static void emit_post_inc_dec(Node *node, char *op) {
SAVE;
emit_expr(node->operand);
push("rax");
emit("%s $1, %%rax", op);
emit_store(node->operand);
pop("rax");
}
static void emit_assign(Node *node) {
SAVE;
if (node->left->ctype->type == CTYPE_STRUCT &&
node->left->ctype->size > 8) {
emit_copy_struct(node->left, node->right);
} else {
emit_expr(node->right);
emit_load_convert(node->ctype, node->right->ctype);
emit_store(node->left);
}
}
static void emit_assign(Node *node) {
SAVE;
if (node->left->ty->kind == KIND_STRUCT &&
node->left->ty->size > 8) {
emit_copy_struct(node->left, node->right);
} else {
emit_expr(node->right);
emit_load_convert(node->ty, node->right->ty);
emit_store(node->left);
}
}
void emit_store_dst(jitdata *jd, instruction *iptr, s4 d)
{
emit_store(jd, iptr, VAROP(iptr->dst), d);
}
static void emit_pre_inc_dec(Node *node, char *op) {
emit_expr(node->operand);
emit("%s $1, %%rax", op);
emit_store(node->operand);
}
/**
* Emit LLVM for one TGSI instruction.
* \param return TRUE for success, FALSE otherwise
*/
static boolean
emit_instruction(
struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_instruction *inst,
const struct tgsi_opcode_info *info,
int *pc)
{
LLVMValueRef src0, src1, src2;
LLVMValueRef tmp0, tmp1;
LLVMValueRef dst0;
/*
* Stores and write masks are handled in a general fashion after the long
* instruction opcode switch statement.
*
* Although not stricitly necessary, we avoid generating instructions for
* channels which won't be stored, in cases where's that easy. For some
* complex instructions, like texture sampling, it is more convenient to
* assume a full writemask and then let LLVM optimization passes eliminate
* redundant code.
*/
(*pc)++;
assert(info->num_dst <= 1);
if (info->num_dst) {
dst0 = bld->base.undef;
}
switch (inst->Instruction.Opcode) {
case TGSI_OPCODE_ARL:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_floor(&bld->base, src0);
break;
case TGSI_OPCODE_MOV:
dst0 = emit_fetch(bld, inst, 0);
break;
case TGSI_OPCODE_LIT:
return FALSE;
case TGSI_OPCODE_RCP:
/* TGSI_OPCODE_RECIP */
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_rcp(&bld->base, src0);
break;
case TGSI_OPCODE_RSQ:
/* TGSI_OPCODE_RECIPSQRT */
src0 = emit_fetch(bld, inst, 0);
tmp0 = lp_build_abs(&bld->base, src0);
dst0 = lp_build_rsqrt(&bld->base, tmp0);
break;
case TGSI_OPCODE_EXP:
return FALSE;
case TGSI_OPCODE_LOG:
return FALSE;
case TGSI_OPCODE_MUL:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_mul(&bld->base, src0, src1);
break;
case TGSI_OPCODE_ADD:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_add(&bld->base, src0, src1);
break;
case TGSI_OPCODE_DP3:
/* TGSI_OPCODE_DOT3 */
return FALSE;
case TGSI_OPCODE_DP4:
/* TGSI_OPCODE_DOT4 */
return FALSE;
case TGSI_OPCODE_DST:
return FALSE;
case TGSI_OPCODE_MIN:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_max(&bld->base, src0, src1);
break;
case TGSI_OPCODE_MAX:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_max(&bld->base, src0, src1);
break;
case TGSI_OPCODE_SLT:
/* TGSI_OPCODE_SETLT */
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_SGE:
/* TGSI_OPCODE_SETGE */
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_MAD:
/* TGSI_OPCODE_MADD */
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp0 = lp_build_mul(&bld->base, src0, src1);
dst0 = lp_build_add(&bld->base, tmp0, src2);
break;
case TGSI_OPCODE_SUB:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_sub(&bld->base, src0, src1);
break;
case TGSI_OPCODE_LRP:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp0 = lp_build_sub(&bld->base, src1, src2);
