#include "tgsi/tgsi_info.h"
#include "tgsi/tgsi_parse.h"
#include "tgsi/tgsi_util.h"
-#include "tgsi/tgsi_exec.h"
#include "tgsi/tgsi_scan.h"
#include "lp_bld_type.h"
#include "lp_bld_const.h"
#include "lp_bld_arit.h"
+#include "lp_bld_bitarit.h"
+#include "lp_bld_gather.h"
#include "lp_bld_logic.h"
#include "lp_bld_swizzle.h"
#include "lp_bld_flow.h"
+#include "lp_bld_quad.h"
#include "lp_bld_tgsi.h"
#include "lp_bld_limits.h"
#include "lp_bld_debug.h"
+#include "lp_bld_printf.h"
#define FOR_EACH_CHANNEL( CHAN )\
#define CHAN_Y 1
#define CHAN_Z 2
#define CHAN_W 3
-
-#define QUAD_TOP_LEFT 0
-#define QUAD_TOP_RIGHT 1
-#define QUAD_BOTTOM_LEFT 2
-#define QUAD_BOTTOM_RIGHT 3
+#define NUM_CHANNELS 4
#define LP_MAX_INSTRUCTIONS 256
{
struct lp_build_context base;
- /* Builder for integer masks and indices */
- struct lp_build_context int_bld;
+ /* Builder for vector integer masks and indices */
+ struct lp_build_context uint_bld;
+
+ /* Builder for scalar elements of shader's data type (float) */
+ struct lp_build_context elem_bld;
LLVMValueRef consts_ptr;
const LLVMValueRef *pos;
LLVMValueRef addr[LP_MAX_TGSI_ADDRS][NUM_CHANNELS];
LLVMValueRef preds[LP_MAX_TGSI_PREDS][NUM_CHANNELS];
- /* we allocate an array of temps if we have indirect
- * addressing and then the temps above is unused */
+ /* We allocate/use this array of temps if (1 << TGSI_FILE_TEMPORARY) is
+ * set in the indirect_files field.
+ * The temps[] array above is unused then.
+ */
LLVMValueRef temps_array;
- boolean has_indirect_addressing;
+
+ /* We allocate/use this array of output if (1 << TGSI_FILE_OUTPUT) is
+ * set in the indirect_files field.
+ * The outputs[] array above is unused then.
+ */
+ LLVMValueRef outputs_array;
+
+ /* We allocate/use this array of inputs if (1 << TGSI_FILE_INPUT) is
+ * set in the indirect_files field.
+ * The inputs[] array above is unused then.
+ */
+ LLVMValueRef inputs_array;
+
+ const struct tgsi_shader_info *info;
+ /** bitmask indicating which register files are accessed indirectly */
+ unsigned indirect_files;
struct lp_build_mask_context *mask;
struct lp_exec_mask exec_mask;
uint max_instructions;
};
-static const unsigned char
-swizzle_left[4] = {
- QUAD_TOP_LEFT, QUAD_TOP_LEFT,
- QUAD_BOTTOM_LEFT, QUAD_BOTTOM_LEFT
-};
-
-static const unsigned char
-swizzle_right[4] = {
- QUAD_TOP_RIGHT, QUAD_TOP_RIGHT,
- QUAD_BOTTOM_RIGHT, QUAD_BOTTOM_RIGHT
-};
-
-static const unsigned char
-swizzle_top[4] = {
- QUAD_TOP_LEFT, QUAD_TOP_RIGHT,
- QUAD_TOP_LEFT, QUAD_TOP_RIGHT
-};
-
-static const unsigned char
-swizzle_bottom[4] = {
- QUAD_BOTTOM_LEFT, QUAD_BOTTOM_RIGHT,
- QUAD_BOTTOM_LEFT, QUAD_BOTTOM_RIGHT
-};
-
static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
{
mask->bld = bld;
}
mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
assert(LLVMTypeOf(val) == mask->int_vec_type);
- mask->cond_mask = val;
-
+ mask->cond_mask = LLVMBuildAnd(mask->bld->builder,
+ mask->cond_mask,
+ val,
+ "");
lp_exec_mask_update(mask);
}
lp_exec_mask_update(mask);
}
+
+/**
+ * Return pointer to a temporary register channel (src or dest).
+ * Note that indirect addressing cannot be handled here.
+ * \param index which temporary register
+ * \param chan which channel of the temp register.
+ */
static LLVMValueRef
-emit_ddx(struct lp_build_tgsi_soa_context *bld,
- LLVMValueRef src)
+get_temp_ptr(struct lp_build_tgsi_soa_context *bld,
+ unsigned index,
+ unsigned chan)
{
- LLVMValueRef src_left = lp_build_swizzle1_aos(&bld->base, src, swizzle_left);
- LLVMValueRef src_right = lp_build_swizzle1_aos(&bld->base, src, swizzle_right);
- return lp_build_sub(&bld->base, src_right, src_left);
+ assert(chan < 4);
+ if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
+ LLVMValueRef lindex = lp_build_const_int32(index * 4 + chan);
+ return LLVMBuildGEP(bld->base.builder, bld->temps_array, &lindex, 1, "");
+ }
+ else {
+ return bld->temps[index][chan];
+ }
}
-
+/**
+ * Return pointer to a output register channel (src or dest).
