-#if defined(USE_X86_ASM) || defined(SLANG_X86)
+#ifdef USE_X86_ASM
+#if defined(__i386__) || defined(__386__)
-#include "imports.h"
+#include <stdio.h>
+#include <string.h>
+#include <assert.h>
+
+#include "main/execmem.h"
#include "x86sse.h"
#define DISASSEM 0
#define X86_TWOB 0x0f
-/* Emit bytes to the instruction stream:
- */
-static void emit_1b( struct x86_function *p, GLbyte b0 )
+#if 0
+static unsigned char *cptr( void (*label)() )
{
- *(GLbyte *)(p->csr++) = b0;
+ return (unsigned char *)(unsigned long)label;
}
+#endif
+
-static void emit_1i( struct x86_function *p, GLint i0 )
+static void do_realloc( struct x86_function *p )
{
- *(GLint *)(p->csr) = i0;
- p->csr += 4;
+ if (p->size == 0) {
+ p->size = 1024;
+ p->store = _mesa_exec_malloc(p->size);
+ p->csr = p->store;
+ }
+ else {
+ unsigned used = p->csr - p->store;
+ unsigned char *tmp = p->store;
+ p->size *= 2;
+ p->store = _mesa_exec_malloc(p->size);
+ memcpy(p->store, tmp, used);
+ p->csr = p->store + used;
+ _mesa_exec_free(tmp);
+ }
}
-static void disassem( struct x86_function *p, const char *fn )
+/* Emit bytes to the instruction stream:
+ */
+static unsigned char *reserve( struct x86_function *p, int bytes )
{
-#if DISASSEM && 0
- if (fn && fn != p->fn) {
- _mesa_printf("0x%x: %s\n", p->csr, fn);
- p->fn = fn;
+ if (p->csr + bytes - p->store > p->size)
+ do_realloc(p);
+
+ {
+ unsigned char *csr = p->csr;
+ p->csr += bytes;
+ return csr;
}
-#endif
}
-static void emit_1ub_fn( struct x86_function *p, GLubyte b0, const char *fn )
+
+
+static void emit_1b( struct x86_function *p, char b0 )
{
- disassem(p, fn);
- *(p->csr++) = b0;
+ char *csr = (char *)reserve(p, 1);
+ *csr = b0;
}
-static void emit_2ub_fn( struct x86_function *p, GLubyte b0, GLubyte b1, const char *fn )
+static void emit_1i( struct x86_function *p, int i0 )
{
- disassem(p, fn);
- *(p->csr++) = b0;
- *(p->csr++) = b1;
+ int *icsr = (int *)reserve(p, sizeof(i0));
+ *icsr = i0;
}
-static void emit_3ub_fn( struct x86_function *p, GLubyte b0, GLubyte b1, GLubyte b2, const char *fn )
+static void emit_1ub( struct x86_function *p, unsigned char b0 )
{
- disassem(p, fn);
- *(p->csr++) = b0;
- *(p->csr++) = b1;
- *(p->csr++) = b2;
+ unsigned char *csr = reserve(p, 1);
+ *csr++ = b0;
}
-#define emit_1ub(p, b0) emit_1ub_fn(p, b0, __FUNCTION__)
-#define emit_2ub(p, b0, b1) emit_2ub_fn(p, b0, b1, __FUNCTION__)
-#define emit_3ub(p, b0, b1, b2) emit_3ub_fn(p, b0, b1, b2, __FUNCTION__)
+static void emit_2ub( struct x86_function *p, unsigned char b0, unsigned char b1 )
+{
+ unsigned char *csr = reserve(p, 2);
+ *csr++ = b0;
+ *csr++ = b1;
+}
+static void emit_3ub( struct x86_function *p, unsigned char b0, unsigned char b1, unsigned char b2 )
+{
+ unsigned char *csr = reserve(p, 3);
+ *csr++ = b0;
+ *csr++ = b1;
+ *csr++ = b2;
+}
/* Build a modRM byte + possible displacement. No treatment of SIB
* indexing. BZZT - no way to encode an absolute address.
*/
-static void emit_modrm( struct x86_function *p,
- struct x86_reg reg,
+static void emit_modrm( struct x86_function *p,
+ struct x86_reg reg,
struct x86_reg regmem )
{
- GLubyte val = 0;
-
+ unsigned char val = 0;
+
assert(reg.mod == mod_REG);
-
+
val |= regmem.mod << 6; /* mod field */
val |= reg.idx << 3; /* reg field */
val |= regmem.idx; /* r/m field */
-
- emit_1ub_fn(p, val, 0);
+
+ emit_1ub(p, val);
/* Oh-oh we've stumbled into the SIB thing.