tmp0 = lp_build_mul(&bld->base, src0, tmp0);
dst0 = lp_build_add(&bld->base, tmp0, src2);
break;
case TGSI_OPCODE_CND:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp1 = lp_build_const_vec(bld->base.type, 0.5);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src2, tmp1);
dst0 = lp_build_select(&bld->base, tmp0, src0, src1);
break;
case TGSI_OPCODE_DP2A:
return FALSE;
case TGSI_OPCODE_FRC:
src0 = emit_fetch(bld, inst, 0);
tmp0 = lp_build_floor(&bld->base, src0);
dst0 = lp_build_sub(&bld->base, src0, tmp0);
break;
case TGSI_OPCODE_CLAMP:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp0 = lp_build_max(&bld->base, src0, src1);
dst0 = lp_build_min(&bld->base, tmp0, src2);
break;
case TGSI_OPCODE_FLR:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_floor(&bld->base, src0);
break;
case TGSI_OPCODE_ROUND:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_round(&bld->base, src0);
break;
case TGSI_OPCODE_EX2:
src0 = emit_fetch(bld, inst, 0);
tmp0 = lp_build_swizzle_scalar_aos(&bld->base, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_exp2(&bld->base, tmp0);
break;
case TGSI_OPCODE_LG2:
src0 = emit_fetch(bld, inst, 0);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_log2(&bld->base, tmp0);
break;
case TGSI_OPCODE_POW:
src0 = emit_fetch(bld, inst, 0);
src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
src1 = emit_fetch(bld, inst, 1);
src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
dst0 = lp_build_pow(&bld->base, src0, src1);
break;
case TGSI_OPCODE_XPD:
return FALSE;
case TGSI_OPCODE_ABS:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_abs(&bld->base, src0);
break;
case TGSI_OPCODE_RCC:
/* deprecated? */
assert(0);
return FALSE;
case TGSI_OPCODE_DPH:
return FALSE;
case TGSI_OPCODE_COS:
src0 = emit_fetch(bld, inst, 0);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_cos(&bld->base, tmp0);
break;
case TGSI_OPCODE_DDX:
return FALSE;
case TGSI_OPCODE_DDY:
return FALSE;
case TGSI_OPCODE_KILP:
/* predicated kill */
return FALSE;
case TGSI_OPCODE_KIL:
/* conditional kill */
return FALSE;
case TGSI_OPCODE_PK2H:
return FALSE;
break;
case TGSI_OPCODE_PK2US:
return FALSE;
break;
case TGSI_OPCODE_PK4B:
return FALSE;
break;
case TGSI_OPCODE_PK4UB:
return FALSE;
case TGSI_OPCODE_RFL:
return FALSE;
case TGSI_OPCODE_SEQ:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_EQUAL, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_SFL:
dst0 = bld->base.zero;
break;
case TGSI_OPCODE_SGT:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_SIN:
src0 = emit_fetch(bld, inst, 0);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_sin(&bld->base, tmp0);
break;
case TGSI_OPCODE_SLE:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LEQUAL, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_SNE:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_STR:
dst0 = bld->base.one;
break;
case TGSI_OPCODE_TEX:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
break;
case TGSI_OPCODE_TXD:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
break;
case TGSI_OPCODE_UP2H:
/* deprecated */
assert (0);
return FALSE;
break;
case TGSI_OPCODE_UP2US:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_UP4B:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_UP4UB:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_X2D:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ARA:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ARR:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_round(&bld->base, src0);