+ * Note that indirect addressing cannot be handled here.
+ * \param index which output register
+ * \param chan which channel of the output register.
+ */
static LLVMValueRef
-emit_ddy(struct lp_build_tgsi_soa_context *bld,
- LLVMValueRef src)
+get_output_ptr(struct lp_build_tgsi_soa_context *bld,
+ unsigned index,
+ unsigned chan)
{
- LLVMValueRef src_top = lp_build_swizzle1_aos(&bld->base, src, swizzle_top);
- LLVMValueRef src_bottom = lp_build_swizzle1_aos(&bld->base, src, swizzle_bottom);
- return lp_build_sub(&bld->base, src_top, src_bottom);
+ assert(chan < 4);
+ if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
+ LLVMValueRef lindex = lp_build_const_int32(index * 4 + chan);
+ return LLVMBuildGEP(bld->base.builder, bld->outputs_array, &lindex, 1, "");
+ }
+ else {
+ return bld->outputs[index][chan];
+ }
}
+/**
+ * Gather vector.
+ * XXX the lp_build_gather() function should be capable of doing this
+ * with a little work.
+ */
static LLVMValueRef
-get_temp_ptr(struct lp_build_tgsi_soa_context *bld,
- unsigned index,
- unsigned chan,
- boolean is_indirect,
- LLVMValueRef addr)
+build_gather(struct lp_build_tgsi_soa_context *bld,
+ LLVMValueRef base_ptr,
+ LLVMValueRef indexes)
{
- assert(chan < 4);
- if (!bld->has_indirect_addressing) {
- return bld->temps[index][chan];
- } else {
- LLVMValueRef lindex =
- LLVMConstInt(LLVMInt32Type(), index * 4 + chan, 0);
- if (is_indirect)
- lindex = lp_build_add(&bld->base, lindex, addr);
- return LLVMBuildGEP(bld->base.builder, bld->temps_array, &lindex, 1, "");
+ LLVMValueRef res = bld->base.undef;
+ unsigned i;
+
+ /*
+ * Loop over elements of index_vec, load scalar value, insert it into 'res'.
+ */
+ for (i = 0; i < bld->base.type.length; i++) {
+ LLVMValueRef ii = LLVMConstInt(LLVMInt32Type(), i, 0);
+ LLVMValueRef index = LLVMBuildExtractElement(bld->base.builder,
+ indexes, ii, "");
+ LLVMValueRef scalar_ptr = LLVMBuildGEP(bld->base.builder, base_ptr,
+ &index, 1, "gather_ptr");
+ LLVMValueRef scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
+
+ res = LLVMBuildInsertElement(bld->base.builder, res, scalar, ii, "");
+ }
+
+ return res;
+}
+
+
+/**
+ * Scatter/store vector.
+ */
+static void
+emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
+ LLVMValueRef base_ptr,
+ LLVMValueRef indexes,
+ LLVMValueRef values,
+ struct lp_exec_mask *mask,
+ LLVMValueRef pred)
+{
+ LLVMBuilderRef builder = bld->base.builder;
+ unsigned i;
+
+ /* Mix the predicate and execution mask */
+ if (mask->has_mask) {
+ if (pred) {
+ pred = LLVMBuildAnd(mask->bld->builder, pred, mask->exec_mask, "");
+ }
+ else {
+ pred = mask->exec_mask;
+ }
+ }
+
+ /*
+ * Loop over elements of index_vec, store scalar value.
+ */
+ for (i = 0; i < bld->base.type.length; i++) {
+ LLVMValueRef ii = LLVMConstInt(LLVMInt32Type(), i, 0);
+ LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
+ LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
+ LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
+ LLVMValueRef scalar_pred = pred ?
+ LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
+
+ if (0)
+ lp_build_printf(builder, "scatter %d: val %f at %d %p\n",
+ ii, val, index, scalar_ptr);
+
+ if (scalar_pred) {
+ LLVMValueRef real_val, dst_val;
+ dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
+ real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
+ LLVMBuildStore(builder, real_val, scalar_ptr);
+ }
+ else {
+ LLVMBuildStore(builder, val, scalar_ptr);
+ }
}
}
+
+/**
+ * Read the current value of the ADDR register, convert the floats to
+ * ints, add the base index and return the vector of offsets.
+ * The offsets will be used to index into the constant buffer or
+ * temporary register file.
