Mesa (gallium-0.2): mesa: rework GLSL array code generation
Alan Hourihane
alanh at kemper.freedesktop.org
Thu Nov 20 13:44:24 UTC 2008
Module: Mesa
Branch: gallium-0.2
Commit: ae0ff8097b85d3537a7be1674d55a44a9bd6018e
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=ae0ff8097b85d3537a7be1674d55a44a9bd6018e
Author: Brian Paul <brian.paul at tungstengraphics.com>
Date: Wed Nov 19 14:12:25 2008 -0700
mesa: rework GLSL array code generation
We now express arrays in terms of indirect addressing. For example:
dst = a[i];
becomes:
MOV dst, TEMP[1 + TEMP[2].y];
At instruction-emit time indirect addressing is converted into ARL/
ADDR-relative form:
ARL ADDR.x, TEMP[2].y;
MOV dst, TEMP[1 + ADDR.x];
This fixes a number of array-related issues. Arrays of arrays and complex
array/struct nesting works now.
There may be some regressions, but more work is coming.
---
src/mesa/shader/slang/slang_codegen.c | 32 +--
src/mesa/shader/slang/slang_emit.c | 382 +++++++++++++++++++++-----------
src/mesa/shader/slang/slang_ir.c | 50 ++++
src/mesa/shader/slang/slang_ir.h | 57 +++++-
src/mesa/shader/slang/slang_vartable.h | 10 +-
5 files changed, 371 insertions(+), 160 deletions(-)
diff --git a/src/mesa/shader/slang/slang_codegen.c b/src/mesa/shader/slang/slang_codegen.c
index 66e2d1f..851335d 100644
--- a/src/mesa/shader/slang/slang_codegen.c
+++ b/src/mesa/shader/slang/slang_codegen.c
@@ -3116,7 +3116,7 @@ _slang_gen_struct_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 = -1, swz;
+ GLint fieldSize, fieldOffset = -1;
/* type of field */
slang_typeinfo_construct(&field_ti);
@@ -3149,22 +3149,12 @@ _slang_gen_struct_field(slang_assemble_ctx * A, slang_operation *oper)
if (!n)
return NULL;
-
- /* 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);
+
+ /* Store the field's offset in storage->Index */
+ n->Store = _slang_new_ir_storage(base->Store->File,
+ fieldOffset,
+ fieldSize);
return n;
}
@@ -3257,12 +3247,12 @@ _slang_gen_array_element(slang_assemble_ctx * A, slang_operation *oper)
}
elem = new_node2(IR_ELEMENT, array, index);
- elem->Store = _slang_new_ir_storage_relative(constIndex,
- elemSize,
- array->Store);
- assert(elem->Store->Parent);
- /* XXX try to do some array bounds checking here */
+ /* The storage info here will be updated during code emit */
+ elem->Store = _slang_new_ir_storage(array->Store->File,
+ array->Store->Index,
+ elemSize);
+
return elem;
}
else {
diff --git a/src/mesa/shader/slang/slang_emit.c b/src/mesa/shader/slang/slang_emit.c
index 11699c3..0a3ab39 100644
--- a/src/mesa/shader/slang/slang_emit.c
+++ b/src/mesa/shader/slang/slang_emit.c
@@ -261,13 +261,16 @@ fix_swizzle(GLuint swizzle)
static void
storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st)
{
+ const GLboolean relAddr = st->RelAddr;
const GLint size = st->Size;
GLint index = st->Index;
GLuint swizzle = st->Swizzle;
+ assert(index >= 0);
/* if this is storage relative to some parent storage, walk up the tree */
while (st->Parent) {
st = st->Parent;
+ assert(st->Index >= 0);
index += st->Index;
swizzle = _slang_swizzle_swizzle(st->Swizzle, swizzle);
}
@@ -304,6 +307,8 @@ storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st)
}
dst->WriteMask = writemask;
}
+
+ dst->RelAddr = relAddr;
}
@@ -318,8 +323,10 @@ storage_to_src_reg(struct prog_src_register *src, const slang_ir_storage *st)
GLuint swizzle = st->Swizzle;
/* if this is storage relative to some parent storage, walk up the tree */
+ assert(index >= 0);
while (st->Parent) {
st = st->Parent;
+ assert(st->Index >= 0);
index += st->Index;
swizzle = _slang_swizzle_swizzle(fix_swizzle(st->Swizzle), swizzle);
}
@@ -413,19 +420,131 @@ new_instruction(slang_emit_info *emitInfo, gl_inst_opcode opcode)
}
+static struct prog_instruction *
+emit_arl_load(slang_emit_info *emitInfo,
+ enum register_file file, GLint index, GLuint swizzle)
+{
+ struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_ARL);
+ inst->SrcReg[0].File = file;
+ inst->SrcReg[0].Index = index;
+ inst->SrcReg[0].Swizzle = swizzle;
+ inst->DstReg.File = PROGRAM_ADDRESS;
+ inst->DstReg.Index = 0;
+ inst->DstReg.WriteMask = WRITEMASK_X;
+ return inst;
+}
+
+
/**
* Emit a new instruction with given opcode, operands.
