// CHECK-NEXT:      shape: [ 5 ],
// CHECK-NEXT:      buffer: 3,
// CHECK-NEXT:      name: "arith.constant",
// CHECK-NEXT:      quantization: {
// CHECK-EMPTY:
// CHECK-NEXT:      },
// CHECK-NEXT:      has_rank: true
// CHECK-NEXT:    }, {
// CHECK-NEXT:      shape: [ 5, 3 ],
// CHECK-NEXT:      buffer: 4,
// CHECK-NEXT:      name: "arith.constant1",
// CHECK-NEXT:      quantization: {
// CHECK-EMPTY:
// CHECK-NEXT:      },

#include "textflag.h"

// Register definitions
#define Rtable	R0	// Pointer to MD5 constants table
#define Rdata	R1	// Pointer to data to hash
#define Ra	R2	// MD5 accumulator
#define Rb	R3	// MD5 accumulator
#define Rc	R4	// MD5 accumulator
#define Rd	R5	// MD5 accumulator
#define Rc0	R6	// MD5 constant
#define Rc1	R7	// MD5 constant
#define Rc2	R8	// MD5 constant
// r9, r10 are forbidden

  }
};

// Inserts a "quant.qcast" and "quant.dcast" op pair (QDQs) in place of the
// tf.FakeQyantWithMinMax{Vars|VarsPerChannel|Args}Op
// before the op being constant folded. Since the constant
// folding logic will use a "arith.constant" op to replace the
// "tf.FakeQuantWithMinMaxVarsOp", the "quant.qcast" op is used to preserve
// the quantization parameters as a TypeAttr and "quant.dcast" op used to

    if (dim.size() < 0) {
      return errors::Internal("Unknown dimension size in constant tensor ",
                              n->name());
    }
    total_elements *= dim.size();
  }

  // TODO(sanjoy): It may make sense to combine this threshold with XLA's "large
  // constant" threshold, if there is one.
  const int kSmallTensorThreshold = 16;
  return total_elements < kSmallTensorThreshold;
}

					cause = check.sprintf("constant %s overflows %s (in %s)", x.val, u, T)
				} else {
					cause = check.sprintf("cannot convert %s to type %s (in %s)", x, u, T)
				}
				return false
			}
			return true
		})
		x.mode = value // type parameters are not constants
	case x.convertibleTo(check, T, &cause):
		// non-constant conversion
		ok = true
		x.mode = value
	}

	if !ok {

// This code was written in Go based on the original C implementation.

// This is a tool needed to generate the appropriate constants needed for
// the vpmsum algorithm.  It is included to generate new constant tables if
// new polynomial values are included in the future.

package main

import (
	"bytes"
	"fmt"
	"os"
)

var blocking = 32 * 1024

		return constant.MakeFloat64(f)
	}
	return nil
}

func fitsFloat64(x constant.Value) bool {
	f, _ := constant.Float64Val(x)
	return !math.IsInf(f, 0)
}

func roundFloat64(x constant.Value) constant.Value {
	f, _ := constant.Float64Val(x)
	if !math.IsInf(f, 0) {
		return constant.MakeFloat64(f)
	}
	return nil
}

// representable checks that a constant operand is representable in the given

// RUN: tf-opt %s -unfold-large-splat-constant | FileCheck %s

// CHECK-LABEL: @unfold_large_constant_splat
func.func @unfold_large_constant_splat() -> (tensor<10x10xf32>, tensor<1000x1000xf32>) {
  %0 = arith.constant dense<0.00000e+00> : tensor<10x10xf32>
  %1 = arith.constant dense<1.00000e+00> : tensor<1000x1000xf32>
  func.return %0, %1 : tensor<10x10xf32>, tensor<1000x1000xf32>

  // CHECK-DAG: %cst = arith.constant dense<0.000000e+00> : tensor<10x10xf32>

## Adding New Feature Constants

For consistency, features must be registered in [features.yaml](features.yaml), or your test will fail.  Each entry in this file will be equivalent to a dot delimited feature label.  For instance:

```yaml
    usability:

// to compute types and values, but in addition to Check, as these
// functions ignore the context in which an expression is used (e.g., an
// assignment). Thus, top-level untyped constants will return an
// untyped type rather than the respective context-specific type.
func CheckExpr(fset *token.FileSet, pkg *Package, pos token.Pos, expr ast.Expr, info *Info) (err error) {
	// determine scope
	var scope *Scope