floor_riscv64.s raw

   1  // Copyright 2024 The Go Authors. All rights reserved.
   2  // Use of this source code is governed by a BSD-style
   3  // license that can be found in the LICENSE file.
   4  
   5  #include "textflag.h"
   6  
   7  // RISC-V offered floating-point (FP) rounding by FP conversion instructions (FCVT)
   8  // with rounding mode field.
   9  // As Go spec expects FP rounding result in FP, we have to use FCVT integer
  10  // back to FP (fp -> int -> fp).
  11  // RISC-V only set Inexact flag during invalid FP-integer conversion without changing any data,
  12  // on the other hand, RISC-V sets out of integer represent range yet valid FP into NaN.
  13  // When it comes to integer-FP conversion, invalid FP like NaN, +-Inf will be
  14  // converted into the closest valid FP, for example:
  15  //
  16  // `Floor(-Inf) -> int64(0x7fffffffffffffff) -> float64(9.22e+18)`
  17  // `Floor(18446744073709549568.0) -> int64(0x7fffffffffffffff) -> float64(9.22e+18)`
  18  //
  19  // This ISA conversion limitation requires we skip all invalid or out of range FP
  20  // before any normal rounding operations.
  21  
  22  #define ROUNDFN(NAME, MODE) 	\
  23  TEXT NAME(SB),NOSPLIT,$0; 	\
  24  	MOVD	x+0(FP), F10; 	\
  25  	FMVXD	F10, X10;	\
  26  	/* Drop all fraction bits */;\
  27  	SRL	$52, X10, X12;	\
  28  	/* Remove sign bit */;	\
  29  	AND	$0x7FF, X12, X12;\
  30  	/* Return either input is +-Inf, NaN(0x7FF) or out of precision limitation */;\
  31  	/* 1023: bias of exponent, [-2^53, 2^53]: exactly integer represent range */;\
  32  	MOV	$1023+53, X11;	\
  33  	BLTU	X11, X12, 4(PC);\
  34  	FCVTLD.MODE F10, X11;	\
  35  	FCVTDL	X11, F11;	\
  36  	/* RISC-V rounds negative values to +0, restore original sign */;\
  37  	FSGNJD	F10, F11, F10;	\
  38  	MOVD	F10, ret+8(FP); \
  39  	RET
  40  
  41  // func archFloor(x float64) float64
  42  ROUNDFN(·archFloor, RDN)
  43  
  44  // func archCeil(x float64) float64
  45  ROUNDFN(·archCeil, RUP)
  46  
  47  // func archTrunc(x float64) float64
  48  ROUNDFN(·archTrunc, RTZ)
  49