+ * At this point the instruction may have multiple indirect register
+ * loads/stores. We convert those into ARL loads and address-relative
+ * operands. See comments inside.
+ * At some point in the future we could directly emit indirectly addressed
+ * registers in Mesa GPU instructions.
*/
static struct prog_instruction *
emit_instruction(slang_emit_info *emitInfo,
gl_inst_opcode opcode,
const slang_ir_storage *dst,
+ const slang_ir_storage *src0,
const slang_ir_storage *src1,
- const slang_ir_storage *src2,
- const slang_ir_storage *src3)
+ const slang_ir_storage *src2)
{
struct prog_instruction *inst;
+ GLuint numIndirect = 0;
+ const slang_ir_storage *src[3];
+ slang_ir_storage newSrc[3], newDst;
+ GLuint i;
+ GLboolean isTemp[3];
+
+ isTemp[0] = isTemp[1] = isTemp[2] = GL_FALSE;
+
+ src[0] = src0;
+ src[1] = src1;
+ src[2] = src2;
+
+ /* count up how many operands are indirect loads */
+ for (i = 0; i < 3; i++) {
+ if (src[i] && src[i]->IsIndirect)
+ numIndirect++;
+ }
+ if (dst && dst->IsIndirect)
+ numIndirect++;
+
+ /* Take special steps for indirect register loads.
+ * If we had multiple address registers this would be simpler.
+ * For example, this GLSL code:
+ * x[i] = y[j] + z[k];
+ * would translate into something like:
+ * ARL ADDR.x, i;
+ * ARL ADDR.y, j;
+ * ARL ADDR.z, k;
+ * ADD TEMP[ADDR.x+5], TEMP[ADDR.y+9], TEMP[ADDR.z+4];
+ * But since we currently only have one address register we have to do this:
+ * ARL ADDR.x, i;
+ * MOV t1, TEMP[ADDR.x+9];
+ * ARL ADDR.x, j;
+ * MOV t2, TEMP[ADDR.x+4];
+ * ARL ADDR.x, k;
+ * ADD TEMP[ADDR.x+5], t1, t2;
+ * The code here figures this out...