*/
if (regmem.file == file_REG32 &&
regmem.idx == reg_SP) {
- emit_1ub_fn(p, 0x24, 0); /* simplistic! */
+ emit_1ub(p, 0x24); /* simplistic! */
}
switch (regmem.mod) {
static void emit_modrm_noreg( struct x86_function *p,
- GLuint op,
+ unsigned op,
struct x86_reg regmem )
{
struct x86_reg dummy = x86_make_reg(file_REG32, op);
* the arguments presented.
*/
static void emit_op_modrm( struct x86_function *p,
- GLubyte op_dst_is_reg,
- GLubyte op_dst_is_mem,
+ unsigned char op_dst_is_reg,
+ unsigned char op_dst_is_mem,
struct x86_reg dst,
struct x86_reg src )
-{
+{
switch (dst.mod) {
case mod_REG:
- emit_1ub_fn(p, op_dst_is_reg, 0);
+ emit_1ub(p, op_dst_is_reg);
emit_modrm(p, dst, src);
break;
case mod_INDIRECT:
case mod_DISP32:
case mod_DISP8:
assert(src.mod == mod_REG);
- emit_1ub_fn(p, op_dst_is_mem, 0);
+ emit_1ub(p, op_dst_is_mem);
emit_modrm(p, src, dst);
break;
default:
}
struct x86_reg x86_make_disp( struct x86_reg reg,
- GLint disp )
+ int disp )
{
assert(reg.file == file_REG32);
return x86_make_reg( reg.file, reg.idx );
}
-GLubyte *x86_get_label( struct x86_function *p )
+unsigned char *x86_get_label( struct x86_function *p )
{
return p->csr;
}
void x86_jcc( struct x86_function *p,
enum x86_cc cc,
- GLubyte *label )
+ unsigned char *label )
{
- GLint offset = label - (x86_get_label(p) + 2);
-
+ int offset = label - (x86_get_label(p) + 2);
+
if (offset <= 127 && offset >= -128) {
emit_1ub(p, 0x70 + cc);
- emit_1b(p, (GLbyte) offset);
+ emit_1b(p, (char) offset);
}
else {
offset = label - (x86_get_label(p) + 6);
/* Always use a 32bit offset for forward jumps:
*/
-GLubyte *x86_jcc_forward( struct x86_function *p,
+unsigned char *x86_jcc_forward( struct x86_function *p,
enum x86_cc cc )
{
emit_2ub(p, 0x0f, 0x80 + cc);
return x86_get_label(p);
}
-GLubyte *x86_jmp_forward( struct x86_function *p)
+unsigned char *x86_jmp_forward( struct x86_function *p)
{
emit_1ub(p, 0xe9);
emit_1i(p, 0);
return x86_get_label(p);
}
-GLubyte *x86_call_forward( struct x86_function *p)
+unsigned char *x86_call_forward( struct x86_function *p)
{
emit_1ub(p, 0xe8);
emit_1i(p, 0);
/* Fixup offset from forward jump:
*/
void x86_fixup_fwd_jump( struct x86_function *p,
- GLubyte *fixup )
+ unsigned char *fixup )
{
*(int *)(fixup - 4) = x86_get_label(p) - fixup;
}
-void x86_jmp( struct x86_function *p, GLubyte *label)
+void x86_jmp( struct x86_function *p, unsigned char *label)
{
emit_1ub(p, 0xe9);
emit_1i(p, label - x86_get_label(p) - 4);
}
-void x86_call( struct x86_function *p, GLubyte *label)
+#if 0
+/* This doesn't work once we start reallocating & copying the
+ * generated code on buffer fills, because the call is relative to the
+ * current pc.
+ */
+void x86_call( struct x86_function *p, void (*label)())
{
emit_1ub(p, 0xe8);
- emit_1i(p, label - x86_get_label(p) - 4);
+ emit_1i(p, cptr(label) - x86_get_label(p) - 4);
}
+#else
+void x86_call( struct x86_function *p, struct x86_reg reg)
+{
+ emit_1ub(p, 0xff);
+ emit_modrm_noreg(p, 2, reg);
+}
+#endif
+
/* michal:
* Temporary. As I need immediate operands, and dont want to mess with the codegen,
* I load the immediate into general purpose register and use it.
