+@node MIPS FP ABI History
+@subsection History of FP ABIs
+@cindex @code{.gnu_attribute 4, @var{n}} directive, MIPS
+@cindex @code{.gnu_attribute Tag_GNU_MIPS_ABI_FP, @var{n}} directive, MIPS
+The MIPS ABIs support a variety of different floating-point extensions
+where calling-convention and register sizes vary for floating-point data.
+The extensions exist to support a wide variety of optional architecture
+features. The resulting ABI variants are generally incompatible with each
+other and must be tracked carefully.
+
+Traditionally the use of an explicit @code{.gnu_attribute 4, @var{n}}
+directive is used to indicate which ABI is in use by a specific module.
+It was then left to the user to ensure that command line options and the
+selected ABI were compatible with some potential for inconsistencies.
+
+@node MIPS FP ABI Variants
+@subsection Supported FP ABIs
+The supported floating-point ABI variants are:
+
+@table @code
+@item 0 - No floating-point
+This variant is used to indicate that floating-point is not used within
+the module at all and therefore has no impact on the ABI. This is the
+default.
+
+@item 1 - Double-precision
+This variant indicates that double-precision support is used. For 64-bit
+ABIs this means that 64-bit wide floating-point registers are required.
+For 32-bit ABIs this means that 32-bit wide floating-point registers are
+required and double-precision operations use pairs of registers.
+
+@item 2 - Single-precision
+This variant indicates that single-precision support is used. Double
+precision operations will be supported via soft-float routines.
+
+@item 3 - Soft-float
+This variant indicates that although floating-point support is used all
+operations are emulated in software. This means the ABI is modified to
+pass all floating-point data in general-purpose registers.
+
+@item 4 - Deprecated
+This variant existed as an initial attempt at supporting 64-bit wide
+floating-point registers for O32 ABI on a MIPS32r2 cpu. This has been
+superceded by @value{5}, @value{6} and @value{7}.
+
+@item 5 - Double-precision 32-bit CPU, 32-bit or 64-bit FPU
+This variant is used by 32-bit ABIs to indicate that the floating-point
+code in the module has been designed to operate correctly with either
+32-bit wide or 64-bit wide floating-point registers. Double-precision
+support is used. Only O32 currently supports this variant and requires
+a minimum architecture of MIPS II.
+
+@item 6 - Double-precision 32-bit FPU, 64-bit FPU
+This variant is used by 32-bit ABIs to indicate that the floating-point
+code in the module requires 64-bit wide floating-point registers.
+Double-precision support is used. Only O32 currently supports this
+variant and requires a minimum architecture of MIPS32r2.
+
+@item 7 - Double-precision compat 32-bit FPU, 64-bit FPU
+This variant is used by 32-bit ABIs to indicate that the floating-point
+code in the module requires 64-bit wide floating-point registers.
+Double-precision support is used. This differs from the previous ABI
+as it restricts use of odd-numbered single-precision registers. Only
+O32 currently supports this variant and requires a minimum architecture
+of MIPS32r2.
+@end table
+
+@node MIPS FP ABI Selection
+@subsection Automatic selection of FP ABI
+@cindex @code{.module fp=@var{nn}} directive, MIPS
+In order to simplify and add safety to the process of selecting the
+correct floating-point ABI, the assembler will automatically infer the
+correct @code{.gnu_attribute 4, @var{n}} directive based on command line
+options and @code{.module} overrides. Where an explicit
+@code{.gnu_attribute 4, @var{n}} directive has been seen then a warning
+will be raised if it does not match an inferred setting.
+
+The floating-point ABI is inferred as follows. If @samp{-msoft-float}
+has been used the module will be marked as soft-float. If
+@samp{-msingle-float} has been used then the module will be marked as
+single-precision. The remaining ABIs are then selected based
+on the FP register width. Double-precision is selected if the width
+of GP and FP registers match and the special double-precision variants
+for 32-bit ABIs are then selected depending on @samp{-mfpxx},
+@samp{-mfp64} and @samp{-mno-odd-spreg}.
+
+@node MIPS FP ABI Compatibility
+@subsection Linking different FP ABI variants
+Modules using the default FP ABI (no floating-point) can be linked with
+any other (singular) FP ABI variant.
+
+Special compatibility support exists for O32 with the four
+double-precision FP ABI variants. The @samp{-mfpxx} FP ABI is specifically
+designed to be compatible with the standard double-precision ABI and the
+@samp{-mfp64} FP ABIs. This makes it desirable for O32 modules to be
+built as @samp{-mfpxx} to ensure the maximum compatibility with other
+modules produced for more specific needs. The only FP ABIs which cannot
+be linked together are the standard double-precision ABI and the full
+@samp{-mfp64} ABI with @samp{-modd-spreg}.
+
+@node MIPS NaN Encodings
+@section Directives to record which NaN encoding is being used
+
+@cindex MIPS IEEE 754 NaN data encoding selection
+@cindex @code{.nan} directive, MIPS
+The IEEE 754 floating-point standard defines two types of not-a-number
+(NaN) data: ``signalling'' NaNs and ``quiet'' NaNs. The original version
+of the standard did not specify how these two types should be
+distinguished. Most implementations followed the i387 model, in which
+the first bit of the significand is set for quiet NaNs and clear for
+signalling NaNs. However, the original MIPS implementation assigned the
+opposite meaning to the bit, so that it was set for signalling NaNs and
+clear for quiet NaNs.
+
+The 2008 revision of the standard formally suggested the i387 choice
+and as from Sep 2012 the current release of the MIPS architecture
+therefore optionally supports that form. Code that uses one NaN encoding
+would usually be incompatible with code that uses the other NaN encoding,
+so MIPS ELF objects have a flag (@code{EF_MIPS_NAN2008}) to record which
+encoding is being used.
+
+Assembly files can use the @code{.nan} directive to select between the
+two encodings. @samp{.nan 2008} says that the assembly file uses the
+IEEE 754-2008 encoding while @samp{.nan legacy} says that the file uses
+the original MIPS encoding. If several @code{.nan} directives are given,
+the final setting is the one that is used.
+
+The command-line options @option{-mnan=legacy} and @option{-mnan=2008}
+can be used instead of @samp{.nan legacy} and @samp{.nan 2008}
+respectively. However, any @code{.nan} directive overrides the
+command-line setting.
+
+@samp{.nan legacy} is the default if no @code{.nan} directive or
+@option{-mnan} option is given.
+
+Note that @sc{gnu} @code{@value{AS}} does not produce NaNs itself and
+therefore these directives do not affect code generation. They simply
+control the setting of the @code{EF_MIPS_NAN2008} flag.
+
+Traditional MIPS assemblers do not support these directives.
+
+@node MIPS Option Stack