/**************************************************************************
*
* Copyright 2009 VMware, Inc.
- * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * Copyright 2007-2008 VMware, Inc.
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
- * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
* ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
#include "tgsi/tgsi_parse.h"
#include "tgsi/tgsi_util.h"
#include "tgsi/tgsi_scan.h"
+#include "tgsi/tgsi_strings.h"
#include "lp_bld_tgsi_action.h"
#include "lp_bld_type.h"
#include "lp_bld_const.h"
#include "lp_bld_sample.h"
#include "lp_bld_struct.h"
+/* SM 4.0 says that subroutines can nest 32 deep and
+ * we need one more for our main function */
+#define LP_MAX_NUM_FUNCS 33
-static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
+#define DUMP_GS_EMITS 0
+
+/*
+ * If non-zero, the generated LLVM IR will print intermediate results on every TGSI
+ * instruction.
+ *
+ * TODO:
+ * - take execution masks in consideration
+ * - debug control-flow instructions
+ */
+#define DEBUG_EXECUTION 0
+
+
+/*
+ * Emit code to print a register value.
+ */
+static void
+emit_dump_reg(struct gallivm_state *gallivm,
+ unsigned file,
+ unsigned index,
+ unsigned chan,
+ LLVMValueRef value)
{
- LLVMTypeRef int_type = LLVMInt32TypeInContext(bld->gallivm->context);
- LLVMBuilderRef builder = bld->gallivm->builder;
+ char buf[32];
+
+ util_snprintf(buf, sizeof buf, " %s[%u].%c = ",
+ tgsi_file_name(file),
+ index, "xyzw"[chan]);
+
+ lp_build_print_value(gallivm, buf, value);
+}
+
+/*
+ * Return the context for the current function.
+ * (always 'main', if shader doesn't do any function calls)
+ */
+static INLINE struct function_ctx *
+func_ctx(struct lp_exec_mask *mask)
+{
+ assert(mask->function_stack_size > 0);
+ assert(mask->function_stack_size <= LP_MAX_NUM_FUNCS);
+ return &mask->function_stack[mask->function_stack_size - 1];
+}
+
+/*
+ * Returns true if we're in a loop.
+ * It's global, meaning that it returns true even if there's
+ * no loop inside the current function, but we were inside
+ * a loop inside another function, from which this one was called.
+ */
+static INLINE boolean
+mask_has_loop(struct lp_exec_mask *mask)
+{
+ int i;
+ for (i = mask->function_stack_size - 1; i >= 0; --i) {
+ const struct function_ctx *ctx = &mask->function_stack[i];
+ if (ctx->loop_stack_size > 0)
+ return TRUE;
+ }
+ return FALSE;
+}
+
+/*
+ * Returns true if we're inside a switch statement.
+ * It's global, meaning that it returns true even if there's
+ * no switch in the current function, but we were inside
+ * a switch inside another function, from which this one was called.
+ */
+static INLINE boolean
+mask_has_switch(struct lp_exec_mask *mask)
+{
+ int i;
+ for (i = mask->function_stack_size - 1; i >= 0; --i) {
+ const struct function_ctx *ctx = &mask->function_stack[i];
+ if (ctx->switch_stack_size > 0)
+ return TRUE;
+ }
+ return FALSE;
+}
+
+/*
+ * Returns true if we're inside a conditional.
+ * It's global, meaning that it returns true even if there's
+ * no conditional in the current function, but we were inside
+ * a conditional inside another function, from which this one was called.
+ */
+static INLINE boolean
+mask_has_cond(struct lp_exec_mask *mask)
+{
+ int i;
+ for (i = mask->function_stack_size - 1; i >= 0; --i) {
+ const struct function_ctx *ctx = &mask->function_stack[i];
+ if (ctx->cond_stack_size > 0)
+ return TRUE;
+ }
+ return FALSE;
+}
+
+
+/*
+ * Initialize a function context at the specified index.
+ */
+static void
+lp_exec_mask_function_init(struct lp_exec_mask *mask, int function_idx)
+{
+ LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
+ LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = &mask->function_stack[function_idx];
+
+ ctx->cond_stack_size = 0;
+ ctx->loop_stack_size = 0;
+ ctx->switch_stack_size = 0;
+
+ if (function_idx == 0) {
+ ctx->ret_mask = mask->ret_mask;
+ }
+
+ ctx->loop_limiter = lp_build_alloca(mask->bld->gallivm,
+ int_type, "looplimiter");
+ LLVMBuildStore(
+ builder,
+ LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
+ ctx->loop_limiter);
+}
+static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
+{
mask->bld = bld;
mask->has_mask = FALSE;
mask->ret_in_main = FALSE;
- mask->cond_stack_size = 0;
- mask->loop_stack_size = 0;
- mask->call_stack_size = 0;
+ /* For the main function */
+ mask->function_stack_size = 1;
mask->int_vec_type = lp_build_int_vec_type(bld->gallivm, mask->bld->type);
- mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask = mask->cond_mask =
+ mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask =
+ mask->cond_mask = mask->switch_mask =
LLVMConstAllOnes(mask->int_vec_type);
- mask->loop_limiter = lp_build_alloca(bld->gallivm, int_type, "looplimiter");
+ mask->function_stack = CALLOC(LP_MAX_NUM_FUNCS,
+ sizeof(mask->function_stack[0]));
+ lp_exec_mask_function_init(mask, 0);
+}
- LLVMBuildStore(
- builder,
- LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
- mask->loop_limiter);
+static void
+lp_exec_mask_fini(struct lp_exec_mask *mask)
+{
+ FREE(mask->function_stack);
}
static void lp_exec_mask_update(struct lp_exec_mask *mask)
{
LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ boolean has_loop_mask = mask_has_loop(mask);
+ boolean has_cond_mask = mask_has_cond(mask);
+ boolean has_switch_mask = mask_has_switch(mask);
+ boolean has_ret_mask = mask->function_stack_size > 1 ||
+ mask->ret_in_main;
- if (mask->loop_stack_size) {
+ if (has_loop_mask) {
/*for loops we need to update the entire mask at runtime */
LLVMValueRef tmp;
assert(mask->break_mask);
} else
mask->exec_mask = mask->cond_mask;
- if (mask->call_stack_size || mask->ret_in_main) {
+ if (has_switch_mask) {
+ mask->exec_mask = LLVMBuildAnd(builder,
+ mask->exec_mask,
+ mask->switch_mask,
+ "switchmask");
+ }
+
+ if (has_ret_mask) {
mask->exec_mask = LLVMBuildAnd(builder,
mask->exec_mask,
mask->ret_mask,
"callmask");
}
- mask->has_mask = (mask->cond_stack_size > 0 ||
- mask->loop_stack_size > 0 ||
- mask->call_stack_size > 0 ||
- mask->ret_in_main);
+ mask->has_mask = (has_cond_mask ||
+ has_loop_mask ||
+ has_switch_mask ||
+ has_ret_mask);
}
static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
LLVMValueRef val)
{
LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
- assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
- if (mask->cond_stack_size == 0) {
+ if (ctx->cond_stack_size >= LP_MAX_TGSI_NESTING) {
+ ctx->cond_stack_size++;
+ return;
+ }
+ if (ctx->cond_stack_size == 0 && mask->function_stack_size == 1) {
assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
}
- mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
+ ctx->cond_stack[ctx->cond_stack_size++] = mask->cond_mask;
assert(LLVMTypeOf(val) == mask->int_vec_type);
mask->cond_mask = LLVMBuildAnd(builder,
mask->cond_mask,
static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
{
LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
LLVMValueRef prev_mask;
LLVMValueRef inv_mask;
- assert(mask->cond_stack_size);
- prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
- if (mask->cond_stack_size == 1) {
+ assert(ctx->cond_stack_size);
+ if (ctx->cond_stack_size >= LP_MAX_TGSI_NESTING)
+ return;
+ prev_mask = ctx->cond_stack[ctx->cond_stack_size - 1];
+ if (ctx->cond_stack_size == 1 && mask->function_stack_size == 1) {
assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
}
static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
{
- assert(mask->cond_stack_size);
- mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
+ struct function_ctx *ctx = func_ctx(mask);
+ assert(ctx->cond_stack_size);
+ --ctx->cond_stack_size;
+ if (ctx->cond_stack_size >= LP_MAX_TGSI_NESTING)
+ return;
+ mask->cond_mask = ctx->cond_stack[ctx->cond_stack_size];
lp_exec_mask_update(mask);
}
static void lp_exec_bgnloop(struct lp_exec_mask *mask)
{
LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
- if (mask->loop_stack_size == 0) {
- assert(mask->loop_block == NULL);
- assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
- assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
- assert(mask->break_var == NULL);
+ if (ctx->loop_stack_size >= LP_MAX_TGSI_NESTING) {
+ ++ctx->loop_stack_size;
+ return;
}
- assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
+ ctx->break_type_stack[ctx->loop_stack_size + ctx->switch_stack_size] =
+ ctx->break_type;
+ ctx->break_type = LP_EXEC_MASK_BREAK_TYPE_LOOP;
- mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
- mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
- mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
- mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
- ++mask->loop_stack_size;
+ ctx->loop_stack[ctx->loop_stack_size].loop_block = ctx->loop_block;
+ ctx->loop_stack[ctx->loop_stack_size].cont_mask = mask->cont_mask;
+ ctx->loop_stack[ctx->loop_stack_size].break_mask = mask->break_mask;
+ ctx->loop_stack[ctx->loop_stack_size].break_var = ctx->break_var;
+ ++ctx->loop_stack_size;
- mask->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
- LLVMBuildStore(builder, mask->break_mask, mask->break_var);
+ ctx->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
+ LLVMBuildStore(builder, mask->break_mask, ctx->break_var);
- mask->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
+ ctx->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
- LLVMBuildBr(builder, mask->loop_block);
- LLVMPositionBuilderAtEnd(builder, mask->loop_block);
+ LLVMBuildBr(builder, ctx->loop_block);
+ LLVMPositionBuilderAtEnd(builder, ctx->loop_block);
- mask->break_mask = LLVMBuildLoad(builder, mask->break_var, "");
+ mask->break_mask = LLVMBuildLoad(builder, ctx->break_var, "");
lp_exec_mask_update(mask);
}
-static void lp_exec_break(struct lp_exec_mask *mask)
+static void lp_exec_break(struct lp_exec_mask *mask,
+ struct lp_build_tgsi_context * bld_base)
{
LLVMBuilderRef builder = mask->bld->gallivm->builder;
- LLVMValueRef exec_mask = LLVMBuildNot(builder,
+ struct function_ctx *ctx = func_ctx(mask);
+
+ if (ctx->break_type == LP_EXEC_MASK_BREAK_TYPE_LOOP) {
+ LLVMValueRef exec_mask = LLVMBuildNot(builder,
+ mask->exec_mask,
+ "break");
+
+ mask->break_mask = LLVMBuildAnd(builder,
+ mask->break_mask,
+ exec_mask, "break_full");
+ }
+ else {
+ unsigned opcode = bld_base->instructions[bld_base->pc + 1].Instruction.Opcode;
+ boolean break_always = (opcode == TGSI_OPCODE_ENDSWITCH ||
+ opcode == TGSI_OPCODE_CASE);
+
+
+ if (ctx->switch_in_default) {
+ /*
+ * stop default execution but only if this is an unconditional switch.
+ * (The condition here is not perfect since dead code after break is
+ * allowed but should be sufficient since false negatives are just
+ * unoptimized - so we don't have to pre-evaluate that).
