[Mesa-dev] [PATCH 41/50] glsl: Add a lowering pass for 64-bit float min()

[Dave Airlie]

From: Elie Tournier <tournier.elie at gmail.com>

[airlied: update to handle min(dvec, double) case.

Signed-off-by: Elie Tournier <elie.tournier at collabora.com>
---
 src/compiler/glsl/ir_optimization.h      |  1 +
 src/compiler/glsl/lower_instructions.cpp | 33 ++++++++++++++++++++++++++++++++
 2 files changed, 34 insertions(+)

diff --git a/src/compiler/glsl/ir_optimization.h b/src/compiler/glsl/ir_optimization.h
index 3c99ae0..f9b688a 100644
--- a/src/compiler/glsl/ir_optimization.h
+++ b/src/compiler/glsl/ir_optimization.h
@@ -54,6 +54,7 @@
 #define DDIV_TO_MUL_RCP           0x100000
 #define DIV_TO_MUL_RCP            (FDIV_TO_MUL_RCP | DDIV_TO_MUL_RCP)
 #define SQRT_TO_ABS_SQRT          0x200000
+#define DMIN_DMAX_TO_LESS         0x400000
 
 /* Opertaions for lower_64bit_integer_instructions() */
 #define MUL64                     (1U << 0)
diff --git a/src/compiler/glsl/lower_instructions.cpp b/src/compiler/glsl/lower_instructions.cpp
index 91f71b3..8c3d623 100644
--- a/src/compiler/glsl/lower_instructions.cpp
+++ b/src/compiler/glsl/lower_instructions.cpp
@@ -43,6 +43,7 @@
  * - BORROW_TO_ARITH
  * - SAT_TO_CLAMP
  * - DOPS_TO_DFRAC
+ * - DMIN_DMAX_TO_LESS
  *
  * SUB_TO_ADD_NEG:
  * ---------------
@@ -115,6 +116,12 @@
  * DOPS_TO_DFRAC:
  * --------------
  * Converts double trunc, ceil, floor, round to fract
+ *
+ * DMIN_DMAX_TO_LESS:
+ * ----------------
+ * Converts double min, max into less.
+ * min(x,y) = less(x,y) ? x, y;
+ * max(x,y) = less(x,y) ? y, x;
  */
 
 #include "c99_math.h"
@@ -169,6 +176,7 @@ private:
    void find_msb_to_float_cast(ir_expression *ir);
    void imul_high_to_mul(ir_expression *ir);
    void sqrt_to_abs_sqrt(ir_expression *ir);
+   void dmin_to_less(ir_expression *ir);
 
    ir_expression *_carry(operand a, operand b);
 };
@@ -1666,6 +1674,25 @@ lower_instructions_visitor::sqrt_to_abs_sqrt(ir_expression *ir)
    this->progress = true;
 }
 
+void
+lower_instructions_visitor::dmin_to_less(ir_expression *ir)
+{
+   const unsigned vec_elem = ir->type->vector_elements;
+   ir_rvalue *x_clone = ir->operands[0]->clone(ir, NULL);
+   ir_rvalue *y_clone = ir->operands[1]->clone(ir, NULL);
+   ir->operation = ir_triop_csel;
+   ir->init_num_operands();
+   if (ir->operands[1]->type->vector_elements == 1 && vec_elem > 1) {
+	   ir->operands[0] = less(ir->operands[0], swizzle(ir->operands[1], SWIZZLE_XXXX, vec_elem));
+	   ir->operands[2] = swizzle(y_clone, SWIZZLE_XXXX, vec_elem);
+   } else {
+	   ir->operands[0] = less(ir->operands[0], ir->operands[1]);
+	   ir->operands[2] = y_clone;
+   }
+   ir->operands[1] = x_clone;
+   this->progress = true;
+}
+
 ir_visitor_status
 lower_instructions_visitor::visit_leave(ir_expression *ir)
 {
@@ -1809,6 +1836,12 @@ lower_instructions_visitor::visit_leave(ir_expression *ir)
          sqrt_to_abs_sqrt(ir);
       break;
 
+   case ir_binop_min:
+      if (lowering(DMIN_DMAX_TO_LESS) &&
+          ir->type->is_double())
+         dmin_to_less(ir);
+      break;
+
    default:
       return visit_continue;
    }
-- 
2.9.5