From 4d6d219b58d43f6e14a6f9d7e49a00d437d65343 Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Thu, 23 Jun 2016 16:57:23 -0700 Subject: [PATCH] glsl: Add lowering pass for ir_triop_bitfield_extract Signed-off-by: Ian Romanick Reviewed-by: Matt Turner --- src/compiler/glsl/ir_optimization.h | 1 + src/compiler/glsl/lower_instructions.cpp | 80 ++++++++++++++++++++++++++++++++ 2 files changed, 81 insertions(+) diff --git a/src/compiler/glsl/ir_optimization.h b/src/compiler/glsl/ir_optimization.h index c856003..9550052 100644 --- a/src/compiler/glsl/ir_optimization.h +++ b/src/compiler/glsl/ir_optimization.h @@ -43,6 +43,7 @@ #define DOPS_TO_DFRAC 0x800 #define DFREXP_DLDEXP_TO_ARITH 0x1000 #define BIT_COUNT_TO_MATH 0x02000 +#define EXTRACT_TO_SHIFTS 0x04000 /** * \see class lower_packing_builtins_visitor diff --git a/src/compiler/glsl/lower_instructions.cpp b/src/compiler/glsl/lower_instructions.cpp index fc87605..0d103d2 100644 --- a/src/compiler/glsl/lower_instructions.cpp +++ b/src/compiler/glsl/lower_instructions.cpp @@ -160,6 +160,7 @@ private: void dtrunc_to_dfrac(ir_expression *); void dsign_to_csel(ir_expression *); void bit_count_to_math(ir_expression *); + void extract_to_shifts(ir_expression *); }; } /* anonymous namespace */ @@ -1001,6 +1002,80 @@ lower_instructions_visitor::bit_count_to_math(ir_expression *ir) this->progress = true; } +void +lower_instructions_visitor::extract_to_shifts(ir_expression *ir) +{ + ir_variable *bits = + new(ir) ir_variable(ir->operands[0]->type, "bits", ir_var_temporary); + + base_ir->insert_before(bits); + base_ir->insert_before(assign(bits, ir->operands[2])); + + if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) { + ir_constant *c1 = + new(ir) ir_constant(1u, ir->operands[0]->type->vector_elements); + ir_constant *c32 = + new(ir) ir_constant(32u, ir->operands[0]->type->vector_elements); + ir_constant *cFFFFFFFF = + new(ir) ir_constant(0xFFFFFFFFu, ir->operands[0]->type->vector_elements); + + /* At least some hardware treats (x << y) as (x << (y%32)). This means + * we'd get a mask of 0 when bits is 32. Special case it. + * + * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u; + */ + ir_expression *mask = csel(equal(bits, c32), + cFFFFFFFF, + sub(lshift(c1, bits), c1->clone(ir, NULL))); + + /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says: + * + * If bits is zero, the result will be zero. + * + * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional + * select as in the signed integer case. + * + * (value >> offset) & mask; + */ + ir->operation = ir_binop_bit_and; + ir->operands[0] = rshift(ir->operands[0], ir->operands[1]); + ir->operands[1] = mask; + ir->operands[2] = NULL; + } else { + ir_constant *c0 = + new(ir) ir_constant(int(0), ir->operands[0]->type->vector_elements); + ir_constant *c32 = + new(ir) ir_constant(int(32), ir->operands[0]->type->vector_elements); + ir_variable *temp = + new(ir) ir_variable(ir->operands[0]->type, "temp", ir_var_temporary); + + /* temp = 32 - bits; */ + base_ir->insert_before(temp); + base_ir->insert_before(assign(temp, sub(c32, bits))); + + /* expr = value << (temp - offset)) >> temp; */ + ir_expression *expr = + rshift(lshift(ir->operands[0], sub(temp, ir->operands[1])), temp); + + /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says: + * + * If bits is zero, the result will be zero. + * + * Due to the (x << (y%32)) behavior mentioned before, the (value << + * (32-0)) doesn't "erase" all of the data as we would like, so finish + * up with: + * + * (bits == 0) ? 0 : e; + */ + ir->operation = ir_triop_csel; + ir->operands[0] = equal(c0, bits); + ir->operands[1] = c0->clone(ir, NULL); + ir->operands[2] = expr; + } + + this->progress = true; +} + ir_visitor_status lower_instructions_visitor::visit_leave(ir_expression *ir) { @@ -1108,6 +1183,11 @@ lower_instructions_visitor::visit_leave(ir_expression *ir) bit_count_to_math(ir); break; + case ir_triop_bitfield_extract: + if (lowering(EXTRACT_TO_SHIFTS)) + extract_to_shifts(ir); + break; + default: return visit_continue; } -- 2.7.4