RISC-V: Introduce the -mlong-double-{64,128} option

Similarly to some other targets, this patch adds the ability to
change the size of long double.  This is to support the ABI deviation
permitting long double to be 64-bit wide.

Add a testcase checking that, whenever -mlong-double-64 is given,
64-bit double-precision instructions are used instead of calls to
libgcc's 128-bit softfloat functions.

FIXME: This patch is incomplete and unsafe until a mechanism is added
that would propagate this option to the assembler and linker to ensure
that 64-bit long double and 128-bit long double code can never be linked
together.

Signed-off-by: Artemiy Volkov <artemiy@synopsys.com>

gcc/config/riscv/riscv.h

@@ -190,7 +190,7 @@ ASM_MISA_SPEC
 
 #define FLOAT_TYPE_SIZE 32
 #define DOUBLE_TYPE_SIZE 64
-#define LONG_DOUBLE_TYPE_SIZE 128
+#define LONG_DOUBLE_TYPE_SIZE (TARGET_LONG_DOUBLE_64 ? 64 : 128)
 
 /* Allocation boundary (in *bits*) for storing arguments in argument list.  */
 #define PARM_BOUNDARY BITS_PER_WORD

gcc/config/riscv/riscv.opt

@@ -637,3 +637,11 @@ Enum(arcv_mpy_option) String(10c) Value(ARCV_MPY_OPTION_10C)
 -param=arcv-mpy-option=
 Target RejectNegative Joined Enum(arcv_mpy_option) Var(arcv_mpy_option) Init(ARCV_MPY_OPTION_2C)
 The type of MPY unit used by the RMX-100 core (to be used in combination with -mtune=rmx100) (default: 2c).
+
+mlong-double-64
+Target RejectNegative Negative(mlong-double-128) Mask(LONG_DOUBLE_64)
+Use 64-bit long double.
+
+mlong-double-128
+Target RejectNegative Negative(mlong-double-64) InverseMask(LONG_DOUBLE_64)
+Use 128-bit long double.

gcc/testsuite/gcc.target/riscv/arcv-long-double-64.c

@@ -0,0 +1,17 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target rv32 } */
+/* { dg-options "-march=rv32gc -mabi=ilp32d -mlong-double-64" } */
+
+int __attribute__((noinline))
+g (long double ld)
+{
+  return (int)ld;
+}
+
+int
+f (int x)
+{
+  return (long double) 3.141592;
+}
+
+/* { dg-final { scan-assembler-times "__fixtfsi" 0 } } */