break;
case TGSI_OPCODE_BRA:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CAL:
return FALSE;
case TGSI_OPCODE_RET:
return FALSE;
case TGSI_OPCODE_END:
*pc = -1;
break;
case TGSI_OPCODE_SSG:
/* TGSI_OPCODE_SGN */
tmp0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_sgn(&bld->base, tmp0);
break;
case TGSI_OPCODE_CMP:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, bld->base.zero);
dst0 = lp_build_select(&bld->base, tmp0, src1, src2);
break;
case TGSI_OPCODE_SCS:
return FALSE;
case TGSI_OPCODE_TXB:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
break;
case TGSI_OPCODE_NRM:
/* fall-through */
case TGSI_OPCODE_NRM4:
return FALSE;
case TGSI_OPCODE_DIV:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_DP2:
return FALSE;
case TGSI_OPCODE_TXL:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
break;
case TGSI_OPCODE_TXP:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
break;
case TGSI_OPCODE_BRK:
return FALSE;
case TGSI_OPCODE_IF:
return FALSE;
case TGSI_OPCODE_BGNLOOP:
return FALSE;
case TGSI_OPCODE_BGNSUB:
return FALSE;
case TGSI_OPCODE_ELSE:
return FALSE;
case TGSI_OPCODE_ENDIF:
return FALSE;
case TGSI_OPCODE_ENDLOOP:
return FALSE;
case TGSI_OPCODE_ENDSUB:
return FALSE;
case TGSI_OPCODE_PUSHA:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_POPA:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CEIL:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_ceil(&bld->base, src0);
break;
case TGSI_OPCODE_I2F:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_NOT:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TRUNC:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_trunc(&bld->base, src0);
break;
case TGSI_OPCODE_SHL:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ISHR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_AND:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_OR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_MOD:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_XOR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_SAD:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TXF:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TXQ:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CONT:
return FALSE;
case TGSI_OPCODE_EMIT:
return FALSE;
break;
case TGSI_OPCODE_ENDPRIM:
return FALSE;
break;
case TGSI_OPCODE_NOP:
break;
default:
return FALSE;
}
if (info->num_dst) {
emit_store(bld, inst, 0, dst0);
}
return TRUE;
}
/* Lots of hardcoding
*
* EAX -- pointer to current output vertex
* ECX -- pointer to current attribute
*
*/
static GLboolean build_vertex_emit( struct x86_program *p )
{
struct gl_context *ctx = p->ctx;
TNLcontext *tnl = TNL_CONTEXT(ctx);
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
GLuint j = 0;
struct x86_reg vertexEAX = x86_make_reg(file_REG32, reg_AX);
struct x86_reg srcECX = x86_make_reg(file_REG32, reg_CX);
struct x86_reg countEBP = x86_make_reg(file_REG32, reg_BP);
struct x86_reg vtxESI = x86_make_reg(file_REG32, reg_SI);
struct x86_reg temp = x86_make_reg(file_XMM, 0);
struct x86_reg vp0 = x86_make_reg(file_XMM, 1);
struct x86_reg vp1 = x86_make_reg(file_XMM, 2);
struct x86_reg temp2 = x86_make_reg(file_XMM, 3);
GLubyte *fixup, *label;
/* Push a few regs?
*/
x86_push(&p->func, countEBP);
x86_push(&p->func, vtxESI);
/* Get vertex count, compare to zero
*/
x86_xor(&p->func, srcECX, srcECX);
x86_mov(&p->func, countEBP, x86_fn_arg(&p->func, 2));
x86_cmp(&p->func, countEBP, srcECX);
fixup = x86_jcc_forward(&p->func, cc_E);
/* Initialize destination register.