+ */
+ if (numIndirect > 0) {
+ for (i = 0; i < 3; i++) {
+ if (src[i] && src[i]->IsIndirect) {
+ /* load the ARL register with the indirect register */
+ emit_arl_load(emitInfo,
+ src[i]->IndirectFile,
+ src[i]->IndirectIndex,
+ src[i]->IndirectSwizzle);
+
+ if (numIndirect > 1) {
+ /* Need to load src[i] into a temporary register */
+ slang_ir_storage srcRelAddr;
+ alloc_local_temp(emitInfo, &newSrc[i], src[i]->Size);
+ isTemp[i] = GL_TRUE;
+
+ /* set RelAddr flag on src register */
+ srcRelAddr = *src[i];
+ srcRelAddr.RelAddr = GL_TRUE;
+ srcRelAddr.IsIndirect = GL_FALSE; /* not really needed */
+
+ /* MOV newSrc, srcRelAddr; */
+ inst = emit_instruction(emitInfo,
+ OPCODE_MOV,
+ &newSrc[i],
+ &srcRelAddr,
+ NULL,
+ NULL);
+
+ src[i] = &newSrc[i];
+ }
+ else {
+ /* just rewrite the src[i] storage to be ARL-relative */
+ newSrc[i] = *src[i];
+ newSrc[i].RelAddr = GL_TRUE;
+ newSrc[i].IsIndirect = GL_FALSE; /* not really needed */
+ src[i] = &newSrc[i];
+ }
+ }
+ }
+ }
+
+ /* Take special steps for indirect dest register write */
+ if (dst && dst->IsIndirect) {
+ /* load the ARL register with the indirect register */
+ emit_arl_load(emitInfo,
+ dst->IndirectFile,
+ dst->IndirectIndex,
+ dst->IndirectSwizzle);
+ newDst = *dst;
+ newDst.RelAddr = GL_TRUE;
+ newDst.IsIndirect = GL_FALSE;
+ dst = &newDst;
+ }
+ /* OK, emit the instruction and its dst, src regs */
inst = new_instruction(emitInfo, opcode);
if (!inst)
return NULL;
@@ -433,33 +552,17 @@ emit_instruction(slang_emit_info *emitInfo,
if (dst)
storage_to_dst_reg(&inst->DstReg, dst);
- if (src1)
- storage_to_src_reg(&inst->SrcReg[0], src1);
- if (src2)
- storage_to_src_reg(&inst->SrcReg[1], src2);
- if (src3)
- storage_to_src_reg(&inst->SrcReg[2], src3);
-
- return inst;
-}
-
+ for (i = 0; i < 3; i++) {
+ if (src[i])
+ storage_to_src_reg(&inst->SrcReg[i], src[i]);
+ }
-/**
- * Emit an ARL instruction.
- */
-static struct prog_instruction *
-emit_arl_instruction(slang_emit_info *emitInfo,
- GLint addrReg,
- const slang_ir_storage *src)
-{
- struct prog_instruction *inst;
+ /* Free any temp registers that we allocated above */
+ for (i = 0; i < 3; i++) {
+ if (isTemp[i])
+ _slang_free_temp(emitInfo->vt, &newSrc[i]);
+ }
- assert(addrReg == 0); /* only one addr reg at this time */
- inst = new_instruction(emitInfo, OPCODE_ARL);
- storage_to_src_reg(&inst->SrcReg[0], src);
- inst->DstReg.File = PROGRAM_ADDRESS;
- inst->DstReg.Index = addrReg;
- inst->DstReg.WriteMask = WRITEMASK_X;
return inst;
}
@@ -1220,7 +1323,8 @@ emit_copy(slang_emit_info *emitInfo, slang_ir_node *n)
if (inst &&
_slang_is_temp(emitInfo->vt, n->Children[1]->Store) &&
(inst->DstReg.File == n->Children[1]->Store->File) &&
- (inst->DstReg.Index == n->Children[1]->Store->Index)) {
+ (inst->DstReg.Index == n->Children[1]->Store->Index) &&
+ !n->Children[0]->Store->IsIndirect) {
/* Peephole optimization:
* The Right-Hand-Side has its results in a temporary place.
* Modify the RHS (and the prev instruction) to store its results
@@ -1678,80 +1782,38 @@ emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
inst = emit(emitInfo, n->Children[0]);
- /* setup storage info, if needed */
- if (!n->Store->Parent)
- n->Store->Parent = n->Children[0]->Store;
-
+#if 0
assert(n->Store->Parent);
-
+ /* Apply this node's swizzle to parent's storage */
+ GLuint swizzle = n->Store->Swizzle;
+ _slang_copy_ir_storage(n->Store, n->Store->Parent);
+ n->Store->Swizzle = _slang_swizzle_swizzle(n->Store->Swizzle, swizzle);
+ assert(!n->Store->Parent);
+#endif
return inst;
}
/**
- * Move a block registers from src to dst (or move a single register).
- * \param size size of block, in floats (<=4 means one register)
+ * Dereference array element: element == array[index]
+ * This basically involves emitting code for computing the array index
+ * and updating the node/element's storage info.
