X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fshader%2Fslang%2Fslang_codegen.c;h=f9199a917f6706e7502bb1614b852322f94568ff;hb=6b05708ce2c4120921f43d82f1751e039306e0d6;hp=0c14e64cf97d9c7966a22985f3670ed0c8d2896e;hpb=ec6c8f86f3fa0fa19d34dfa1ee1f9e4610e6e9d6;p=mesa.git diff --git a/src/mesa/shader/slang/slang_codegen.c b/src/mesa/shader/slang/slang_codegen.c index 0c14e64cf97..f9199a917f6 100644 --- a/src/mesa/shader/slang/slang_codegen.c +++ b/src/mesa/shader/slang/slang_codegen.c @@ -1,6 +1,6 @@ /* * Mesa 3-D graphics library - * Version: 6.5.3 + * Version: 7.1 * * Copyright (C) 2005-2007 Brian Paul All Rights Reserved. * @@ -37,17 +37,19 @@ -#include "imports.h" -#include "macros.h" -#include "mtypes.h" -#include "program.h" -#include "prog_instruction.h" -#include "prog_parameter.h" -#include "prog_statevars.h" +#include "main/imports.h" +#include "main/macros.h" +#include "main/mtypes.h" +#include "shader/program.h" +#include "shader/prog_instruction.h" +#include "shader/prog_parameter.h" +#include "shader/prog_print.h" +#include "shader/prog_statevars.h" #include "slang_typeinfo.h" #include "slang_codegen.h" #include "slang_compile.h" #include "slang_label.h" +#include "slang_mem.h" #include "slang_simplify.h" #include "slang_emit.h" #include "slang_vartable.h" @@ -107,42 +109,84 @@ _slang_field_offset(const slang_type_specifier *spec, slang_atom field) } +/** + * Return the size (in floats) of the given type specifier. + * If the size is greater than 4, the size should be a multiple of 4 + * so that the correct number of 4-float registers are allocated. + * For example, a mat3x2 is size 12 because we want to store the + * 3 columns in 3 float[4] registers. + */ GLuint _slang_sizeof_type_specifier(const slang_type_specifier *spec) { + GLuint sz; switch (spec->type) { case SLANG_SPEC_VOID: - return 0; + sz = 0; + break; case SLANG_SPEC_BOOL: - return 1; + sz = 1; + break; case SLANG_SPEC_BVEC2: - return 2; + sz = 2; + break; case SLANG_SPEC_BVEC3: - return 3; + sz = 3; + break; case SLANG_SPEC_BVEC4: - return 4; + sz = 4; + break; case SLANG_SPEC_INT: - return 1; + sz = 1; + break; case SLANG_SPEC_IVEC2: - return 2; + sz = 2; + break; case SLANG_SPEC_IVEC3: - return 3; + sz = 3; + break; case SLANG_SPEC_IVEC4: - return 4; + sz = 4; + break; case SLANG_SPEC_FLOAT: - return 1; + sz = 1; + break; case SLANG_SPEC_VEC2: - return 2; + sz = 2; + break; case SLANG_SPEC_VEC3: - return 3; + sz = 3; + break; case SLANG_SPEC_VEC4: - return 4; + sz = 4; + break; case SLANG_SPEC_MAT2: - return 2 * 2; + sz = 2 * 4; /* 2 columns (regs) */ + break; case SLANG_SPEC_MAT3: - return 3 * 3; + sz = 3 * 4; + break; case SLANG_SPEC_MAT4: - return 4 * 4; + sz = 4 * 4; + break; + case SLANG_SPEC_MAT23: + sz = 2 * 4; /* 2 columns (regs) */ + break; + case SLANG_SPEC_MAT32: + sz = 3 * 4; /* 3 columns (regs) */ + break; + case SLANG_SPEC_MAT24: + sz = 2 * 4; + break; + case SLANG_SPEC_MAT42: + sz = 4 * 4; /* 4 columns (regs) */ + break; + case SLANG_SPEC_MAT34: + sz = 3 * 4; + break; + case SLANG_SPEC_MAT43: + sz = 4 * 4; /* 4 columns (regs) */ + break; case SLANG_SPEC_SAMPLER1D: case SLANG_SPEC_SAMPLER2D: case SLANG_SPEC_SAMPLER3D: @@ -151,16 +195,27 @@ _slang_sizeof_type_specifier(const slang_type_specifier *spec) case SLANG_SPEC_SAMPLER2DSHADOW: case SLANG_SPEC_SAMPLER2DRECT: case SLANG_SPEC_SAMPLER2DRECTSHADOW: - return 1; /* a sampler is basically just an integer index */ + sz = 1; /* a sampler is basically just an integer index */ + break; case SLANG_SPEC_STRUCT: - return _slang_field_offset(spec, 0); /* special use */ + sz = _slang_field_offset(spec, 0); /* special use */ + if (sz > 4) { + sz = (sz + 3) & ~0x3; /* round up to multiple of four */ + } + break; case SLANG_SPEC_ARRAY: - return _slang_sizeof_type_specifier(spec->_array); + sz = _slang_sizeof_type_specifier(spec->_array); + break; default: _mesa_problem(NULL, "Unexpected type in _slang_sizeof_type_specifier()"); - return 0; + sz = 0; + } + + if (sz > 4) { + /* if size is > 4, it should be a multiple of four */ + assert((sz & 0x3) == 0); } - return 0; + return sz; } @@ -189,7 +244,12 @@ _slang_attach_storage(slang_ir_node *n, slang_variable *var) } else { /* alloc new storage info */ - n->Store = _slang_new_ir_storage(PROGRAM_UNDEFINED, -1, -5); + n->Store = _slang_new_ir_storage(PROGRAM_UNDEFINED, -7, -5); +#if 0 + printf("%s var=%s Store=%p Size=%d\n", __FUNCTION__, + (char*) var->a_name, + (void*) n->Store, n->Store->Size); +#endif if (n->Var) n->Var->aux = n->Store; assert(n->Var->aux); @@ -228,6 +288,8 @@ sampler_to_texture_index(const slang_type_specifier_type type) } +#define SWIZZLE_ZWWW MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_W, SWIZZLE_W) + /** * Return the VERT_ATTRIB_* or FRAG_ATTRIB_* value that corresponds to * a vertex or fragment program input variable. Return -1 if the input @@ -262,9 +324,11 @@ _slang_input_index(const char *name, GLenum target, GLuint *swizzleOut) { "gl_FragCoord", FRAG_ATTRIB_WPOS, SWIZZLE_NOOP }, { "gl_Color", FRAG_ATTRIB_COL0, SWIZZLE_NOOP }, { "gl_SecondaryColor", FRAG_ATTRIB_COL1, SWIZZLE_NOOP }, - { "gl_FogFragCoord", FRAG_ATTRIB_FOGC, SWIZZLE_XXXX }, { "gl_TexCoord", FRAG_ATTRIB_TEX0, SWIZZLE_NOOP }, + /* note: we're packing several quantities into the fogcoord vector */ + { "gl_FogFragCoord", FRAG_ATTRIB_FOGC, SWIZZLE_XXXX }, { "gl_FrontFacing", FRAG_ATTRIB_FOGC, SWIZZLE_YYYY }, /*XXX*/ + { "gl_PointCoord", FRAG_ATTRIB_FOGC, SWIZZLE_ZWWW }, { NULL, 0, SWIZZLE_NOOP } }; GLuint i; @@ -361,6 +425,7 @@ static slang_asm_info AsmInfo[] = { { "vec4_sle", IR_SLE, 1, 2 }, { "vec4_slt", IR_SLT, 1, 2 }, /* vec4 unary */ + { "vec4_move", IR_MOVE, 1, 1 }, { "vec4_floor", IR_FLOOR, 1, 1 }, { "vec4_frac", IR_FRAC, 1, 1 }, { "vec4_abs", IR_ABS, 1, 1 }, @@ -384,8 +449,8 @@ static slang_asm_info AsmInfo[] = { { "vec4_texp_rect", IR_TEX, 1, 2 },/* rectangle w/ projection */ /* unary op */ - { "int_to_float", IR_I_TO_F, 1, 1 }, - { "float_to_int", IR_F_TO_I, 1, 1 }, + { "ivec4_to_vec4", IR_I_TO_F, 1, 1 }, /* int[4] to float[4] */ + { "vec4_to_ivec4", IR_F_TO_I, 1, 1 }, /* float[4] to int[4] */ { "float_exp", IR_EXP, 1, 1 }, { "float_exp2", IR_EXP2, 1, 1 }, { "float_log2", IR_LOG2, 1, 1 }, @@ -406,7 +471,7 @@ static slang_ir_node * new_node3(slang_ir_opcode op, slang_ir_node *c0, slang_ir_node *c1, slang_ir_node *c2) { - slang_ir_node *n = (slang_ir_node *) calloc(1, sizeof(slang_ir_node)); + slang_ir_node *n = (slang_ir_node *) _slang_alloc(sizeof(slang_ir_node)); if (n) { n->Opcode = op; n->Children[0] = c0; @@ -437,6 +502,9 @@ new_node0(slang_ir_opcode op) } +/** + * Create sequence of two nodes. + */ static slang_ir_node * new_seq(slang_ir_node *left, slang_ir_node *right) { @@ -477,12 +545,12 @@ new_not(slang_ir_node *n) /** - * Inlined subroutine. + * Non-inlined function call. */ static slang_ir_node * -new_inlined_function_call(slang_ir_node *code, slang_label *name) +new_function_call(slang_ir_node *code, slang_label *name) { - slang_ir_node *n = new_node1(IR_FUNC, code); + slang_ir_node *n = new_node1(IR_CALL, code); assert(name); if (n) n->Label = name; @@ -590,11 +658,16 @@ new_var(slang_assemble_ctx *A, slang_operation *oper, slang_atom name) if (!var) return NULL; + assert(var->declared); + assert(!oper->var || oper->var == var); n = new_node0(IR_VAR); if (n) { _slang_attach_storage(n, var); + /* + printf("new_var %s store=%p\n", (char*)name, (void*) n->Store); + */ } return n; } @@ -647,33 +720,60 @@ _slang_find_node_type(slang_operation *oper, slang_operation_type type) /** - * Produce inline code for a call to an assembly instruction. - * XXX Note: children are passed as asm args in-order, not by name! + * Count the number of operations of the given time rooted at 'oper'. */ -static slang_operation * -slang_inline_asm_function(slang_assemble_ctx *A, - slang_function *fun, slang_operation *oper) +static GLuint +_slang_count_node_type(slang_operation *oper, slang_operation_type type) { - const GLuint numArgs = oper->num_children; - const slang_operation *args = oper->children; - GLuint i; - slang_operation *inlined = slang_operation_new(1); + GLuint i, count = 0; + if (oper->type == type) { + return 1; + } + for (i = 0; i < oper->num_children; i++) { + count += _slang_count_node_type(&oper->children[i], type); + } + return count; +} - /*assert(oper->type == SLANG_OPER_CALL); or vec4_add, etc */ - /* - printf("Inline asm %s\n", (char*) fun->header.a_name); - */ - inlined->type = fun->body->children[0].type; - inlined->a_id = fun->body->children[0].a_id; - inlined->num_children = numArgs; - inlined->children = slang_operation_new(numArgs); - inlined->locals->outer_scope = oper->locals->outer_scope; - for (i = 0; i < numArgs; i++) { - slang_operation_copy(inlined->children + i, args + i); +/** + * Check if the 'return' statement found under 'oper' is a "tail return" + * that can be no-op'd. For example: + * + * void func(void) + * { + * .. do something .. + * return; // this is a no-op + * } + * + * This is used when determining if a function can be inlined. If the + * 'return' is not the last statement, we can't inline the function since + * we still need the semantic behaviour of the 'return' but we don't want + * to accidentally return from the _calling_ function. We'd need to use an + * unconditional branch, but we don't have such a GPU instruction (not + * always, at least). + */ +static GLboolean +_slang_is_tail_return(const slang_operation *oper) +{ + GLuint k = oper->num_children; + + while (k > 0) { + const slang_operation *last = &oper->children[k - 1]; + if (last->type == SLANG_OPER_RETURN) + return GL_TRUE; + else if (last->type == SLANG_OPER_IDENTIFIER || + last->type == SLANG_OPER_LABEL) + k--; /* try prev child */ + else if (last->type == SLANG_OPER_BLOCK_NO_NEW_SCOPE || + last->type == SLANG_OPER_BLOCK_NEW_SCOPE) + /* try sub-children */ + return _slang_is_tail_return(last); + else + break; } - return inlined; + return GL_FALSE; } @@ -836,6 +936,88 @@ slang_substitute(slang_assemble_ctx *A, slang_operation *oper, } +/** + * Produce inline code for a call