*/
-void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, GLint imm )
+void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm )
{
assert(dst.mod == mod_REG);
emit_1ub(p, 0xb8 + dst.idx);
emit_op_modrm(p, 0x2b, 0x29, dst, src );
}
+void x86_or( struct x86_function *p,
+ struct x86_reg dst,
+ struct x86_reg src )
+{
+ emit_op_modrm( p, 0x0b, 0x09, dst, src );
+}
+
+void x86_and( struct x86_function *p,
+ struct x86_reg dst,
+ struct x86_reg src )
+{
+ emit_op_modrm( p, 0x23, 0x21, dst, src );
+}
+
/***********************************************************************
emit_modrm( p, dst, src );
}
+void sse_andnps( struct x86_function *p,
+ struct x86_reg dst,
+ struct x86_reg src )
+{
+ emit_2ub(p, X86_TWOB, 0x55);
+ emit_modrm( p, dst, src );
+}
+
void sse_andps( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
emit_modrm( p, dst, src );
}
+void sse_rsqrtps( struct x86_function *p,
+ struct x86_reg dst,
+ struct x86_reg src )
+{
+ emit_2ub(p, X86_TWOB, 0x52);
+ emit_modrm( p, dst, src );
+}
void sse_rsqrtss( struct x86_function *p,
struct x86_reg dst,
emit_modrm( p, dst, src );
}
+void sse_orps( struct x86_function *p,
+ struct x86_reg dst,
+ struct x86_reg src )
+{
+ emit_2ub(p, X86_TWOB, 0x56);
+ emit_modrm( p, dst, src );
+}
+
+void sse_xorps( struct x86_function *p,
+ struct x86_reg dst,
+ struct x86_reg src )
+{
+ emit_2ub(p, X86_TWOB, 0x57);
+ emit_modrm( p, dst, src );
+}
void sse_cvtps2pi( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
{
- assert(dst.file == file_MMX &&
+ assert(dst.file == file_MMX &&
(src.file == file_XMM || src.mod != mod_REG));
p->need_emms = 1;
void sse_shufps( struct x86_function *p,
struct x86_reg dest,
struct x86_reg arg0,
- GLubyte shuf)
+ unsigned char shuf)
{
emit_2ub(p, X86_TWOB, 0xC6);
emit_modrm(p, dest, arg0);
- emit_1ub(p, shuf);
+ emit_1ub(p, shuf);
}
void sse_cmpps( struct x86_function *p,
struct x86_reg dest,
struct x86_reg arg0,
- GLubyte cc)
+ unsigned char cc)
{
emit_2ub(p, X86_TWOB, 0xC2);
emit_modrm(p, dest, arg0);
- emit_1ub(p, cc);
+ emit_1ub(p, cc);
+}
+
+void sse_pmovmskb( struct x86_function *p,
+ struct x86_reg dest,
+ struct x86_reg src)
+{
+ emit_3ub(p, 0x66, X86_TWOB, 0xD7);
+ emit_modrm(p, dest, src);
}
/***********************************************************************
void sse2_pshufd( struct x86_function *p,
struct x86_reg dest,
struct x86_reg arg0,
- GLubyte shuf)
+ unsigned char shuf)
{
emit_3ub(p, 0x66, X86_TWOB, 0x70);
emit_modrm(p, dest, arg0);
- emit_1ub(p, shuf);
+ emit_1ub(p, shuf);
+}
+
+void sse2_cvttps2dq( struct x86_function *p,
+ struct x86_reg dst,
+ struct x86_reg src )
+{
+ emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
+ emit_modrm( p, dst, src );
}
void sse2_cvtps2dq( struct x86_function *p,
emit_modrm( p, dst, src );
}
+void sse2_rcpps( struct x86_function *p,
+ struct x86_reg dst,
+ struct x86_reg src )
+{
+ emit_2ub(p, X86_TWOB, 0x53);
+ emit_modrm( p, dst, src );
+}
+
void sse2_rcpss( struct x86_function *p,
struct x86_reg dst,
struct x86_reg src )
static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
- GLubyte dst0ub0,
- GLubyte dst0ub1,
- GLubyte arg0ub0,
- GLubyte arg0ub1,
- GLubyte argmem_noreg)
+ unsigned char dst0ub0,
+ unsigned char dst0ub1,
+ unsigned char arg0ub0,
+ unsigned char arg0ub1,
+ unsigned char argmem_noreg)
{
assert(dst.file == file_x87);
if (arg.file == file_x87) {
- if (dst.idx == 0)
+ if (dst.idx == 0)
emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
- else if (arg.idx == 0)
+ else if (arg.idx == 0)
emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
else
assert(0);
}
else if (dst.idx == 0) {
- assert(arg.file = file_REG32);
+ assert(arg.file == file_REG32);
emit_1ub(p, 0xd8);