+ */
+ if(break_always && ctx->switch_pc) {
+ bld_base->pc = ctx->switch_pc;
+ return;
+ }
+ }
+
+ if (break_always) {
+ mask->switch_mask = LLVMConstNull(mask->bld->int_vec_type);
+ }
+ else {
+ LLVMValueRef exec_mask = LLVMBuildNot(builder,
+ mask->exec_mask,
+ "break");
+ mask->switch_mask = LLVMBuildAnd(builder,
+ mask->switch_mask,
+ exec_mask, "break_switch");
+ }
+ }
+
+ lp_exec_mask_update(mask);
+}
+
+static void lp_exec_break_condition(struct lp_exec_mask *mask,
+ LLVMValueRef cond)
+{
+ LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
+ LLVMValueRef cond_mask = LLVMBuildAnd(builder,
mask->exec_mask,
- "break");
+ cond, "cond_mask");
+ cond_mask = LLVMBuildNot(builder, cond_mask, "break_cond");
- mask->break_mask = LLVMBuildAnd(builder,
- mask->break_mask,
- exec_mask, "break_full");
+ if (ctx->break_type == LP_EXEC_MASK_BREAK_TYPE_LOOP) {
+ mask->break_mask = LLVMBuildAnd(builder,
+ mask->break_mask,
+ cond_mask, "breakc_full");
+ }
+ else {
+ mask->switch_mask = LLVMBuildAnd(builder,
+ mask->switch_mask,
+ cond_mask, "breakc_switch");
+ }
lp_exec_mask_update(mask);
}
struct lp_exec_mask *mask)
{
LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
LLVMBasicBlockRef endloop;
LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
assert(mask->break_mask);
+
+ assert(ctx->loop_stack_size);
+ if (ctx->loop_stack_size > LP_MAX_TGSI_NESTING) {
+ --ctx->loop_stack_size;
+ return;
+ }
+
/*
* Restore the cont_mask, but don't pop
*/
- assert(mask->loop_stack_size);
- mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
+ mask->cont_mask = ctx->loop_stack[ctx->loop_stack_size - 1].cont_mask;
lp_exec_mask_update(mask);
/*
* Unlike the continue mask, the break_mask must be preserved across loop
* iterations
*/
- LLVMBuildStore(builder, mask->break_mask, mask->break_var);
+ LLVMBuildStore(builder, mask->break_mask, ctx->break_var);
/* Decrement the loop limiter */
- limiter = LLVMBuildLoad(builder, mask->loop_limiter, "");
+ limiter = LLVMBuildLoad(builder, ctx->loop_limiter, "");
limiter = LLVMBuildSub(
builder,
LLVMConstInt(int_type, 1, false),
"");
- LLVMBuildStore(builder, limiter, mask->loop_limiter);
+ LLVMBuildStore(builder, limiter, ctx->loop_limiter);
/* i1cond = (mask != 0) */
i1cond = LLVMBuildICmp(
builder,
LLVMIntNE,
LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
- LLVMConstNull(reg_type), "");
+ LLVMConstNull(reg_type), "i1cond");
/* i2cond = (looplimiter > 0) */
i2cond = LLVMBuildICmp(
builder,
LLVMIntSGT,
limiter,
- LLVMConstNull(int_type), "");
+ LLVMConstNull(int_type), "i2cond");
/* if( i1cond && i2cond ) */
icond = LLVMBuildAnd(builder, i1cond, i2cond, "");
endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
LLVMBuildCondBr(builder,
- icond, mask->loop_block, endloop);
+ icond, ctx->loop_block, endloop);
LLVMPositionBuilderAtEnd(builder, endloop);
- assert(mask->loop_stack_size);
- --mask->loop_stack_size;
- mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
- mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
- mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
- mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
+ assert(ctx->loop_stack_size);
+ --ctx->loop_stack_size;
+ mask->cont_mask = ctx->loop_stack[ctx->loop_stack_size].cont_mask;
+ mask->break_mask = ctx->loop_stack[ctx->loop_stack_size].break_mask;
+ ctx->loop_block = ctx->loop_stack[ctx->loop_stack_size].loop_block;
+ ctx->break_var = ctx->loop_stack[ctx->loop_stack_size].break_var;
+ ctx->break_type = ctx->break_type_stack[ctx->loop_stack_size +
+ ctx->switch_stack_size];
+
+ lp_exec_mask_update(mask);
+}
+
+static void lp_exec_switch(struct lp_exec_mask *mask,
+ LLVMValueRef switchval)
+{
+ struct function_ctx *ctx = func_ctx(mask);
+
+ if (ctx->switch_stack_size >= LP_MAX_TGSI_NESTING ||
+ ctx->loop_stack_size > LP_MAX_TGSI_NESTING) {
+ ctx->switch_stack_size++;
+ return;
+ }
+
+ ctx->break_type_stack[ctx->loop_stack_size + ctx->switch_stack_size] =
+ ctx->break_type;
+ ctx->break_type = LP_EXEC_MASK_BREAK_TYPE_SWITCH;
+
+ ctx->switch_stack[ctx->switch_stack_size].switch_mask = mask->switch_mask;
+ ctx->switch_stack[ctx->switch_stack_size].switch_val = ctx->switch_val;
+ ctx->switch_stack[ctx->switch_stack_size].switch_mask_default = ctx->switch_mask_default;
+ ctx->switch_stack[ctx->switch_stack_size].switch_in_default = ctx->switch_in_default;
+ ctx->switch_stack[ctx->switch_stack_size].switch_pc = ctx->switch_pc;
+ ctx->switch_stack_size++;
+
+ mask->switch_mask = LLVMConstNull(mask->int_vec_type);
+ ctx->switch_val = switchval;
+ ctx->switch_mask_default = LLVMConstNull(mask->int_vec_type);
+ ctx->switch_in_default = false;
+ ctx->switch_pc = 0;
+
+ lp_exec_mask_update(mask);
+}
+
+static void lp_exec_endswitch(struct lp_exec_mask *mask,
+ struct lp_build_tgsi_context * bld_base)
+{
+ LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
+
+ if (ctx->switch_stack_size > LP_MAX_TGSI_NESTING) {
+ ctx->switch_stack_size--;
+ return;
+ }
+
+ /* check if there's deferred default if so do it now */
+ if (ctx->switch_pc && !ctx->switch_in_default) {
+ LLVMValueRef prevmask, defaultmask;
+ unsigned tmp_pc;
+ prevmask = ctx->switch_stack[ctx->switch_stack_size - 1].switch_mask;
+ defaultmask = LLVMBuildNot(builder, ctx->switch_mask_default, "sw_default_mask");
+ mask->switch_mask = LLVMBuildAnd(builder, prevmask, defaultmask, "sw_mask");
+ ctx->switch_in_default = true;
+
+ lp_exec_mask_update(mask);
+
+ assert(bld_base->instructions[ctx->switch_pc - 1].Instruction.Opcode ==
+ TGSI_OPCODE_DEFAULT);
+
+ tmp_pc = bld_base->pc;
+ bld_base->pc = ctx->switch_pc;
+ /*
+ * re-purpose switch_pc to point to here again, since we stop execution of
+ * the deferred default after next break.
+ */
+ ctx->switch_pc = tmp_pc - 1;
+
+ return;
+ }
+
+ else if (ctx->switch_pc && ctx->switch_in_default) {
+ assert(bld_base->pc == ctx->switch_pc + 1);
+ }
+
+ ctx->switch_stack_size--;
+ mask->switch_mask = ctx->switch_stack[ctx->switch_stack_size].switch_mask;
+ ctx->switch_val = ctx->switch_stack[ctx->switch_stack_size].switch_val;
+ ctx->switch_mask_default = ctx->switch_stack[ctx->switch_stack_size].switch_mask_default;
+ ctx->switch_in_default = ctx->switch_stack[ctx->switch_stack_size].switch_in_default;
+ ctx->switch_pc = ctx->switch_stack[ctx->switch_stack_size].switch_pc;
+
+ ctx->break_type = ctx->break_type_stack[ctx->loop_stack_size + ctx->switch_stack_size];
lp_exec_mask_update(mask);
}
-/* stores val into an address pointed to by dst.
+static void lp_exec_case(struct lp_exec_mask *mask,
+ LLVMValueRef caseval)
+{
+ LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
+
+ LLVMValueRef casemask, prevmask;
+
+ if (ctx->switch_stack_size > LP_MAX_TGSI_NESTING) {
+ return;
+ }
+
+ /* skipping case mask evaluation here is NOT optional (not in all cases anyway). */
+ if (!ctx->switch_in_default) {
+ prevmask = ctx->switch_stack[ctx->switch_stack_size - 1].switch_mask;
+ casemask = lp_build_cmp(mask->bld, PIPE_FUNC_EQUAL, caseval, ctx->switch_val);
+ ctx->switch_mask_default = LLVMBuildOr(builder, casemask,
+ ctx->switch_mask_default, "sw_default_mask");
+ casemask = LLVMBuildOr(builder, casemask, mask->switch_mask, "");
+ mask->switch_mask = LLVMBuildAnd(builder, casemask, prevmask, "sw_mask");
+
+ lp_exec_mask_update(mask);
+ }
+}
+
+/*
+ * Analyse default statement in a switch.
+ * \return true if default is last statement, false otherwise
+ * \param default_pc_start contains pc of instruction to jump to
+ * if default wasn't last but there's no
+ * fallthrough into default.
+ */
+static boolean default_analyse_is_last(struct lp_exec_mask *mask,
+ struct lp_build_tgsi_context * bld_base,
+ int *default_pc_start)
+{
+ unsigned pc = bld_base->pc;
+ struct function_ctx *ctx = func_ctx(mask);
+ unsigned curr_switch_stack = ctx->switch_stack_size;
+
+ if (ctx->switch_stack_size > LP_MAX_TGSI_NESTING) {
+ return false;
+ }
+
+ /* skip over case statements which are together with default */
+ while (bld_base->instructions[pc].Instruction.Opcode == TGSI_OPCODE_CASE) {
+ pc++;
+ }
+
+ while (pc != -1 && pc < bld_base->num_instructions) {
+ unsigned opcode = bld_base->instructions[pc].Instruction.Opcode;
+ switch (opcode) {
+ case TGSI_OPCODE_CASE:
+ if (curr_switch_stack == ctx->switch_stack_size) {
+ *default_pc_start = pc - 1;
+ return false;
+ }
+ break;
+ case TGSI_OPCODE_SWITCH:
+ curr_switch_stack++;
+ break;
+ case TGSI_OPCODE_ENDSWITCH:
+ if (curr_switch_stack == ctx->switch_stack_size) {
+ *default_pc_start = pc - 1;
+ return true;
+ }
+ curr_switch_stack--;
+ break;
+ }
+ pc++;
+ }
+ /* should never arrive here */
+ assert(0);
+ return true;
+}
+
+static void lp_exec_default(struct lp_exec_mask *mask,
+ struct lp_build_tgsi_context * bld_base)
+{
+ LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
+
+ int default_exec_pc;
+ boolean default_is_last;
+
+ if (ctx->switch_stack_size > LP_MAX_TGSI_NESTING) {
+ return;
+ }
+
+ /*
+ * This is a messy opcode, because it may not be always at the end and
+ * there can be fallthrough in and out of it.
+ */
+
+ default_is_last = default_analyse_is_last(mask, bld_base, &default_exec_pc);
+ /*
+ * If it is last statement in switch (note that case statements appearing
+ * "at the same time" as default don't change that) everything is just fine,
+ * update switch mask and go on. This means we can handle default with
+ * fallthrough INTO it without overhead, if it is last.
+ */
+ if (default_is_last) {
+ LLVMValueRef prevmask, defaultmask;
+ prevmask = ctx->switch_stack[ctx->switch_stack_size - 1].switch_mask;
+ defaultmask = LLVMBuildNot(builder, ctx->switch_mask_default, "sw_default_mask");
+ defaultmask = LLVMBuildOr(builder, defaultmask, mask->switch_mask, "");
+ mask->switch_mask = LLVMBuildAnd(builder, prevmask, defaultmask, "sw_mask");
+ ctx->switch_in_default = true;
+
+ lp_exec_mask_update(mask);
+ }
+ else {
+ /*
+ * Technically, "case" immediately before default isn't really a
+ * fallthrough, however we still have to count them as such as we
+ * already have updated the masks.
+ * If that happens in practice could add a switch optimizer pass
+ * which just gets rid of all case statements appearing together with
+ * default (or could do switch analysis at switch start time instead).
+ */
+ unsigned opcode = bld_base->instructions[bld_base->pc - 1].Instruction.Opcode;
+ boolean ft_into = (opcode != TGSI_OPCODE_BRK &&
+ opcode != TGSI_OPCODE_SWITCH);
+ /*
+ * If it is not last statement and there was no fallthrough into it,
+ * we record the PC and continue execution at next case (again, those
+ * case encountered at the same time don't count). At endswitch
+ * time, we update switchmask, and go back executing the code we skipped
+ * until the next break (possibly re-executing some code with changed mask
+ * if there was a fallthrough out of default).
+ * Finally, if it is not last statement and there was a fallthrough into it,
+ * do the same as with the former case, except instead of skipping the code
+ * just execute it without updating the mask, then go back and re-execute.
