RISC-V: Add support for the Zfa extension

This patch adds support for the RISC-V Zfa extension,
which introduces additional floating-point instructions:
* fli (load-immediate) with pre-defined immediates
* fminm/fmaxm (like fmin/fmax but with different NaN behaviour)
* fround/froundmx (round to integer)
* fcvtmod.w.d (Modular Convert-to-Integer)
* fmv* to access high bits of FP registers in case XLEN < FLEN
* fleq/fltq (quiet comparison instructions)

Zfa defines its instructions in combination with the following
extensions:
* single-precision floating-point (F)
* double-precision floating-point (D)
* quad-precision floating-point (Q)
* half-precision floating-point (Zfh)

This patch is based on an earlier version from Tsukasa OI:
  https://sourceware.org/pipermail/binutils/2022-September/122939.html
Most significant change to that commit is the switch from the rs1-field
value to the actual floating-point value in the last operand of the fli*
instructions. Everything that strtof() can parse is accepted and
the '%a' printf specifier is used to output hex floating-point literals
in the disassembly.

The Zfa specification is frozen (and has passed public review).  It is
available as a chapter in "The RISC-V Instruction Set Manual: Volume 1":
  https://github.com/riscv/riscv-isa-manual/releases

bfd/ChangeLog:

	* elfxx-riscv.c (riscv_multi_subset_supports): Add instruction
	class support for 'Zfa' extension.
	(riscv_multi_subset_supports_ext): Likewise.
	(riscv_implicit_subsets): Add 'Zfa' -> 'F' dependency.

gas/ChangeLog:

	* config/tc-riscv.c (flt_lookup): New helper to lookup a float value
	in an array.
	(validate_riscv_insn): Add 'Wfv' as new format string directive.
	(riscv_ip): Likewise.
	* doc/c-riscv.texi: Add floating-point chapter and describe
	limiations of the Zfa FP literal parsing.
	* testsuite/gas/riscv/zfa-32.d: New test.
	* testsuite/gas/riscv/zfa-32.s: New test.
	* testsuite/gas/riscv/zfa-64.d: New test.
	* testsuite/gas/riscv/zfa-64.s: New test.
	* testsuite/gas/riscv/zfa-fail.d: New test.
	* testsuite/gas/riscv/zfa-fail.l: New test.
	* testsuite/gas/riscv/zfa-fail.s: New test.
	* testsuite/gas/riscv/zfa.d: New test.
	* testsuite/gas/riscv/zfa.s: New test.
	* testsuite/gas/riscv/zfa.s: New test.

	* opcode/riscv-opc.h (MATCH_FLI_H): New.
	(MASK_FLI_H): New.
	(MATCH_FMINM_H): New.
	(MASK_FMINM_H): New.
	(MATCH_FMAXM_H): New.
	(MASK_FMAXM_H): New.
	(MATCH_FROUND_H): New.
	(MASK_FROUND_H): New.
	(MATCH_FROUNDNX_H): New.
	(MASK_FROUNDNX_H): New.
	(MATCH_FLTQ_H): New.
	(MASK_FLTQ_H): New.
	(MATCH_FLEQ_H): New.
	(MASK_FLEQ_H): New.
	(MATCH_FLI_S): New.
	(MASK_FLI_S): New.
	(MATCH_FMINM_S): New.
	(MASK_FMINM_S): New.
	(MATCH_FMAXM_S): New.
	(MASK_FMAXM_S): New.
	(MATCH_FROUND_S): New.
	(MASK_FROUND_S): New.
	(MATCH_FROUNDNX_S): New.
	(MASK_FROUNDNX_S): New.
	(MATCH_FLTQ_S): New.
	(MASK_FLTQ_S): New.
	(MATCH_FLEQ_S): New.
	(MASK_FLEQ_S): New.
	(MATCH_FLI_D): New.
	(MASK_FLI_D): New.
	(MATCH_FMINM_D): New.
	(MASK_FMINM_D): New.
	(MATCH_FMAXM_D): New.
	(MASK_FMAXM_D): New.
	(MATCH_FROUND_D): New.
	(MASK_FROUND_D): New.
	(MATCH_FROUNDNX_D): New.
	(MASK_FROUNDNX_D): New.
	(MATCH_FLTQ_D): New.
	(MASK_FLTQ_D): New.
	(MATCH_FLEQ_D): New.
	(MASK_FLEQ_D): New.
	(MATCH_FLI_Q): New.
	(MASK_FLI_Q): New.
	(MATCH_FMINM_Q): New.
	(MASK_FMINM_Q): New.
	(MATCH_FMAXM_Q): New.
	(MASK_FMAXM_Q): New.
	(MATCH_FROUND_Q): New.
	(MASK_FROUND_Q): New.
	(MATCH_FROUNDNX_Q): New.
	(MASK_FROUNDNX_Q): New.
	(MATCH_FLTQ_Q): New.
	(MASK_FLTQ_Q): New.
	(MATCH_FLEQ_Q): New.
	(MASK_FLEQ_Q): New.
	(MATCH_FCVTMOD_W_D): New.
	(MASK_FCVTMOD_W_D): New.
	(MATCH_FMVH_X_D): New.
	(MASK_FMVH_X_D): New.
	(MATCH_FMVH_X_Q): New.
	(MASK_FMVH_X_Q): New.
	(MATCH_FMVP_D_X): New.
	(MASK_FMVP_D_X): New.
	(MATCH_FMVP_Q_X): New.
	(MASK_FMVP_Q_X): New.
	(DECLARE_INSN): New.
	* opcode/riscv.h (enum riscv_insn_class): Add instruction
	classes for the Zfa extension.