+ */
+static LLVMValueRef
+get_indirect_index(struct lp_build_tgsi_soa_context *bld,
+ unsigned reg_file, unsigned reg_index,
+ const struct tgsi_src_register *indirect_reg)
+{
+ struct lp_build_context *uint_bld = &bld->uint_bld;
+ /* always use X component of address register */
+ unsigned swizzle = indirect_reg->SwizzleX;
+ LLVMValueRef base;
+ LLVMValueRef rel;
+ LLVMValueRef max_index;
+ LLVMValueRef index;
+
+ assert(bld->indirect_files & (1 << reg_file));
+
+ base = lp_build_const_int_vec(uint_bld->type, reg_index);
+
+ assert(swizzle < 4);
+ rel = LLVMBuildLoad(bld->base.builder,
+ bld->addr[indirect_reg->Index][swizzle],
+ "load addr reg");
+
+ /* for indexing we want integers */
+ rel = LLVMBuildFPToSI(bld->base.builder,
+ rel,
+ uint_bld->vec_type, "");
+
+ index = lp_build_add(uint_bld, base, rel);
+
+ max_index = lp_build_const_int_vec(uint_bld->type,
+ bld->info->file_max[reg_file]);
+
+ assert(!uint_bld->type.sign);
+ index = lp_build_min(uint_bld, index, max_index);
+
+ return index;
+}
+
+
/**
* Register fetch.
*/
emit_fetch(
struct lp_build_tgsi_soa_context *bld,
const struct tgsi_full_instruction *inst,
- unsigned index,
+ unsigned src_op,
const unsigned chan_index )
{
- const struct tgsi_full_src_register *reg = &inst->Src[index];
+ struct lp_build_context *uint_bld = &bld->uint_bld;
+ const struct tgsi_full_src_register *reg = &inst->Src[src_op];
const unsigned swizzle =
tgsi_util_get_full_src_register_swizzle(reg, chan_index);
LLVMValueRef res;
- LLVMValueRef addr = NULL;
+ LLVMValueRef indirect_index = NULL;
if (swizzle > 3) {
assert(0 && "invalid swizzle in emit_fetch()");
}
if (reg->Register.Indirect) {
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->base.type);
- unsigned swizzle = tgsi_util_get_src_register_swizzle( ®->Indirect, chan_index );
- addr = LLVMBuildLoad(bld->base.builder,
- bld->addr[reg->Indirect.Index][swizzle],
- "");
- /* for indexing we want integers */
- addr = LLVMBuildFPToSI(bld->base.builder, addr,
- int_vec_type, "");
- addr = LLVMBuildExtractElement(bld->base.builder,
- addr, LLVMConstInt(LLVMInt32Type(), 0, 0),
- "");
- addr = lp_build_mul(&bld->base, addr, LLVMConstInt(LLVMInt32Type(), 4, 0));
+ indirect_index = get_indirect_index(bld,
+ reg->Register.File,
+ reg->Register.Index,
+ ®->Indirect);
+ } else {
+ assert(reg->Register.Index <= bld->info->file_max[reg->Register.File]);
}
switch (reg->Register.File) {
case TGSI_FILE_CONSTANT:
- {
- LLVMValueRef index = LLVMConstInt(LLVMInt32Type(),
- reg->Register.Index*4 + swizzle, 0);
+ if (reg->Register.Indirect) {
+ LLVMValueRef swizzle_vec =
+ lp_build_const_int_vec(uint_bld->type, swizzle);
+ LLVMValueRef index_vec; /* index into the const buffer */
+
+ /* index_vec = indirect_index * 4 + swizzle */
+ index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
+ index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
+
+ /* Gather values from the constant buffer */
+ res = build_gather(bld, bld->consts_ptr, index_vec);
+ }
+ else {
+ LLVMValueRef index; /* index into the const buffer */
LLVMValueRef scalar, scalar_ptr;
- if (reg->Register.Indirect) {
- /*lp_build_printf(bld->base.builder,
- "\taddr = %d\n", addr);*/
- index = lp_build_add(&bld->base, index, addr);
- }
+ index = lp_build_const_int32(reg->Register.Index*4 + swizzle);
+
scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr,
&index, 1, "");
scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
break;
case TGSI_FILE_INPUT:
- res = bld->inputs[reg->Register.Index][swizzle];
+ if (reg->Register.Indirect) {
+ LLVMValueRef swizzle_vec =
+ lp_build_const_int_vec(uint_bld->type, swizzle);
+ LLVMValueRef length_vec =
+ lp_build_const_int_vec(uint_bld->type, bld->base.type.length);
+ LLVMValueRef index_vec; /* index into the const buffer */
+ LLVMValueRef inputs_array;