to an assembly instruction. + * This is typically used to compile a call to a built-in function like this: + * + * vec4 mix(const vec4 x, const vec4 y, const vec4 a) + * { + * __asm vec4_lrp __retVal, a, y, x; + * } + * + * + * A call to + * r = mix(p1, p2, p3); + * + * Becomes: + * + * mov + * / \ + * r vec4_lrp + * / | \ + * p3 p2 p1 + * + * We basically translate a SLANG_OPER_CALL into a SLANG_OPER_ASM. + */ +static slang_operation * +slang_inline_asm_function(slang_assemble_ctx *A, + slang_function *fun, slang_operation *oper) +{ + const GLuint numArgs = oper->num_children; + GLuint i; + slang_operation *inlined; + const GLboolean haveRetValue = _slang_function_has_return_value(fun); + slang_variable **substOld; + slang_operation **substNew; + + ASSERT(slang_is_asm_function(fun)); + ASSERT(fun->param_count == numArgs + haveRetValue); + + /* + printf("Inline %s as %s\n", + (char*) fun->header.a_name, + (char*) fun->body->children[0].a_id); + */ + + /* + * We'll substitute formal params with actual args in the asm call. + */ + substOld = (slang_variable **) + _slang_alloc(numArgs * sizeof(slang_variable *)); + substNew = (slang_operation **) + _slang_alloc(numArgs * sizeof(slang_operation *)); + for (i = 0; i < numArgs; i++) { + substOld[i] = fun->parameters->variables[i]; + substNew[i] = oper->children + i; + } + + /* make a copy of the code to inline */ + inlined = slang_operation_new(1); + slang_operation_copy(inlined, &fun->body->children[0]); + if (haveRetValue) { + /* get rid of the __retVal child */ + inlined->num_children--; + for (i = 0; i < inlined->num_children; i++) { + inlined->children[i] = inlined->children[i + 1]; + } + } + + /* now do formal->actual substitutions */ + slang_substitute(A, inlined, numArgs, substOld, substNew, GL_FALSE); + + _slang_free(substOld); + _slang_free(substNew); + +#if 0 + printf("+++++++++++++ inlined asm function %s +++++++++++++\n", + (char *) fun->header.a_name); + slang_print_tree(inlined, 3); + printf("+++++++++++++++++++++++++++++++++++++++++++++++++++\n"); +#endif + + return inlined; +} + /** * Inline the given function call operation. @@ -860,6 +1042,7 @@ slang_inline_function_call(slang_assemble_ctx * A, slang_function *fun, slang_operation **substNew; GLuint substCount, numCopyIn, i; slang_function *prevFunction; + slang_variable_scope *newScope = NULL; /* save / push */ prevFunction = A->CurFunction; @@ -870,14 +1053,14 @@ slang_inline_function_call(slang_assemble_ctx * A, slang_function *fun, /* allocate temporary arrays */ paramMode = (ParamMode *) - _mesa_calloc(totalArgs * sizeof(ParamMode)); + _slang_alloc(totalArgs * sizeof(ParamMode)); substOld = (slang_variable **) - _mesa_calloc(totalArgs * sizeof(slang_variable *)); + _slang_alloc(totalArgs * sizeof(slang_variable *)); substNew = (slang_operation **) - _mesa_calloc(totalArgs * sizeof(slang_operation *)); + _slang_alloc(totalArgs * sizeof(slang_operation *)); #if 0 - printf("Inline call to %s (total vars=%d nparams=%d)\n", + printf("\nInline call to %s (total vars=%d nparams=%d)\n", (char *) fun->header.a_name, fun->parameters->num_variables, numArgs); #endif @@ -996,7 +1179,8 @@ slang_inline_function_call(slang_assemble_ctx * A, slang_function *fun, slang_operation_copy(inlined, fun->body); /*** XXX review this */ - assert(inlined->type = SLANG_OPER_BLOCK_NO_NEW_SCOPE); + assert(inlined->type == SLANG_OPER_BLOCK_NO_NEW_SCOPE || + inlined->type == SLANG_OPER_BLOCK_NEW_SCOPE); inlined->type = SLANG_OPER_BLOCK_NEW_SCOPE; #if 0 @@ -1027,9 +1211,7 @@ slang_inline_function_call(slang_assemble_ctx * A, slang_function *fun, slang_operation *decl = slang_operation_insert(&inlined->num_children, &inlined->children, numCopyIn); - /* - printf("COPY_IN %s from expr\n", (char*)p->a_name); - */ + decl->type = SLANG_OPER_VARIABLE_DECL; assert(decl->locals); decl->locals->outer_scope = inlined->locals; @@ -1040,10 +1222,29 @@ slang_inline_function_call(slang_assemble_ctx * A, slang_function *fun, /* child[0] is the var's initializer */ slang_operation_copy(&decl->children[0], args + i); + /* add parameter 'p' to the local variable scope here */ + { + slang_variable *pCopy = slang_variable_scope_grow(inlined->locals); + pCopy->type = p->type; + pCopy->a_name = p->a_name; + pCopy->array_len = p->array_len; + } + + newScope = inlined->locals; numCopyIn++; } } + /* Now add copies of the function's local vars to the new variable scope */ + for (i = totalArgs; i < fun->parameters->num_variables; i++) { + slang_variable *p = fun->parameters->variables[i]; + slang_variable *pCopy = slang_variable_scope_grow(inlined->locals); + pCopy->type = p->type; + pCopy->a_name = p->a_name; + pCopy->array_len = p->array_len; + } + + /* New epilog statements: * 1. Create end of function label to jump to from return statements. * 2. Copy the 'out' parameter vars @@ -1075,12 +1276,19 @@ slang_inline_function_call(slang_assemble_ctx * A, slang_function *fun, } } - _mesa_free(paramMode); - _mesa_free(substOld); - _mesa_free(substNew); + _slang_free(paramMode); + _slang_free(substOld); + _slang_free(substNew); + + /* Update scoping to use the new local vars instead of the + * original function's vars. This is especially important + * for nested inlining. + */ + if (newScope) + slang_replace_scope(inlined, fun->parameters, newScope); #if 0 - printf("Done Inline call to %s (total vars=%d nparams=%d)\n", + printf("Done Inline call to %s (total vars=%d nparams=%d)\n\n", (char *) fun->header.a_name, fun->parameters->num_variables, numArgs); slang_print_tree(top, 0); @@ -1113,29 +1321,65 @@ _slang_gen_function_call(slang_assemble_ctx *A, slang_function *fun, } else { /* non-assembly function */ + /* We always generate an "inline-able" block of code here. + * We may either: + * 1. insert the inline code + * 2. Generate a call to the "inline" code as a subroutine + */ + + + slang_operation *ret = NULL; + inlined = slang_inline_function_call(A, fun, oper, dest); - if (inlined && _slang_find_node_type(inlined, SLANG_OPER_RETURN)) { - /* This inlined function has one or more 'return' statements. - * So, we can't truly inline this function because we need to - * implement 'return' with RET (and CAL). - * XXX check if there's one 'return' and if it's the very last - * statement in the function - we can optimize that case. - */ - assert(inlined->type == SLANG_OPER_BLOCK_NEW_SCOPE || - inlined->type == SLANG_OPER_SEQUENCE); - inlined->type = SLANG_OPER_INLINED_CALL; - inlined->fun = fun; - inlined->label = _slang_label_new_unique((char*) fun->header.a_name); + if (!inlined) + return NULL; + + ret = _slang_find_node_type(inlined, SLANG_OPER_RETURN); + if (ret) { + /* check if this is a "tail" return */ + if (_slang_count_node_type(inlined, SLANG_OPER_RETURN) == 1 && + _slang_is_tail_return(inlined)) { + /* The only RETURN is the last stmt in the function, no-op it + * and inline the function body. + */ + ret->type = SLANG_OPER_NONE; + } + else { + slang_operation *callOper; + /* The function we're calling has one or more 'return' statements. + * So, we can't truly inline this function because we need to + * implement 'return' with RET (and CAL). + * Nevertheless, we performed "inlining" to make a new instance + * of the function body to deal with static register allocation. + * + * XXX check if there's one 'return' and if it's the very last + * statement in the function - we can optimize that case. + */ + assert(inlined->type == SLANG_OPER_BLOCK_NEW_SCOPE || + inlined->type == SLANG_OPER_SEQUENCE); + + if (_slang_function_has_return_value(fun) && !dest) { + assert(inlined->children[0].type == SLANG_OPER_VARIABLE_DECL); + assert(inlined->children[2].type == SLANG_OPER_IDENTIFIER); + callOper = &inlined->children[1]; + } + else { + callOper = inlined; + } + callOper->type = SLANG_OPER_NON_INLINED_CALL; + callOper->fun = fun; + callOper->label = _slang_label_new_unique((char*) fun->header.a_name); + } } } if (!inlined) return NULL; - /* Replace the function call with the inlined block */ + /* Replace the function call with the inlined block (or new CALL stmt) */ slang_operation_destruct(oper); *oper = *inlined; - /* XXX slang_operation_destruct(inlined) ??? */ + _slang_free(inlined); #if 0 assert(inlined->locals); @@ -1149,7 +1393,6 @@ _slang_gen_function_call(slang_assemble_ctx *A, slang_function *fun, /*_slang_label_delete(A->curFuncEndLabel);*/ A->curFuncEndLabel = prevFuncEndLabel; - assert(A->curFuncEndLabel); return n; } @@ -1168,109 +1411,291 @@ slang_find_asm_info(const char *name) } +/** + * Return the default swizzle mask for accessing a variable of the + * given size (in floats). If size = 1, comp is used to identify + * which component [0..3] of the register holds the variable. + */ static GLuint -make_writemask(const char *field) -{ - GLuint mask = 0x0; - while (*field) { - switch (*field) { - case 'x': - mask |= WRITEMASK_X; - break; - case 'y': - mask |= WRITEMASK_Y; - break; - case 'z': - mask |= WRITEMASK_Z; - break; - case 'w': - mask |= WRITEMASK_W; - break; - default: - _mesa_problem(NULL, "invalid writemask in make_writemask()"); - return 0; - } - field++; +_slang_var_swizzle(GLint size, GLint comp) +{ + switch (size) { + case 1: + return MAKE_SWIZZLE4(comp, comp, comp, comp); + case 2: + return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_NIL, SWIZZLE_NIL); + case 3: + return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_NIL); + default: + return SWIZZLE_XYZW; } - if (mask == 0x0) - return WRITEMASK_XYZW; - else - return mask; } /** - * Generate IR tree for an asm instruction/operation such as: - * __asm vec4_dot __retVal.x, v1, v2; + * Some write-masked assignments are simple, but others are hard. + * Simple example: + * vec3 v; + * v.xy = vec2(a, b); + * Hard example: + * vec3 v; + * v.zy = vec2(a, b); + * this gets transformed/swizzled into: + * v.zy = vec2(a, b).