*/
x86_mov(&p->func, vertexEAX, x86_fn_arg(&p->func, 3));
/* Dereference ctx to get tnl, then vtx:
*/
x86_mov(&p->func, vtxESI, x86_fn_arg(&p->func, 1));
x86_mov(&p->func, vtxESI, x86_make_disp(vtxESI, get_offset(ctx, &ctx->swtnl_context)));
vtxESI = x86_make_disp(vtxESI, get_offset(tnl, &tnl->clipspace));
/* Possibly load vp0, vp1 for viewport calcs:
*/
if (vtx->need_viewport) {
sse_movups(&p->func, vp0, x86_make_disp(vtxESI, get_offset(vtx, &vtx->vp_scale[0])));
sse_movups(&p->func, vp1, x86_make_disp(vtxESI, get_offset(vtx, &vtx->vp_xlate[0])));
}
/* always load, needed or not:
*/
sse_movups(&p->func, p->chan0, x86_make_disp(vtxESI, get_offset(vtx, &vtx->chan_scale[0])));
sse_movups(&p->func, p->identity, x86_make_disp(vtxESI, get_offset(vtx, &vtx->identity[0])));
/* Note address for loop jump */
label = x86_get_label(&p->func);
/* Emit code for each of the attributes. Currently routes
* everything through SSE registers, even when it might be more
* efficient to stick with regular old x86. No optimization or
* other tricks - enough new ground to cover here just getting
* things working.
*/
while (j < vtx->attr_count) {
struct tnl_clipspace_attr *a = &vtx->attr[j];
struct x86_reg dest = x86_make_disp(vertexEAX, a->vertoffset);
/* Now, load an XMM reg from src, perhaps transform, then save.
* Could be shortcircuited in specific cases:
*/
switch (a->format) {
case EMIT_1F:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 1, x86_deref(srcECX), a->inputsize);
emit_store(p, dest, 1, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_2F:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 2, x86_deref(srcECX), a->inputsize);
emit_store(p, dest, 2, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_3F:
/* Potentially the worst case - hardcode 2+1 copying:
*/
if (0) {
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 3, x86_deref(srcECX), a->inputsize);
emit_store(p, dest, 3, temp);
update_src_ptr(p, srcECX, vtxESI, a);
}
else {
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 2, x86_deref(srcECX), a->inputsize);
emit_store(p, dest, 2, temp);
if (a->inputsize > 2) {
emit_load(p, temp, 1, x86_make_disp(srcECX, 8), 1);
emit_store(p, x86_make_disp(dest,8), 1, temp);
}
else {
sse_movss(&p->func, x86_make_disp(dest,8), get_identity(p));
}
update_src_ptr(p, srcECX, vtxESI, a);
}
break;
case EMIT_4F:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
emit_store(p, dest, 4, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_2F_VIEWPORT:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 2, x86_deref(srcECX), a->inputsize);
sse_mulps(&p->func, temp, vp0);
sse_addps(&p->func, temp, vp1);
emit_store(p, dest, 2, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_3F_VIEWPORT:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 3, x86_deref(srcECX), a->inputsize);
sse_mulps(&p->func, temp, vp0);
sse_addps(&p->func, temp, vp1);
emit_store(p, dest, 3, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_4F_VIEWPORT:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
sse_mulps(&p->func, temp, vp0);
sse_addps(&p->func, temp, vp1);
emit_store(p, dest, 4, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_3F_XYW:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
sse_shufps(&p->func, temp, temp, SHUF(X,Y,W,Z));
emit_store(p, dest, 3, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_1UB_1F:
/* Test for PAD3 + 1UB:
*/
if (j > 0 &&
a[-1].vertoffset + a[-1].vertattrsize <= a->vertoffset - 3)
{
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 1, x86_deref(srcECX), a->inputsize);
sse_shufps(&p->func, temp, temp, SHUF(X,X,X,X));
