11 |
|
|
12 |
|
#ifdef OPTIMIZED_FLAGS |
13 |
|
|
14 |
< |
#if defined(__i386__) && defined(X86_ASSEMBLY) |
14 |
> |
#if (defined(__i386__) && defined(X86_ASSEMBLY)) || (defined(__x86_64__) && defined(X86_64_ASSEMBLY)) |
15 |
|
|
16 |
|
#ifndef SAHF_SETO_PROFITABLE |
17 |
|
|
18 |
+ |
/* PUSH/POP instructions are naturally 64-bit sized on x86-64, thus |
19 |
+ |
unsigned long hereunder is either 64-bit or 32-bit wide depending |
20 |
+ |
on the target. */ |
21 |
|
struct flag_struct { |
22 |
< |
uae_u32 cznv; |
23 |
< |
uae_u32 x; |
22 |
> |
unsigned long cznv; |
23 |
> |
unsigned long x; |
24 |
|
}; |
25 |
|
|
26 |
|
#define FLAGVAL_Z 0x40 |
27 |
|
#define FLAGVAL_N 0x80 |
28 |
|
|
29 |
< |
#define SET_ZFLG(y) (regflags.cznv = (regflags.cznv & ~0x40) | (((y) & 1) << 6)) |
30 |
< |
#define SET_CFLG(y) (regflags.cznv = (regflags.cznv & ~1) | ((y) & 1)) |
31 |
< |
#define SET_VFLG(y) (regflags.cznv = (regflags.cznv & ~0x800) | (((y) & 1) << 11)) |
32 |
< |
#define SET_NFLG(y) (regflags.cznv = (regflags.cznv & ~0x80) | (((y) & 1) << 7)) |
29 |
> |
#define SET_ZFLG(y) (regflags.cznv = (((uae_u32)regflags.cznv) & ~0x40) | (((y) & 1) << 6)) |
30 |
> |
#define SET_CFLG(y) (regflags.cznv = (((uae_u32)regflags.cznv) & ~1) | ((y) & 1)) |
31 |
> |
#define SET_VFLG(y) (regflags.cznv = (((uae_u32)regflags.cznv) & ~0x800) | (((y) & 1) << 11)) |
32 |
> |
#define SET_NFLG(y) (regflags.cznv = (((uae_u32)regflags.cznv) & ~0x80) | (((y) & 1) << 7)) |
33 |
|
#define SET_XFLG(y) (regflags.x = (y)) |
34 |
|
|
35 |
|
#define GET_ZFLG ((regflags.cznv >> 6) & 1) |
77 |
|
|
78 |
|
#define optflag_testl(v) \ |
79 |
|
__asm__ __volatile__ ("andl %1,%1\n\t" \ |
80 |
< |
"pushfl\n\t" \ |
81 |
< |
"popl %0\n\t" \ |
80 |
> |
"pushf\n\t" \ |
81 |
> |
"pop %0\n\t" \ |
82 |
|
: "=r" (regflags.cznv) : "r" (v) : "cc") |
83 |
|
|
84 |
|
#define optflag_testw(v) \ |
85 |
|
__asm__ __volatile__ ("andw %w1,%w1\n\t" \ |
86 |
< |
"pushfl\n\t" \ |
87 |
< |
"popl %0\n\t" \ |
86 |
> |
"pushf\n\t" \ |
87 |
> |
"pop %0\n\t" \ |
88 |
|
: "=r" (regflags.cznv) : "r" (v) : "cc") |
89 |
|
|
90 |
|
#define optflag_testb(v) \ |
91 |
|
__asm__ __volatile__ ("andb %b1,%b1\n\t" \ |
92 |
< |
"pushfl\n\t" \ |
93 |
< |
"popl %0\n\t" \ |
92 |
> |
"pushf\n\t" \ |
93 |
> |
"pop %0\n\t" \ |
94 |
|
: "=r" (regflags.cznv) : "q" (v) : "cc") |
95 |
|
|
96 |
|
#define optflag_addl(v, s, d) do { \ |
97 |
|
__asm__ __volatile__ ("addl %k2,%k1\n\t" \ |
98 |
< |
"pushfl\n\t" \ |
99 |
< |
"popl %0\n\t" \ |
98 |
> |
"pushf\n\t" \ |
99 |
> |
"pop %0\n\t" \ |
100 |
|
: "=r" (regflags.cznv), "=r" (v) : "rmi" (s), "1" (d) : "cc"); \ |
101 |
|
COPY_CARRY; \ |
102 |
|
} while (0) |
103 |
|
|
104 |
|
#define optflag_addw(v, s, d) do { \ |
105 |
|
__asm__ __volatile__ ("addw %w2,%w1\n\t" \ |
106 |
< |
"pushfl\n\t" \ |
107 |
< |
"popl %0\n\t" \ |
106 |
> |
"pushf\n\t" \ |
107 |
> |
"pop %0\n\t" \ |
108 |
|
: "=r" (regflags.cznv), "=r" (v) : "rmi" (s), "1" (d) : "cc"); \ |