*yx* (* = don't care) + * This function helps to determine simple vs. non-simple. */ -static slang_ir_node * -_slang_gen_asm(slang_assemble_ctx *A, slang_operation *oper, - slang_operation *dest) +static GLboolean +_slang_simple_writemask(GLuint writemask, GLuint swizzle) { - const slang_asm_info *info; - slang_ir_node *kids[3], *n; - GLuint j, firstOperand; - - assert(oper->type == SLANG_OPER_ASM); - - info = slang_find_asm_info((char *) oper->a_id); - if (!info) { - _mesa_problem(NULL, "undefined __asm function %s\n", - (char *) oper->a_id); - assert(info); - } - assert(info->NumParams <= 3); - - if (info->NumParams == oper->num_children) { - /* Storage for result is not specified. - * Children[0], [1] are the operands. - */ - firstOperand = 0; - } - else { - /* Storage for result (child[0]) is specified. - * Children[1], [2] are the operands. - */ - firstOperand = 1; - } - - /* assemble child(ren) */ - kids[0] = kids[1] = kids[2] = NULL; - for (j = 0; j < info->NumParams; j++) { - kids[j] = _slang_gen_operation(A, &oper->children[firstOperand + j]); - if (!kids[j]) - return NULL; + switch (writemask) { + case WRITEMASK_X: + return GET_SWZ(swizzle, 0) == SWIZZLE_X; + case WRITEMASK_Y: + return GET_SWZ(swizzle, 1) == SWIZZLE_Y; + case WRITEMASK_Z: + return GET_SWZ(swizzle, 2) == SWIZZLE_Z; + case WRITEMASK_W: + return GET_SWZ(swizzle, 3) == SWIZZLE_W; + case WRITEMASK_XY: + return (GET_SWZ(swizzle, 0) == SWIZZLE_X) + && (GET_SWZ(swizzle, 1) == SWIZZLE_Y); + case WRITEMASK_XYZ: + return (GET_SWZ(swizzle, 0) == SWIZZLE_X) + && (GET_SWZ(swizzle, 1) == SWIZZLE_Y) + && (GET_SWZ(swizzle, 2) == SWIZZLE_Z); + case WRITEMASK_XYZW: + return swizzle == SWIZZLE_NOOP; + default: + return GL_FALSE; } +} - n = new_node3(info->Opcode, kids[0], kids[1], kids[2]); - if (firstOperand) { - /* Setup n->Store to be a particular location. Otherwise, storage - * for the result (a temporary) will be allocated later. - */ - GLuint writemask = WRITEMASK_XYZW; - slang_operation *dest_oper; - slang_ir_node *n0; +/** + * Convert the given swizzle into a writemask. In some cases this + * is trivial, in other cases, we'll need to also swizzle the right + * hand side to put components in the right places. + * \param swizzle the incoming swizzle + * \param writemaskOut returns the writemask + * \param swizzleOut swizzle to apply to the right-hand-side + * \return GL_FALSE for simple writemasks, GL_TRUE for non-simple + */ +static GLboolean +swizzle_to_writemask(slang_assemble_ctx *A, GLuint swizzle, + GLuint *writemaskOut, GLuint *swizzleOut) +{ + GLuint mask = 0x0, newSwizzle[4]; + GLint i, size; - dest_oper = &oper->children[0]; - while (dest_oper->type == SLANG_OPER_FIELD) { - /* writemask */ - writemask &= make_writemask((char*) dest_oper->a_id); - dest_oper = &dest_oper->children[0]; + /* make new dst writemask, compute size */ + for (i = 0; i < 4; i++) { + const GLuint swz = GET_SWZ(swizzle, i); + if (swz == SWIZZLE_NIL) { + /* end */ + break; } + assert(swz >= 0 && swz <= 3); - n0 = _slang_gen_operation(A, dest_oper); - assert(n0->Var); - assert(n0->Store); - assert(!n->Store); - n->Store = n0->Store; - n->Writemask = writemask; + if (swizzle != SWIZZLE_XXXX && + swizzle != SWIZZLE_YYYY && + swizzle != SWIZZLE_ZZZZ && + swizzle != SWIZZLE_WWWW && + (mask & (1 << swz))) { + /* a channel can't be specified twice (ex: ".xyyz") */ + slang_info_log_error(A->log, "Invalid writemask '%s'", + _mesa_swizzle_string(swizzle, 0, 0)); + return GL_FALSE; + } - free(n0); + mask |= (1 << swz); } + assert(mask <= 0xf); + size = i; /* number of components in mask/swizzle */ - return n; -} + *writemaskOut = mask; + + /* make new src swizzle, by inversion */ + for (i = 0; i < 4; i++) { + newSwizzle[i] = i; /*identity*/ + } + for (i = 0; i < size; i++) { + const GLuint swz = GET_SWZ(swizzle, i); + newSwizzle[swz] = i; + } + *swizzleOut = MAKE_SWIZZLE4(newSwizzle[0], + newSwizzle[1], + newSwizzle[2], + newSwizzle[3]); + + if (_slang_simple_writemask(mask, *swizzleOut)) { + if (size >= 1) + assert(GET_SWZ(*swizzleOut, 0) == SWIZZLE_X); + if (size >= 2) + assert(GET_SWZ(*swizzleOut, 1) == SWIZZLE_Y); + if (size >= 3) + assert(GET_SWZ(*swizzleOut, 2) == SWIZZLE_Z); + if (size >= 4) + assert(GET_SWZ(*swizzleOut, 3) == SWIZZLE_W); + return GL_TRUE; + } + else + return GL_FALSE; +} + + +/** + * Recursively traverse 'oper' to produce a swizzle mask in the event + * of any vector subscripts and swizzle suffixes. + * Ex: for "vec4 v", "v[2].x" resolves to v.z + */ +static GLuint +resolve_swizzle(const slang_operation *oper) +{ + if (oper->type == SLANG_OPER_FIELD) { + /* writemask from .xyzw suffix */ + slang_swizzle swz; + if (_slang_is_swizzle((char*) oper->a_id, 4, &swz)) { + GLuint swizzle = MAKE_SWIZZLE4(swz.swizzle[0], + swz.swizzle[1], + swz.swizzle[2], + swz.swizzle[3]); + GLuint child_swizzle = resolve_swizzle(&oper->children[0]); + GLuint s = _slang_swizzle_swizzle(child_swizzle, swizzle); + return s; + } + else + return SWIZZLE_XYZW; + } + else if (oper->type == SLANG_OPER_SUBSCRIPT && + oper->children[1].type == SLANG_OPER_LITERAL_INT) { + /* writemask from [index] */ + GLuint child_swizzle = resolve_swizzle(&oper->children[0]); + GLuint i = (GLuint) oper->children[1].literal[0]; + GLuint swizzle; + GLuint s; + switch (i) { + case 0: + swizzle = SWIZZLE_XXXX; + break; + case 1: + swizzle = SWIZZLE_YYYY; + break; + case 2: + swizzle = SWIZZLE_ZZZZ; + break; + case 3: + swizzle = SWIZZLE_WWWW; + break; + default: + swizzle = SWIZZLE_XYZW; + } + s = _slang_swizzle_swizzle(child_swizzle, swizzle); + return s; + } + else { + return SWIZZLE_XYZW; + } +} + + +/** + * As above, but produce a writemask. + */ +static GLuint +resolve_writemask(slang_assemble_ctx *A, const slang_operation *oper) +{ + GLuint swizzle = resolve_swizzle(oper); + GLuint writemask, swizzleOut; + swizzle_to_writemask(A, swizzle, &writemask, &swizzleOut); + return writemask; +} + + +/** + * Recursively descend through swizzle nodes to find the node's storage info. + */ +static slang_ir_storage * +get_store(const slang_ir_node *n) +{ + if (n->Opcode == IR_SWIZZLE) { + return get_store(n->Children[0]); + } + return n->Store; +} + + + +/** + * Generate IR tree for an asm instruction/operation such as: + * __asm vec4_dot __retVal.x, v1, v2; + */ +static slang_ir_node * +_slang_gen_asm(slang_assemble_ctx *A, slang_operation *oper, + slang_operation *dest) +{ + const slang_asm_info *info; + slang_ir_node *kids[3], *n; + GLuint j, firstOperand; + + assert(oper->type == SLANG_OPER_ASM); + + info = slang_find_asm_info((char *) oper->a_id); + if (!info) { + _mesa_problem(NULL, "undefined __asm function %s\n", + (char *) oper->a_id); + assert(info); + } + assert(info->NumParams <= 3); + + if (info->NumParams == oper->num_children) { + /* Storage for result is not specified. + * Children[0], [1], [2] are the operands. + */ + firstOperand = 0; + } + else { + /* Storage for result (child[0]) is specified. + * Children[1], [2], [3] are the operands. + */ + firstOperand = 1; + } + + /* assemble child(ren) */ + kids[0] = kids[1] = kids[2] = NULL; + for (j = 0; j < info->NumParams; j++) { + kids[j] = _slang_gen_operation(A, &oper->children[firstOperand + j]); + if (!kids[j]) + return NULL; + } + + n = new_node3(info->Opcode, kids[0], kids[1], kids[2]); + + if (firstOperand) { + /* Setup n->Store to be a particular location. Otherwise, storage + * for the result (a temporary) will be allocated later. + */ + GLuint writemask = WRITEMASK_XYZW; + slang_operation *dest_oper; + slang_ir_node *n0; + + dest_oper = &oper->children[0]; + + writemask = resolve_writemask(A, dest_oper); + + n0 = _slang_gen_operation(A, dest_oper); + if (!n0) + return NULL; + + assert(!n->Store); + n->Store = n0->Store; + + assert(n->Store->File != PROGRAM_UNDEFINED || n->Store->Parent); + + _slang_free(n0); + } + + return n; +} static void @@ -1289,7 +1714,7 @@ print_funcs(struct slang_function_scope_ *scope, const char *name) /** - * Return first function in the scope that has the given name. + * Find a function of the given name, taking 'numArgs' arguments. * This is the function we'll try to call when there is no exact match * between function parameters and call arguments. * @@ -1297,20 +1722,238 @@ print_funcs(struct slang_function_scope_ *scope, const char *name) * all of them... */ static slang_function * -_slang_first_function(struct slang_function_scope_ *scope, const char *name) +_slang_find_function_by_argc(slang_function_scope *scope, + const char *name, int numArgs) { - GLuint i; - for (i = 0; i < scope->num_functions; i++) { - slang_function *f = &scope->functions[i]; - if (strcmp(name, (char*) f->header.a_name) == 0) - return f; + while (scope) { + GLuint i; + for (i = 0; i < scope->num_functions; i++) { + slang_function *f = &scope->functions[i]; + if (strcmp(name, (char*) f->header.a_name) == 0) { + int haveRetValue = _slang_function_has_return_value(f); + if (numArgs == f->param_count - haveRetValue) + return f; + } + } + scope = scope->outer_scope; } - if (scope->outer_scope) - return _slang_first_function(scope->outer_scope, name); + return NULL; } +static slang_function * +_slang_find_function_by_max_argc(slang_function_scope *scope, + const char *name) +{ + slang_function *maxFunc = NULL; + GLuint maxArgs = 0; + + while (scope) { + GLuint i; + for (i = 0; i < scope->num_functions; i++) { + slang_function *f = &scope->functions[i]; + if (strcmp(name, (char*) f->header.a_name) == 0) { + if (f->param_count > maxArgs) { + maxArgs = f->param_count; + maxFunc = f; + } + } + } + scope = scope->outer_scope; + } + + return maxFunc; +} + + +/** + * Generate a new slang_function which is a constructor for a user-defined + * struct type. + */ +static slang_function * +_slang_make_constructor(slang_assemble_ctx *A, slang_struct *str) +{ + const GLint numFields = str->fields->num_variables; + + slang_function *fun = (slang_function *) _mesa_malloc(sizeof(slang_function)); + if (!fun) + return NULL; + + slang_function_construct(fun); + + /* function header (name, return type) */ + fun->kind = SLANG_FUNC_CONSTRUCTOR; + fun->header.a_name = str->a_name; + fun->header.type.qualifier = SLANG_QUAL_NONE; + fun->header.type.specifier.type = SLANG_SPEC_STRUCT; + fun->header.type.specifier._struct = str; + + /* function