the output is an unsigned integer.

first macrostage of module `DivSqrtRecF64_mulAddZ31'.

This builds correctly for me

Chisel 3 compatibility.

They differ in their effects on width inference.

x >> 3 results in a node that's 3 bits narrower than x.
x >> UInt(3) results in a node with the same width as x.

This reverts commit 5141a03a, which
depends on a new version of Chisel.  Unfortunately, that version of
Chisel has a bug that breaks the FMA unit.  Although I've fixed the
bug, It hasn't yet made it into a release.  At that time, we can
re-incorporate this patch.

These bugs cropped up because tests were failing to fail.

This reverts commit a56caadf.

While we should consider getting rid of priorityEncode, this change is
not functionally equivalent and breaks the rocket-chip integration tests.

Chisel 3 support

This avoids some internal casts that are a bit ugly.

Not sure why this was here, but it appears to be a poor implementation of
Chisel's PriorityEncoder.  I've also removed some defunct arguments to the
priorityEncode wrappers.

These were all SInt/UInt constants or cast changes.  I believe these changes
should do nothing in Chisel 2 -- they're all just things that have been defined
as illegal in Chisel 3 because they were implementation defined in Chisel 2.  I
changed these to explicitly match the Chisel 2 implementation's behavior.

"make test-c" still passes.

I'm not actually sure what happened here, but it appears one of my patches has
been lost.  The RISCV-specific patch is still there, and it's in "git log", but
it's not in the current Makefile.  "git blame" points to the patch that
contains the code I changed, which I wouldn't expect.

This was originally part of b8cfc323.

converting from small floating-point formats to large integer sizes.  (Such
conversions are still largely untested at this time.)

module `DivSqrtRecF64_mulAddZ31'.

greater-than), and associated tests.

thorough testing.

No need to handle exponent and significant separately.

We failed to kill the excess precision.

Support for autometad regression testing

This runs the tests for a fixed amount of time, as opposed to just
running them until they find 20 failures.  This makes the test harness
suitable for running automatically.

In order to run the standalone tests we need to be able to run
"testfloat_gen", part of the TestFloat-3 package.  Since this is
probably going to be something that's not commonly installed, I added
some Makefile targets that download and build "testfloat_gen" so it
can be run without messing around with any system dependencies.

Note that TestFloat-3 by default builds in 8086 mode, so there's a
patch here that makes it build in RISC-V mode instead.