This patch ensures that the result of a shift is an undefined value
if the right operand is an undefined value. It also improves the
code generated for right shifts where the right operand is constant
and optimises away shifts where the right operand is a constant 0.
It also fixes a few places where the expression type (signed/unsigned)
was not being set correctly.
This patch adds a check on the expression type before adjusting an
unsized expression width up to integer_width. This adjustment should
only be performed for vectorable expressions.
The patch also fixes a few compiler warnings that only show up on
32-bit builds.
This patch modifies the compiler and the ivl interface to pass the
type of indexed part select that is being represented (up/down) in
a procedural L-value or R-value. This is needed by any back end that
wants to correctly denormalize the (zero based) base expression.
Constant user functions aren't currently supported. A suitable
error message should be output when a user function call is
encountered in a parameter expression. This got lost in the
parameter expression rework.
The compiler currently performs parameter expression elaboration before
performing parameter overrides. This means that the information needed
to correctly determine the expression type and width may not be available
at the time elaboration is performed. This patch reworks the code to
delay elaboration until after all overrides have been performed. It
also provides a new -g option that controls how the width of parameter
expressions is calculated when the parameter itself is unsized.
This gets the enumeration type through to the ivl_target API so
that code generators can do something with it. Generate stub
output with tgt-stub, and generate the proper vvp run time to
make simple enumerations work from end to end.
This patch covers more than it should. It removes many of the -Wextra
warnings in the main ivl directory. It also makes some minor code
improvements, adds support for constant logicals in eval_tree (&&/||),
adds support for correctly sign extending === and !==, it starts to
standardize the eval_tree debug messages and fixes a strength bug
in the target interface (found with -Wextra). The rest of the warnings
and eval_tree() rework will need to come as a second patch.
This patch changes all the iterator code to use a prefix ++ instead
of postfix since it is more efficient (no need for a temporary). It
is likely that the compiler could optimize this away, but lets make
it efficient from the start.
I'm adding more uses of the make_range_from_width function, so
it seems like time to get rid of its use of the svector template.
This thread led to a lot of other uses of svector that had to
also be removed.
An unsized signed constant that is smaller than integer_width needs
to be resized up to either the integer_width or the expression width
whichever is smaller.
This patch mimics what was done for vectors, but is simpler since
arrays don't use the endian information. It also needs to address
the fact that .array/port assumes the expression is unsigned so
any signed expression must be padded to make it larger than the
maximum array word when it is converted to unsigned.
This patch modifies all the variable bit and indexed part selects
to use a common routine. This routine determines the minimum
width needed to calculate the result correctly, pads the expression
if needed and then converts the expression to signed if required to
make the calculation correct.
Added 'expr_wid' parameter to calls to 'eval_expr()' within
PEBinary::elaborate_eval_expr_base_()'. This makes a specific problem
go away and may even be the correct thing to do.
We only need to set the width of a multiple if the width is greater
than zero and if the expression is vectorable. This matches what is
done for addition.
Some new shadow issues have crept in. This patch fixes these new
issues and adds -Wshadow to the normal warning flags to keep any
new occurrences from happening.
Comparison expressions have sorta-self-determined arguments.
Handle the special cause that some of the arguments may be
themselves unsized, and so expecting to be even wider then
otherwise.
Operands to reduction unary operators are self determined, so
evaluate the operands that way. But this means that binary expressions
in this context should take pains to use their test_width tested
expression width.
This exposed a case where the test_width methods were not called
for self-determined expressions. Fix that too.
This patch enhances elaboration to drop/ignore zero replication
count constants. Not doing this was causing problems later in
the compiler. We still pass non-constant expressions since
both user and system functions must be run for their possible
side effects. Constants can never have a side effect so just
dropping them is acceptable.
When handling the $signed/$unsigned system functions, the compiler
was applying the new signed/unsigned property to the NetExpr object
representing the input argument. This caused the input argument to
be evaluated incorrectly. This patch fixes this by applying the new
property to the NetExpr object created to pad the result to the
required size.
In testing this fix, it was also discovered that the width of the
input argument expression was not being calculated correctly. This
patch also fixes this issue.
Creation of implicit nets requires knowledge of whether an identifier
has been declared before it is used. Currently implicit nets are
created during elaboration, but by this stage the order of declaration
and use is not known. This patch moves the creation of implicit nets
into the parser stage.
Under some situation the optimized repeat expression can get out
of sync while it is being processed this causes the code to
reference freed memory. This happens when the repeat expression
is converted to an unsigned integer, but the repeat expression
is used in a different path. It was easier to just remove the
optimization then track down the convoluted call sequence that
was causing this and then figure out what needed to be fixed.
The repeat expression must be constant, so is likely not too
complicated so this optimization is very minor and should not
be missed.
This patch is a major rewrite of the indexed part selects (+: and -:).
It made the following enhancements:
1. Make indexed part selects work correctly with both big and little
endian vectors.
2. Add a warning flag that warns about constant out of bounds/or 'bx
indexed selects.
3. Moved the -: parameter code to its own routine.
4. Added support for straddling before part selects in a CA.
5. Added more assert(! number_is_unknown) statements.
6. Add warning for &PV<> select with a signed index signal that is
less than the width of an int. This will be fixed later.
7. Add support for loading a 'bx/'bz constant into a numeric register.
8. Add a number of signed value fixes to the compiler/code generator.
9. Major fix of draw_select_expr() in the code generator.
In 1364-2005 it is an explicit error to take the select of a scalar
or real value. We added the checks for real a while ago. This patch
adds the functionality for scalar values. In the future we may want
to push the scalar property to the run time.