+ */
+ ctx->switch_pc = bld_base->pc;
+ if (!ft_into) {
+ bld_base->pc = default_exec_pc;
+ }
+ }
+}
+
+
+/* stores val into an address pointed to by dst_ptr.
* mask->exec_mask is used to figure out which bits of val
* should be stored into the address
* (0 means don't store this bit, 1 means do store).
struct lp_build_context *bld_store,
LLVMValueRef pred,
LLVMValueRef val,
- LLVMValueRef dst)
+ LLVMValueRef dst_ptr)
{
LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ assert(lp_check_value(bld_store->type, val));
+ assert(LLVMGetTypeKind(LLVMTypeOf(dst_ptr)) == LLVMPointerTypeKind);
+ assert(LLVMGetElementType(LLVMTypeOf(dst_ptr)) == LLVMTypeOf(val));
+
/* Mix the predicate and execution mask */
if (mask->has_mask) {
if (pred) {
}
if (pred) {
- LLVMValueRef real_val, dst_val;
-
- dst_val = LLVMBuildLoad(builder, dst, "");
- real_val = lp_build_select(bld_store,
- pred,
- val, dst_val);
+ LLVMValueRef res, dst;
- LLVMBuildStore(builder, real_val, dst);
+ dst = LLVMBuildLoad(builder, dst_ptr, "");
+ res = lp_build_select(bld_store, pred, val, dst);
+ LLVMBuildStore(builder, res, dst_ptr);
} else
- LLVMBuildStore(builder, val, dst);
+ LLVMBuildStore(builder, val, dst_ptr);
}
static void lp_exec_mask_call(struct lp_exec_mask *mask,
int func,
int *pc)
{
- assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
- mask->call_stack[mask->call_stack_size].pc = *pc;
- mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
- mask->call_stack_size++;
+ if (mask->function_stack_size >= LP_MAX_NUM_FUNCS) {
+ return;
+ }
+
+ lp_exec_mask_function_init(mask, mask->function_stack_size);
+ mask->function_stack[mask->function_stack_size].pc = *pc;
+ mask->function_stack[mask->function_stack_size].ret_mask = mask->ret_mask;
+ mask->function_stack_size++;
*pc = func;
}
static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
{
LLVMBuilderRef builder = mask->bld->gallivm->builder;
+ struct function_ctx *ctx = func_ctx(mask);
LLVMValueRef exec_mask;
- if (mask->cond_stack_size == 0 &&
- mask->loop_stack_size == 0 &&
- mask->call_stack_size == 0) {
+ if (ctx->cond_stack_size == 0 &&
+ ctx->loop_stack_size == 0 &&
+ ctx->switch_stack_size == 0 &&
+ mask->function_stack_size == 1) {
/* returning from main() */
*pc = -1;
return;
}
- if (mask->call_stack_size == 0) {
+ if (mask->function_stack_size == 1) {
/*
* This requires special handling since we need to ensure
* we don't drop the mask even if we have no call stack
static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
{
- assert(mask->call_stack_size);
- mask->call_stack_size--;
- *pc = mask->call_stack[mask->call_stack_size].pc;
- mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
+ struct function_ctx *ctx;
+
+ assert(mask->function_stack_size > 1);
+ assert(mask->function_stack_size <= LP_MAX_NUM_FUNCS);
+
+ ctx = func_ctx(mask);
+ mask->function_stack_size--;
+
+ *pc = ctx->pc;
+ mask->ret_mask = ctx->ret_mask;
+
lp_exec_mask_update(mask);
}
+static LLVMValueRef
+get_file_ptr(struct lp_build_tgsi_soa_context *bld,
+ unsigned file,
+ unsigned index,
+ unsigned chan)
+{
+ LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
+ LLVMValueRef (*array_of_vars)[TGSI_NUM_CHANNELS];
+ LLVMValueRef var_of_array;
+
+ switch (file) {
+ case TGSI_FILE_TEMPORARY:
+ array_of_vars = bld->temps;
+ var_of_array = bld->temps_array;
+ break;
+ case TGSI_FILE_OUTPUT:
+ array_of_vars = bld->outputs;
+ var_of_array = bld->outputs_array;
+ break;
+ default:
+ assert(0);
+ return NULL;
+ }
+
+ assert(chan < 4);
+
+ if (bld->indirect_files & (1 << file)) {
+ LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
+ return LLVMBuildGEP(builder, var_of_array, &lindex, 1, "");
+ }
+ else {
+ assert(index <= bld->bld_base.info->file_max[file]);
+ return array_of_vars[index][chan];
+ }
+}
+
+
/**
* Return pointer to a temporary register channel (src or dest).
* Note that indirect addressing cannot be handled here.
unsigned index,
unsigned chan)
{
- LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
- assert(chan < 4);
- if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
- LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
- return LLVMBuildGEP(builder, bld->temps_array, &lindex, 1, "");
- }
- else {
- return bld->temps[index][chan];
- }
+ return get_file_ptr(bld, TGSI_FILE_TEMPORARY, index, chan);
}
/**
unsigned index,
unsigned chan)
{
- LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
- assert(chan < 4);
- if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
- LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm,
- index * 4 + chan);
- return LLVMBuildGEP(builder, bld->outputs_array, &lindex, 1, "");
- }
- else {
- return bld->outputs[index][chan];
+ return get_file_ptr(bld, TGSI_FILE_OUTPUT, index, chan);
+}
+
+/*
+ * If we have indirect addressing in outputs copy our alloca array
+ * to the outputs slots specified by the caller to make sure
+ * our outputs are delivered consistently via the same interface.
+ */
+static void
+gather_outputs(struct lp_build_tgsi_soa_context * bld)
+{
+ if ((bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
+ unsigned index, chan;
+ assert(bld->bld_base.info->num_outputs <=
+ bld->bld_base.info->file_max[TGSI_FILE_OUTPUT] + 1);
+ for (index = 0; index < bld->bld_base.info->num_outputs; ++index) {
+ for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
+ bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
+ }
+ }
}
}
static LLVMValueRef
build_gather(struct lp_build_context *bld,
LLVMValueRef base_ptr,
- LLVMValueRef indexes)
+ LLVMValueRef indexes,
+ LLVMValueRef *overflow_mask)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef res = bld->undef;
unsigned i;
+ LLVMValueRef temp_ptr;
+
+ if (overflow_mask) {
+ temp_ptr = lp_build_alloca(
+ bld->gallivm,
+ lp_build_vec_type(bld->gallivm, bld->type), "");
+ }
/*
* Loop over elements of index_vec, load scalar value, insert it into 'res'.
LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
LLVMValueRef index = LLVMBuildExtractElement(builder,
indexes, ii, "");
- LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr,
- &index, 1, "gather_ptr");
- LLVMValueRef scalar = LLVMBuildLoad(builder, scalar_ptr, "");
+ LLVMValueRef scalar_ptr, scalar;
+ LLVMValueRef overflow;
+ struct lp_build_if_state if_ctx;
+
+ /*
+ * overflow_mask is a boolean vector telling us which channels
+ * in the vector overflowed. We use the overflow behavior for
+ * constant buffers which is defined as:
+ * Out of bounds access to constant buffer returns 0 in all
+ * componenets. Out of bounds behavior is always with respect
+ * to the size of the buffer bound at that slot.
+ */
+ if (overflow_mask) {
+ overflow = LLVMBuildExtractElement(builder, *overflow_mask,
+ ii, "");
+ lp_build_if(&if_ctx, bld->gallivm, overflow);
+ {
+ LLVMValueRef val = LLVMBuildLoad(builder, temp_ptr, "");
+ val = LLVMBuildInsertElement(
+ builder, val,
+ LLVMConstNull(LLVMFloatTypeInContext(bld->gallivm->context)),
+ ii, "");
+ LLVMBuildStore(builder, val, temp_ptr);
+ }
+ lp_build_else(&if_ctx);
+ {
+ LLVMValueRef val = LLVMBuildLoad(builder, temp_ptr, "");
+
+ scalar_ptr = LLVMBuildGEP(builder, base_ptr,
+ &index, 1, "gather_ptr");
+ scalar = LLVMBuildLoad(builder, scalar_ptr, "");
- res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
+ val = LLVMBuildInsertElement(builder, val, scalar, ii, "");
+
+ LLVMBuildStore(builder, val, temp_ptr);
+ }
+ lp_build_endif(&if_ctx);
+ } else {
+ scalar_ptr = LLVMBuildGEP(builder, base_ptr,
+ &index, 1, "gather_ptr");
+ scalar = LLVMBuildLoad(builder, scalar_ptr, "");
+
+ res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
+ }
+ }
+
+ if (overflow_mask) {
+ res = LLVMBuildLoad(builder, temp_ptr, "gather_val");
}
return res;
index = lp_build_add(uint_bld, base, rel);
- max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
- uint_bld->type,
- bld->bld_base.info->file_max[reg_file]);
+ /*
+ * emit_fetch_constant handles constant buffer overflow so this code
+ * is pointless for them.
+ * Furthermore the D3D10 spec in section 6.5 says:
+ * If the constant buffer bound to a slot is larger than the size
+ * declared in the shader for that slot, implementations are allowed
+ * to return incorrect data (not necessarily 0) for indices that are
+ * larger than the declared size but smaller than the buffer size.
+ */
+ if (reg_file != TGSI_FILE_CONSTANT) {
+ max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
+ uint_bld->type,
+ bld->bld_base.info->file_max[reg_file]);
- assert(!uint_bld->type.sign);
- index = lp_build_min(uint_bld, index, max_index);
+ assert(!uint_bld->type.sign);
+ index = lp_build_min(uint_bld, index, max_index);
+ }
return index;
}
return bld_fetch;
}
+static LLVMValueRef
+get_soa_array_offsets(struct lp_build_context *uint_bld,
+ LLVMValueRef indirect_index,
+ unsigned chan_index,
+ boolean need_perelement_offset)
+{
+ struct gallivm_state *gallivm = uint_bld->gallivm;
+ LLVMValueRef chan_vec =
+ lp_build_const_int_vec(uint_bld->gallivm, uint_bld->type, chan_index);
+ LLVMValueRef length_vec =
+ lp_build_const_int_vec(gallivm, uint_bld->type, uint_bld->type.length);
+ LLVMValueRef index_vec;
+
+ /* 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);
+
+ if (need_perelement_offset) {
+ LLVMValueRef pixel_offsets;
+ int i;
+ /* build pixel offset vector: {0, 1, 2, 3, ...} */
+ pixel_offsets = uint_bld->undef;
+ for (i = 0; i < uint_bld->type.length; i++) {
+ LLVMValueRef ii = lp_build_const_int32(gallivm, i);
+ pixel_offsets = LLVMBuildInsertElement(gallivm->builder, pixel_offsets,
+ ii, ii, "");
+ }
+ index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
+ }
+ return index_vec;
+}
+
static LLVMValueRef
emit_fetch_constant(
struct lp_build_tgsi_context * bld_base,
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_context *uint_bld = &bld_base->uint_bld;
- LLVMValueRef indirect_index = NULL;
unsigned dimension = 0;
- LLVMValueRef dimension_index;
LLVMValueRef consts_ptr;
+ LLVMValueRef num_consts;
LLVMValueRef res;
/* XXX: Handle fetching xyzw components as a vector */
assert(dimension < LP_MAX_TGSI_CONST_BUFFERS);
}
- dimension_index = lp_build_const_int32(gallivm, dimension);
- consts_ptr = lp_build_array_get(gallivm, bld->consts_ptr, dimension_index);
+ consts_ptr = bld->consts[dimension];
+ num_consts = bld->consts_sizes[dimension];
if (reg->Register.Indirect) {
+ LLVMValueRef indirect_index;
+ LLVMValueRef swizzle_vec =
+ lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
+ LLVMValueRef index_vec; /* index into the const buffer */
+ LLVMValueRef overflow_mask;
+
indirect_index = get_indirect_index(bld,
reg->Register.File,
reg->Register.Index,
®->Indirect);
- }
- if (reg->Register.Indirect) {
- LLVMValueRef swizzle_vec =
- lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
- LLVMValueRef index_vec; /* index into the const buffer */
+ /* All fetches are from the same constant buffer, so
+ * we need to propagate the size to a vector to do a
+ * vector comparison */
+ num_consts = lp_build_broadcast_scalar(uint_bld, num_consts);
+ /* Construct a boolean vector telling us which channels
+ * overflow the bound constant buffer */
+ overflow_mask = LLVMBuildICmp(builder, LLVMIntUGE,
+ indirect_index,
+ num_consts, "");
/* 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_base->base, consts_ptr, index_vec);
+ res = build_gather(&bld_base->base, consts_ptr, index_vec,
+ &overflow_mask);
}
else {
LLVMValueRef index; /* index into the const buffer */
LLVMValueRef scalar, scalar_ptr;
- index = lp_build_const_int32(gallivm, reg->Register.Index*4 + swizzle);
+ index = lp_build_const_int32(gallivm, reg->Register.Index * 4 + swizzle);
scalar_ptr = LLVMBuildGEP(builder, consts_ptr,
&index, 1, "");
struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
}
+
return res;
}
unsigned swizzle)
{
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
- LLVMValueRef res = bld->immediates[reg->Register.Index][swizzle];
- assert(res);
+ struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef res = NULL;
+
+ if (bld->use_immediates_array || reg->Register.Indirect) {
+ LLVMValueRef imms_array;
+ LLVMTypeRef fptr_type;
+
+ /* cast imms_array pointer to float* */
+ fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
+ imms_array = LLVMBuildBitCast(builder, bld->imms_array, fptr_type, "");
+
+ if (reg->Register.Indirect) {
+ LLVMValueRef indirect_index;
+ LLVMValueRef index_vec; /* index into the immediate register array */
+
+ indirect_index = get_indirect_index(bld,
+ reg->Register.File,
+ reg->Register.Index,
+ ®->Indirect);
+ /*
+ * Unlike for other reg classes, adding pixel offsets is unnecessary -
+ * immediates are stored as full vectors (FIXME??? - might be better
+ * to store them the same as constants) but all elements are the same
+ * in any case.