109 |
|
COPY_CARRY; \ |
110 |
|
} while (0) |
111 |
|
|
112 |
|
#define optflag_addb(v, s, d) do { \ |
113 |
|
__asm__ __volatile__ ("addb %b2,%b1\n\t" \ |
114 |
< |
"pushfl\n\t" \ |
115 |
< |
"popl %0\n\t" \ |
114 |
> |
"pushf\n\t" \ |
115 |
> |
"pop %0\n\t" \ |
116 |
|
: "=r" (regflags.cznv), "=q" (v) : "qmi" (s), "1" (d) : "cc"); \ |
117 |
|
COPY_CARRY; \ |
118 |
|
} while (0) |
119 |
|
|
120 |
|
#define optflag_subl(v, s, d) do { \ |
121 |
|
__asm__ __volatile__ ("subl %k2,%k1\n\t" \ |
122 |
< |
"pushfl\n\t" \ |
123 |
< |
"popl %0\n\t" \ |
122 |
> |
"pushf\n\t" \ |
123 |
> |
"pop %0\n\t" \ |
124 |
|
: "=r" (regflags.cznv), "=r" (v) : "rmi" (s), "1" (d) : "cc"); \ |
125 |
|
COPY_CARRY; \ |
126 |
|
} while (0) |
127 |
|
|
128 |
|
#define optflag_subw(v, s, d) do { \ |
129 |
|
__asm__ __volatile__ ("subw %w2,%w1\n\t" \ |
130 |
< |
"pushfl\n\t" \ |
131 |
< |
"popl %0\n\t" \ |
130 |
> |
"pushf\n\t" \ |
131 |
> |
"pop %0\n\t" \ |
132 |
|
: "=r" (regflags.cznv), "=r" (v) : "rmi" (s), "1" (d) : "cc"); \ |
133 |
|
COPY_CARRY; \ |
134 |
|
} while (0) |
135 |
|
|
136 |
|
#define optflag_subb(v, s, d) do { \ |
137 |
|
__asm__ __volatile__ ("subb %b2,%b1\n\t" \ |
138 |
< |
"pushfl\n\t" \ |
139 |
< |
"popl %0\n\t" \ |
138 |
> |
"pushf\n\t" \ |
139 |
> |
"pop %0\n\t" \ |
140 |
|
: "=r" (regflags.cznv), "=q" (v) : "qmi" (s), "1" (d) : "cc"); \ |
141 |
|
COPY_CARRY; \ |
142 |
|
} while (0) |
143 |
|
|
144 |
|
#define optflag_cmpl(s, d) \ |
145 |
|
__asm__ __volatile__ ("cmpl %k1,%k2\n\t" \ |
146 |
< |
"pushfl\n\t" \ |
147 |
< |
"popl %0\n\t" \ |
146 |
> |
"pushf\n\t" \ |
147 |
> |
"pop %0\n\t" \ |
148 |
|
: "=r" (regflags.cznv) : "rmi" (s), "r" (d) : "cc") |
149 |
|
|
150 |
|
#define optflag_cmpw(s, d) \ |
151 |
|
__asm__ __volatile__ ("cmpw %w1,%w2\n\t" \ |
152 |
< |
"pushfl\n\t" \ |
153 |
< |
"popl %0\n\t" \ |
152 |
> |
"pushf\n\t" \ |
153 |
> |
"pop %0\n\t" \ |
154 |
|
: "=r" (regflags.cznv) : "rmi" (s), "r" (d) : "cc") |
155 |
|
|
156 |
|
#define optflag_cmpb(s, d) \ |
157 |
|
__asm__ __volatile__ ("cmpb %b1,%b2\n\t" \ |
158 |
< |
"pushfl\n\t" \ |
159 |
< |
"popl %0\n\t" \ |
158 |
> |
"pushf\n\t" \ |
159 |
> |
"pop %0\n\t" \ |
160 |
|
: "=r" (regflags.cznv) : "qmi" (s), "q" (d) : "cc") |
161 |
|
|
162 |
|
#else |
219 |
|
return 0; |
220 |
|
} |
221 |
|
|
222 |
+ |
/* Manually emit LAHF instruction so that 64-bit assemblers can grok it */ |
223 |
+ |
#if defined __x86_64__ && defined __GNUC__ |
224 |
+ |
#define ASM_LAHF ".byte 0x9f" |
225 |
+ |
#else |
226 |
+ |
#define ASM_LAHF "lahf" |
227 |
+ |
#endif |
228 |
+ |
|
229 |
|
/* Is there any way to do this without declaring *all* memory clobbered? |
230 |
|
I.e. any way to tell gcc that some byte-sized value is in %al? */ |
231 |
|
#define optflag_testl(v) \ |
232 |
|
__asm__ __volatile__ ("andl %0,%0\n\t" \ |
233 |
< |
"lahf\n\t" \ |
233 |
> |
ASM_LAHF "\n\t" \ |
234 |
|
"seto %%al\n\t" \ |
235 |
|
"movb %%al,regflags\n\t" \ |
236 |
|