+ LLVMTypeRef float4_ptr_type;
+
+ /* index_vec = (indirect_index * 4 + swizzle) * length */
+ index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
+ index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
+ index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
+
+ /* cast inputs_array pointer to float* */
+ float4_ptr_type = LLVMPointerType(LLVMFloatType(), 0);
+ inputs_array = LLVMBuildBitCast(uint_bld->builder, bld->inputs_array,
+ float4_ptr_type, "");
+
+ /* Gather values from the temporary register array */
+ res = build_gather(bld, inputs_array, index_vec);
+ } else {
+ if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
+ LLVMValueRef lindex = lp_build_const_int32(reg->Register.Index * 4 + swizzle);
+ LLVMValueRef input_ptr = LLVMBuildGEP(bld->base.builder,
+ bld->inputs_array, &lindex, 1, "");
+ res = LLVMBuildLoad(bld->base.builder, input_ptr, "");
+ }
+ else {
+ res = bld->inputs[reg->Register.Index][swizzle];
+ }
+ }
assert(res);
break;
case TGSI_FILE_TEMPORARY:
- {
- LLVMValueRef temp_ptr = get_temp_ptr(bld, reg->Register.Index,
- swizzle,
- reg->Register.Indirect,
- addr);
+ if (reg->Register.Indirect) {
+ LLVMValueRef swizzle_vec =
+ lp_build_const_int_vec(uint_bld->type, swizzle);
+ LLVMValueRef length_vec =
+ lp_build_const_int_vec(uint_bld->type, bld->base.type.length);
+ LLVMValueRef index_vec; /* index into the const buffer */
+ LLVMValueRef temps_array;
+ LLVMTypeRef float4_ptr_type;
+
+ /* index_vec = (indirect_index * 4 + swizzle) * length */
+ index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
+ index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
+ index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
+
+ /* cast temps_array pointer to float* */
+ float4_ptr_type = LLVMPointerType(LLVMFloatType(), 0);
+ temps_array = LLVMBuildBitCast(uint_bld->builder, bld->temps_array,
+ float4_ptr_type, "");
+
+ /* Gather values from the temporary register array */
+ res = build_gather(bld, temps_array, index_vec);
+ }
+ else {
+ LLVMValueRef temp_ptr;
+ temp_ptr = get_temp_ptr(bld, reg->Register.Index, swizzle);
res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
- if(!res)
+ if (!res)
return bld->base.undef;
}
break;
break;
case TGSI_UTIL_SIGN_SET:
- /* TODO: Use bitwese OR for floating point */
res = lp_build_abs( &bld->base, res );
- res = LLVMBuildNeg( bld->base.builder, res, "" );
- break;
-
+ /* fall through */
case TGSI_UTIL_SIGN_TOGGLE:
- res = LLVMBuildNeg( bld->base.builder, res, "" );
+ res = lp_build_negate( &bld->base, res );
break;
case TGSI_UTIL_SIGN_KEEP:
/* TODO: use interpolation coeffs for inputs */
if(ddx)
- *ddx = emit_ddx(bld, src);
+ *ddx = lp_build_ddx(&bld->base, src);
if(ddy)
- *ddy = emit_ddy(bld, src);
+ *ddy = lp_build_ddy(&bld->base, src);
}
LLVMValueRef value)
{
const struct tgsi_full_dst_register *reg = &inst->Dst[index];
- LLVMValueRef addr = NULL;
+ struct lp_build_context *uint_bld = &bld->uint_bld;
+ LLVMValueRef indirect_index = NULL;
switch( inst->Instruction.Saturate ) {
case TGSI_SAT_NONE:
}
if (reg->Register.Indirect) {
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->base.type);
- unsigned swizzle = tgsi_util_get_src_register_swizzle( ®->Indirect, chan_index );
- addr = LLVMBuildLoad(bld->base.builder,
- bld->addr[reg->Indirect.Index][swizzle],
- "");
- /* for indexing we want integers */
- addr = LLVMBuildFPToSI(bld->base.builder, addr,
- int_vec_type, "");
- addr = LLVMBuildExtractElement(bld->base.builder,
- addr, LLVMConstInt(LLVMInt32Type(), 0, 0),
- "");
- addr = lp_build_mul(&bld->base, addr, LLVMConstInt(LLVMInt32Type(), 4, 0));
+ indirect_index = get_indirect_index(bld,
+ reg->Register.File,
+ reg->Register.Index,
+ ®->Indirect);
+ } else {
+ assert(reg->Register.Index <= bld->info->file_max[reg->Register.File]);
}
switch( reg->Register.File ) {