+ */
+ index_vec = get_soa_array_offsets(&bld_base->uint_bld,
+ indirect_index,
+ swizzle,
+ FALSE);
+
+ /* Gather values from the immediate register array */
+ res = build_gather(&bld_base->base, imms_array, index_vec, NULL);
+ } else {
+ LLVMValueRef lindex = lp_build_const_int32(gallivm,
+ reg->Register.Index * 4 + swizzle);
+ LLVMValueRef imms_ptr = LLVMBuildGEP(builder,
+ bld->imms_array, &lindex, 1, "");
+ res = LLVMBuildLoad(builder, imms_ptr, "");
+ }
+ }
+ else {
+ res = bld->immediates[reg->Register.Index][swizzle];
+ }
if (stype == TGSI_TYPE_UNSIGNED) {
- res = LLVMConstBitCast(res, bld_base->uint_bld.vec_type);
+ res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
} else if (stype == TGSI_TYPE_SIGNED) {
- res = LLVMConstBitCast(res, bld_base->int_bld.vec_type);
+ res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
}
return res;
}
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
- struct lp_build_context *uint_bld = &bld_base->uint_bld;
- LLVMValueRef indirect_index = NULL;
LLVMValueRef res;
if (reg->Register.Indirect) {
+ LLVMValueRef indirect_index;
+ LLVMValueRef index_vec; /* index into the input reg array */
+ LLVMValueRef inputs_array;
+ LLVMTypeRef fptr_type;
+
indirect_index = get_indirect_index(bld,
reg->Register.File,
reg->Register.Index,
®->Indirect);
- }
-
- if (reg->Register.Indirect) {
- LLVMValueRef swizzle_vec =
- lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
- LLVMValueRef length_vec =
- lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.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);
+ index_vec = get_soa_array_offsets(&bld_base->uint_bld,
+ indirect_index,
+ swizzle,
+ TRUE);
/* cast inputs_array pointer to float* */
- float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
- inputs_array = LLVMBuildBitCast(builder, bld->inputs_array,
- float4_ptr_type, "");
+ fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
+ inputs_array = LLVMBuildBitCast(builder, bld->inputs_array, fptr_type, "");
- /* Gather values from the temporary register array */
- res = build_gather(&bld_base->base, inputs_array, index_vec);
+ /* Gather values from the input register array */
+ res = build_gather(&bld_base->base, inputs_array, index_vec, NULL);
} else {
if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
LLVMValueRef lindex = lp_build_const_int32(gallivm,
return res;
}
+
+static LLVMValueRef
+emit_fetch_gs_input(
+ struct lp_build_tgsi_context * bld_base,
+ const struct tgsi_full_src_register * reg,
+ enum tgsi_opcode_type stype,
+ unsigned swizzle)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+ struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+ const struct tgsi_shader_info *info = bld->bld_base.info;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef attrib_index = NULL;
+ LLVMValueRef vertex_index = NULL;
+ LLVMValueRef swizzle_index = lp_build_const_int32(gallivm, swizzle);
+ LLVMValueRef res;
+
+ if (info->input_semantic_name[reg->Register.Index] == TGSI_SEMANTIC_PRIMID) {
+ /* This is really a system value not a regular input */
+ assert(!reg->Register.Indirect);
+ assert(!reg->Dimension.Indirect);
+ res = bld->system_values.prim_id;
+ if (stype != TGSI_TYPE_UNSIGNED && stype != TGSI_TYPE_SIGNED) {
+ res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
+ }
+ return res;
+ }
+
+ if (reg->Register.Indirect) {
+ attrib_index = get_indirect_index(bld,
+ reg->Register.File,
+ reg->Register.Index,
+ ®->Indirect);
+ } else {
+ attrib_index = lp_build_const_int32(gallivm, reg->Register.Index);
+ }
+
+ if (reg->Dimension.Indirect) {
+ vertex_index = get_indirect_index(bld,
+ reg->Register.File,
+ reg->Dimension.Index,
+ ®->DimIndirect);
+ } else {
+ vertex_index = lp_build_const_int32(gallivm, reg->Dimension.Index);
+ }
+
+ res = bld->gs_iface->fetch_input(bld->gs_iface, bld_base,
+ reg->Dimension.Indirect,
+ vertex_index,
+ reg->Register.Indirect,
+ attrib_index,
+ swizzle_index);
+
+ assert(res);
+
+ if (stype == TGSI_TYPE_UNSIGNED) {
+ res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
+ } else if (stype == TGSI_TYPE_SIGNED) {
+ res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
+ }
+
+ return res;
+}
+
static LLVMValueRef
emit_fetch_temporary(
struct lp_build_tgsi_context * bld_base,
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
- struct lp_build_context *uint_bld = &bld_base->uint_bld;
- LLVMValueRef indirect_index = NULL;
LLVMValueRef res;
if (reg->Register.Indirect) {
+ LLVMValueRef indirect_index;
+ LLVMValueRef index_vec; /* index into the temp reg array */
+ LLVMValueRef temps_array;
+ LLVMTypeRef fptr_type;
+
indirect_index = get_indirect_index(bld,
reg->Register.File,
reg->Register.Index,
®->Indirect);
- }
-
- if (reg->Register.Indirect) {
- LLVMValueRef swizzle_vec =
- lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
- LLVMValueRef length_vec =
- lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
- bld->bld_base.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);
+ index_vec = get_soa_array_offsets(&bld_base->uint_bld,
+ indirect_index,
+ swizzle,
+ TRUE);
/* cast temps_array pointer to float* */
- float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
- temps_array = LLVMBuildBitCast(builder, bld->temps_array,
- float4_ptr_type, "");
+ fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
+ temps_array = LLVMBuildBitCast(builder, bld->temps_array, fptr_type, "");
/* Gather values from the temporary register array */
- res = build_gather(&bld_base->base, temps_array, index_vec);
+ res = build_gather(&bld_base->base, temps_array, index_vec, NULL);
}
else {
LLVMValueRef temp_ptr;
atype = TGSI_TYPE_UNSIGNED;
break;
+ case TGSI_SEMANTIC_PRIMID:
+ res = bld->system_values.prim_id;
+ atype = TGSI_TYPE_UNSIGNED;
+ break;
+
default:
assert(!"unexpected semantic in emit_fetch_system_value");
res = bld_base->base.zero;
}
}
+
/**
* Register store.
*/
LLVMValueRef value)
{
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
- struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_dst_register *reg = &inst->Dst[index];
- struct lp_build_context *uint_bld = &bld_base->uint_bld;
+ struct lp_build_context *float_bld = &bld_base->base;
+ struct lp_build_context *int_bld = &bld_base->int_bld;
LLVMValueRef indirect_index = NULL;
- struct lp_build_context *bld_store;
enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
- switch (dtype) {
- default:
- case TGSI_TYPE_FLOAT:
- case TGSI_TYPE_UNTYPED:
- bld_store = &bld_base->base;
- break;
- case TGSI_TYPE_UNSIGNED:
- bld_store = &bld_base->uint_bld;
- break;
- case TGSI_TYPE_SIGNED:
- bld_store = &bld_base->int_bld;
- break;
- case TGSI_TYPE_DOUBLE:
- case TGSI_TYPE_VOID:
- assert(0);
- bld_store = NULL;
- break;
- }
-
+ /*
+ * Apply saturation.
+ *
+ * It is always assumed to be float.
+ */
switch( inst->Instruction.Saturate ) {
case TGSI_SAT_NONE:
break;
case TGSI_SAT_ZERO_ONE:
- value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
- value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
+ assert(dtype == TGSI_TYPE_FLOAT ||
+ dtype == TGSI_TYPE_UNTYPED);
+ value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
+ value = lp_build_clamp_zero_one_nanzero(float_bld, value);
break;
case TGSI_SAT_MINUS_PLUS_ONE:
- value = lp_build_max(&bld->bld_base.base, value, lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, -1.0));
- value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
+ assert(dtype == TGSI_TYPE_FLOAT ||
+ dtype == TGSI_TYPE_UNTYPED);
+ value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
+ /* This will give -1.0 for NaN which is probably not what we want. */
+ value = lp_build_max_ext(float_bld, value,
+ lp_build_const_vec(gallivm, float_bld->type, -1.0),
+ GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
+ value = lp_build_min(float_bld, value, float_bld->one);
break;
default:
®->Indirect);
} else {
assert(reg->Register.Index <=
- bld->bld_base.info->file_max[reg->Register.File]);
+ bld_base->info->file_max[reg->Register.File]);
+ }
+
+ if (DEBUG_EXECUTION) {
+ emit_dump_reg(gallivm, reg->Register.File, reg->Register.Index, chan_index, value);
}
switch( reg->Register.File ) {
case TGSI_FILE_OUTPUT:
+ /* Outputs are always stored as floats */
+ value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
+
if (reg->Register.Indirect) {
- LLVMValueRef chan_vec =
- lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
- LLVMValueRef length_vec =
- lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
- LLVMValueRef index_vec; /* indexes into the temp registers */
+ LLVMValueRef index_vec; /* indexes into the output 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->bld_base.base.type.length; i++) {
- LLVMValueRef ii = lp_build_const_int32(gallivm, i);
- pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
- ii, ii, "");
- }
+ LLVMTypeRef fptr_type;
- /* 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);
+ index_vec = get_soa_array_offsets(&bld_base->uint_bld,
+ indirect_index,
+ chan_index,
+ TRUE);
- float_ptr_type =
- LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
- outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
- float_ptr_type, "");
+ fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
+ outputs_array = LLVMBuildBitCast(builder, bld->outputs_array, fptr_type, "");
- /* Scatter store values into temp registers */
+ /* Scatter store values into output registers */
emit_mask_scatter(bld, outputs_array, index_vec, value,
&bld->exec_mask, pred);
}
else {
LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
- chan_index);
- lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, out_ptr);
+ chan_index);
+ lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, out_ptr);
}
break;
case TGSI_FILE_TEMPORARY:
+ /* Temporaries are always stored as floats */
+ value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
+
if (reg->Register.Indirect) {
- LLVMValueRef chan_vec =
- lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
- LLVMValueRef length_vec =
- lp_build_const_int_vec(gallivm, uint_bld->type,
- bld->bld_base.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->bld_base.base.type.length; i++) {
- LLVMValueRef ii = lp_build_const_int32(gallivm, i);
- pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
- ii, ii, "");
- }
+ LLVMTypeRef fptr_type;
- /* 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);
+ index_vec = get_soa_array_offsets(&bld_base->uint_bld,
+ indirect_index,
+ chan_index,
+ TRUE);
- float_ptr_type =
- LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
- temps_array = LLVMBuildBitCast(builder, bld->temps_array,
- float_ptr_type, "");
+ fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
+ temps_array = LLVMBuildBitCast(builder, bld->temps_array, fptr_type, "");
/* Scatter store values into temp registers */
emit_mask_scatter(bld, temps_array, index_vec, value,
}
else {
LLVMValueRef temp_ptr;
-
- switch (dtype) {
- case TGSI_TYPE_UNSIGNED:
- case TGSI_TYPE_SIGNED: {
- LLVMTypeRef itype = bld_base->int_bld.vec_type;
- LLVMTypeRef ivtype = LLVMPointerType(itype, 0);
- LLVMValueRef tint_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
- chan_index);
- LLVMValueRef temp_value_ptr;
-
- temp_ptr = LLVMBuildBitCast(builder, tint_ptr, ivtype, "");
- temp_value_ptr = LLVMBuildBitCast(builder, value, itype, "");
- value = temp_value_ptr;
- break;
- }
- default:
- case TGSI_TYPE_FLOAT:
- case TGSI_TYPE_UNTYPED:
- temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
- chan_index);
- break;
- }
-
- lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value, temp_ptr);
+ temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, chan_index);
+ lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, temp_ptr);
}
break;
case TGSI_FILE_ADDRESS:
assert(dtype == TGSI_TYPE_SIGNED);
- assert(LLVMTypeOf(value) == bld_base->base.int_vec_type);
- lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
+ assert(LLVMTypeOf(value) == int_bld->vec_type);
+ value = LLVMBuildBitCast(builder, value, int_bld->vec_type, "");
+ lp_exec_mask_store(&bld->exec_mask, int_bld, pred, value,
bld->addr[reg->Register.Index][chan_index]);
break;
case TGSI_FILE_PREDICATE:
- lp_exec_mask_store(&bld->exec_mask, bld_store, pred, value,
+ assert(LLVMTypeOf(value) == float_bld->vec_type);
+ value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
+ lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value,
bld->preds[reg->Register.Index][chan_index]);
break;
default:
assert( 0 );
}
+
+ (void)dtype;
+}
+
+/*
+ * Called at the beginning of the translation of each TGSI instruction, to
+ * emit some debug code.
