ref22.txt
File diff suppressed because it is too large
Load Diff
BIN
tests/test22
BIN
tests/test22
Binary file not shown.
101
tests/test22.c
101
tests/test22.c
@ -5,6 +5,8 @@
|
||||
#include <stdint.h>
|
||||
#include <math.h>
|
||||
|
||||
// Build with `gcc -march=core2 -mfpmath=387 -O2 test22.c -o test22`
|
||||
|
||||
#if defined(__x86_64__)
|
||||
uint64_t _fucomip_(double a, double b)
|
||||
{
|
||||
@ -30,7 +32,7 @@ uint64_t _fucompp_(double a, double b)
|
||||
:"=a" (ret):"m"(a), "m"(b):"cc");
|
||||
return ret;
|
||||
}
|
||||
uint64_t _fistpw_(double a)
|
||||
uint64_t _fistpw_(double a, uint8_t rd)
|
||||
{
|
||||
uint16_t ret;
|
||||
uint16_t t1, t2;
|
||||
@ -38,15 +40,31 @@ uint64_t _fistpw_(double a)
|
||||
"fldl %1\n"
|
||||
"fstcw %2\n"
|
||||
"mov %2, %%ax\n"
|
||||
"mov $0x0c, %%ah\n"
|
||||
"mov %4, %%ah\n"
|
||||
"mov %%ax, %3\n"
|
||||
"fldcw %3\n"
|
||||
"fistp %0\n"
|
||||
"fldcw %2\n"
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2):"cc");
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2), "m"(rd):"cc");
|
||||
return ret;
|
||||
}
|
||||
uint64_t _fistpl_(double a)
|
||||
uint64_t _fisttpw_(double a, uint8_t rd)
|
||||
{
|
||||
uint16_t ret;
|
||||
uint16_t t1, t2;
|
||||
asm volatile (
|
||||
"fldl %1\n"
|
||||
"fstcw %2\n"
|
||||
"mov %2, %%ax\n"
|
||||
"mov %4, %%ah\n"
|
||||
"mov %%ax, %3\n"
|
||||
"fldcw %3\n"
|
||||
"fisttp %0\n"
|
||||
"fldcw %2\n"
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2), "m"(rd):"cc");
|
||||
return ret;
|
||||
}
|
||||
uint64_t _fistpl_(double a, uint8_t rd)
|
||||
{
|
||||
uint32_t ret;
|
||||
uint16_t t1, t2;
|
||||
@ -54,15 +72,31 @@ uint64_t _fistpl_(double a)
|
||||
"fldl %1\n"
|
||||
"fstcw %2\n"
|
||||
"mov %2, %%ax\n"
|
||||
"mov $0x0c, %%ah\n"
|
||||
"mov %4, %%ah\n"
|
||||
"mov %%ax, %3\n"
|
||||
"fldcw %3\n"
|
||||
"fistpl %0\n"
|
||||
"fldcw %2\n"
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2):"cc");
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2), "m"(rd):"cc");
|
||||
return ret;
|
||||
}
|
||||
uint64_t _fistpq_(double a)
|
||||
uint64_t _fisttpl_(double a, uint8_t rd)
|
||||
{
|
||||
uint32_t ret;
|
||||
uint16_t t1, t2;
|
||||
asm volatile (
|
||||
"fldl %1\n"
|
||||
"fstcw %2\n"
|
||||
"mov %2, %%ax\n"
|
||||
"mov %4, %%ah\n"
|
||||
"mov %%ax, %3\n"
|
||||
"fldcw %3\n"
|
||||
"fisttpl %0\n"
|
||||
"fldcw %2\n"
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2), "m"(rd):"cc");
|
||||
return ret;
|
||||
}
|
||||
uint64_t _fistpq_(double a, uint8_t rd)
|
||||
{
|
||||
uint64_t ret;
|
||||
uint16_t t1, t2;
|
||||
@ -70,15 +104,15 @@ uint64_t _fistpq_(double a)
|
||||
"fldl %1\n"
|
||||
"fstcw %2\n"
|
||||
"mov %2, %%ax\n"
|
||||
"mov $0x0c, %%ah\n"
|
||||
"mov %4, %%ah\n"
|
||||
"mov %%ax, %3\n"
|
||||
"fldcw %3\n"
|
||||
"fistpq %0\n"
|
||||
"fldcw %2\n"
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2):"cc");
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2), "m"(rd):"cc");
|
||||
return ret;
|
||||
}
|
||||
uint64_t _frndint_(double a)
|
||||
uint64_t _fisttpq_(double a, uint8_t rd)
|
||||
{
|
||||
uint64_t ret;
|
||||
uint16_t t1, t2;
|
||||
@ -86,13 +120,29 @@ uint64_t _frndint_(double a)
|
||||
"fldl %1\n"
|
||||
"fstcw %2\n"
|
||||
"mov %2, %%ax\n"
|
||||
"mov $0x0c, %%ah\n"
|
||||
"mov %4, %%ah\n"
|
||||