case TGSI_FILE_OUTPUT:
- lp_exec_mask_store(&bld->exec_mask, pred, value,
- bld->outputs[reg->Register.Index][chan_index]);
+ if (reg->Register.Indirect) {
+ LLVMBuilderRef builder = bld->base.builder;
+ LLVMValueRef chan_vec =
+ lp_build_const_int_vec(uint_bld->type, chan_index);
+ LLVMValueRef length_vec =
+ lp_build_const_int_vec(uint_bld->type, bld->base.type.length);
+ LLVMValueRef index_vec; /* indexes into the temp registers */
+ LLVMValueRef outputs_array;
+ LLVMValueRef pixel_offsets;
+ LLVMTypeRef float_ptr_type;
+ int i;
+
+ /* build pixel offset vector: {0, 1, 2, 3, ...} */
+ pixel_offsets = uint_bld->undef;
+ for (i = 0; i < bld->base.type.length; i++) {
+ LLVMValueRef ii = lp_build_const_int32(i);
+ pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
+ ii, ii, "");
+ }
+
+ /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
+ index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
+ index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
+ index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
+ index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
+
+ float_ptr_type = LLVMPointerType(LLVMFloatType(), 0);
+ outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
+ float_ptr_type, "");
+
+ /* Scatter store values into temp registers */
+ emit_mask_scatter(bld, outputs_array, index_vec, value,
+ &bld->exec_mask, pred);
+ }
+ else {
+ LLVMValueRef out_ptr = get_output_ptr(bld, reg->Register.Index,
+ chan_index);
+ lp_exec_mask_store(&bld->exec_mask, pred, value, out_ptr);
+ }
break;
- case TGSI_FILE_TEMPORARY: {
- LLVMValueRef temp_ptr = get_temp_ptr(bld, reg->Register.Index,
- chan_index,
- reg->Register.Indirect,
- addr);
- lp_exec_mask_store(&bld->exec_mask, pred, value, temp_ptr);
+ case TGSI_FILE_TEMPORARY:
+ if (reg->Register.Indirect) {
+ LLVMBuilderRef builder = bld->base.builder;
+ LLVMValueRef chan_vec =
+ lp_build_const_int_vec(uint_bld->type, chan_index);
+ LLVMValueRef length_vec =
+ lp_build_const_int_vec(uint_bld->type, bld->base.type.length);
+ LLVMValueRef index_vec; /* indexes into the temp registers */
+ LLVMValueRef temps_array;
+ LLVMValueRef pixel_offsets;
+ LLVMTypeRef float_ptr_type;
+ int i;
+
+ /* build pixel offset vector: {0, 1, 2, 3, ...} */
+ pixel_offsets = uint_bld->undef;
+ for (i = 0; i < bld->base.type.length; i++) {
+ LLVMValueRef ii = lp_build_const_int32(i);
+ pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
+ ii, ii, "");
+ }
+
+ /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
+ index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
+ index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
+ index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
+ index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
+
+ float_ptr_type = LLVMPointerType(LLVMFloatType(), 0);
+ temps_array = LLVMBuildBitCast(builder, bld->temps_array,
+ float_ptr_type, "");
+
+ /* Scatter store values into temp registers */
+ emit_mask_scatter(bld, temps_array, index_vec, value,
+ &bld->exec_mask, pred);
+ }
+ else {
+ LLVMValueRef temp_ptr = get_temp_ptr(bld, reg->Register.Index,
+ chan_index);
+ lp_exec_mask_store(&bld->exec_mask, pred, value, temp_ptr);
+ }
break;
- }
case TGSI_FILE_ADDRESS:
lp_exec_mask_store(&bld->exec_mask, pred, value,
case TGSI_FILE_PREDICATE:
lp_exec_mask_store(&bld->exec_mask, pred, value,
- bld->preds[index][chan_index]);
+ bld->preds[reg->Register.Index][chan_index]);
break;
default:
* High-level instruction translators.
*/
-enum tex_modifier {
- TEX_MODIFIER_NONE = 0,
- TEX_MODIFIER_PROJECTED,
- TEX_MODIFIER_LOD_BIAS,
- TEX_MODIFIER_EXPLICIT_LOD,
- TEX_MODIFIER_EXPLICIT_DERIV
-};
-
static void
emit_tex( struct lp_build_tgsi_soa_context *bld,
const struct tgsi_full_instruction *inst,
- enum tex_modifier modifier,
+ enum lp_build_tex_modifier modifier,
LLVMValueRef *texel)