+ */
+static void
+emit_debug(
+ struct lp_build_tgsi_context * bld_base,
+ const struct tgsi_full_instruction * inst,
+ const struct tgsi_opcode_info * info)
+
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ if (DEBUG_EXECUTION) {
+ /*
+ * Dump the TGSI instruction.
+ */
+
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ char buf[512];
+ buf[0] = '$';
+ buf[1] = ' ';
+ tgsi_dump_instruction_str(inst, bld_base->pc, &buf[2], sizeof buf - 2);
+ lp_build_printf(gallivm, buf);
+
+ /* Dump the execution mask.
+ */
+ if (bld->exec_mask.has_mask) {
+ lp_build_print_value(gallivm, " mask = ", bld->exec_mask.exec_mask);
+ }
+ }
}
static void
}
}
+static unsigned
+tgsi_to_pipe_tex_target(unsigned tgsi_target)
+{
+ switch (tgsi_target) {
+ case TGSI_TEXTURE_BUFFER:
+ return PIPE_BUFFER;
+ case TGSI_TEXTURE_1D:
+ case TGSI_TEXTURE_SHADOW1D:
+ return PIPE_TEXTURE_1D;
+ case TGSI_TEXTURE_2D:
+ case TGSI_TEXTURE_SHADOW2D:
+ case TGSI_TEXTURE_2D_MSAA:
+ return PIPE_TEXTURE_2D;
+ case TGSI_TEXTURE_3D:
+ return PIPE_TEXTURE_3D;
+ case TGSI_TEXTURE_CUBE:
+ case TGSI_TEXTURE_SHADOWCUBE:
+ return PIPE_TEXTURE_CUBE;
+ case TGSI_TEXTURE_RECT:
+ case TGSI_TEXTURE_SHADOWRECT:
+ return PIPE_TEXTURE_RECT;
+ case TGSI_TEXTURE_1D_ARRAY:
+ case TGSI_TEXTURE_SHADOW1D_ARRAY:
+ return PIPE_TEXTURE_1D_ARRAY;
+ case TGSI_TEXTURE_2D_ARRAY:
+ case TGSI_TEXTURE_SHADOW2D_ARRAY:
+ case TGSI_TEXTURE_2D_ARRAY_MSAA:
+ return PIPE_TEXTURE_2D_ARRAY;
+ case TGSI_TEXTURE_CUBE_ARRAY:
+ case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
+ return PIPE_TEXTURE_CUBE_ARRAY;
+ default:
+ assert(0);
+ return PIPE_BUFFER;
+ }
+}
+
+
+static enum lp_sampler_lod_property
+lp_build_lod_property(
+ struct lp_build_tgsi_context *bld_base,
+ const struct tgsi_full_instruction *inst,
+ unsigned src_op)
+{
+ const struct tgsi_full_src_register *reg = &inst->Src[src_op];
+ enum lp_sampler_lod_property lod_property;
+
+ /*
+ * Not much we can do here. We could try catching inputs declared
+ * with constant interpolation but not sure it's worth it - since for
+ * TEX opcodes as well as FETCH/LD the lod comes from same reg as
+ * the coords, so it could only work for SAMPLE/TXQ/SVIEWINFO), just
+ * like the constant/immediate recognition below.
+ * What seems to be of more value would be to recognize temps holding
+ * broadcasted scalars but no way we can do it.
+ * Tried asking llvm but without any success (using LLVMIsConstant
+ * even though this isn't exactly what we'd need), even as simple as
+ * IMM[0] UINT32 (0,-1,0,0)
+ * MOV TEMP[0] IMM[0].yyyy
+ * SVIEWINFO TEMP[1], TEMP[0].xxxx, SVIEWINFO[0]
+ * doesn't work.
+ * This means there's ZERO chance this will ever catch a scalar lod
+ * with traditional tex opcodes as well as texel fetches, since the lod
+ * comes from the same reg as coords (except some test shaders using
+ * constant coords maybe).
+ * There's at least hope for sample opcodes as well as size queries.
+ */
+ if (reg->Register.File == TGSI_FILE_CONSTANT ||
+ reg->Register.File == TGSI_FILE_IMMEDIATE) {
+ lod_property = LP_SAMPLER_LOD_SCALAR;
+ }
+ else if (bld_base->info->processor == TGSI_PROCESSOR_FRAGMENT) {
+ if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
+ lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
+ }
+ else {
+ lod_property = LP_SAMPLER_LOD_PER_QUAD;
+ }
+ }
+ else {
+ /* never use scalar (per-quad) lod the results are just too wrong. */
+ lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
+ }
+ return lod_property;
+}
+
+
/**
* High-level instruction translators.
*/
emit_tex( struct lp_build_tgsi_soa_context *bld,
const struct tgsi_full_instruction *inst,
enum lp_build_tex_modifier modifier,
- LLVMValueRef *texel)
+ LLVMValueRef *texel,
+ unsigned sampler_reg)
{
- unsigned unit;
+ unsigned unit = inst->Src[sampler_reg].Register.Index;
LLVMValueRef lod_bias, explicit_lod;
LLVMValueRef oow = NULL;
- LLVMValueRef coords[4];
+ LLVMValueRef coords[5];
LLVMValueRef offsets[3] = { NULL };
struct lp_derivatives derivs;
struct lp_derivatives *deriv_ptr = NULL;
- unsigned num_coords;
- unsigned dims;
- unsigned i;
+ enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
+ unsigned num_derivs, num_offsets, i;
+ unsigned shadow_coord = 0;
+ unsigned layer_coord = 0;
if (!bld->sampler) {
_debug_printf("warning: found texture instruction but no sampler generator supplied\n");
}
switch (inst->Texture.Texture) {
- case TGSI_TEXTURE_1D:
- num_coords = 1;
- dims = 1;
- break;
case TGSI_TEXTURE_1D_ARRAY:
- num_coords = 2;
- dims = 1;
+ layer_coord = 1;
+ /* fallthrough */
+ case TGSI_TEXTURE_1D:
+ num_offsets = 1;
+ num_derivs = 1;
break;
+ case TGSI_TEXTURE_2D_ARRAY:
+ layer_coord = 2;
+ /* fallthrough */
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_RECT:
- num_coords = 2;
- dims = 2;
+ num_offsets = 2;
+ num_derivs = 2;
break;
- case TGSI_TEXTURE_SHADOW1D:
case TGSI_TEXTURE_SHADOW1D_ARRAY:
- num_coords = 3;
- dims = 1;
+ layer_coord = 1;
+ /* fallthrough */
+ case TGSI_TEXTURE_SHADOW1D:
+ shadow_coord = 2;
+ num_offsets = 1;
+ num_derivs = 1;
+ break;
+ case TGSI_TEXTURE_SHADOW2D_ARRAY:
+ layer_coord = 2;
+ shadow_coord = 3;
+ num_offsets = 2;
+ num_derivs = 2;
break;
case TGSI_TEXTURE_SHADOW2D:
case TGSI_TEXTURE_SHADOWRECT:
- case TGSI_TEXTURE_2D_ARRAY:
+ shadow_coord = 2;
+ num_offsets = 2;
+ num_derivs = 2;
+ break;
case TGSI_TEXTURE_CUBE:
- num_coords = 3;
- dims = 2;
+ num_offsets = 2;
+ num_derivs = 3;
break;
case TGSI_TEXTURE_3D:
- num_coords = 3;
- dims = 3;
+ num_offsets = 3;
+ num_derivs = 3;
break;
- case TGSI_TEXTURE_SHADOW2D_ARRAY:
case TGSI_TEXTURE_SHADOWCUBE:
- num_coords = 4;
- dims = 2;
+ shadow_coord = 3;
+ num_offsets = 2;
+ num_derivs = 3;
break;
+ case TGSI_TEXTURE_CUBE_ARRAY:
+ num_offsets = 2;
+ num_derivs = 3;
+ layer_coord = 3;
+ break;
+ case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
+ num_offsets = 2;
+ num_derivs = 3;
+ layer_coord = 3;
+ shadow_coord = 4; /* shadow coord special different reg */
+ break;
+ case TGSI_TEXTURE_2D_MSAA:
+ case TGSI_TEXTURE_2D_ARRAY_MSAA:
default:
assert(0);
return;
}
/* Note lod and especially projected are illegal in a LOT of cases */
- if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
- assert(num_coords < 4);
- lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
- explicit_lod = NULL;
- }
- else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
- assert(num_coords < 4);
- lod_bias = NULL;
- explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
+ modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
+ LLVMValueRef lod;
+ if (inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
+ inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY) {
+ /* note that shadow cube array with bias/explicit lod does not exist */
+ lod = lp_build_emit_fetch(&bld->bld_base, inst, 1, 0);
+ }
+ else {
+ lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
+ }
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
+ lod_bias = lod;
+ explicit_lod = NULL;
+ }
+ else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
+ lod_bias = NULL;
+ explicit_lod = lod;
+ }
+ lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
}
else {
lod_bias = NULL;
}
if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
- assert(num_coords < 4);
- oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
+ oow = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
oow = lp_build_rcp(&bld->bld_base.base, oow);
}
- for (i = 0; i < num_coords; i++) {
- coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
+ for (i = 0; i < num_derivs; i++) {
+ coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
}
- for (i = num_coords; i < 4; i++) {
+ for (i = num_derivs; i < 5; i++) {
coords[i] = bld->bld_base.base.undef;
}
+ /* Layer coord always goes into 3rd slot, except for cube map arrays */
+ if (layer_coord) {
+ if (layer_coord == 3) {
+ coords[3] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
+ }
+ else {
+ coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
+ }
+ if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
+ coords[2] = lp_build_mul(&bld->bld_base.base, coords[2], oow);
+ }
+ /* Shadow coord occupies always 5th slot. */
+ if (shadow_coord) {
+ if (shadow_coord == 4) {
+ coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 1, 0);
+ }
+ else {
+ coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 0, shadow_coord);
+ }
+ if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
+ coords[4] = lp_build_mul(&bld->bld_base.base, coords[4], oow);
+ }
+
if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
unsigned dim;
- for (dim = 0; dim < dims; ++dim) {
- derivs.ddx[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 1, dim );
- derivs.ddy[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 2, dim );
+ for (dim = 0; dim < num_derivs; ++dim) {
+ derivs.ddx[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 1, dim);
+ derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 2, dim);
}
deriv_ptr = &derivs;
- unit = inst->Src[3].Register.Index;
- } else {
- unit = inst->Src[1].Register.Index;
+ /*
+ * could also check all src regs if constant but I doubt such
+ * cases exist in practice.