emit_pack_store_4ub(p, x86_make_disp(dest, -3), temp); /* overkill! */
update_src_ptr(p, srcECX, vtxESI, a);
}
else {
printf("Can't emit 1ub %x %x %d\n", a->vertoffset, a[-1].vertoffset, a[-1].vertattrsize );
return GL_FALSE;
}
break;
case EMIT_3UB_3F_RGB:
case EMIT_3UB_3F_BGR:
/* Test for 3UB + PAD1:
*/
if (j == vtx->attr_count - 1 ||
a[1].vertoffset >= a->vertoffset + 4) {
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 3, x86_deref(srcECX), a->inputsize);
if (a->format == EMIT_3UB_3F_BGR)
sse_shufps(&p->func, temp, temp, SHUF(Z,Y,X,W));
emit_pack_store_4ub(p, dest, temp);
update_src_ptr(p, srcECX, vtxESI, a);
}
/* Test for 3UB + 1UB:
*/
else if (j < vtx->attr_count - 1 &&
a[1].format == EMIT_1UB_1F &&
a[1].vertoffset == a->vertoffset + 3) {
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 3, x86_deref(srcECX), a->inputsize);
update_src_ptr(p, srcECX, vtxESI, a);
/* Make room for incoming value:
*/
sse_shufps(&p->func, temp, temp, SHUF(W,X,Y,Z));
get_src_ptr(p, srcECX, vtxESI, &a[1]);
emit_load(p, temp2, 1, x86_deref(srcECX), a[1].inputsize);
sse_movss(&p->func, temp, temp2);
update_src_ptr(p, srcECX, vtxESI, &a[1]);
/* Rearrange and possibly do BGR conversion:
*/
if (a->format == EMIT_3UB_3F_BGR)
sse_shufps(&p->func, temp, temp, SHUF(W,Z,Y,X));
else
sse_shufps(&p->func, temp, temp, SHUF(Y,Z,W,X));
emit_pack_store_4ub(p, dest, temp);
j++; /* NOTE: two attrs consumed */
}
else {
printf("Can't emit 3ub\n");
return GL_FALSE; /* add this later */
}
break;
case EMIT_4UB_4F_RGBA:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
emit_pack_store_4ub(p, dest, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_4UB_4F_BGRA:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
sse_shufps(&p->func, temp, temp, SHUF(Z,Y,X,W));
emit_pack_store_4ub(p, dest, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_4UB_4F_ARGB:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
sse_shufps(&p->func, temp, temp, SHUF(W,X,Y,Z));
emit_pack_store_4ub(p, dest, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_4UB_4F_ABGR:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
sse_shufps(&p->func, temp, temp, SHUF(W,Z,Y,X));
emit_pack_store_4ub(p, dest, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case EMIT_4CHAN_4F_RGBA:
switch (CHAN_TYPE) {
case GL_UNSIGNED_BYTE:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
emit_pack_store_4ub(p, dest, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case GL_FLOAT:
get_src_ptr(p, srcECX, vtxESI, a);
emit_load(p, temp, 4, x86_deref(srcECX), a->inputsize);
emit_store(p, dest, 4, temp);
update_src_ptr(p, srcECX, vtxESI, a);
break;
case GL_UNSIGNED_SHORT:
default:
printf("unknown CHAN_TYPE %s\n", _mesa_lookup_enum_by_nr(CHAN_TYPE));
return GL_FALSE;
}
break;
default:
printf("unknown a[%d].format %d\n", j, a->format);
return GL_FALSE; /* catch any new opcodes */
}
/* Increment j by at least 1 - may have been incremented above also:
*/
j++;
}
/* Next vertex:
*/
x86_lea(&p->func, vertexEAX, x86_make_disp(vertexEAX, vtx->vertex_size));
/* decr count, loop if not zero
*/
x86_dec(&p->func, countEBP);
x86_test(&p->func, countEBP, countEBP);
x86_jcc(&p->func, cc_NZ, label);
/* Exit mmx state?
*/
if (p->func.need_emms)
mmx_emms(&p->func);
/* Land forward jump here:
*/
x86_fixup_fwd_jump(&p->func, fixup);
/* Pop regs and return
*/
x86_pop(&p->func, x86_get_base_reg(vtxESI));
x86_pop(&p->func, countEBP);
x86_ret(&p->func);
assert(!vtx->emit);
vtx->emit = (tnl_emit_func)x86_get_func(&p->func);
assert( (char *) p->func.csr - (char *) p->func.store <= MAX_SSE_CODE_SIZE );
return GL_TRUE;
}