*/
static struct prog_instruction *
-move_block(slang_emit_info *emitInfo,
- GLuint size, GLboolean relAddr,
- const slang_ir_storage *dst,
- const slang_ir_storage *src)
+emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
{
+ slang_ir_storage *arrayStore, *indexStore;
+ const int elemSize = n->Store->Size; /* number of floats */
+ const GLint elemSizeVec = (elemSize + 3) / 4; /* number of vec4 */
struct prog_instruction *inst;
- if (size > 4) {
- /* move matrix/struct etc (block of registers) */
- slang_ir_storage dstStore = *dst;
- slang_ir_storage srcStore = *src;
-
- dstStore.Size = 4;
- srcStore.Size = 4;
- while (size >= 4) {
- inst = emit_instruction(emitInfo, OPCODE_MOV,
- &dstStore,
- &srcStore,
- NULL,
- NULL);
- inst->SrcReg[0].RelAddr = relAddr;
- inst_comment(inst, "IR_COPY block");
- srcStore.Index++;
- dstStore.Index++;
- size -= 4;
- }
- }
- else {
- /* single register move */
- inst = emit_instruction(emitInfo,
- OPCODE_MOV,
- dst,
- src,
- NULL,
- NULL);
- inst->SrcReg[0].RelAddr = relAddr;
- }
- return inst;
-}
-
-
-
-/**
- * Dereference array element. Just resolve storage for the array
- * element represented by this node.
- * This is typically where Indirect addressing comes into play.
- * See comments on struct slang_ir_storage.
- */
-static struct prog_instruction *
-emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
-{
assert(n->Opcode == IR_ELEMENT);
- assert(n->Store);
- assert(n->Store->File == PROGRAM_UNDEFINED);
- assert(n->Store->Parent);
- assert(n->Store->Size > 0);
+ assert(elemSize > 0);
+ /* special case for built-in state variables, like light state */
{
slang_ir_storage *root = n->Store;
+ assert(!root->Parent);
while (root->Parent)
root = root->Parent;
@@ -1762,69 +1824,98 @@ emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
}
}
- /* do codegen for array */
+ /* do codegen for array itself */
emit(emitInfo, n->Children[0]);
+ arrayStore = n->Children[0]->Store;
+
+ /* The initial array element storage is the array's storage,
+ * then modified below.
+ */
+ _slang_copy_ir_storage(n->Store, arrayStore);
+
if (n->Children[1]->Opcode == IR_FLOAT) {
- /* Constant array index.
- * Set Store's index to be the offset of the array element in
- * the register file.
- */
+ /* Constant array index */
const GLint element = (GLint) n->Children[1]->Value[0];
- const GLint sz = (n->Store->Size + 3) / 4; /* size in slots/registers */
- n->Store->Index = sz * element;
- assert(n->Store->Parent);
+ /* this element's storage is the array's storage, plus constant offset */
+ n->Store->Index += elemSizeVec * element;
}
else {
/* Variable array index */
- struct prog_instruction *inst;
/* do codegen for array index expression */
emit(emitInfo, n->Children[1]);
+ indexStore = n->Children[1]->Store;
+
+ if (indexStore->IsIndirect) {
+ /* need to put the array index into a temporary since we can't
+ * directly support a[b[i]] constructs.