{
unsigned unit;
return;
}
- if (modifier == TEX_MODIFIER_LOD_BIAS) {
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
lod_bias = emit_fetch( bld, inst, 0, 3 );
explicit_lod = NULL;
}
- else if (modifier == TEX_MODIFIER_EXPLICIT_LOD) {
+ else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
lod_bias = NULL;
explicit_lod = emit_fetch( bld, inst, 0, 3 );
}
explicit_lod = NULL;
}
- if (modifier == TEX_MODIFIER_PROJECTED) {
+ if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
oow = emit_fetch( bld, inst, 0, 3 );
oow = lp_build_rcp(&bld->base, oow);
}
for (i = 0; i < num_coords; i++) {
coords[i] = emit_fetch( bld, inst, 0, i );
- if (modifier == TEX_MODIFIER_PROJECTED)
+ if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
coords[i] = lp_build_mul(&bld->base, coords[i], oow);
}
for (i = num_coords; i < 3; i++) {
coords[i] = bld->base.undef;
}
- if (modifier == TEX_MODIFIER_EXPLICIT_DERIV) {
+ if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
+ LLVMTypeRef i32t = LLVMInt32Type();
+ LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
for (i = 0; i < num_coords; i++) {
- ddx[i] = emit_fetch( bld, inst, 1, i );
- ddy[i] = emit_fetch( bld, inst, 2, i );
+ LLVMValueRef src1 = emit_fetch( bld, inst, 1, i );
+ LLVMValueRef src2 = emit_fetch( bld, inst, 2, i );
+ ddx[i] = LLVMBuildExtractElement(bld->base.builder, src1, index0, "");
+ ddy[i] = LLVMBuildExtractElement(bld->base.builder, src2, index0, "");
}
unit = inst->Src[3].Register.Index;
} else {
for (i = 0; i < num_coords; i++) {
- ddx[i] = emit_ddx( bld, coords[i] );
- ddy[i] = emit_ddy( bld, coords[i] );
+ ddx[i] = lp_build_scalar_ddx( &bld->base, coords[i] );
+ ddy[i] = lp_build_scalar_ddy( &bld->base, coords[i] );
}
unit = inst->Src[1].Register.Index;
}
for (i = num_coords; i < 3; i++) {
- ddx[i] = bld->base.undef;
- ddy[i] = bld->base.undef;
+ ddx[i] = LLVMGetUndef(bld->base.elem_type);
+ ddy[i] = LLVMGetUndef(bld->base.elem_type);
}
bld->sampler->emit_fetch_texel(bld->sampler,
texel);
}
+static boolean
+near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
+ int pc)
+{
+ int i;
+
+ for (i = 0; i < 5; i++) {
+ unsigned opcode;
+
+ if (pc + i >= bld->info->num_instructions)
+ return TRUE;
+
+ opcode = bld->instructions[pc + i].Instruction.Opcode;
+
+ if (opcode == TGSI_OPCODE_END)
+ return TRUE;
+
+ if (opcode == TGSI_OPCODE_TEX ||
+ opcode == TGSI_OPCODE_TXP ||
+ opcode == TGSI_OPCODE_TXD ||
+ opcode == TGSI_OPCODE_TXB ||
+ opcode == TGSI_OPCODE_TXL ||
+ opcode == TGSI_OPCODE_TXF ||
+ opcode == TGSI_OPCODE_TXQ ||
+ opcode == TGSI_OPCODE_CAL ||
+ opcode == TGSI_OPCODE_CALLNZ ||
+ opcode == TGSI_OPCODE_IF ||
+ opcode == TGSI_OPCODE_IFC ||
+ opcode == TGSI_OPCODE_BGNLOOP ||
+ opcode == TGSI_OPCODE_SWITCH)
+ return FALSE;
+ }
+
+ return TRUE;
+}
+
+
/**
* Kill fragment if any of the src register values are negative.
static void
emit_kil(
struct lp_build_tgsi_soa_context *bld,
- const struct tgsi_full_instruction *inst )
+ const struct tgsi_full_instruction *inst,
+ int pc)
{
const struct tgsi_full_src_register *reg = &inst->Src[0];
LLVMValueRef terms[NUM_CHANNELS];
}
}
- if(mask)
+ if(mask) {
lp_build_mask_update(bld->mask, mask);
+
+ if (!near_end_of_shader(bld, pc))
+ lp_build_mask_check(bld->mask);
+ }
}
*/
static void
emit_kilp(struct lp_build_tgsi_soa_context *bld,
- const struct tgsi_full_instruction *inst)
+ const struct tgsi_full_instruction *inst,
+ int pc)
{
LLVMValueRef mask;
mask = LLVMBuildNot(bld->base.builder, bld->exec_mask.exec_mask, "kilp");
}
else {
- mask = bld->base.zero;
+ LLVMValueRef zero = LLVMConstNull(bld->base.int_vec_type);
+ mask = zero;
}
lp_build_mask_update(bld->mask, mask);
+
+ if (!near_end_of_shader(bld, pc))
+ lp_build_mask_check(bld->mask);
}
+
+/**
+ * Emit code which will dump the value of all the temporary registers
+ * to stdout.