parameters (= struct's fields) */ + { + GLint i; + for (i = 0; i < numFields; i++) { + /* + printf("Field %d: %s\n", i, (char*) str->fields->variables[i]->a_name); + */ + slang_variable *p = slang_variable_scope_grow(fun->parameters); + *p = *str->fields->variables[i]; /* copy the type */ + p->type.qualifier = SLANG_QUAL_CONST; + } + fun->param_count = fun->parameters->num_variables; + } + + /* Add __retVal to params */ + { + slang_variable *p = slang_variable_scope_grow(fun->parameters); + slang_atom a_retVal = slang_atom_pool_atom(A->atoms, "__retVal"); + assert(a_retVal); + p->a_name = a_retVal; + p->type = fun->header.type; + p->type.qualifier = SLANG_QUAL_OUT; + fun->param_count++; + } + + /* function body is: + * block: + * declare T; + * T.f1 = p1; + * T.f2 = p2; + * ... + * T.fn = pn; + * return T; + */ + { + slang_variable_scope *scope; + slang_variable *var; + GLint i; + + fun->body = slang_operation_new(1); + fun->body->type = SLANG_OPER_BLOCK_NEW_SCOPE; + fun->body->num_children = numFields + 2; + fun->body->children = slang_operation_new(numFields + 2); + + scope = fun->body->locals; + scope->outer_scope = fun->parameters; + + /* create local var 't' */ + var = slang_variable_scope_grow(scope); + var->a_name = slang_atom_pool_atom(A->atoms, "t"); + var->type = fun->header.type; + + /* declare t */ + { + slang_operation *decl; + + decl = &fun->body->children[0]; + decl->type = SLANG_OPER_VARIABLE_DECL; + decl->locals = _slang_variable_scope_new(scope); + decl->a_id = var->a_name; + } + + /* assign params to fields of t */ + for (i = 0; i < numFields; i++) { + slang_operation *assign = &fun->body->children[1 + i]; + + assign->type = SLANG_OPER_ASSIGN; + assign->locals = _slang_variable_scope_new(scope); + assign->num_children = 2; + assign->children = slang_operation_new(2); + + { + slang_operation *lhs = &assign->children[0]; + + lhs->type = SLANG_OPER_FIELD; + lhs->locals = _slang_variable_scope_new(scope); + lhs->num_children = 1; + lhs->children = slang_operation_new(1); + lhs->a_id = str->fields->variables[i]->a_name; + + lhs->children[0].type = SLANG_OPER_IDENTIFIER; + lhs->children[0].a_id = var->a_name; + lhs->children[0].locals = _slang_variable_scope_new(scope); + +#if 0 + lhs->children[1].num_children = 1; + lhs->children[1].children = slang_operation_new(1); + lhs->children[1].children[0].type = SLANG_OPER_IDENTIFIER; + lhs->children[1].children[0].a_id = str->fields->variables[i]->a_name; + lhs->children[1].children->locals = _slang_variable_scope_new(scope); +#endif + } + + { + slang_operation *rhs = &assign->children[1]; + + rhs->type = SLANG_OPER_IDENTIFIER; + rhs->locals = _slang_variable_scope_new(scope); + rhs->a_id = str->fields->variables[i]->a_name; + } + } + + /* return t; */ + { + slang_operation *ret = &fun->body->children[numFields + 1]; + + ret->type = SLANG_OPER_RETURN; + ret->locals = _slang_variable_scope_new(scope); + ret->num_children = 1; + ret->children = slang_operation_new(1); + ret->children[0].type = SLANG_OPER_IDENTIFIER; + ret->children[0].a_id = var->a_name; + ret->children[0].locals = _slang_variable_scope_new(scope); + + } + } + /* + slang_print_function(fun, 1); + */ + return fun; +} + + +/** + * Find/create a function (constructor) for the given structure name. + */ +static slang_function * +_slang_locate_struct_constructor(slang_assemble_ctx *A, const char *name) +{ + unsigned int i; + for (i = 0; i < A->space.structs->num_structs; i++) { + slang_struct *str = &A->space.structs->structs[i]; + if (strcmp(name, (const char *) str->a_name) == 0) { + /* found a structure type that matches the function name */ + if (!str->constructor) { + /* create the constructor function now */ + str->constructor = _slang_make_constructor(A, str); + } + return str->constructor; + } + } + return NULL; +} + + + +static GLboolean +_slang_is_vec_mat_type(const char *name) +{ + static const char *vecmat_types[] = { + "float", "int", "bool", + "vec2", "vec3", "vec4", + "ivec2", "ivec3", "ivec4", + "bvec2", "bvec3", "bvec4", + "mat2", "mat3", "mat4", + "mat2x3", "mat2x4", "mat3x2", "mat3x4", "mat4x2", "mat4x3", + NULL + }; + int i; + for (i = 0; vecmat_types[i]; i++) + if (_mesa_strcmp(name, vecmat_types[i]) == 0) + return GL_TRUE; + return GL_FALSE; +} + /** * Assemble a function call, given a particular function name. @@ -1324,29 +1967,83 @@ _slang_gen_function_call_name(slang_assemble_ctx *A, const char *name, const GLuint param_count = oper->num_children; slang_atom atom; slang_function *fun; + GLboolean error; + slang_ir_node *n; atom = slang_atom_pool_atom(A->atoms, name); if (atom == SLANG_ATOM_NULL) return NULL; /* - * Use 'name' to find the function to call + * First, try to find function by name and exact argument type matching. */ fun = _slang_locate_function(A->space.funcs, atom, params, param_count, - &A->space, A->atoms, A->log); + &A->space, A->atoms, A->log, &error); + + if (error) { + slang_info_log_error(A->log, + "Function '%s' not found (check argument types)", + name); + return NULL; + } + if (!fun) { - /* A function with exactly the right parameters/types was not found. - * Try adapting the parameters. + /* Next, try locating a constructor function for a user-defined type */ + fun = _slang_locate_struct_constructor(A, name); + } + + /* + * At this point, some heuristics are used to try to find a function + * that matches the calling signature by means of casting or "unrolling" + * of constructors. + */ + + if (!fun && _slang_is_vec_mat_type(name)) { + /* Next, if this call looks like a vec() or mat() constructor call, + * try "unwinding" the args to satisfy a constructor. */ - fun = _slang_first_function(A->space.funcs, name); - if (!_slang_adapt_call(oper, fun, &A->space, A->atoms, A->log)) { - slang_info_log_error(A->log, "Function '%s' not found (check argument types)", name); + fun = _slang_find_function_by_max_argc(A->space.funcs, name); + if (fun) { + if (!_slang_adapt_call(oper, fun, &A->space, A->atoms, A->log)) { + slang_info_log_error(A->log, + "Function '%s' not found (check argument types)", + name); + return NULL; + } + } + } + + if (!fun && _slang_is_vec_mat_type(name)) { + /* Next, try casting args to the types of the formal parameters */ + int numArgs = oper->num_children; + fun = _slang_find_function_by_argc(A->space.funcs, name, numArgs); + if (!fun || !_slang_cast_func_params(oper, fun, &A->space, A->atoms, A->log)) { + slang_info_log_error(A->log, + "Function '%s' not found (check argument types)", + name); return NULL; } assert(fun); } - return _slang_gen_function_call(A, fun, oper, dest); + if (!fun) { + slang_info_log_error(A->log, + "Function '%s' not found (check argument types)", + name); + return NULL; + } + + n = _slang_gen_function_call(A, fun, oper, dest); + + if (n && !n->Store && !dest + && fun->header.type.specifier.type != SLANG_SPEC_VOID) { + /* setup n->Store for the result of the function call */ + GLint size = _slang_sizeof_type_specifier(&fun->header.type.specifier); + n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, size); + /*printf("Alloc storage for function result, size %d \n", size);*/ + } + + return n; } @@ -1387,6 +2084,23 @@ _slang_is_scalar_or_boolean(slang_assemble_ctx *A, slang_operation *oper) } +/** + * Test if an operation is boolean. + */ +static GLboolean +_slang_is_boolean(slang_assemble_ctx *A, slang_operation *oper) +{ + slang_typeinfo type; + GLboolean isBool; + + slang_typeinfo_construct(&type); + _slang_typeof_operation(A, oper, &type); + isBool = (type.spec.type == SLANG_SPEC_BOOL); + slang_typeinfo_destruct(&type); + return isBool; +} + + /** * Generate loop code using high-level IR_LOOP instruction */ @@ -1402,7 +2116,7 @@ _slang_gen_while(slang_assemble_ctx * A, const slang_operation *oper) GLboolean isConst, constTrue; /* type-check expression */ - if (!_slang_is_scalar_or_boolean(A, &oper->children[0])) { + if (!_slang_is_boolean(A, &oper->children[0])) { slang_info_log_error(A->log, "scalar/boolean expression expected for 'while'"); return NULL; } @@ -1465,7 +2179,7 @@ _slang_gen_do(slang_assemble_ctx * A, const slang_operation *oper) GLboolean isConst, constTrue; /* type-check expression */ - if (!_slang_is_scalar_or_boolean(A, &oper->children[1])) { + if (!_slang_is_boolean(A, &oper->children[1])) { slang_info_log_error(A->log, "scalar/boolean expression expected for 'do/while'"); return NULL; } @@ -1561,7 +2275,7 @@ _slang_gen_continue(slang_assemble_ctx * A, const slang_operation *oper) * Determine if the given operation is of a specific type. */ static GLboolean -is_operation_type(const const slang_operation *oper, slang_operation_type type) +is_operation_type(const slang_operation *oper, slang_operation_type type) { if (oper->type == type) return GL_TRUE; @@ -1594,6 +2308,11 @@ _slang_gen_if(slang_assemble_ctx * A, const slang_operation *oper) GLboolean isConst, constTrue; /* type-check expression */ + if (!_slang_is_boolean(A, &oper->children[0])) { + slang_info_log_error(A->log, "boolean expression expected for 'if'"); + return NULL; + } + if (!