"mov %%ax, %3\n"
|
||||
"fldcw %3\n"
|
||||
"fisttpq %0\n"
|
||||
"fldcw %2\n"
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2), "m"(rd):"cc");
|
||||
return ret;
|
||||
}
|
||||
uint64_t _frndint_(double a, uint8_t rd)
|
||||
{
|
||||
uint64_t ret;
|
||||
uint16_t t1, t2;
|
||||
asm volatile (
|
||||
"fldl %1\n"
|
||||
"fstcw %2\n"
|
||||
"mov %2, %%ax\n"
|
||||
"mov %4, %%ah\n"
|
||||
"mov %%ax, %3\n"
|
||||
"fldcw %3\n"
|
||||
"frndint\n"
|
||||
"fstpl %0\n"
|
||||
"fldcw %2\n"
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2):"cc");
|
||||
:"=m" (ret):"m"(a), "m"(t1), "m"(t2), "m"(rd):"cc");
|
||||
return ret;
|
||||
}
|
||||
#endif
|
||||
@ -107,6 +157,8 @@ int main(int argc, const char** argv)
|
||||
{0x0, 0x4082c00000000000LL},
|
||||
{0x8000000000000000LL, 0x4082c00000000000LL},
|
||||
{0x8000000000000000LL, 0x0},
|
||||
{0x40dfffc000000000LL, 0x40e0002000000000LL}, // 0x7fff and 0x8001 as double
|
||||
{0xc0e0002000000000LL, 0xc0dfffc000000000LL}, // -0x8001 and -0x7fff as double
|
||||
{0x8000000000000000LL, 0x3ff0000000000000LL},
|
||||
{0x3ff0000000000000LL, 0x3fe89d9000000000LL},
|
||||
{0x3ff0000000000000LL, 0x7ff0000000000000LL},
|
||||
@ -133,14 +185,23 @@ int main(int argc, const char** argv)
|
||||
za = (flags>>(8+6))&1?'Z':'-';
|
||||
pa = (flags>>(8+2))&1?'P':'-';
|
||||
printf("%c%c%c\n", za, pa, ca);
|
||||
printf("FRNDINT 0x%llx => 0x%llx\n", *(uint64_t*)&a, _frndint_(a));
|
||||
printf("FRNDINT 0x%llx => 0x%llx\n", *(uint64_t*)&b, _frndint_(b));
|
||||
printf("FISTP 0x%llx => word: %x\n", *(uint64_t*)&a, _fistpw_(a));
|
||||
printf("FISTP 0x%llx => word: %x\n", *(uint64_t*)&b, _fistpw_(b));
|
||||
printf("FISTP 0x%llx => long: %x\n", *(uint64_t*)&a, _fistpl_(a));
|
||||
printf("FISTP 0x%llx => long: %x\n", *(uint64_t*)&b, _fistpl_(b));
|
||||
printf("FISTP 0x%llx => quad: %llx\n", *(uint64_t*)&a, _fistpq_(a));
|
||||
printf("FISTP 0x%llx => quad: %llx\n", *(uint64_t*)&b, _fistpq_(b));
|
||||
for(int rd=0; rd<3; ++rd) {
|
||||
printf("Rounding %d\n", rd);
|
||||
printf(" FRNDINT 0x%llx => 0x%llx\n", *(uint64_t*)&a, _frndint_(a, rd<<2));
|
||||
printf(" FRNDINT 0x%llx => 0x%llx\n", *(uint64_t*)&b, _frndint_(b, rd<<2));
|
||||
printf(" FISTP 0x%llx => word: %x\n", *(uint64_t*)&a, _fistpw_(a, rd<<2));
|
||||
printf(" FISTP 0x%llx => word: %x\n", *(uint64_t*)&b, _fistpw_(b, rd<<2));
|
||||
printf(" FISTP 0x%llx => long: %x\n", *(uint64_t*)&a, _fistpl_(a, rd<<2));
|
||||
printf(" FISTP 0x%llx => long: %x\n", *(uint64_t*)&b, _fistpl_(b, rd<<2));
|
||||
printf(" FISTP 0x%llx => quad: %llx\n", *(uint64_t*)&a, _fistpq_(a, rd<<2));
|
||||
printf(" FISTP 0x%llx => quad: %llx\n", *(uint64_t*)&b, _fistpq_(b, rd<<2));
|
||||
printf(" FISTTP 0x%llx => word: %x\n", *(uint64_t*)&a, _fisttpw_(a, rd<<2));
|
||||
printf(" FISTTP 0x%llx => word: %x\n", *(uint64_t*)&b, _fisttpw_(b, rd<<2));
|
||||
printf(" FISTTP 0x%llx => long: %x\n", *(uint64_t*)&a, _fisttpl_(a, rd<<2));
|
||||
printf(" FISTTP 0x%llx => long: %x\n", *(uint64_t*)&b, _fisttpl_(b, rd<<2));
|
||||
printf(" FISTTP 0x%llx => quad: %llx\n", *(uint64_t*)&a, _fisttpq_(a, rd<<2));
|
||||
printf(" FISTTP 0x%llx => quad: %llx\n", *(uint64_t*)&b, _fisttpq_(b, rd<<2));
|
||||
}
|
||||
}
|
||||
printf("\nDone\n");
|
||||
}
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user