emit_modrm_noreg(p, argmem_noreg, arg);
}
void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
- x87_arith_op(p, dst, arg,
+ x87_arith_op(p, dst, arg,
0xd8, 0xc8,
0xdc, 0xc8,
4);
void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
- x87_arith_op(p, dst, arg,
+ x87_arith_op(p, dst, arg,
0xd8, 0xe0,
0xdc, 0xe8,
4);
void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
- x87_arith_op(p, dst, arg,
+ x87_arith_op(p, dst, arg,
0xd8, 0xe8,
0xdc, 0xe0,
5);
void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
- x87_arith_op(p, dst, arg,
+ x87_arith_op(p, dst, arg,
0xd8, 0xc0,
0xdc, 0xc0,
0);
void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
- x87_arith_op(p, dst, arg,
+ x87_arith_op(p, dst, arg,
0xd8, 0xf0,
0xdc, 0xf8,
6);
void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
- x87_arith_op(p, dst, arg,
+ x87_arith_op(p, dst, arg,
0xd8, 0xf8,
0xdc, 0xf0,
7);
/* st1 = st1 * log2(st0 + 1.0);
* pop_stack;
*
- * A fast operation, with restrictions: -.29 < st0 < .29
+ * A fast operation, with restrictions: -.29 < st0 < .29
*/
void x87_fyl2xp1( struct x86_function *p )
{
void x87_fld( struct x86_function *p, struct x86_reg arg )
{
- if (arg.file == file_x87)
+ if (arg.file == file_x87)
emit_2ub(p, 0xd9, 0xc0 + arg.idx);
else {
emit_1ub(p, 0xd9);
void x87_fst( struct x86_function *p, struct x86_reg dst )
{
- if (dst.file == file_x87)
+ if (dst.file == file_x87)
emit_2ub(p, 0xdd, 0xd0 + dst.idx);
else {
emit_1ub(p, 0xd9);
void x87_fstp( struct x86_function *p, struct x86_reg dst )
{
- if (dst.file == file_x87)
+ if (dst.file == file_x87)
emit_2ub(p, 0xdd, 0xd8 + dst.idx);
else {
emit_1ub(p, 0xd9);
void x87_fcom( struct x86_function *p, struct x86_reg dst )
{
- if (dst.file == file_x87)
+ if (dst.file == file_x87)
emit_2ub(p, 0xd8, 0xd0 + dst.idx);
else {
emit_1ub(p, 0xd8);
void x87_fcomp( struct x86_function *p, struct x86_reg dst )
{
- if (dst.file == file_x87)
+ if (dst.file == file_x87)
emit_2ub(p, 0xd8, 0xd8 + dst.idx);
else {
emit_1ub(p, 0xd8);
assert(dst.file == file_REG32);
if (dst.idx == reg_AX &&
- dst.mod == mod_REG)
+ dst.mod == mod_REG)
emit_2ub(p, 0xdf, 0xe0);
else {
emit_1ub(p, 0xdd);
struct x86_reg dst,
struct x86_reg src )
{
- assert(dst.file == file_MMX &&
+ assert(dst.file == file_MMX &&
(src.file == file_MMX || src.mod != mod_REG));
p->need_emms = 1;
struct x86_reg dst,
struct x86_reg src )
{
- assert(dst.file == file_MMX &&
+ assert(dst.file == file_MMX &&
(src.file == file_MMX || src.mod != mod_REG));
p->need_emms = 1;
* account any push/pop activity:
*/
struct x86_reg x86_fn_arg( struct x86_function *p,
- GLuint arg )
+ unsigned arg )
{
- return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
+ return x86_make_disp(x86_make_reg(file_REG32, reg_SP),
p->stack_offset + arg * 4); /* ??? */
}
void x86_init_func( struct x86_function *p )
{
- x86_init_func_size(p, 1024);
+ p->size = 0;
+ p->store = NULL;
+ p->csr = p->store;
}
-void x86_init_func_size( struct x86_function *p, GLuint code_size )
+int x86_init_func_size( struct x86_function *p, unsigned code_size )
{
+ p->size = code_size;
p->store = _mesa_exec_malloc(code_size);
p->csr = p->store;
+ return p->store != NULL;
}
void x86_release_func( struct x86_function *p )
{
_mesa_exec_free(p->store);
+ p->store = NULL;
+ p->csr = NULL;
+ p->size = 0;
}
void (*x86_get_func( struct x86_function *p ))(void)
{
- if (DISASSEM)
- _mesa_printf("disassemble %p %p\n", p->store, p->csr);
- return (void (*)(void))p->store;
+ if (DISASSEM && p->store)
+ printf("disassemble %p %p\n", p->store, p->csr);
+ return (void (*)(void)) (unsigned long) p->store;
}
#else
}
#endif
+
+#else /* USE_X86_ASM */
+
+int x86sse_c_dummy_var; /* silence warning */
+
+#endif /* USE_X86_ASM */