_slang_is_scalar_or_boolean(A, &oper->children[0])) { slang_info_log_error(A->log, "scalar/boolean expression expected for 'if'"); return NULL; @@ -1614,22 +2333,16 @@ _slang_gen_if(slang_assemble_ctx * A, const slang_operation *oper) cond = _slang_gen_operation(A, &oper->children[0]); cond = new_cond(cond); - if (is_operation_type(&oper->children[1], SLANG_OPER_BREAK)) { + if (is_operation_type(&oper->children[1], SLANG_OPER_BREAK) + && !haveElseClause) { /* Special case: generate a conditional break */ ifBody = new_break_if_true(A->CurLoop, cond); - if (haveElseClause) { - elseBody = _slang_gen_operation(A, &oper->children[2]); - return new_seq(ifBody, elseBody); - } return ifBody; } - else if (is_operation_type(&oper->children[1], SLANG_OPER_CONTINUE)) { + else if (is_operation_type(&oper->children[1], SLANG_OPER_CONTINUE) + && !haveElseClause) { /* Special case: generate a conditional break */ ifBody = new_cont_if_true(A->CurLoop, cond); - if (haveElseClause) { - elseBody = _slang_gen_operation(A, &oper->children[2]); - return new_seq(ifBody, elseBody); - } return ifBody; } else { @@ -1699,16 +2412,16 @@ static slang_ir_node * _slang_gen_temporary(GLint size) { slang_ir_storage *store; - slang_ir_node *n; + slang_ir_node *n = NULL; - store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, size); + store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -2, size); if (store) { n = new_node0(IR_VAR_DECL); if (n) { n->Store = store; } else { - free(store); + _slang_free(store); } } return n; @@ -1722,19 +2435,73 @@ static slang_ir_node * _slang_gen_var_decl(slang_assemble_ctx *A, slang_variable *var) { slang_ir_node *n; - assert(!is_sampler_type(&var->type)); + + /*assert(!var->declared);*/ + var->declared = GL_TRUE; + n = new_node0(IR_VAR_DECL); if (n) { _slang_attach_storage(n, var); - assert(var->aux); assert(n->Store == var->aux); assert(n->Store); assert(n->Store->Index < 0); - n->Store->File = PROGRAM_TEMPORARY; + if (is_sampler_type(&var->type)) { + n->Store->File = PROGRAM_SAMPLER; + } + else { + n->Store->File = PROGRAM_TEMPORARY; + } + n->Store->Size = _slang_sizeof_type_specifier(&n->Var->type.specifier); + + if (n->Store->Size <= 0) { + slang_info_log_error(A->log, "invalid declaration for '%s'", + (char*) var->a_name); + return NULL; + } +#if 0 + printf("%s var %p %s store=%p index=%d size=%d\n", + __FUNCTION__, (void *) var, (char *) var->a_name, + (void *) n->Store, n->Store->Index, n->Store->Size); +#endif + + if (var->array_len > 0) { + /* this is an array */ + /* cannot be const-qualified */ + if (var->type.qualifier == SLANG_QUAL_CONST) { + slang_info_log_error(A->log, "array '%s' cannot be const", + (char*) var->a_name); + return NULL; + } + else { + /* round up element size to mult of 4 */ + GLint sz = (n->Store->Size + 3) & ~3; + /* mult by array size */ + sz *= var->array_len; + n->Store->Size = sz; + } + } + assert(n->Store->Size > 0); + + /* setup default swizzle for storing the variable */ + switch (n->Store->Size) { + case 2: + n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, + SWIZZLE_NIL, SWIZZLE_NIL); + break; + case 3: + n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, + SWIZZLE_Z, SWIZZLE_NIL); + break; + default: + /* Note that float-sized vars may be allocated in any x/y/z/w + * slot, but that won't be determined until code emit time. + */ + n->Store->Swizzle = SWIZZLE_NOOP; + } } return n; } @@ -1751,16 +2518,36 @@ _slang_gen_select(slang_assemble_ctx *A, slang_operation *oper) { slang_ir_node *cond, *ifNode, *trueExpr, *falseExpr, *trueNode, *falseNode; slang_ir_node *tmpDecl, *tmpVar, *tree; - slang_typeinfo type; - int size; + slang_typeinfo type0, type1, type2; + int size, isBool, isEqual; assert(oper->type == SLANG_OPER_SELECT); assert(oper->num_children == 3); + /* type of children[0] must be boolean */ + slang_typeinfo_construct(&type0); + _slang_typeof_operation(A, &oper->children[0], &type0); + isBool = (type0.spec.type == SLANG_SPEC_BOOL); + slang_typeinfo_destruct(&type0); + if (!isBool) { + slang_info_log_error(A->log, "selector type is not boolean"); + return NULL; + } + + slang_typeinfo_construct(&type1); + slang_typeinfo_construct(&type2); + _slang_typeof_operation(A, &oper->children[1], &type1); + _slang_typeof_operation(A, &oper->children[2], &type2); + isEqual = slang_type_specifier_equal(&type1.spec, &type2.spec); + slang_typeinfo_destruct(&type1); + slang_typeinfo_destruct(&type2); + if (!isEqual) { + slang_info_log_error(A->log, "incompatible types for ?: operator"); + return NULL; + } + /* size of x or y's type */ - slang_typeinfo_construct(&type); - _slang_typeof_operation(A, &oper->children[1], &type); - size = _slang_sizeof_type_specifier(&type.spec); + size = _slang_sizeof_type_specifier(&type1.spec); assert(size > 0); /* temporary var */ @@ -1774,13 +2561,13 @@ _slang_gen_select(slang_assemble_ctx *A, slang_operation *oper) tmpVar = new_node0(IR_VAR); tmpVar->Store = tmpDecl->Store; trueExpr = _slang_gen_operation(A, &oper->children[1]); - trueNode = new_node2(IR_MOVE, tmpVar, trueExpr); + trueNode = new_node2(IR_COPY, tmpVar, trueExpr); /* if-false body (child 2) */ tmpVar = new_node0(IR_VAR); tmpVar->Store = tmpDecl->Store; falseExpr = _slang_gen_operation(A, &oper->children[2]); - falseNode = new_node2(IR_MOVE, tmpVar, falseExpr); + falseNode = new_node2(IR_COPY, tmpVar, falseExpr); ifNode = new_if(cond, trueNode, falseNode); @@ -1818,11 +2605,6 @@ _slang_gen_logical_and(slang_assemble_ctx *A, slang_operation *oper) select->children[2].literal_size = 1; n = _slang_gen_select(A, select); - - /* xxx wrong */ - free(select->children); - free(select); - return n; } @@ -1849,11 +2631,6 @@ _slang_gen_logical_or(slang_assemble_ctx *A, slang_operation *oper) slang_operation_copy(&select->children[2], &oper->children[1]); n = _slang_gen_select(A, select); - - /* xxx wrong */ - free(select->children); - free(select); - return n; } @@ -1865,8 +2642,7 @@ static slang_ir_node * _slang_gen_return(slang_assemble_ctx * A, slang_operation *oper) { const GLboolean haveReturnValue - = (oper->num_children == 1 && - oper->children[0].type != SLANG_OPER_VOID); + = (oper->num_children == 1 && oper->children[0].type != SLANG_OPER_VOID); /* error checking */ assert(A->CurFunction); @@ -1876,7 +2652,7 @@ _slang_gen_return(slang_assemble_ctx * A, slang_operation *oper) return NULL; } else if (!haveReturnValue && - A->CurFunction->header.type.specifier.type != SLANG_SPEC_VOID) { + A->CurFunction->header.type.specifier.type != SLANG_SPEC_VOID) { slang_info_log_error(A->log, "return statement requires an expression"); return NULL; } @@ -1932,6 +2708,92 @@ _slang_gen_return(slang_assemble_ctx * A, slang_operation *oper) } +/** + * Determine if the given operation/expression is const-valued. + */ +static GLboolean +_slang_is_constant_expr(const slang_operation *oper) +{ + slang_variable *var; + GLuint i; + + switch (oper->type) { + case SLANG_OPER_IDENTIFIER: + var = _slang_locate_variable(oper->locals, oper->a_id, GL_TRUE); + if (var && var->type.qualifier == SLANG_QUAL_CONST) + return GL_TRUE; + return GL_FALSE; + default: + for (i = 0; i < oper->num_children; i++) { + if (!_slang_is_constant_expr(&oper->children[i])) + return GL_FALSE; + } + return GL_TRUE; + } +} + + +/** + * Check if an assignment of type t1 to t0 is legal. + * XXX more cases needed. + */ +static GLboolean +_slang_assignment_compatible(slang_assemble_ctx *A, + slang_operation *op0, + slang_operation *op1) +{ + slang_typeinfo t0, t1; + GLuint sz0, sz1; + + if (op0->type == SLANG_OPER_POSTINCREMENT || + op0->type == SLANG_OPER_POSTDECREMENT) { + return GL_FALSE; + } + + slang_typeinfo_construct(&t0); + _slang_typeof_operation(A, op0, &t0); + + slang_typeinfo_construct(&t1); + _slang_typeof_operation(A, op1, &t1); + + sz0 = _slang_sizeof_type_specifier(&t0.spec); + sz1 = _slang_sizeof_type_specifier(&t1.spec); + +#if 1 + if (sz0 != sz1) { + /*printf("assignment size mismatch %u vs %u\n", sz0, sz1);*/ + return GL_FALSE; + } +#endif + + if (t0.spec.type == SLANG_SPEC_STRUCT && + t1.spec.type == SLANG_SPEC_STRUCT && + t0.spec._struct->a_name != t1.spec._struct->a_name) + return GL_FALSE; + + if (t0.spec.type == SLANG_SPEC_FLOAT && + t1.spec.type == SLANG_SPEC_BOOL) + return GL_FALSE; + +#if 0 /* not used just yet - causes problems elsewhere */ + if (t0.spec.type == SLANG_SPEC_INT && + t1.spec.type == SLANG_SPEC_FLOAT) + return GL_FALSE; +#endif + + if (t0.spec.type == SLANG_SPEC_BOOL && + t1.spec.type == SLANG_SPEC_FLOAT) + return GL_FALSE; + + if (t0.spec.type == SLANG_SPEC_BOOL && + t1.spec.type == SLANG_SPEC_INT) + return GL_FALSE; + + return GL_TRUE; +} + + + /** * Generate IR tree for a variable declaration. */ @@ -1942,63 +2804,105 @@ _slang_gen_declaration(slang_assemble_ctx *A, slang_operation *oper) slang_ir_node *varDecl; slang_variable *v; const char *varName = (char *) oper->a_id; + slang_operation *initializer; + + assert(oper->type == SLANG_OPER_VARIABLE_DECL); + assert(oper->num_children <= 1); + + v = _slang_locate_variable(oper->locals, oper->a_id, GL_TRUE); + if (!v) + return NULL; /* "shouldn't happen" */ + + if (v->type.qualifier == SLANG_QUAL_ATTRIBUTE || + v->type.qualifier == SLANG_QUAL_VARYING || + v->type.qualifier == SLANG_QUAL_UNIFORM) { + /* can't declare attribute/uniform vars inside functions */ + slang_info_log_error(A->log, + "local variable '%s' cannot be an attribute/uniform/varying", + varName); + return NULL; + } + +#if 0 + if (v->declared) { + slang_info_log_error(A->log, "variable '%s' redeclared", varName); + return NULL; + } +#endif + + varDecl = _slang_gen_var_decl(A, v); + if (!varDecl) + return NULL; + + /* check if the var has an initializer */ + if (oper->num_children > 0) { + assert(oper->num_children == 1); + initializer = &oper->children[0]; + } + else if (v->initializer) { + initializer = v->initializer; + } + else { + initializer = NULL; + } - assert(oper->num_children == 0 || oper->num_children == 1); + if (v->type.qualifier == SLANG_QUAL_CONST && !initializer) { + slang_info_log_error(A->log, + "const-qualified variable '%s' requires initializer", + varName); + return NULL; + } - v = _slang_locate_variable(oper->locals, oper->a_id, GL_TRUE); - assert(v); - varDecl = _slang_gen_var_decl(A, v); + if (initializer) { + slang_ir_node *var, *init; - if (oper->num_children > 0) { - /* child is initializer */ - slang_ir_node *var, *init, *rhs; - assert(oper->num_children == 1); - var = new_var(A, oper, oper->a_id); - if (!var) { - slang_info_log_error(A->log, "undefined variable '%s'", varName); + /* type check/compare var and initializer */ + if (!