+ */
+
- /* allocate temp storage for the array element */
- assert(n->Store->Index < 0);
- n->Store->File = PROGRAM_TEMPORARY;
- n->Store->Parent = NULL;
- alloc_node_storage(emitInfo, n, -1);
+ /*indexStore = tempstore();*/
+ }
- if (n->Store->Size > 4) {
- /* need to multiply the index by the element size */
- const GLint elemSize = (n->Store->Size + 3) / 4;
- slang_ir_storage indexTemp, elemSizeStore;
- /* constant containing the element size */
- constant_to_storage(emitInfo, (float) elemSize, &elemSizeStore);
+ if (elemSize > 4) {
+ /* need to multiply array index by array element size */
+ struct prog_instruction *inst;
+ slang_ir_storage *indexTemp;
+ slang_ir_storage elemSizeStore;
/* allocate 1 float indexTemp */
- alloc_local_temp(emitInfo, &indexTemp, 1);
+ indexTemp = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1);
+ _slang_alloc_temp(emitInfo->vt, indexTemp);
+
+ /* allocate a constant containing the element size */
+ constant_to_storage(emitInfo, (float) elemSizeVec, &elemSizeStore);
- /* MUL temp, index, elemSize */
- inst = emit_instruction(emitInfo, OPCODE_MUL,
- &indexTemp, /* dest */
- n->Children[1]->Store, /* the index */
+ /* multiply array index by element size */
+ inst = emit_instruction(emitInfo,
+ OPCODE_MUL,
+ indexTemp, /* dest */
+ indexStore, /* the index */
&elemSizeStore,
NULL);
- /* load ADDR[0].X = temp */
- inst = emit_arl_instruction(emitInfo, 0, &indexTemp);
-
- _slang_free_temp(emitInfo->vt, &indexTemp);
+ indexStore = indexTemp;
}
- else {
- /* simply load address reg w/ array index */
- inst = emit_arl_instruction(emitInfo, 0, n->Children[1]->Store);
+
+ if (arrayStore->IsIndirect) {
+ /* ex: in a[i][j], a[i] (the arrayStore) is indirect */
+ /* Need to add indexStore to arrayStore->Indirect store */
+ slang_ir_storage indirectArray;
+ slang_ir_storage *indexTemp;
+
+ _slang_init_ir_storage(&indirectArray,
+ arrayStore->IndirectFile,
+ arrayStore->IndirectIndex,
+ 1,
+ arrayStore->IndirectSwizzle);
+
+ /* allocate 1 float indexTemp */
+ indexTemp = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1);
+ _slang_alloc_temp(emitInfo->vt, indexTemp);
+
+ inst = emit_instruction(emitInfo,
+ OPCODE_ADD,
+ indexTemp, /* dest */
+ indexStore, /* the index */
+ &indirectArray, /* indirect array base */
+ NULL);
+
+ indexStore = indexTemp;
}
- /* copy from array element to temp storage */
- move_block(emitInfo, n->Store->Size, GL_TRUE,
- n->Store, n->Children[0]->Store);
+ /* update the array element storage info */
+ n->Store->IsIndirect = GL_TRUE;
+ n->Store->IndirectFile = indexStore->File;
+ n->Store->IndirectIndex = indexStore->Index;
+ n->Store->IndirectSwizzle = indexStore->Swizzle;
}
- /* if array element size is one, make sure we only access X */
- if (n->Store->Size == 1)
- n->Store->Swizzle = SWIZZLE_XXXX;
+ n->Store->Size = elemSize;
return NULL; /* no instruction */
}
@@ -1837,9 +1928,11 @@ static struct prog_instruction *
emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
{
slang_ir_storage *root = n->Store;
+ GLint fieldOffset, fieldSize;
assert(n->Opcode == IR_FIELD);
+ assert(!root->Parent);
while (root->Parent)
root = root->Parent;
@@ -1855,12 +1948,45 @@ emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
slang_info_log_error(emitInfo->log, "Error parsing state variable");
return NULL;
}
+ return NULL;
}
else {
/* do codegen for struct */
emit(emitInfo, n->Children[0]);
+ assert(n->Children[0]->Store->Index >= 0);
}
+ fieldOffset = n->Store->Index;
+ fieldSize = n->Store->Size;
+
+ _slang_copy_ir_storage(n->Store, n->Children[0]->Store);
+
+ n->Store->Index = n->Children[0]->Store->Index + fieldOffset / 4;
+ /* XXX test this:
+ n->Store->Index += fieldOffset / 4;
+ */
+
+ switch (fieldSize) {
+ case 1:
+ {
+ GLint swz = fieldOffset % 4;
+ n->Store->Swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz);
+ }
+ break;
+ 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:
+ n->Store->Swizzle = SWIZZLE_XYZW;
+ }
+
+ assert(n->Store->Index >= 0);
+
return NULL; /* no instruction */
}
@@ -1917,7 +2043,7 @@ emit_var_decl(slang_emit_info *emitInfo, slang_ir_node *n)
/**
* Emit code for a reference to a variable.
- * Actually, no code is generated but we may do some memory alloation.
+ * Actually, no code is generated but we may do some memory allocation.
* In particular, state vars (uniforms) are allocated on an as-needed basis.