+ */
+ if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
+ if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
+ lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
+ }
+ else {
+ lod_property = LP_SAMPLER_LOD_PER_QUAD;
+ }
+ }
+ else {
+ lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
+ }
}
/* some advanced gather instructions (txgo) would require 4 offsets */
if (inst->Texture.NumOffsets == 1) {
unsigned dim;
- for (dim = 0; dim < dims; dim++) {
- offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
+ for (dim = 0; dim < num_offsets; dim++) {
+ offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
}
}
coords,
offsets,
deriv_ptr,
- lod_bias, explicit_lod,
+ lod_bias, explicit_lod, lod_property,
texel);
}
struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
unsigned texture_unit, sampler_unit;
LLVMValueRef lod_bias, explicit_lod;
- LLVMValueRef coords[4];
+ LLVMValueRef coords[5];
LLVMValueRef offsets[3] = { NULL };
struct lp_derivatives derivs;
struct lp_derivatives *deriv_ptr = NULL;
- unsigned num_coords, dims;
- unsigned i;
+ enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
+
+ unsigned num_offsets, num_derivs, i;
+ unsigned layer_coord = 0;
if (!bld->sampler) {
_debug_printf("warning: found texture instruction but no sampler generator supplied\n");
*/
switch (bld->sv[texture_unit].Resource) {
case TGSI_TEXTURE_1D:
- num_coords = 1;
- dims = 1;
+ num_offsets = 1;
+ num_derivs = 1;
break;
case TGSI_TEXTURE_1D_ARRAY:
- num_coords = 2;
- dims = 1;
+ layer_coord = 1;
+ num_offsets = 1;
+ num_derivs = 1;
break;
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_RECT:
- num_coords = 2;
- dims = 2;
+ num_offsets = 2;
+ num_derivs = 2;
break;
case TGSI_TEXTURE_2D_ARRAY:
+ layer_coord = 2;
+ num_offsets = 2;
+ num_derivs = 2;
+ break;
case TGSI_TEXTURE_CUBE:
- num_coords = 3;
- dims = 2;
+ num_offsets = 2;
+ num_derivs = 3;
break;
case TGSI_TEXTURE_3D:
- num_coords = 3;
- dims = 3;
+ num_offsets = 3;
+ num_derivs = 3;
break;
case TGSI_TEXTURE_CUBE_ARRAY:
- num_coords = 4;
- dims = 3;
+ layer_coord = 3;
+ num_offsets = 2;
+ num_derivs = 3;
break;
default:
assert(0);
return;
}
- /*
- * unlike old-style tex opcodes the texture/sampler indices
- * always come from src1 and src2 respectively.
- */
- texture_unit = inst->Src[1].Register.Index;
- sampler_unit = inst->Src[2].Register.Index;
-
- if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
- lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
- explicit_lod = NULL;
- }
- else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
- lod_bias = NULL;
- explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS ||
+ modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
+ LLVMValueRef lod = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
+ if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
+ lod_bias = lod;
+ explicit_lod = NULL;
+ }
+ else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
+ lod_bias = NULL;
+ explicit_lod = lod;
+ }
+ lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
}
else if (modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO) {
lod_bias = NULL;
explicit_lod = NULL;
}
- for (i = 0; i < num_coords; i++) {
- coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
+ for (i = 0; i < num_derivs; i++) {
+ coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
}
- for (i = num_coords; i < 4; i++) {
+ for (i = num_derivs; i < 5; i++) {
coords[i] = bld->bld_base.base.undef;
}
- /*
- * XXX: whack shadow comparison value into place.
- * Should probably fix the interface for separate value
- * (it will not work for cube arrays if it is part of coords).
- */
+
+ /* Layer coord always goes into 3rd slot, except for cube map arrays */
+ if (layer_coord) {
+ if (layer_coord == 3)
+ coords[3] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
+ else
+ coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
+ }
+ /* Shadow coord occupies always 5th slot. */
if (compare) {
- unsigned c_coord = num_coords > 2 ? 3 : 2;
- assert(num_coords < 4);
- coords[c_coord] = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
+ coords[4] = lp_build_emit_fetch(&bld->bld_base, inst, 3, 0);
}
if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
unsigned dim;
- for (dim = 0; dim < dims; ++dim) {
- derivs.ddx[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 3, dim );
- derivs.ddy[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 4, dim );
+ for (dim = 0; dim < num_derivs; ++dim) {
+ derivs.ddx[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 3, dim);
+ derivs.ddy[dim] = lp_build_emit_fetch(&bld->bld_base, inst, 4, dim);
}
deriv_ptr = &derivs;
+ /*
+ * could also check all src regs if constant but I doubt such
+ * cases exist in practice.
+ */
+ if (bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT) {
+ if (gallivm_debug & GALLIVM_DEBUG_NO_QUAD_LOD) {
+ lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
+ }
+ else {
+ lod_property = LP_SAMPLER_LOD_PER_QUAD;
+ }
+ }
+ else {
+ lod_property = LP_SAMPLER_LOD_PER_ELEMENT;
+ }
}
/* some advanced gather instructions (txgo) would require 4 offsets */
if (inst->Texture.NumOffsets == 1) {
unsigned dim;
- for (dim = 0; dim < dims; dim++) {
- offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
+ for (dim = 0; dim < num_offsets; dim++) {
+ offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
}
}
coords,
offsets,
deriv_ptr,
- lod_bias, explicit_lod,
+ lod_bias, explicit_lod, lod_property,
texel);
+
+ if (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
+ inst->Src[1].Register.SwizzleY != PIPE_SWIZZLE_GREEN ||
+ inst->Src[1].Register.SwizzleZ != PIPE_SWIZZLE_BLUE ||
+ inst->Src[1].Register.SwizzleW != PIPE_SWIZZLE_ALPHA) {
+ unsigned char swizzles[4];
+ swizzles[0] = inst->Src[1].Register.SwizzleX;
+ swizzles[1] = inst->Src[1].Register.SwizzleY;
+ swizzles[2] = inst->Src[1].Register.SwizzleZ;
+ swizzles[3] = inst->Src[1].Register.SwizzleW;
+
+ lp_build_swizzle_soa_inplace(&bld->bld_base.base, texel, swizzles);
+ }
}
static void
LLVMValueRef explicit_lod = NULL;
LLVMValueRef coords[3];
LLVMValueRef offsets[3] = { NULL };
- unsigned num_coords;
- unsigned dims;
- unsigned i;
+ enum lp_sampler_lod_property lod_property = LP_SAMPLER_LOD_SCALAR;
+ unsigned dims, i;
+ unsigned layer_coord = 0;
if (!bld->sampler) {
_debug_printf("warning: found texture instruction but no sampler generator supplied\n");
switch (target) {
case TGSI_TEXTURE_1D:
case TGSI_TEXTURE_BUFFER:
- num_coords = 1;
dims = 1;
break;
case TGSI_TEXTURE_1D_ARRAY:
- num_coords = 2;
+ layer_coord = 1;
dims = 1;
break;
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_RECT:
- num_coords = 2;
dims = 2;
break;
case TGSI_TEXTURE_2D_ARRAY:
- num_coords = 3;
+ layer_coord = 2;
dims = 2;
break;
case TGSI_TEXTURE_3D:
- num_coords = 3;
dims = 3;
break;
default:
/* always have lod except for buffers ? */
if (target != TGSI_TEXTURE_BUFFER) {
- explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
+ explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 3);
+ lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
}
- for (i = 0; i < num_coords; i++) {
- coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
+ for (i = 0; i < dims; i++) {
+ coords[i] = lp_build_emit_fetch(&bld->bld_base, inst, 0, i);
}
- for (i = num_coords; i < 3; i++) {
+ for (i = dims; i < 3; i++) {
coords[i] = coord_undef;
}
+ if (layer_coord)
+ coords[2] = lp_build_emit_fetch(&bld->bld_base, inst, 0, layer_coord);
if (inst->Texture.NumOffsets == 1) {
unsigned dim;
for (dim = 0; dim < dims; dim++) {
- offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
+ offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim);
}
}
coords,
offsets,
NULL,
- NULL, explicit_lod,
+ NULL, explicit_lod, lod_property,
texel);
+
+ if (is_samplei &&
+ (inst->Src[1].Register.SwizzleX != PIPE_SWIZZLE_RED ||
+ inst->Src[1].Register.SwizzleY != PIPE_SWIZZLE_GREEN ||
+ inst->Src[1].Register.SwizzleZ != PIPE_SWIZZLE_BLUE ||
+ inst->Src[1].Register.SwizzleW != PIPE_SWIZZLE_ALPHA)) {
+ unsigned char swizzles[4];
+ swizzles[0] = inst->Src[1].Register.SwizzleX;
+ swizzles[1] = inst->Src[1].Register.SwizzleY;
+ swizzles[2] = inst->Src[1].Register.SwizzleZ;
+ swizzles[3] = inst->Src[1].Register.SwizzleW;
+
+ lp_build_swizzle_soa_inplace(&bld->bld_base.base, texel, swizzles);
+ }
}
static void
boolean is_sviewinfo)
{
LLVMValueRef explicit_lod;
+ enum lp_sampler_lod_property lod_property;
unsigned has_lod;
unsigned i;
unsigned unit = inst->Src[1].Register.Index;
- unsigned target;
+ unsigned target, pipe_target;
if (is_sviewinfo) {
target = bld->sv[unit].Resource;
return;
}
- if (has_lod)
- explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 0 );
- else
+ if (has_lod) {
+ explicit_lod = lp_build_emit_fetch(&bld->bld_base, inst, 0, 0);
+ lod_property = lp_build_lod_property(&bld->bld_base, inst, 0);
+ }
+ else {
explicit_lod = NULL;
+ lod_property = LP_SAMPLER_LOD_SCALAR;
+ }
+
+
+ pipe_target = tgsi_to_pipe_tex_target(target);
bld->sampler->emit_size_query(bld->sampler,
bld->bld_base.base.gallivm,
bld->bld_base.int_bld.type,
- unit,
- is_sviewinfo,
+ unit, pipe_target,
+ TRUE,
+ lod_property,
explicit_lod,
sizes_out);
}
static boolean
near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
- int pc)
+ int pc)
{
int i;
unsigned opcode;
if (pc + i >= bld->bld_base.info->num_instructions)
- return TRUE;
+ return TRUE;
opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
if (opcode == TGSI_OPCODE_END)
- return TRUE;
+ 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;
+ 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_TEX2 ||
+ opcode == TGSI_OPCODE_TXB2 ||
+ opcode == TGSI_OPCODE_TXL2 ||
+ opcode == TGSI_OPCODE_SAMPLE ||
+ opcode == TGSI_OPCODE_SAMPLE_B ||
+ opcode == TGSI_OPCODE_SAMPLE_C ||
+ opcode == TGSI_OPCODE_SAMPLE_C_LZ ||
+ opcode == TGSI_OPCODE_SAMPLE_D ||
+ opcode == TGSI_OPCODE_SAMPLE_I ||
+ opcode == TGSI_OPCODE_SAMPLE_L ||
+ opcode == TGSI_OPCODE_SVIEWINFO ||
+ opcode == TGSI_OPCODE_CAL ||
+ opcode == TGSI_OPCODE_CALLNZ ||
+ opcode == TGSI_OPCODE_IF ||
+ opcode == TGSI_OPCODE_UIF ||
+ 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(
+emit_kill_if(
struct lp_build_tgsi_soa_context *bld,
const struct tgsi_full_instruction *inst,
int pc)
}
}
- if(mask) {
- lp_build_mask_update(bld->mask, mask);
-
- if (!near_end_of_shader(bld, pc))
- lp_build_mask_check(bld->mask);
+ if (bld->exec_mask.has_mask) {
+ LLVMValueRef invmask;
+ invmask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
+ mask = LLVMBuildOr(builder, mask, invmask, "");
}
+
+ lp_build_mask_update(bld->mask, mask);
+ if (!near_end_of_shader(bld, pc))
+ lp_build_mask_check(bld->mask);
}
/**
- * Predicated fragment kill.
- * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
+ * Unconditional fragment kill.
* The only predication is the execution mask which will apply if
* we're inside a loop or conditional.
*/
static void
-emit_kilp(struct lp_build_tgsi_soa_context *bld,
+emit_kill(struct lp_build_tgsi_soa_context *bld,
int pc)
{
LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
* to stdout.
*/
static void
-emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
+emit_dump_file(struct lp_build_tgsi_soa_context *bld,
+ unsigned file)
{
+ const struct tgsi_shader_info *info = bld->bld_base.info;
struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
- LLVMValueRef temp_ptr;
- LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
- LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
- LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
- LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
+ LLVMValueRef reg_ptr;
int index;
- int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
+ int max_index = info->file_max[file];
- for (index = 0; index < n; index++) {
- LLVMValueRef idx = lp_build_const_int32(gallivm, index);
- LLVMValueRef v[4][4], res;
+ /*
+ * Some register files, particularly constants, can be very large,
+ * and dumping everything could make this unusably slow.
+ */
+ max_index = MIN2(max_index, 32);
+
+ for (index = 0; index <= max_index; index++) {
+ LLVMValueRef res;
+ unsigned mask;
int chan;
- lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
+ if (index < 8 * sizeof(unsigned) &&
+ (info->file_mask[file] & (1 << index)) == 0) {
+ /* This was not declared.*/
+ continue;
+ }
+
+ if (file == TGSI_FILE_INPUT) {
+ mask = info->input_usage_mask[index];
+ } else {
+ mask = TGSI_WRITEMASK_XYZW;
+ }
for (chan = 0; chan < 4; chan++) {
- temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
- res = LLVMBuildLoad(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, "");
- }
+ if ((mask & (1 << chan)) == 0) {
+ /* This channel is not used.*/
+ continue;
+ }
+
+ if (file == TGSI_FILE_CONSTANT) {
+ struct tgsi_full_src_register reg;
+ memset(®, 0, sizeof reg);
+ reg.Register.File = file;
+ reg.Register.Index = index;
+ reg.Register.SwizzleX = 0;
+ reg.Register.SwizzleY = 1;
+ reg.Register.SwizzleZ = 2;
+ reg.Register.SwizzleW = 3;
+
+ res = bld->bld_base.emit_fetch_funcs[file](&bld->bld_base, ®, TGSI_TYPE_FLOAT, chan);
+ if (!res) {
+ continue;
+ }
+ } else if (file == TGSI_FILE_INPUT) {
+ res = bld->inputs[index][chan];
+ if (!res) {
+ continue;
+ }
+ } else if (file == TGSI_FILE_TEMPORARY) {
+ reg_ptr = lp_get_temp_ptr_soa(bld, index, chan);
+ assert(reg_ptr);
+ res = LLVMBuildLoad(builder, reg_ptr, "");
+ } else if (file == TGSI_FILE_OUTPUT) {
+ reg_ptr = lp_get_output_ptr(bld, index, chan);
+ assert(reg_ptr);
+ res = LLVMBuildLoad(builder, reg_ptr, "");
+ } else {
+ assert(0);
+ continue;
+ }
- lp_build_printf(gallivm, " X: %f %f %f %f\n",
- v[0][0], v[0][1], v[0][2], v[0][3]);
- lp_build_printf(gallivm, " Y: %f %f %f %f\n",
- v[1][0], v[1][1], v[1][2], v[1][3]);
- lp_build_printf(gallivm, " Z: %f %f %f %f\n",
- v[2][0], v[2][1], v[2][2], v[2][3]);
- lp_build_printf(gallivm, " W: %f %f %f %f\n",
- v[3][0], v[3][1], v[3][2], v[3][3]);
+ emit_dump_reg(gallivm, file, index, chan, res);
+ }
}
}
const unsigned last = decl->Range.Last;
unsigned idx, i;
- for (idx = first; idx <= last; ++idx) {
- assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
- switch (decl->Declaration.File) {
- case TGSI_FILE_TEMPORARY:
- assert(idx < LP_MAX_TGSI_TEMPS);
- if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
+ assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
+
+ switch (decl->Declaration.File) {
+ case TGSI_FILE_TEMPORARY:
+ if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
+ assert(last < LP_MAX_INLINED_TEMPS);
+ for (idx = first; idx <= last; ++idx) {
for (i = 0; i < TGSI_NUM_CHANNELS; i++)
bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
}
- break;
+ }
+ break;
- case TGSI_FILE_OUTPUT:
- if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
+ case TGSI_FILE_OUTPUT:
+ if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
+ for (idx = first; idx <= last; ++idx) {
for (i = 0; i < TGSI_NUM_CHANNELS; i++)
bld->outputs[idx][i] = lp_build_alloca(gallivm,
vec_type, "output");
}
- break;
+ }
+ break;
- case TGSI_FILE_ADDRESS:
- /* ADDR registers are the only allocated with an integer LLVM IR type,
- * as they are guaranteed to always have integers.
- * XXX: Not sure if this exception is worthwhile (or the whole idea of
- * an ADDR register for that matter).
- */
+ case TGSI_FILE_ADDRESS:
+ /* ADDR registers are only allocated with an integer LLVM IR type,
+ * as they are guaranteed to always have integers.
+ * XXX: Not sure if this exception is worthwhile (or the whole idea of
+ * an ADDR register for that matter).
+ */
+ assert(last < LP_MAX_TGSI_ADDRS);
+ for (idx = first; idx <= last; ++idx) {
assert(idx < LP_MAX_TGSI_ADDRS);
for (i = 0; i < TGSI_NUM_CHANNELS; i++)
bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
- break;
+ }
+ break;
- case TGSI_FILE_PREDICATE:
- assert(idx < LP_MAX_TGSI_PREDS);
+ case TGSI_FILE_PREDICATE:
+ assert(last < LP_MAX_TGSI_PREDS);
+ for (idx = first; idx <= last; ++idx) {
for (i = 0; i < TGSI_NUM_CHANNELS; i++)
bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
"predicate");
- break;
+ }
+ break;
- case TGSI_FILE_SAMPLER_VIEW:
- /*
- * The target stored here MUST match whatever there actually
- * is in the set sampler views (what about return type?).
- */
- assert(idx < PIPE_MAX_SHADER_SAMPLER_VIEWS);
+ case TGSI_FILE_SAMPLER_VIEW:
+ /*
+ * The target stored here MUST match whatever there actually
+ * is in the set sampler views (what about return type?).
+ */
+ assert(last < PIPE_MAX_SHADER_SAMPLER_VIEWS);
+ for (idx = first; idx <= last; ++idx) {
bld->sv[idx] = decl->SamplerView;
- break;
-
- default:
- /* don't need to declare other vars */
- break;
}
+ break;
+
+ case TGSI_FILE_CONSTANT:
+ {
+ /*
+ * We could trivially fetch the per-buffer pointer when fetching the
+ * constant, relying on llvm to figure out it's always the same pointer
+ * anyway. However, doing so results in a huge (more than factor of 10)
+ * slowdown in llvm compilation times for some (but not all) shaders
+ * (more specifically, the IR optimization spends way more time in
+ * DominatorTree::dominates). At least with llvm versions 3.1, 3.3.
+ */
+ unsigned idx2D = decl->Dim.Index2D;
+ LLVMValueRef index2D = lp_build_const_int32(gallivm, idx2D);
+ assert(idx2D < LP_MAX_TGSI_CONST_BUFFERS);
+ bld->consts[idx2D] =
+ lp_build_array_get(gallivm, bld->consts_ptr, index2D);
+ bld->consts_sizes[idx2D] =
+ lp_build_array_get(gallivm, bld->const_sizes_ptr, index2D);
+ }
+ break;
+
+ default:
+ /* don't need to declare other vars */
+ break;
}
}
{
struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
struct gallivm_state * gallivm = bld_base->base.gallivm;
-
- /* simply copy the immediate values into the next immediates[] slot */
+ LLVMValueRef imms[4];
unsigned i;
const uint size = imm->Immediate.NrTokens - 1;
assert(size <= 4);
- assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
switch (imm->Immediate.DataType) {
case TGSI_IMM_FLOAT32:
for( i = 0; i < size; ++i )
- bld->immediates[bld->num_immediates][i] =
- lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
+ imms[i] =
+ lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
break;
case TGSI_IMM_UINT32:
for( i = 0; i < size; ++i ) {
LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
- bld->immediates[bld->num_immediates][i] =
- LLVMConstBitCast(tmp, bld_base->base.vec_type);
+ imms[i] = LLVMConstBitCast(tmp, bld_base->base.vec_type);
}
break;
case TGSI_IMM_INT32:
for( i = 0; i < size; ++i ) {
LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
- bld->immediates[bld->num_immediates][i] =
- LLVMConstBitCast(tmp, bld_base->base.vec_type);
+ imms[i] = LLVMConstBitCast(tmp, bld_base->base.vec_type);
}
-
+
break;
}
for( i = size; i < 4; ++i )
- bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
+ imms[i] = bld_base->base.undef;
+
+ if (bld->use_immediates_array) {
+ unsigned index = bld->num_immediates;
+ struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+
+ assert(bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE));
+ for (i = 0; i < 4; ++i ) {
+ LLVMValueRef lindex = lp_build_const_int32(
+ bld->bld_base.base.gallivm, index * 4 + i);
+ LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
+ bld->imms_array, &lindex, 1, "");
+ LLVMBuildStore(builder, imms[i], imm_ptr);
+ }
+ } else {
+ /* simply copy the immediate values into the next immediates[] slot */
+ unsigned i;
+ const uint size = imm->Immediate.NrTokens - 1;
+ assert(size <= 4);
+ assert(bld->num_immediates < LP_MAX_INLINED_IMMEDIATES);
+
+ for(i = 0; i < 4; ++i )
+ bld->immediates[bld->num_immediates][i] = imms[i];
+
+ if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
+ unsigned index = bld->num_immediates;
+ struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ for (i = 0; i < 4; ++i ) {
+ LLVMValueRef lindex = lp_build_const_int32(
+ bld->bld_base.base.gallivm, index * 4 + i);
+ LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
+ bld->imms_array, &lindex, 1, "");
+ LLVMBuildStore(builder,
+ bld->immediates[index][i],
+ imm_ptr);
+ }
+ }
+ }
bld->num_immediates++;
}
}
static void
-kilp_emit(
+kill_emit(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
- emit_kilp(bld, bld_base->pc - 1);
+ emit_kill(bld, bld_base->pc - 1);
}
static void
-kil_emit(
+kill_if_emit(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
- emit_kil(bld, emit_data->inst, bld_base->pc - 1);
+ emit_kill_if(bld, emit_data->inst, bld_base->pc - 1);
}
static void
{
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
- emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
+ emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
+ emit_data->output, 1);
+}
+
+static void
+tex2_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
+ emit_data->output, 2);
}
static void
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
- emit_data->output);
+ emit_data->output, 1);
+}
+
+static void
+txb2_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
+ emit_data->output, 2);
}
static void
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
- emit_data->output);
+ emit_data->output, 3);
}
static void
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
- emit_data->output);
+ emit_data->output, 1);
+}
+
+static void
+txl2_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
+ emit_data->output, 2);
}
static void
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
- emit_data->output);
+ emit_data->output, 1);
}
static void
emit_size_query(bld, emit_data->inst, emit_data->output, TRUE);
}
+static LLVMValueRef
+mask_vec(struct lp_build_tgsi_context *bld_base)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+ LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
+ struct lp_exec_mask *exec_mask = &bld->exec_mask;
+
+ if (!exec_mask->has_mask) {
+ return lp_build_mask_value(bld->mask);
+ }
+ return LLVMBuildAnd(builder, lp_build_mask_value(bld->mask),
+ exec_mask->exec_mask, "");
+}
+
+static void
+increment_vec_ptr_by_mask(struct lp_build_tgsi_context * bld_base,
+ LLVMValueRef ptr,
+ LLVMValueRef mask)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
+
+ current_vec = LLVMBuildSub(builder, current_vec, mask, "");
+
+ LLVMBuildStore(builder, current_vec, ptr);
+}
+
+static void
+clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context * bld_base,
+ LLVMValueRef ptr,
+ LLVMValueRef mask)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
+
+ current_vec = lp_build_select(&bld_base->uint_bld,
+ mask,
+ bld_base->uint_bld.zero,
+ current_vec);
+
+ LLVMBuildStore(builder, current_vec, ptr);
+}
+
+static LLVMValueRef
+clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context * bld,
+ LLVMValueRef current_mask_vec,
+ LLVMValueRef total_emitted_vertices_vec)
+{
+ LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
+ struct lp_build_context *int_bld = &bld->bld_base.int_bld;
+ LLVMValueRef max_mask = lp_build_cmp(int_bld, PIPE_FUNC_LESS,
+ total_emitted_vertices_vec,
+ bld->max_output_vertices_vec);
+
+ return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
+}
+
+static void
+emit_vertex(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+ LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
+
+ if (bld->gs_iface->emit_vertex) {
+ LLVMValueRef mask = mask_vec(bld_base);
+ LLVMValueRef total_emitted_vertices_vec =
+ LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
+ mask = clamp_mask_to_max_output_vertices(bld, mask,
+ total_emitted_vertices_vec);
+ gather_outputs(bld);
+ bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base,
+ bld->outputs,
+ total_emitted_vertices_vec);
+ increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr,
+ mask);
+ increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr,
+ mask);
+#if DUMP_GS_EMITS
+ lp_build_print_value(bld->bld_base.base.gallivm,
+ " +++ emit vertex masked ones = ",
+ mask);
+ lp_build_print_value(bld->bld_base.base.gallivm,
+ " +++ emit vertex emitted = ",
+ total_emitted_vertices_vec);
+#endif
+ }
+}
+
+
+static void
+end_primitive_masked(struct lp_build_tgsi_context * bld_base,
+ LLVMValueRef mask)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+ LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
+
+ if (bld->gs_iface->end_primitive) {
+ struct lp_build_context *uint_bld = &bld_base->uint_bld;
+ LLVMValueRef emitted_vertices_vec =
+ LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr, "");
+ LLVMValueRef emitted_prims_vec =
+ LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
+
+ LLVMValueRef emitted_mask = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
+ emitted_vertices_vec,
+ uint_bld->zero);
+ /* We need to combine the current execution mask with the mask
+ telling us which, if any, execution slots actually have
+ unemitted primitives, this way we make sure that end_primitives
+ executes only on the paths that have unflushed vertices */
+ mask = LLVMBuildAnd(builder, mask, emitted_mask, "");
+
+ bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base,
+ emitted_vertices_vec,
+ emitted_prims_vec);
+
+#if DUMP_GS_EMITS
+ lp_build_print_value(bld->bld_base.base.gallivm,
+ " +++ end prim masked ones = ",
+ mask);
+ lp_build_print_value(bld->bld_base.base.gallivm,
+ " +++ end prim emitted verts1 = ",
+ emitted_vertices_vec);
+ lp_build_print_value(bld->bld_base.base.gallivm,
+ " +++ end prim emitted prims1 = ",
+ LLVMBuildLoad(builder,
+ bld->emitted_prims_vec_ptr, ""));
+#endif
+ increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr,
+ mask);
+ clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr,
+ mask);
+#if DUMP_GS_EMITS
+ lp_build_print_value(bld->bld_base.base.gallivm,
+ " +++ end prim emitted verts2 = ",
+ LLVMBuildLoad(builder,
+ bld->emitted_vertices_vec_ptr, ""));
+#endif
+ }
+
+}
+
+static void
+end_primitive(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ if (bld->gs_iface->end_primitive) {
+ LLVMValueRef mask = mask_vec(bld_base);
+ end_primitive_masked(bld_base, mask);
+ }
+}
+
static void
cal_emit(
const struct lp_build_tgsi_action * action,
{
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
- lp_exec_break(&bld->exec_mask);
+ lp_exec_break(&bld->exec_mask, bld_base);
+}
+
+static void
+breakc_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ struct lp_build_context *uint_bld = &bld_base->uint_bld;
+ LLVMValueRef unsigned_cond =
+ LLVMBuildBitCast(builder, emit_data->args[0], uint_bld->vec_type, "");
+ LLVMValueRef cond = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
+ unsigned_cond,
+ uint_bld->zero);
+
+ lp_exec_break_condition(&bld->exec_mask, cond);
}
static void
lp_exec_mask_cond_push(&bld->exec_mask, tmp);
}
+static void
+uif_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ LLVMValueRef tmp;
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+ struct lp_build_context *uint_bld = &bld_base->uint_bld;
+
+ tmp = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
+ emit_data->args[0], uint_bld->zero);
+ lp_exec_mask_cond_push(&bld->exec_mask, tmp);
+}
+
+static void
+case_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ lp_exec_case(&bld->exec_mask, emit_data->args[0]);
+}
+
+static void
+default_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ lp_exec_default(&bld->exec_mask, bld_base);
+}
+
+static void
+switch_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ lp_exec_switch(&bld->exec_mask, emit_data->args[0]);
+}
+
+static void
+endswitch_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+
+ lp_exec_endswitch(&bld->exec_mask, bld_base);
+}
+
static void
bgnloop_emit(
const struct lp_build_tgsi_action * action,
"output_array");
}
+ if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
+ LLVMValueRef array_size =
+ lp_build_const_int32(gallivm,
+ bld_base->info->file_max[TGSI_FILE_IMMEDIATE] * 4 + 4);
+ bld->imms_array = lp_build_array_alloca(gallivm,
+ bld_base->base.vec_type, array_size,
+ "imms_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)) {
+ if (bld->indirect_files & (1 << TGSI_FILE_INPUT) && !bld->gs_iface) {
unsigned index, chan;
LLVMTypeRef vec_type = bld_base->base.vec_type;
LLVMValueRef array_size = lp_build_const_int32(gallivm,
}
}
}
+
+ if (bld->gs_iface) {
+ struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
+ bld->emitted_prims_vec_ptr =
+ lp_build_alloca(gallivm,
+ uint_bld->vec_type,
+ "emitted_prims_ptr");
+ bld->emitted_vertices_vec_ptr =
+ lp_build_alloca(gallivm,
+ uint_bld->vec_type,
+ "emitted_vertices_ptr");
+ bld->total_emitted_vertices_vec_ptr =
+ lp_build_alloca(gallivm,
+ uint_bld->vec_type,
+ "total_emitted_vertices_ptr");
+
+ LLVMBuildStore(gallivm->builder, uint_bld->zero,
+ bld->emitted_prims_vec_ptr);
+ LLVMBuildStore(gallivm->builder, uint_bld->zero,
+ bld->emitted_vertices_vec_ptr);
+ LLVMBuildStore(gallivm->builder, uint_bld->zero,
+ bld->total_emitted_vertices_vec_ptr);
+ }
+
+ if (DEBUG_EXECUTION) {
+ lp_build_printf(gallivm, "\n");
+ emit_dump_file(bld, TGSI_FILE_CONSTANT);
+ if (!bld->gs_iface)
+ emit_dump_file(bld, TGSI_FILE_INPUT);
+ }
}
static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
{
struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- if (0) {
+ if (DEBUG_EXECUTION) {
/* for debugging */
- emit_dump_temps(bld);
+ if (0) {
+ emit_dump_file(bld, TGSI_FILE_TEMPORARY);
+ }
+ emit_dump_file(bld, TGSI_FILE_OUTPUT);
+ lp_build_printf(bld_base->base.gallivm, "\n");
}
/* 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(bld_base->info->num_outputs <=
- bld_base->info->file_max[TGSI_FILE_OUTPUT] + 1);
- for (index = 0; index < bld_base->info->num_outputs; ++index) {
- for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
- bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
- }
- }
+ * to the outputs slots specified by the caller */
+ if (bld->gs_iface) {
+ LLVMValueRef total_emitted_vertices_vec;
+ LLVMValueRef emitted_prims_vec;
+ /* implicit end_primitives, needed in case there are any unflushed
+ vertices in the cache. Note must not call end_primitive here
+ since the exec_mask is not valid at this point. */
+ end_primitive_masked(bld_base, lp_build_mask_value(bld->mask));
+
+ total_emitted_vertices_vec =
+ LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
+ emitted_prims_vec =
+ LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
+
+ bld->gs_iface->gs_epilogue(bld->gs_iface,
+ &bld->bld_base,
+ total_emitted_vertices_vec,
+ emitted_prims_vec);
+ } else {
+ gather_outputs(bld);
}
}
struct lp_type type,
struct lp_build_mask_context *mask,
LLVMValueRef consts_ptr,
+ LLVMValueRef const_sizes_ptr,
const struct lp_bld_tgsi_system_values *system_values,
- const LLVMValueRef *pos,
const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
struct lp_build_sampler_soa *sampler,
- const struct tgsi_shader_info *info)
+ const struct tgsi_shader_info *info,
+ const struct lp_build_tgsi_gs_iface *gs_iface)
{
struct lp_build_tgsi_soa_context bld;
lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
bld.mask = mask;
- bld.pos = pos;
bld.inputs = inputs;
bld.outputs = outputs;
bld.consts_ptr = consts_ptr;
+ bld.const_sizes_ptr = const_sizes_ptr;
bld.sampler = sampler;
bld.bld_base.info = info;
bld.indirect_files = info->indirect_files;
+ /*
+ * If the number of temporaries is rather large then we just
+ * allocate them as an array right from the start and treat
+ * like indirect temporaries.
+ */
+ if (info->file_max[TGSI_FILE_TEMPORARY] >= LP_MAX_INLINED_TEMPS) {
+ bld.indirect_files |= (1 << TGSI_FILE_TEMPORARY);
+ }
+ /*
+ * For performance reason immediates are always backed in a static
+ * array, but if their number is too great, we have to use just
+ * a dynamically allocated array.
+ */
+ bld.use_immediates_array =
+ (info->file_max[TGSI_FILE_IMMEDIATE] >= LP_MAX_INLINED_IMMEDIATES);
+ if (bld.use_immediates_array) {
+ bld.indirect_files |= (1 << TGSI_FILE_IMMEDIATE);
+ }
+
+
bld.bld_base.soa = TRUE;
+ bld.bld_base.emit_debug = emit_debug;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_BREAKC].emit = breakc_emit;
bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_CASE].emit = case_emit;
bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_DEFAULT].emit = default_emit;
bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_ENDSWITCH].emit = endswitch_emit;
bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
- bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
- bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_KILL_IF].emit = kill_if_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_KILL].emit = kill_emit;
bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_SWITCH].emit = switch_emit;
bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
bld.bld_base.op_actions[TGSI_OPCODE_TXF].emit = txf_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_TEX2].emit = tex2_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_TXB2].emit = txb2_emit;
+ bld.bld_base.op_actions[TGSI_OPCODE_TXL2].emit = txl2_emit;
/* DX10 sampling ops */
bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE].emit = sample_emit;
bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_B].emit = sample_b_emit;
bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_L].emit = sample_l_emit;
bld.bld_base.op_actions[TGSI_OPCODE_SVIEWINFO].emit = sviewinfo_emit;
- lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.base);
+ if (gs_iface) {
+ /* There's no specific value for this because it should always
+ * be set, but apps using ext_geometry_shader4 quite often
+ * were forgetting so we're using MAX_VERTEX_VARYING from
+ * that spec even though we could debug_assert if it's not
+ * set, but that's a lot uglier. */
+ uint max_output_vertices = 32;
+ uint i = 0;
+ /* inputs are always indirect with gs */
+ bld.indirect_files |= (1 << TGSI_FILE_INPUT);
+ bld.gs_iface = gs_iface;
+ bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_gs_input;
+ bld.bld_base.op_actions[TGSI_OPCODE_EMIT].emit = emit_vertex;
+ bld.bld_base.op_actions[TGSI_OPCODE_ENDPRIM].emit = end_primitive;
+
+ for (i = 0; i < info->num_properties; ++i) {
+ if (info->properties[i].name ==
+ TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES) {
+ max_output_vertices = info->properties[i].data[0];
+ }
+ }
+ bld.max_output_vertices_vec =
+ lp_build_const_int_vec(gallivm, bld.bld_base.int_bld.type,
+ max_output_vertices);
+ }
+
+ lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
bld.system_values = *system_values;
LLVMDumpModule(module);
}
+ lp_exec_mask_fini(&bld.exec_mask);
}