st/glsl_to_tgsi: reduce stack explosion in recursive expression visitor

In optimized builds, visit(ir_expression *) experiences inlining with gcc that
leads the function to have a roughly 32KB stack frame. This is a problem given
that the function is called recursively. In non-optimized builds, the stack
frame is much smaller, hence one gets crashes that happen only in optimized
builds.

Arguably there is a compiler bug or at least severe misfeature here. In any
case, the easy thing to do for now seems to be moving the bulk of the
non-recursive code into a separate function. This is sufficient to convince my
version of gcc not to blow up the stack frame of the recursive part. Just to be
sure, add the gcc-specific noinline attribute to prevent this bug from
reoccuring if inliner heuristics change.

v2: put ATTRIBUTE_NOINLINE into macros.h

Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95133
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95026
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92850
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robdclark@gmail.com>
This commit is contained in:
Nicolai Hähnle
2016-04-25 18:20:50 -05:00
parent 59af21c3e9
commit 98c348d26b
2 changed files with 22 additions and 4 deletions

View File

@@ -450,6 +450,8 @@ public:
virtual void visit(ir_barrier *); virtual void visit(ir_barrier *);
/*@}*/ /*@}*/
void visit_expression(ir_expression *, st_src_reg *) ATTRIBUTE_NOINLINE;
void visit_atomic_counter_intrinsic(ir_call *); void visit_atomic_counter_intrinsic(ir_call *);
void visit_ssbo_intrinsic(ir_call *); void visit_ssbo_intrinsic(ir_call *);
void visit_membar_intrinsic(ir_call *); void visit_membar_intrinsic(ir_call *);
@@ -1535,10 +1537,7 @@ glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction *ir,
void void
glsl_to_tgsi_visitor::visit(ir_expression *ir) glsl_to_tgsi_visitor::visit(ir_expression *ir)
{ {
unsigned int operand;
st_src_reg op[ARRAY_SIZE(ir->operands)]; st_src_reg op[ARRAY_SIZE(ir->operands)];
st_src_reg result_src;
st_dst_reg result_dst;
/* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c) /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
*/ */
@@ -1561,7 +1560,7 @@ glsl_to_tgsi_visitor::visit(ir_expression *ir)
if (ir->operation == ir_quadop_vector) if (ir->operation == ir_quadop_vector)
assert(!"ir_quadop_vector should have been lowered"); assert(!"ir_quadop_vector should have been lowered");
for (operand = 0; operand < ir->get_num_operands(); operand++) { for (unsigned int operand = 0; operand < ir->get_num_operands(); operand++) {
this->result.file = PROGRAM_UNDEFINED; this->result.file = PROGRAM_UNDEFINED;
ir->operands[operand]->accept(this); ir->operands[operand]->accept(this);
if (this->result.file == PROGRAM_UNDEFINED) { if (this->result.file == PROGRAM_UNDEFINED) {
@@ -1578,6 +1577,19 @@ glsl_to_tgsi_visitor::visit(ir_expression *ir)
assert(!ir->operands[operand]->type->is_matrix()); assert(!ir->operands[operand]->type->is_matrix());
} }
visit_expression(ir, op);
}
/* The non-recursive part of the expression visitor lives in a separate
* function and should be prevented from being inlined, to avoid a stack
* explosion when deeply nested expressions are visited.
*/
void
glsl_to_tgsi_visitor::visit_expression(ir_expression* ir, st_src_reg *op)
{
st_src_reg result_src;
st_dst_reg result_dst;
int vector_elements = ir->operands[0]->type->vector_elements; int vector_elements = ir->operands[0]->type->vector_elements;
if (ir->operands[1]) { if (ir->operands[1]) {
vector_elements = MAX2(vector_elements, vector_elements = MAX2(vector_elements,

View File

@@ -214,6 +214,12 @@ do { \
#define MUST_CHECK #define MUST_CHECK
#endif #endif
#if defined(__GNUC__) || (defined(__SUNPRO_C) && (__SUNPRO_C >= 0x590))
#define ATTRIBUTE_NOINLINE __attribute__((noinline))
#else
#define ATTRIBUTE_NOINLINE
#endif
/** Compute ceiling of integer quotient of A divided by B. */ /** Compute ceiling of integer quotient of A divided by B. */
#define DIV_ROUND_UP( A, B ) ( (A) % (B) == 0 ? (A)/(B) : (A)/(B)+1 ) #define DIV_ROUND_UP( A, B ) ( (A) % (B) == 0 ? (A)/(B) : (A)/(B)+1 )