*/
static struct prog_instruction *
@@ -2138,7 +2264,7 @@ _slang_resolve_subroutines(slang_emit_info *emitInfo)
total += emitInfo->Subroutines[i]->NumInstructions;
}
- /* adjust BrancTargets within the functions */
+ /* adjust BranchTargets within the functions */
for (i = 0; i < emitInfo->NumSubroutines; i++) {
struct gl_program *sub = emitInfo->Subroutines[i];
GLuint j;
@@ -2207,7 +2333,7 @@ _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
emitInfo.prog = prog;
emitInfo.Subroutines = NULL;
emitInfo.NumSubroutines = 0;
- emitInfo.MaxInstructions = 0;
+ emitInfo.MaxInstructions = prog->NumInstructions;
emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
emitInfo.EmitCondCodes = ctx->Shader.EmitCondCodes;
diff --git a/src/mesa/shader/slang/slang_ir.c b/src/mesa/shader/slang/slang_ir.c
index 4fe3a2b..c33c80d 100644
--- a/src/mesa/shader/slang/slang_ir.c
+++ b/src/mesa/shader/slang/slang_ir.c
@@ -116,6 +116,20 @@ _slang_ir_info(slang_ir_opcode opcode)
}
+void
+_slang_init_ir_storage(slang_ir_storage *st,
+ enum register_file file, GLint index, GLint size,
+ GLuint swizzle)
+{
+ st->File = file;
+ st->Index = index;
+ st->Size = size;
+ st->Swizzle = swizzle;
+ st->Parent = NULL;
+ st->IsIndirect = GL_FALSE;
+}
+
+
/**
* Return a new slang_ir_storage object.
*/
@@ -130,6 +144,7 @@ _slang_new_ir_storage(enum register_file file, GLint index, GLint size)
st->Size = size;
st->Swizzle = SWIZZLE_NOOP;
st->Parent = NULL;
+ st->IsIndirect = GL_FALSE;
}
return st;
}
@@ -150,6 +165,7 @@ _slang_new_ir_storage_swz(enum register_file file, GLint index, GLint size,
st->Size = size;
st->Swizzle = swizzle;
st->Parent = NULL;
+ st->IsIndirect = GL_FALSE;
}
return st;
}
@@ -170,11 +186,45 @@ _slang_new_ir_storage_relative(GLint index, GLint size,
st->Size = size;
st->Swizzle = SWIZZLE_NOOP;
st->Parent = parent;
+ st->IsIndirect = GL_FALSE;
+ }
+ return st;
+}
+
+
+slang_ir_storage *
+_slang_new_ir_storage_indirect(enum register_file file,
+ GLint index,
+ GLint size,
+ enum register_file indirectFile,
+ GLint indirectIndex,
+ GLuint indirectSwizzle)
+{
+ slang_ir_storage *st;
+ st = (slang_ir_storage *) _slang_alloc(sizeof(slang_ir_storage));
+ if (st) {
+ st->File = file;
+ st->Index = index;
+ st->Size = size;
+ st->Swizzle = SWIZZLE_NOOP;
+ st->IsIndirect = GL_TRUE;
+ st->IndirectFile = indirectFile;
+ st->IndirectIndex = indirectIndex;
+ st->IndirectSwizzle = indirectSwizzle;
}
return st;
}
+/* XXX temporary function */
+void
+_slang_copy_ir_storage(slang_ir_storage *dst, const slang_ir_storage *src)
+{
+ *dst = *src;
+ dst->Parent = NULL;
+}
+
+
static const char *
_slang_ir_name(slang_ir_opcode opcode)
diff --git a/src/mesa/shader/slang/slang_ir.h b/src/mesa/shader/slang/slang_ir.h
index a4552ae..a258e92 100644
--- a/src/mesa/shader/slang/slang_ir.h
+++ b/src/mesa/shader/slang/slang_ir.h
@@ -140,24 +140,53 @@ typedef enum
/**
- * Describes where data storage is allocated.
+ * Describes where data/variables are stored in the various register files.
+ *
+ * In the simple case, the File, Index and Size fields indicate where
+ * a variable is stored. For example, a vec3 variable may be stored
+ * as (File=PROGRAM_TEMPORARY, Index=6, Size=3). Or, File[Index].
+ * Or, a program input like color may be stored as
+ * (File=PROGRAM_INPUT,Index=3,Size=4);
+ *
+ * For single-float values, the Swizzle field indicates which component
+ * of the vector contains the float.