_slang_assignment_compatible(A, oper, initializer)) { + slang_info_log_error(A->log, "incompatible types in assignment"); return NULL; - } - /* XXX make copy of this initializer? */ - rhs = _slang_gen_operation(A, &oper->children[0]); - assert(rhs); - init = new_node2(IR_MOVE, var, rhs); - /*assert(rhs->Opcode != IR_SEQ);*/ - n = new_seq(varDecl, init); - } - else if (v->initializer) { - slang_ir_node *var, *init, *rhs; + } + var = new_var(A, oper, oper->a_id); if (!var) { slang_info_log_error(A->log, "undefined variable '%s'", varName); return NULL; } + + if (v->type.qualifier == SLANG_QUAL_CONST) { + /* if the variable is const, the initializer must be a const + * expression as well. + */ #if 0 - /* XXX make copy of this initializer? */ - { - slang_operation dup; - slang_operation_construct(&dup); - slang_operation_copy(&dup, v->initializer); - _slang_simplify(&dup, &A->space, A->atoms); - rhs = _slang_gen_operation(A, &dup); - } -#else - _slang_simplify(v->initializer, &A->space, A->atoms); - rhs = _slang_gen_operation(A, v->initializer); + if (!_slang_is_constant_expr(initializer)) { + slang_info_log_error(A->log, + "initializer for %s not constant", varName); + return NULL; + } #endif - if (!rhs) + } + + _slang_simplify(initializer, &A->space, A->atoms); + + init = _slang_gen_operation(A, initializer); + if (!init) return NULL; - assert(rhs); - init = new_node2(IR_MOVE, var, rhs); - /* - assert(rhs->Opcode != IR_SEQ); - */ - n = new_seq(varDecl, init); + /*assert(init->Store);*/ + + /* XXX remove this when type checking is added above */ + if (init->Store && var->Store->Size != init->Store->Size) { + slang_info_log_error(A->log, "invalid assignment (wrong types)"); + return NULL; + } + + n = new_node2(IR_COPY, var, init); + n = new_seq(varDecl, n); } else { n = varDecl; } + return n; } @@ -2022,102 +2926,20 @@ _slang_gen_variable(slang_assemble_ctx * A, slang_operation *oper) } -/** - * Some write-masked assignments are simple, but others are hard. - * Simple example: - * vec3 v; - * v.xy = vec2(a, b); - * Hard example: - * vec3 v; - * v.zy = vec2(a, b); - * this gets transformed/swizzled into: - * v.zy = vec2(a, b).*yx* (* = don't care) - * This function helps to determine simple vs. non-simple. - */ -static GLboolean -_slang_simple_writemask(GLuint writemask, GLuint swizzle) -{ - switch (writemask) { - case WRITEMASK_X: - return GET_SWZ(swizzle, 0) == SWIZZLE_X; - case WRITEMASK_Y: - return GET_SWZ(swizzle, 1) == SWIZZLE_Y; - case WRITEMASK_Z: - return GET_SWZ(swizzle, 2) == SWIZZLE_Z; - case WRITEMASK_W: - return GET_SWZ(swizzle, 3) == SWIZZLE_W; - case WRITEMASK_XY: - return (GET_SWZ(swizzle, 0) == SWIZZLE_X) - && (GET_SWZ(swizzle, 1) == SWIZZLE_Y); - case WRITEMASK_XYZ: - return (GET_SWZ(swizzle, 0) == SWIZZLE_X) - && (GET_SWZ(swizzle, 1) == SWIZZLE_Y) - && (GET_SWZ(swizzle, 2) == SWIZZLE_Z); - case WRITEMASK_XYZW: - return swizzle == SWIZZLE_NOOP; - default: - return GL_FALSE; - } -} - /** - * Convert the given swizzle into a writemask. In some cases this - * is trivial, in other cases, we'll need to also swizzle the right - * hand side to put components in the right places. - * \param swizzle the incoming swizzle - * \param writemaskOut returns the writemask - * \param swizzleOut swizzle to apply to the right-hand-side - * \return GL_FALSE for simple writemasks, GL_TRUE for non-simple + * Return the number of components actually named by the swizzle. + * Recall that swizzles may have undefined/don't-care values. */ -static GLboolean -swizzle_to_writemask(GLuint swizzle, - GLuint *writemaskOut, GLuint *swizzleOut) +static GLuint +swizzle_size(GLuint swizzle) { - GLuint mask = 0x0, newSwizzle[4]; - GLint i, size; - - /* make new dst writemask, compute size */ - for (i = 0; i < 4; i++) { - const GLuint swz = GET_SWZ(swizzle, i); - if (swz == SWIZZLE_NIL) { - /* end */ - break; - } - assert(swz >= 0 && swz <= 3); - mask |= (1 << swz); - } - assert(mask <= 0xf); - size = i; /* number of components in mask/swizzle */ - - *writemaskOut = mask; - - /* make new src swizzle, by inversion */ + GLuint size = 0, i; for (i = 0; i < 4; i++) { - newSwizzle[i] = i; /*identity*/ - } - for (i = 0; i < size; i++) { - const GLuint swz = GET_SWZ(swizzle, i); - newSwizzle[swz] = i; - } - *swizzleOut = MAKE_SWIZZLE4(newSwizzle[0], - newSwizzle[1], - newSwizzle[2], - newSwizzle[3]); - - if (_slang_simple_writemask(mask, *swizzleOut)) { - if (size >= 1) - assert(GET_SWZ(*swizzleOut, 0) == SWIZZLE_X); - if (size >= 2) - assert(GET_SWZ(*swizzleOut, 1) == SWIZZLE_Y); - if (size >= 3) - assert(GET_SWZ(*swizzleOut, 2) == SWIZZLE_Z); - if (size >= 4) - assert(GET_SWZ(*swizzleOut, 3) == SWIZZLE_W); - return GL_TRUE; + GLuint swz = GET_SWZ(swizzle, i); + size += (swz >= 0 && swz <= 3); } - else - return GL_FALSE; + return size; } @@ -2127,13 +2949,34 @@ _slang_gen_swizzle(slang_ir_node *child, GLuint swizzle) slang_ir_node *n = new_node1(IR_SWIZZLE, child); assert(child); if (n) { - n->Store = _slang_new_ir_storage(PROGRAM_UNDEFINED, -1, -1); + assert(!n->Store); + n->Store = _slang_new_ir_storage_relative(0, + swizzle_size(swizzle), + child->Store); n->Store->Swizzle = swizzle; } return n; } +static GLboolean +is_store_writable(const slang_assemble_ctx *A, const slang_ir_storage *store) +{ + while (store->Parent) + store = store->Parent; + + if (!(store->File == PROGRAM_OUTPUT || + store->File == PROGRAM_TEMPORARY || + (store->File == PROGRAM_VARYING && + A->program->Target == GL_VERTEX_PROGRAM_ARB))) { + return GL_FALSE; + } + else { + return GL_TRUE; + } +} + + /** * Generate IR tree for an assignment (=). */ @@ -2152,7 +2995,9 @@ _slang_gen_assignment(slang_assemble_ctx * A, slang_operation *oper) } if (var->type.qualifier == SLANG_QUAL_CONST || var->type.qualifier == SLANG_QUAL_ATTRIBUTE || - var->type.qualifier == SLANG_QUAL_UNIFORM) { + var->type.qualifier == SLANG_QUAL_UNIFORM || + (var->type.qualifier == SLANG_QUAL_VARYING && + A->program->Target == GL_FRAGMENT_PROGRAM_ARB)) { slang_info_log_error(A->log, "illegal assignment to read-only variable '%s'", (char *) oper->children[0].a_id); @@ -2177,31 +3022,45 @@ _slang_gen_assignment(slang_assemble_ctx * A, slang_operation *oper) } else { slang_ir_node *n, *lhs, *rhs; + + /* lhs and rhs type checking */ + if (!_slang_assignment_compatible(A, + &oper->children[0], + &oper->children[1])) { + slang_info_log_error(A->log, "incompatible types in assignment"); + return NULL; + } + lhs = _slang_gen_operation(A, &oper->children[0]); + if (!lhs) { + return NULL; + } - if (lhs) { - if (lhs->Store->File != PROGRAM_OUTPUT && - lhs->Store->File != PROGRAM_TEMPORARY && - lhs->Store->File != PROGRAM_VARYING && - lhs->Store->File != PROGRAM_UNDEFINED) { - slang_info_log_error(A->log, - "illegal assignment to read-only l-value"); - return NULL; - } + if (!lhs->Store) { + slang_info_log_error(A->log, + "invalid left hand side for assignment"); + return NULL; + } + + /* check that lhs is writable */ + if (!is_store_writable(A, lhs->Store)) { + slang_info_log_error(A->log, + "illegal assignment to read-only l-value"); + return NULL; } rhs = _slang_gen_operation(A, &oper->children[1]); if (lhs && rhs) { /* convert lhs swizzle into writemask */ GLuint writemask, newSwizzle; - if (!swizzle_to_writemask(lhs->Store->Swizzle, + if (!swizzle_to_writemask(A, lhs->Store->Swizzle, &writemask, &newSwizzle)) { /* Non-simple writemask, need to swizzle right hand side in * order to put components into the right place. */ rhs = _slang_gen_swizzle(rhs, newSwizzle); } - n = new_node2(IR_MOVE, lhs, rhs); + n = new_node2(IR_COPY, lhs, rhs); n->Writemask = writemask; return n; } @@ -2216,7 +3075,7 @@ _slang_gen_assignment(slang_assemble_ctx * A, slang_operation *oper) * Generate IR tree for referencing a field in a struct (or basic vector type) */ static slang_ir_node * -_slang_gen_field(slang_assemble_ctx * A, slang_operation *oper) +_slang_gen_struct_field(slang_assemble_ctx * A, slang_operation *oper) { slang_typeinfo ti; @@ -2232,6 +3091,7 @@ _slang_gen_field(slang_assemble_ctx * A, slang_operation *oper) GLuint swizzle; if (!_slang_is_swizzle((char *) oper->a_id, rows, &swz)) { slang_info_log_error(A->log, "Bad swizzle"); + return NULL; } swizzle = MAKE_SWIZZLE4(swz.swizzle[0], swz.swizzle[1], @@ -2244,7 +3104,9 @@ _slang_gen_field(slang_assemble_ctx * A, slang_operation *oper) n = _slang_gen_swizzle(n, swizzle); return n; } - else if (ti.spec.type == SLANG_SPEC_FLOAT) { + else if ( ti.spec.type == SLANG_SPEC_FLOAT + || ti.spec.type == SLANG_SPEC_INT + || ti.spec.type == SLANG_SPEC_BOOL) { const GLuint rows = 1; slang_swizzle swz; slang_ir_node *n; @@ -2267,19 +3129,25 @@ _slang_gen_field(slang_assemble_ctx * A, slang_operation *oper) /* oper->a_id is the field name */ slang_ir_node *base, *n; slang_typeinfo field_ti; - GLint fieldSize, fieldOffset; + GLint fieldSize, fieldOffset = -1, swz; + /* type of field */ slang_typeinfo_construct(&field_ti); _slang_typeof_operation(A, oper, &field_ti); fieldSize = _slang_sizeof_type_specifier(&field_ti.spec); - fieldOffset = _slang_field_offset(&ti.spec, oper->a_id); + if (fieldSize > 0) + fieldOffset = _slang_field_offset(&ti.spec, oper->a_id); - if (fieldOffset < 0) { + if (fieldSize == 0 || fieldOffset < 0) { + const char *structName; + if (ti.spec._struct) + structName = (char *) ti.spec._struct->a_name; + else + structName = "unknown"; slang_info_log_error(A->log, "\"%s\" is not a member of struct \"%s\"", - (char *) oper->a_id, - (char *) ti.spec._struct->a_name); + (char *) oper->a_id, structName); return NULL; } assert(fieldSize >= 0); @@ -2291,20 +3159,27 @@ _slang_gen_field(slang_assemble_ctx * A, slang_operation *oper) } n = new_node1(IR_FIELD, base); - if (n) { - n->Field = (char *) oper->a_id; - n->FieldOffset = fieldOffset; - assert(n->FieldOffset >= 0); - n->Store = _slang_new_ir_storage(base->Store->File, - base->Store->Index, - fieldSize); - } - return n; + if (!n) + return NULL; -#if 0 - _mesa_problem(NULL, "glsl structs/fields not supported yet"); - return NULL; -#endif + + /* setup the storage info for this node */ + swz = fieldOffset % 4; + + n->Field = (char *) oper->a_id; + n->Store = _slang_new_ir_storage_relative(fieldOffset / 4, + fieldSize, + base->Store); + if (fieldSize == 1) + n->Store->Swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz); + else if (fieldSize == 2) + n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, + SWIZZLE_NIL, SWIZZLE_NIL); + else if (fieldSize == 3) + n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, + SWIZZLE_Z, SWIZZLE_NIL); + + return