+ */
+static void
+emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
+{
+ LLVMBuilderRef builder = bld->base.builder;
+ LLVMValueRef temp_ptr;
+ LLVMValueRef i0 = lp_build_const_int32(0);
+ LLVMValueRef i1 = lp_build_const_int32(1);
+ LLVMValueRef i2 = lp_build_const_int32(2);
+ LLVMValueRef i3 = lp_build_const_int32(3);
+ int index;
+ int n = bld->info->file_max[TGSI_FILE_TEMPORARY];
+
+ for (index = 0; index < n; index++) {
+ LLVMValueRef idx = lp_build_const_int32(index);
+ LLVMValueRef v[4][4], res;
+ int chan;
+
+ lp_build_printf(builder, "TEMP[%d]:\n", idx);
+
+ for (chan = 0; chan < 4; chan++) {
+ temp_ptr = get_temp_ptr(bld, index, chan);
+ res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
+ v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
+ v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
+ v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
+ v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
+ }
+
+ lp_build_printf(builder, " X: %f %f %f %f\n",
+ v[0][0], v[0][1], v[0][2], v[0][3]);
+ lp_build_printf(builder, " Y: %f %f %f %f\n",
+ v[1][0], v[1][1], v[1][2], v[1][3]);
+ lp_build_printf(builder, " Z: %f %f %f %f\n",
+ v[2][0], v[2][1], v[2][2], v[2][3]);
+ lp_build_printf(builder, " W: %f %f %f %f\n",
+ v[3][0], v[3][1], v[3][2], v[3][3]);
+ }
+}
+
+
+
static void
emit_declaration(
struct lp_build_tgsi_soa_context *bld,
const struct tgsi_full_declaration *decl)
{
- LLVMTypeRef vec_type = lp_build_vec_type(bld->base.type);
-
- unsigned first = decl->Range.First;
- unsigned last = decl->Range.Last;
+ LLVMTypeRef vec_type = bld->base.vec_type;
+ const unsigned first = decl->Range.First;
+ const unsigned last = decl->Range.Last;
unsigned idx, i;
for (idx = first; idx <= last; ++idx) {
+ assert(last <= bld->info->file_max[decl->Declaration.File]);
switch (decl->Declaration.File) {
case TGSI_FILE_TEMPORARY:
assert(idx < LP_MAX_TGSI_TEMPS);
- if (bld->has_indirect_addressing) {
- LLVMValueRef array_size = LLVMConstInt(LLVMInt32Type(),
- last*4 + 4, 0);
- bld->temps_array = lp_build_array_alloca(bld->base.builder,
- vec_type, array_size, "");
- } else {
+ if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
for (i = 0; i < NUM_CHANNELS; i++)
bld->temps[idx][i] = lp_build_alloca(bld->base.builder,
- vec_type, "");
+ vec_type, "temp");
}
break;
case TGSI_FILE_OUTPUT:
- for (i = 0; i < NUM_CHANNELS; i++)
- bld->outputs[idx][i] = lp_build_alloca(bld->base.builder,
- vec_type, "");
+ if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
+ for (i = 0; i < NUM_CHANNELS; i++)
+ bld->outputs[idx][i] = lp_build_alloca(bld->base.builder,
+ vec_type, "output");
+ }
break;
case TGSI_FILE_ADDRESS:
assert(idx < LP_MAX_TGSI_ADDRS);
for (i = 0; i < NUM_CHANNELS; i++)
bld->addr[idx][i] = lp_build_alloca(bld->base.builder,
- vec_type, "");
+ vec_type, "addr");
break;
case TGSI_FILE_PREDICATE:
assert(idx < LP_MAX_TGSI_PREDS);
for (i = 0; i < NUM_CHANNELS; i++)
bld->preds[idx][i] = lp_build_alloca(bld->base.builder,
- vec_type, "");
+ vec_type, "predicate");
break;
default:
case TGSI_OPCODE_KILP:
/* predicated kill */
- emit_kilp( bld, inst );
+ emit_kilp( bld, inst, (*pc)-1 );
break;
case TGSI_OPCODE_KIL:
/* conditional kill */
- emit_kil( bld, inst );
+ emit_kil( bld, inst, (*pc)-1 );
break;
case TGSI_OPCODE_PK2H:
break;
case TGSI_OPCODE_TEX:
- emit_tex( bld, inst, TEX_MODIFIER_NONE, dst0 );
+ emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_NONE, dst0 );
break;
case TGSI_OPCODE_TXD:
- emit_tex( bld, inst, TEX_MODIFIER_EXPLICIT_DERIV, dst0 );
+ emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV, dst0 );
break;
case TGSI_OPCODE_UP2H:
break;
case TGSI_OPCODE_END:
+ if (0) {
+ /* for debugging */
+ emit_dump_temps(bld);
+ }
*pc = -1;
break;
break;
case TGSI_OPCODE_TXB:
- emit_tex( bld, inst, TEX_MODIFIER_LOD_BIAS, dst0 );
+ emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS, dst0 );
break;
case TGSI_OPCODE_NRM:
break;
case TGSI_OPCODE_TXL:
- emit_tex( bld, inst, TEX_MODIFIER_EXPLICIT_LOD, dst0 );
+ emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD, dst0 );