+ *
+ * If IsIndirect is set, the storage is accessed through an indirect
+ * register lookup. The value in question will be located at:
+ * File[Index + IndirectFile[IndirectIndex]]
+ *
+ * This is primary used for indexing arrays. For example, consider this
+ * GLSL code:
+ * uniform int i;
+ * float a[10];
+ * float x = a[i];
+ *
+ * here, storage for a[i] would be described by (File=PROGRAM_TEMPORAY,
+ * Index=aPos, IndirectFile=PROGRAM_UNIFORM, IndirectIndex=iPos), which
+ * would mean TEMP[aPos + UNIFORM[iPos]]
*/
-struct _slang_ir_storage
+struct slang_ir_storage_
{
enum register_file File; /**< PROGRAM_TEMPORARY, PROGRAM_INPUT, etc */
GLint Index; /**< -1 means unallocated */
GLint Size; /**< number of floats */
- GLuint Swizzle;
+ GLuint Swizzle; /**< Swizzle AND writemask info */
GLint RefCount; /**< Used during IR tree delete */
- GLboolean RelAddr;
+
+ GLboolean RelAddr; /* we'll remove this eventually */
+
+ GLboolean IsIndirect;
+ enum register_file IndirectFile;
+ GLint IndirectIndex;
+ GLuint IndirectSwizzle;
/** If Parent is non-null, Index is relative to parent.
* The other fields are ignored.
*/
- struct _slang_ir_storage *Parent;
+ struct slang_ir_storage_ *Parent;
};
-typedef struct _slang_ir_storage slang_ir_storage;
+typedef struct slang_ir_storage_ slang_ir_storage;
/**
@@ -199,6 +228,11 @@ extern const slang_ir_info *
_slang_ir_info(slang_ir_opcode opcode);
+extern void
+_slang_init_ir_storage(slang_ir_storage *st,
+ enum register_file file, GLint index, GLint size,
+ GLuint swizzle);
+
extern slang_ir_storage *
_slang_new_ir_storage(enum register_file file, GLint index, GLint size);
@@ -212,6 +246,17 @@ _slang_new_ir_storage_relative(GLint index, GLint size,
slang_ir_storage *parent);
+extern slang_ir_storage *
+_slang_new_ir_storage_indirect(enum register_file file,
+ GLint index,
+ GLint size,
+ enum register_file indirectFile,
+ GLint indirectIndex,
+ GLuint indirectSwizzle);
+
+extern void
+_slang_copy_ir_storage(slang_ir_storage *dst, const slang_ir_storage *src);
+
extern void
_slang_free_ir_tree(slang_ir_node *n);
diff --git a/src/mesa/shader/slang/slang_vartable.h b/src/mesa/shader/slang/slang_vartable.h
index 8a3b992..94bcd63 100644
--- a/src/mesa/shader/slang/slang_vartable.h
+++ b/src/mesa/shader/slang/slang_vartable.h
@@ -2,7 +2,7 @@
#ifndef SLANG_VARTABLE_H
#define SLANG_VARTABLE_H
-struct _slang_ir_storage;
+struct slang_ir_storage_;
typedef struct slang_var_table_ slang_var_table;
@@ -27,16 +27,16 @@ extern struct slang_variable_ *
_slang_find_variable(const slang_var_table *t, slang_atom name);
extern GLboolean
-_slang_alloc_var(slang_var_table *t, struct _slang_ir_storage *store);
+_slang_alloc_var(slang_var_table *t, struct slang_ir_storage_ *store);
extern GLboolean
-_slang_alloc_temp(slang_var_table *t, struct _slang_ir_storage *store);
+_slang_alloc_temp(slang_var_table *t, struct slang_ir_storage_ *store);
extern void
-_slang_free_temp(slang_var_table *t, struct _slang_ir_storage *store);
+_slang_free_temp(slang_var_table *t, struct slang_ir_storage_ *store);
extern GLboolean
-_slang_is_temp(const slang_var_table *t, const struct _slang_ir_storage *store);
+_slang_is_temp(const slang_var_table *t, const struct slang_ir_storage_ *store);
#endif /* SLANG_VARTABLE_H */
More information about the mesa-commit
mailing list