n; } } @@ -2313,7 +3188,7 @@ _slang_gen_field(slang_assemble_ctx * A, slang_operation *oper) * Gen code for array indexing. */ static slang_ir_node * -_slang_gen_subscript(slang_assemble_ctx * A, slang_operation *oper) +_slang_gen_array_element(slang_assemble_ctx * A, slang_operation *oper) { slang_typeinfo array_ti; @@ -2330,7 +3205,7 @@ _slang_gen_subscript(slang_assemble_ctx * A, slang_operation *oper) index = (GLint) oper->children[1].literal[0]; if (oper->children[1].type != SLANG_OPER_LITERAL_INT || - index >= max) { + index >= (GLint) max) { slang_info_log_error(A->log, "Invalid array index for vector type"); return NULL; } @@ -2346,6 +3221,7 @@ _slang_gen_subscript(slang_assemble_ctx * A, slang_operation *oper) /*n->Store = _slang_clone_ir_storage_swz(n->Store, */ n->Writemask = WRITEMASK_X << index; } + assert(n->Store); return n; } else { @@ -2377,21 +3253,30 @@ _slang_gen_subscript(slang_assemble_ctx * A, slang_operation *oper) index = _slang_gen_operation(A, &oper->children[1]); if (array && index) { /* bounds check */ - if (index->Opcode == IR_FLOAT && - ((int) index->Value[0] < 0 || - (int) index->Value[0] >= arrayLen)) { - slang_info_log_error(A->log, + GLint constIndex = -1; + if (index->Opcode == IR_FLOAT) { + constIndex = (int) index->Value[0]; + if (constIndex < 0 || constIndex >= arrayLen) { + slang_info_log_error(A->log, "Array index out of bounds (index=%d size=%d)", - (int) index->Value[0], arrayLen); - _slang_free_ir_tree(array); - _slang_free_ir_tree(index); + constIndex, arrayLen); + _slang_free_ir_tree(array); + _slang_free_ir_tree(index); + return NULL; + } + } + + if (!array->Store) { + slang_info_log_error(A->log, "Invalid array"); return NULL; } elem = new_node2(IR_ELEMENT, array, index); - elem->Store = _slang_new_ir_storage(array->Store->File, - array->Store->Index, - elemSize); + elem->Store = _slang_new_ir_storage_relative(constIndex, + elemSize, + array->Store); + + assert(elem->Store->Parent); /* XXX try to do some array bounds checking here */ return elem; } @@ -2404,47 +3289,79 @@ _slang_gen_subscript(slang_assemble_ctx * A, slang_operation *oper) } -/** - * Look for expressions such as: gl_ModelviewMatrix * gl_Vertex - * and replace with this: gl_Vertex * gl_ModelviewMatrixTranpose - * Since matrices are stored in column-major order, the second form of - * multiplication is much more efficient (just 4 dot products). - */ +static slang_ir_node * +_slang_gen_compare(slang_assemble_ctx *A, slang_operation *oper, + slang_ir_opcode opcode) +{ + slang_typeinfo t0, t1; + slang_ir_node *n; + + slang_typeinfo_construct(&t0); + _slang_typeof_operation(A, &oper->children[0], &t0); + + slang_typeinfo_construct(&t1); + _slang_typeof_operation(A, &oper->children[0], &t1); + + if (t0.spec.type == SLANG_SPEC_ARRAY || + t1.spec.type == SLANG_SPEC_ARRAY) { + slang_info_log_error(A->log, "Illegal array comparison"); + return NULL; + } + + if (oper->type != SLANG_OPER_EQUAL && + oper->type != SLANG_OPER_NOTEQUAL) { + /* <, <=, >, >= can only be used with scalars */ + if ((t0.spec.type != SLANG_SPEC_INT && + t0.spec.type != SLANG_SPEC_FLOAT) || + (t1.spec.type != SLANG_SPEC_INT && + t1.spec.type != SLANG_SPEC_FLOAT)) { + slang_info_log_error(A->log, "Incompatible type(s) for inequality operator"); + return NULL; + } + } + + n = new_node2(opcode, + _slang_gen_operation(A, &oper->children[0]), + _slang_gen_operation(A, &oper->children[1])); + + /* result is a bool (size 1) */ + n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1); + + return n; +} + + +#if 0 static void -_slang_check_matmul_optimization(slang_assemble_ctx *A, slang_operation *oper) -{ - static const struct { - const char *orig; - const char *tranpose; - } matrices[] = { - {"gl_ModelViewMatrix", "gl_ModelViewMatrixTranspose"}, - {"gl_ProjectionMatrix", "gl_ProjectionMatrixTranspose"}, - {"gl_ModelViewProjectionMatrix", "gl_ModelViewProjectionMatrixTranspose"}, - {"gl_TextureMatrix", "gl_TextureMatrixTranspose"}, - {"gl_NormalMatrix", "__NormalMatrixTranspose"}, - { NULL, NULL } - }; +print_vars(slang_variable_scope *s) +{ + int i; + printf("vars: "); + for (i = 0; i < s->num_variables; i++) { + printf("%s %d, \n", + (char*) s->variables[i]->a_name, + s->variables[i]->declared); + } - assert(oper->type == SLANG_OPER_MULTIPLY); - if (oper->children[0].type == SLANG_OPER_IDENTIFIER) { - GLuint i; - for (i = 0; matrices[i].orig; i++) { - if (oper->children[0].a_id - == slang_atom_pool_atom(A->atoms, matrices[i].orig)) { - /* - _mesa_printf("Replace %s with %s\n", - matrices[i].orig, matrices[i].tranpose); - */ - assert(oper->children[0].type == SLANG_OPER_IDENTIFIER); - oper->children[0].a_id - = slang_atom_pool_atom(A->atoms, matrices[i].tranpose); - /* finally, swap the operands */ - _slang_operation_swap(&oper->children[0], &oper->children[1]); - return; - } + printf("\n"); +} +#endif + + +#if 0 +static void +_slang_undeclare_vars(slang_variable_scope *locals) +{ + if (locals->num_variables > 0) { + int i; + for (i = 0; i < locals->num_variables; i++) { + slang_variable *v = locals->variables[i]; + printf("undeclare %s at %p\n", (char*) v->a_name, v); + v->declared = GL_FALSE; } } } +#endif /** @@ -2466,6 +3383,9 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) _slang_pop_var_table(A->vartable); + /*_slang_undeclare_vars(oper->locals);*/ + /*print_vars(oper->locals);*/ + if (n) n = new_node1(IR_SCOPE, n); return n; @@ -2485,29 +3405,9 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) _slang_free_ir_tree(tree); return NULL; /* error must have occured */ } - tree = tree ? new_seq(tree, n) : n; + tree = new_seq(tree, n); } -#if 00 - if (oper->locals->num_variables > 0) { - int i; - /* - printf("\n****** Deallocate vars in scope!\n"); - */ - for (i = 0; i < oper->locals->num_variables; i++) { - slang_variable *v = oper->locals->variables + i; - if (v->aux) { - slang_ir_storage *store = (slang_ir_storage *) v->aux; - /* - printf(" Deallocate var %s\n", (char*) v->a_name); - */ - assert(store->File == PROGRAM_TEMPORARY); - assert(store->Index >= 0); - _slang_free_temp(A->vartable, store->Index, store->Size); - } - } - } -#endif return tree; } else { @@ -2539,29 +3439,17 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) return new_node0(IR_KILL); case SLANG_OPER_EQUAL: - return new_node2(IR_EQUAL, - _slang_gen_operation(A, &oper->children[0]), - _slang_gen_operation(A, &oper->children[1])); + return _slang_gen_compare(A, oper, IR_EQUAL); case SLANG_OPER_NOTEQUAL: - return new_node2(IR_NOTEQUAL, - _slang_gen_operation(A, &oper->children[0]), - _slang_gen_operation(A, &oper->children[1])); + return _slang_gen_compare(A, oper, IR_NOTEQUAL); case SLANG_OPER_GREATER: - return new_node2(IR_SGT, - _slang_gen_operation(A, &oper->children[0]), - _slang_gen_operation(A, &oper->children[1])); + return _slang_gen_compare(A, oper, IR_SGT); case SLANG_OPER_LESS: - return new_node2(IR_SLT, - _slang_gen_operation(A, &oper->children[0]), - _slang_gen_operation(A, &oper->children[1])); + return _slang_gen_compare(A, oper, IR_SLT); case SLANG_OPER_GREATEREQUAL: - return new_node2(IR_SGE, - _slang_gen_operation(A, &oper->children[0]), - _slang_gen_operation(A, &oper->children[1])); + return _slang_gen_compare(A, oper, IR_SGE); case SLANG_OPER_LESSEQUAL: - return new_node2(IR_SLE, - _slang_gen_operation(A, &oper->children[0]), - _slang_gen_operation(A, &oper->children[1])); + return _slang_gen_compare(A, oper, IR_SLE); case SLANG_OPER_ADD: { slang_ir_node *n; @@ -2580,7 +3468,6 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) { slang_ir_node *n; assert(oper->num_children == 2); - _slang_check_matmul_optimization(A, oper); n = _slang_gen_function_call_name(A, "*", oper, NULL); return n; } @@ -2609,14 +3496,14 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) { slang_ir_node *n; assert(oper->num_children == 2); - n = _slang_gen_function_call_name(A, "+=", oper, &oper->children[0]); + n = _slang_gen_function_call_name(A, "+=", oper, NULL); return n; } case SLANG_OPER_SUBASSIGN: { slang_ir_node *n; assert(oper->num_children == 2); - n = _slang_gen_function_call_name(A, "-=", oper, &oper->children[0]); + n = _slang_gen_function_call_name(A, "-=", oper, NULL); return n; } break; @@ -2624,14 +3511,14 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) { slang_ir_node *n; assert(oper->num_children == 2); - n = _slang_gen_function_call_name(A, "*=", oper, &oper->children[0]); + n = _slang_gen_function_call_name(A, "*=", oper, NULL); return n; } case SLANG_OPER_DIVASSIGN: { slang_ir_node *n; assert(oper->num_children == 2); - n = _slang_gen_function_call_name(A, "/=", oper, &oper->children[0]); + n = _slang_gen_function_call_name(A, "/=", oper, NULL); return n; } case SLANG_OPER_LOGICALAND: @@ -2674,9 +3561,9 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) case SLANG_OPER_IF: return _slang_gen_if(A, oper); case SLANG_OPER_FIELD: - return _slang_gen_field(A, oper); + return _slang_gen_struct_field(A, oper); case SLANG_OPER_SUBSCRIPT: - return _slang_gen_subscript(A, oper); + return _slang_gen_array_element(A, oper); case SLANG_OPER_LITERAL_FLOAT: /* fall-through */ case SLANG_OPER_LITERAL_INT: @@ -2713,17 +3600,19 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) return n; } - case SLANG_OPER_INLINED_CALL: + case SLANG_OPER_NON_INLINED_CALL: case SLANG_OPER_SEQUENCE: { slang_ir_node *tree = NULL; GLuint i; for (i = 0; i < oper->num_children; i++) { slang_ir_node *n = _slang_gen_operation(A, &oper->children[i]); - tree = tree ? new_seq(tree, n) : n; + tree = new_seq(tree, n); + if (n) + tree->Store = n->Store; } - if (oper->type == SLANG_OPER_INLINED_CALL) { - tree = new_inlined_function_call(tree, oper->label); + if (oper->type == SLANG_OPER_NON_INLINED_CALL) { + tree = new_function_call(tree, oper->label); } return tree; } @@ -2743,6 +3632,23 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper) } +/** + * Compute total size of array give size of element, number of elements. + */ +static GLint +array_size(GLint baseSize, GLint