break;
case TGSI_OPCODE_TXP:
- emit_tex( bld, inst, TEX_MODIFIER_PROJECTED, dst0 );
+ emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED, dst0 );
break;
case TGSI_OPCODE_BRK:
unsigned i;
int pc = 0;
+ struct lp_type res_type;
+
+ assert(type.length <= LP_MAX_VECTOR_LENGTH);
+ memset(&res_type, 0, sizeof res_type);
+ res_type.width = type.width;
+ res_type.length = type.length;
+ res_type.sign = 1;
+
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.base, builder, type);
- lp_build_context_init(&bld.int_bld, builder, lp_int_type(type));
+ lp_build_context_init(&bld.uint_bld, builder, lp_uint_type(type));
+ lp_build_context_init(&bld.elem_bld, builder, lp_elem_type(type));
bld.mask = mask;
bld.pos = pos;
bld.inputs = inputs;
bld.outputs = outputs;
bld.consts_ptr = consts_ptr;
bld.sampler = sampler;
- bld.has_indirect_addressing = info->opcode_count[TGSI_OPCODE_ARR] > 0 ||
- info->opcode_count[TGSI_OPCODE_ARL] > 0;
+ bld.info = info;
+ bld.indirect_files = info->indirect_files;
bld.instructions = (struct tgsi_full_instruction *)
MALLOC( LP_MAX_INSTRUCTIONS * sizeof(struct tgsi_full_instruction) );
bld.max_instructions = LP_MAX_INSTRUCTIONS;
lp_exec_mask_init(&bld.exec_mask, &bld.base);
+ if (bld.indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
+ LLVMValueRef array_size = LLVMConstInt(LLVMInt32Type(),
+ info->file_max[TGSI_FILE_TEMPORARY]*4 + 4, 0);
+ bld.temps_array = lp_build_array_alloca(bld.base.builder,
+ bld.base.vec_type, array_size,
+ "temp_array");
+ }
+
+ if (bld.indirect_files & (1 << TGSI_FILE_OUTPUT)) {
+ LLVMValueRef array_size = LLVMConstInt(LLVMInt32Type(),
+ info->file_max[TGSI_FILE_OUTPUT]*4 + 4, 0);
+ bld.outputs_array = lp_build_array_alloca(bld.base.builder,
+ bld.base.vec_type, array_size,
+ "output_array");
+ }
+
+ /* If we have indirect addressing in inputs we need to copy them into
+ * our alloca array to be able to iterate over them */
+ if (bld.indirect_files & (1 << TGSI_FILE_INPUT)) {
+ unsigned index, chan;
+ LLVMTypeRef vec_type = bld.base.vec_type;
+ LLVMValueRef array_size = LLVMConstInt(LLVMInt32Type(),
+ info->file_max[TGSI_FILE_INPUT]*4 + 4, 0);
+ bld.inputs_array = lp_build_array_alloca(bld.base.builder,
+ vec_type, array_size,
+ "input_array");
+
+ assert(info->num_inputs <= info->file_max[TGSI_FILE_INPUT] + 1);
+
+ for (index = 0; index < info->num_inputs; ++index) {
+ for (chan = 0; chan < NUM_CHANNELS; ++chan) {
+ LLVMValueRef lindex = lp_build_const_int32(index * 4 + chan);
+ LLVMValueRef input_ptr =
+ LLVMBuildGEP(bld.base.builder, bld.inputs_array,
+ &lindex, 1, "");
+ LLVMValueRef value = bld.inputs[index][chan];
+ if (value)
+ LLVMBuildStore(bld.base.builder, value, input_ptr);
+ }
+ }
+ }
+
tgsi_parse_init( &parse, tokens );
while( !tgsi_parse_end_of_tokens( &parse ) ) {
{
/* save expanded instruction */
if (num_instructions == bld.max_instructions) {
- bld.instructions = REALLOC(bld.instructions,
- bld.max_instructions
- * sizeof(struct tgsi_full_instruction),
- (bld.max_instructions + LP_MAX_INSTRUCTIONS)
- * sizeof(struct tgsi_full_instruction));
+ struct tgsi_full_instruction *instructions;
+ instructions = REALLOC(bld.instructions,
+ bld.max_instructions
+ * sizeof(struct tgsi_full_instruction),
+ (bld.max_instructions + LP_MAX_INSTRUCTIONS)
+ * sizeof(struct tgsi_full_instruction));
+ if (!instructions) {
+ break;
+ }
+ bld.instructions = instructions;
bld.max_instructions += LP_MAX_INSTRUCTIONS;
}
opcode_info->mnemonic);
}
+ /* If we have indirect addressing in outputs we need to copy our alloca array
+ * to the outputs slots specified by the called */
+ if (bld.indirect_files & (1 << TGSI_FILE_OUTPUT)) {
+ unsigned index, chan;
+ assert(info->num_outputs <= info->file_max[TGSI_FILE_OUTPUT] + 1);
+ for (index = 0; index < info->num_outputs; ++index) {
+ for (chan = 0; chan < NUM_CHANNELS; ++chan) {
+ bld.outputs[index][chan] = get_output_ptr(&bld, index, chan);
+ }
+ }
+ }
+
if (0) {
LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
LLVMValueRef function = LLVMGetBasicBlockParent(block);
projects / mesa.git / blobdiff