"movb %%ah,regflags+1\n\t" \ |
238 |
|
|
239 |
|
#define optflag_testw(v) \ |
240 |
|
__asm__ __volatile__ ("andw %w0,%w0\n\t" \ |
241 |
< |
"lahf\n\t" \ |
241 |
> |
ASM_LAHF "\n\t" \ |
242 |
|
"seto %%al\n\t" \ |
243 |
|
"movb %%al,regflags\n\t" \ |
244 |
|
"movb %%ah,regflags+1\n\t" \ |
246 |
|
|
247 |
|
#define optflag_testb(v) \ |
248 |
|
__asm__ __volatile__ ("andb %b0,%b0\n\t" \ |
249 |
< |
"lahf\n\t" \ |
249 |
> |
ASM_LAHF "\n\t" \ |
250 |
|
"seto %%al\n\t" \ |
251 |
|
"movb %%al,regflags\n\t" \ |
252 |
|
"movb %%ah,regflags+1\n\t" \ |
254 |
|
|
255 |
|
#define optflag_addl(v, s, d) do { \ |
256 |
|
__asm__ __volatile__ ("addl %k1,%k0\n\t" \ |
257 |
< |
"lahf\n\t" \ |
257 |
> |
ASM_LAHF "\n\t" \ |
258 |
|
"seto %%al\n\t" \ |
259 |
|
"movb %%al,regflags\n\t" \ |
260 |
|
"movb %%ah,regflags+1\n\t" \ |
264 |
|
|
265 |
|
#define optflag_addw(v, s, d) do { \ |
266 |
|
__asm__ __volatile__ ("addw %w1,%w0\n\t" \ |
267 |
< |
"lahf\n\t" \ |
267 |
> |
ASM_LAHF "\n\t" \ |
268 |
|
"seto %%al\n\t" \ |
269 |
|
"movb %%al,regflags\n\t" \ |
270 |
|
"movb %%ah,regflags+1\n\t" \ |
274 |
|
|
275 |
|
#define optflag_addb(v, s, d) do { \ |
276 |
|
__asm__ __volatile__ ("addb %b1,%b0\n\t" \ |
277 |
< |
"lahf\n\t" \ |
277 |
> |
ASM_LAHF "\n\t" \ |
278 |
|
"seto %%al\n\t" \ |
279 |
|
"movb %%al,regflags\n\t" \ |
280 |
|
"movb %%ah,regflags+1\n\t" \ |
284 |
|
|
285 |
|
#define optflag_subl(v, s, d) do { \ |
286 |
|
__asm__ __volatile__ ("subl %k1,%k0\n\t" \ |
287 |
< |
"lahf\n\t" \ |
287 |
> |
ASM_LAHF "\n\t" \ |
288 |
|
"seto %%al\n\t" \ |
289 |
|
"movb %%al,regflags\n\t" \ |
290 |
|
"movb %%ah,regflags+1\n\t" \ |
294 |
|
|
295 |
|
#define optflag_subw(v, s, d) do { \ |
296 |
|
__asm__ __volatile__ ("subw %w1,%w0\n\t" \ |
297 |
< |
"lahf\n\t" \ |
297 |
> |
ASM_LAHF "\n\t" \ |
298 |
|
"seto %%al\n\t" \ |
299 |
|
"movb %%al,regflags\n\t" \ |
300 |
|
"movb %%ah,regflags+1\n\t" \ |
304 |
|
|
305 |
|
#define optflag_subb(v, s, d) do { \ |
306 |
|
__asm__ __volatile__ ("subb %b1,%b0\n\t" \ |
307 |
< |
"lahf\n\t" \ |
307 |
> |
ASM_LAHF "\n\t" \ |
308 |
|
"seto %%al\n\t" \ |
309 |
|
"movb %%al,regflags\n\t" \ |
310 |
|
"movb %%ah,regflags+1\n\t" \ |
314 |
|
|
315 |
|
#define optflag_cmpl(s, d) \ |
316 |
|
__asm__ __volatile__ ("cmpl %k0,%k1\n\t" \ |
317 |
< |
"lahf\n\t" \ |
317 |
> |
ASM_LAHF "\n\t" \ |
318 |
|
"seto %%al\n\t" \ |
319 |
|
"movb %%al,regflags\n\t" \ |
320 |
|
"movb %%ah,regflags+1\n\t" \ |
322 |
|
|
323 |
|
#define optflag_cmpw(s, d) \ |
324 |
|
__asm__ __volatile__ ("cmpw %w0,%w1\n\t" \ |
325 |
< |
"lahf\n\t" \ |
325 |
> |
ASM_LAHF "\n\t" \ |
326 |
|
"seto %%al\n\t" \ |
327 |
|
"movb %%al,regflags\n\t" \ |
328 |
|
"movb %%ah,regflags+1\n\t" \ |
330 |
|
|
331 |
|
#define optflag_cmpb(s, d) \ |
332 |
|
__asm__ __volatile__ ("cmpb %b0,%b1\n\t" \ |
333 |
< |
"lahf\n\t" \ |
333 |
> |
ASM_LAHF "\n\t" \ |
334 |
|
"seto %%al\n\t" \ |
335 |
|
"movb %%al,regflags\n\t" \ |
336 |
|
"movb %%ah,regflags+1\n\t" \ |