arrayLen) +{ + GLint total; + if (arrayLen > 1) { + /* round up base type to multiple of 4 */ + total = ((baseSize + 3) & ~0x3) * MAX2(arrayLen, 1); + } + else { + total = baseSize; + } + return total; +} + /** * Called by compiler when a global variable has been parsed/compiled. @@ -2764,118 +3670,214 @@ _slang_codegen_global_variable(slang_assemble_ctx *A, slang_variable *var, struct gl_program *prog = A->program; const char *varName = (char *) var->a_name; GLboolean success = GL_TRUE; - GLint texIndex; slang_ir_storage *store = NULL; int dbg = 0; - - texIndex = sampler_to_texture_index(var->type.specifier.type); + const GLenum datatype = _slang_gltype_from_specifier(&var->type.specifier); + const GLint texIndex = sampler_to_texture_index(var->type.specifier.type); + const GLint size = _slang_sizeof_type_specifier(&var->type.specifier); if (texIndex != -1) { - /* Texture sampler: + /* This is a texture sampler variable... * store->File = PROGRAM_SAMPLER - * store->Index = sampler uniform location + * store->Index = sampler number (0..7, typically) * store->Size = texture type index (1D, 2D, 3D, cube, etc) */ - GLint samplerUniform = _mesa_add_sampler(prog->Parameters, varName); - store = _slang_new_ir_storage(PROGRAM_SAMPLER, samplerUniform, texIndex); + if (var->initializer) { + slang_info_log_error(A->log, "illegal assignment to '%s'", varName); + return GL_FALSE; + } +#if FEATURE_es2_glsl /* XXX should use FEATURE_texture_rect */ + /* disallow rect samplers */ + if (var->type.specifier.type == SLANG_SPEC_SAMPLER2DRECT || + var->type.specifier.type == SLANG_SPEC_SAMPLER2DRECTSHADOW) { + slang_info_log_error(A->log, "invalid sampler type for '%s'", varName); + return GL_FALSE; + } +#endif + { + GLint sampNum = _mesa_add_sampler(prog->Parameters, varName, datatype); + store = _slang_new_ir_storage(PROGRAM_SAMPLER, sampNum, texIndex); + } if (dbg) printf("SAMPLER "); } else if (var->type.qualifier == SLANG_QUAL_UNIFORM) { /* Uniform variable */ - const GLint size = _slang_sizeof_type_specifier(&var->type.specifier) - * MAX2(var->array_len, 1); + const GLint totalSize = array_size(size, var->array_len); + const GLuint swizzle = _slang_var_swizzle(totalSize, 0); + if (prog) { /* user-defined uniform */ - GLint uniformLoc = _mesa_add_uniform(prog->Parameters, varName, size); - store = _slang_new_ir_storage(PROGRAM_UNIFORM, uniformLoc, size); + if (datatype == GL_NONE) { + if (var->type.specifier.type == SLANG_SPEC_STRUCT) { + /* temporary work-around */ + GLenum datatype = GL_FLOAT; + GLint uniformLoc = _mesa_add_uniform(prog->Parameters, varName, + totalSize, datatype, NULL); + store = _slang_new_ir_storage_swz(PROGRAM_UNIFORM, uniformLoc, + totalSize, swizzle); + + /* XXX what we need to do is unroll the struct into its + * basic types, creating a uniform variable for each. + * For example: + * struct foo { + * vec3 a; + * vec4 b; + * }; + * uniform foo f; + * + * Should produce uniforms: + * "f.a" (GL_FLOAT_VEC3) + * "f.b" (GL_FLOAT_VEC4) + */ + + if (var->initializer) { + slang_info_log_error(A->log, + "unsupported initializer for uniform '%s'", varName); + return GL_FALSE; + } + } + else { + slang_info_log_error(A->log, + "invalid datatype for uniform variable %s", + varName); + return GL_FALSE; + } + } + else { + GLint uniformLoc; + const GLfloat *initialValues = NULL; + if (var->initializer) { + _slang_simplify(var->initializer, &A->space, A->atoms); + if (var->initializer->type == SLANG_OPER_LITERAL_FLOAT || + var->initializer->type == SLANG_OPER_LITERAL_INT) { + /* simple float/vector initializer */ + initialValues = var->initializer->literal; + } + else { + /* complex initializer */ + slang_info_log_error(A->log, + "unsupported initializer for uniform '%s'", varName); + return GL_FALSE; + } + } + + uniformLoc = _mesa_add_uniform(prog->Parameters, varName, + totalSize, datatype, initialValues); + store = _slang_new_ir_storage_swz(PROGRAM_UNIFORM, uniformLoc, + totalSize, swizzle); + } } else { /* pre-defined uniform, like gl_ModelviewMatrix */ /* We know it's a uniform, but don't allocate storage unless * it's really used. */ - store = _slang_new_ir_storage(PROGRAM_STATE_VAR, -1, size); + store = _slang_new_ir_storage_swz(PROGRAM_STATE_VAR, -1, + totalSize, swizzle); } - if (dbg) printf("UNIFORM (sz %d) ", size); + if (dbg) printf("UNIFORM (sz %d) ", totalSize); } else if (var->type.qualifier == SLANG_QUAL_VARYING) { - const GLint size = 4; /* XXX fix */ + const GLint totalSize = array_size(size, var->array_len); + + /* varyings must be float, vec or mat */ + if (!_slang_type_is_float_vec_mat(var->type.specifier.type) && + var->type.specifier.type != SLANG_SPEC_ARRAY) { + slang_info_log_error(A->log, + "varying '%s' must be float/vector/matrix", + varName); + return GL_FALSE; + } + + if (var->initializer) { + slang_info_log_error(A->log, "illegal initializer for varying '%s'", + varName); + return GL_FALSE; + } + if (prog) { /* user-defined varying */ - GLint varyingLoc = _mesa_add_varying(prog->Varying, varName, size); - store = _slang_new_ir_storage(PROGRAM_VARYING, varyingLoc, size); + GLint varyingLoc = _mesa_add_varying(prog->Varying, varName, totalSize); + GLuint swizzle = _slang_var_swizzle(size, 0); + store = _slang_new_ir_storage_swz(PROGRAM_VARYING, varyingLoc, + totalSize, swizzle); } else { /* pre-defined varying, like gl_Color or gl_TexCoord */ if (type == SLANG_UNIT_FRAGMENT_BUILTIN) { + /* fragment program input */ GLuint swizzle; GLint index = _slang_input_index(varName, GL_FRAGMENT_PROGRAM_ARB, &swizzle); assert(index >= 0); - store = _slang_new_ir_storage(PROGRAM_INPUT, index, size); - store->Swizzle = swizzle; assert(index < FRAG_ATTRIB_MAX); + store = _slang_new_ir_storage_swz(PROGRAM_INPUT, index, + size, swizzle); } else { + /* vertex program output */ GLint index = _slang_output_index(varName, GL_VERTEX_PROGRAM_ARB); + GLuint swizzle = _slang_var_swizzle(size, 0); assert(index >= 0); - assert(type == SLANG_UNIT_VERTEX_BUILTIN); - store = _slang_new_ir_storage(PROGRAM_OUTPUT, index, size); assert(index < VERT_RESULT_MAX); + assert(type == SLANG_UNIT_VERTEX_BUILTIN); + store = _slang_new_ir_storage_swz(PROGRAM_OUTPUT, index, + size, swizzle); } if (dbg) printf("V/F "); } if (dbg) printf("VARYING "); } else if (var->type.qualifier == SLANG_QUAL_ATTRIBUTE) { + GLuint swizzle; + GLint index; + /* attributes must be float, vec or mat */ + if (!_slang_type_is_float_vec_mat(var->type.specifier.type)) { + slang_info_log_error(A->log, + "attribute '%s' must be float/vector/matrix", + varName); + return GL_FALSE; + } + if (prog) { /* user-defined vertex attribute */ - const GLint size = _slang_sizeof_type_specifier(&var->type.specifier); const GLint attr = -1; /* unknown */ - GLint index = _mesa_add_attribute(prog->Attributes, varName, - size, attr); + swizzle = _slang_var_swizzle(size, 0); + index = _mesa_add_attribute(prog->Attributes, varName, + size, datatype, attr); assert(index >= 0); - store = _slang_new_ir_storage(PROGRAM_INPUT, - VERT_ATTRIB_GENERIC0 + index, size); + index = VERT_ATTRIB_GENERIC0 + index; } else { /* pre-defined vertex attrib */ - GLuint swizzle; - GLint index = _slang_input_index(varName, GL_VERTEX_PROGRAM_ARB, - &swizzle); - GLint size = 4; /* XXX? */ + index = _slang_input_index(varName, GL_VERTEX_PROGRAM_ARB, &swizzle); assert(index >= 0); - store = _slang_new_ir_storage(PROGRAM_INPUT, index, size); - store->Swizzle = swizzle; } + store = _slang_new_ir_storage_swz(PROGRAM_INPUT, index, size, swizzle); if (dbg) printf("ATTRIB "); } else if (var->type.qualifier == SLANG_QUAL_FIXEDINPUT) { - GLuint swizzle; + GLuint swizzle = SWIZZLE_XYZW; /* silence compiler warning */ GLint index = _slang_input_index(varName, GL_FRAGMENT_PROGRAM_ARB, &swizzle); - GLint size = 4; /* XXX? */ - store = _slang_new_ir_storage(PROGRAM_INPUT, index, size); - store->Swizzle = swizzle; + store = _slang_new_ir_storage_swz(PROGRAM_INPUT, index, size, swizzle); if (dbg) printf("INPUT "); } else if (var->type.qualifier == SLANG_QUAL_FIXEDOUTPUT) { if (type == SLANG_UNIT_VERTEX_BUILTIN) { GLint index = _slang_output_index(varName, GL_VERTEX_PROGRAM_ARB); - GLint size = 4; /* XXX? */ store = _slang_new_ir_storage(PROGRAM_OUTPUT, index, size); } else { - assert(type == SLANG_UNIT_FRAGMENT_BUILTIN); GLint index = _slang_output_index(varName, GL_FRAGMENT_PROGRAM_ARB); - GLint size = 4; /* XXX? */ - store = _slang_new_ir_storage(PROGRAM_OUTPUT, index, size); + GLint specialSize = 4; /* treat all fragment outputs as float[4] */ + assert(type == SLANG_UNIT_FRAGMENT_BUILTIN); + store = _slang_new_ir_storage(PROGRAM_OUTPUT, index, specialSize); } if (dbg) printf("OUTPUT "); } else if (var->type.qualifier == SLANG_QUAL_CONST && !prog) { /* pre-defined global constant, like gl_MaxLights */ - const GLint size = _slang_sizeof_type_specifier(&var->type.specifier); store = _slang_new_ir_storage(PROGRAM_CONSTANT, -1, size); if (dbg) printf("CONST "); } @@ -2890,7 +3892,7 @@ _slang_codegen_global_variable(slang_assemble_ctx *A, slang_variable *var, if (var->initializer) { slang_ir_node *lhs, *rhs, *init; - /* Generate IR_MOVE instruction to initialize the variable */ + /* Generate IR_COPY instruction to initialize the variable */ lhs = new_node0(IR_VAR); lhs->Var = var; lhs->Store = n->Store; @@ -2900,7 +3902,7 @@ _slang_codegen_global_variable(slang_assemble_ctx *A, slang_variable *var, rhs = _slang_gen_operation(A, var->initializer); assert(rhs); - init = new_node2(IR_MOVE, lhs, rhs); + init = new_node2(IR_COPY, lhs, rhs); n = new_seq(n, init); } @@ -2915,6 +3917,8 @@ _slang_codegen_global_variable(slang_assemble_ctx *A, slang_variable *var, if (store) var->aux = store; /* save var's storage info */ + var->declared = GL_TRUE; + return success; } @@ -2932,7 +3936,7 @@ _slang_codegen_function(slang_assemble_ctx * A, slang_function * fun) if (_mesa_strcmp((char *) fun->header.a_name, "main") != 0) { /* we only really generate code for main, all other functions get - * inlined. + * inlined or codegen'd upon an actual call. */ #if 0 /* do some basic error checking though */