opcodes/ChangeLog:

	* riscv-dis.c (print_insn_args): Add support for
	new format string directive 'Wfv'.
	* riscv-opc.c: Add Zfa instructions.

Co-Developed-by: Tsukasa OI <research_trasio@irq.a4lg.com>
Signed-off-by: Christoph Müllner <christoph.muellner@vrull.eu>
Co-Developed-by: Philipp Tomsich <philipp.tomsich@vrull.eu>
Acked-by: Palmer Dabbelt <palmer@rivosinc.com>

1 files changed, 41 insertions, 0 deletions

diff --git a/gas/doc/c-riscv.texi b/gas/doc/c-riscv.texi
index 98d5d1b..b175ba0 100644
--- a/gas/doc/c-riscv.texi
+++ b/gas/doc/c-riscv.texi
@@ -18,6 +18,7 @@
 * RISC-V-Options::        RISC-V Options
 * RISC-V-Directives::     RISC-V Directives
 * RISC-V-Modifiers::      RISC-V Assembler Modifiers
+* RISC-V-Floating-Point:: RISC-V Floating Point
 * RISC-V-Formats::        RISC-V Instruction Formats
 * RISC-V-ATTRIBUTE::      RISC-V Object Attribute
 * RISC-V-CustomExts::     RISC-V Custom (Vendor-Defined) Extensions
@@ -382,6 +383,46 @@ The pseudo la.tls.gd instruction can be expended to
 
 @end table
 
+@node RISC-V-Floating-Point
+@section RISC-V Floating Point
+@cindex floating point, risc-v (@sc{ieee})
+@cindex RISC-V floating point (@sc{ieee})
+
+The RISC-V architecture uses @sc{ieee} floating-point numbers.
+
+The RISC-V Zfa extension includes a load-immediate instruction
+for floating-point registers, which allows specifying the immediate
+(from a pool of 32 predefined values defined in the specification)
+as operand.
+E.g. to load the value @code{0.0625} as single-precision FP value into
+the FP register @code{ft1} one of the following instructions can be used:
+
+	fli.s ft1, 0.0625 # dec floating-point literal
+	fli.s ft1, 0x1p-4 # hex floating-point literal
+	fli.s ft1, 0x0.8p-3
+	fli.s ft1, 0x1.0p-4
+	fli.s ft1, 0x2p-5
+	fli.s ft1, 0x4p-6
+	...
+
+As can be seen, many valid ways exist to express a floating-point value.
+This is realized by parsing the value operand using strtof() and
+comparing the parsed value against built-in float-constants that
+are written as hex floating-point literals.
+
+This approach works on all machines that use IEEE 754.
+However, there is a chance that this fails on other machines
+with the following error message:
+
+	Error: improper fli value operand
+	Error: illegal operands `fli.s ft1,0.0625
+
+The error indicates that parsing @samp{0x1p-4} and @samp{0.0625}
+to single-precision floating point numbers will not result
+in two equal values on that machine.
+
+If you encounter this problem, then please report it.
+
 @node RISC-V-Formats
 @section RISC-V Instruction Formats
 @cindex instruction formats, risc-v