Only sign extend the operands for division or modulus if
both of them are signed. Previously only the individual
operand signedness was being considered.
This patch fixes the following problems in the right shift code.
The >>> result is only signed if the l-arg is also signed.
The r-arg is always unsigned.
Only sign extend the l-arg when the >>> operator is used.
This relates to an optimization for certain constant shifts.
Removed some obsoleted/outdated code.
This patch adds messages in various places to warn that constant
user functions are not supported. It uses a global variable to
indicate when we are in a constant context. This is a bit of a
kludge, but works well without needing to change a bunch of code.
It is interesting to note that ports are elaborated late enough
that if we had the constant function evaluation code they would
evaluate correctly. This also applies to the function return
range, the concatenation repeat, specparams and initial values.
Signal definitions are early enough that elaboration is what is
failing because the function body is not available (has not been
elaborated). The same thing applies to both parameters and
localparms.
It is possible for an identifier to have multiple index expressions.
(For example, a part select of a memory word.) Make sure all the
index expressions are probed for width and type.
And also, note that the unary ! returns BOOL or LOGIC, not just
the type of its operand. It is slightly different from the other
unary operators in that way.
The test_width for repeat-concatenations is tricky because it
requires the evaluated value for the repeat expression. It should
be OK to call elab_and_eval on that expression even during the
test_width for the containing expression. We'll see.
This patch adds logical support for real values (!, && and ||). It
also prints an appropriate error message when a real value is
incorrectly given to an operator that does not allow real arguments
(bit-wise, reduction, shift and repeat/concatenation).
For most cases just using the value of a parameter is fine, but
a vpi call can access more than the value so we want to use a
parameter reference instead of the value for vpi calls.
Strings were working correctly, integer values need some minor
code generator changes and real values needed to be pushed from
elaboration to the code generators. I also changed the default
real value comment from %g to %#g so that it is more obvious
that the value is a real value.
When doing right shift of unsized expressions, pad the left operand
so that the right shift does not lose. This better accounts for the
lossless expectations of unsized arguments.
Part selects need to be fully defined. If not, then the resulting
expression is 'bx no matter what. The same for bit selects, when
the bit select expression is constant.
When the parameter up index was being reworked someone mistakenly
used the pointer value instead of the actual value in the MSB/LSB
comparison. This obviously could give incorrect results.
This fixes up the elaboration of binary expressions found in
parameter expressions. Parameter expressions are special because
they elaborate early, before all the other parameters are necessarily
completed.
This patch fixes fully out of range constant indexed part selects
to just return 'bx. It also adds support for constant undefined
base values which also just return 'bx.
A bug in the bit width calculation when building an unsized, signed
negative integer value was also fixed (-3 needs 3 bits not 2, etc.)
The power (**) and shift operators are different from other binary
operators because their expression width calculations rely only on
their left operand, with their right operand self-determined. Get
the handling of these operators out of the PEBinary base class to
prevent confusion.
unary expressions that have problems should not assert in the
test_width method. Instead, let the error propagate back and be
handled during expression elaboration. This found a few places
where expression widths/types weren't probed before elaboration.
When padding a signal or when creating a local signal the file and
line information should be related to where the new object was
created not the signal value it is being created from.
This patch modifies the NetE* pad_to_width() routines to take a
LineInfo object to set the location to the correct value.
It fixes some set_line() calls to use the correct location.
It fixes ports to not set the file/line information if it is
already defined. Doing this was causing the definition of
signals to become the instantiation instead of the real
module declaration.
If one of the arguments of a comparison expression has a real value, then
the expression with is 1 no matter the width of the other argument. This
means that the arguments may have different widths in this special case.
Patch is from pr2251119, suggested by Martin Whitaker.
Even though we cannot immediately give a width for a concatenation
that has a repeat expression (the expression must be evaluated first)
we still must scan the test_width of the arguments so that they can
resolve their types.
This includes enough API to get the branch nexus bits and signals
and show them in the dump. This also includes creating the reference
ground for branch access functions that use the implicit ground.
The natures of disciplines were already available, this just brings
the information forward to the ivl_target.h API and exposes them via
access functions.
Signals may have VMA disciplines attached. Make the attached discipline
visible through the ivl_target.h API. Also, re-arrange the internal
handling of the discipline structure so that we can expose disciplines
through the ivl_target C API without creating new structures. The
t-dll-api implementations of the discipline access functions can look
at the elaborated discipline structure directly. This is possible since
the discipline parse and elaboration are very simple.
This patch splits any VVP net functor that needs to access both
statically and automatically allocated state into two sub-classes,
one for handling operations on statically allocated state, the
other for handling operations on automatically allocated state.
This undoes the increase in run-time memory use introduced when
automatic task/function support was first introduced.
This patch also fixes various issues with event handling in automatic
scopes. Event expressions in automatic scopes may now reference either
statically or automatically allocated variables or arrays, or part
selects or word selects thereof. More complex expressions (e.g.
containing arithmetic or logical operators, function calls, etc.) are
not currently supported.
This patch introduces some error checking for language constructs
that may not reference automatically allocated variables. Further
error checking will follow in a subsequent patch.
The arguments of logical and/or are self determined, and the width is
fixed as 1 bit. Account for this special behavior by creating the
PEBLogic class.
The comparison operator operands are self determined, but are forced
to be the width of the wider operand. This means that the operands must
be evaluated with their widths truncated. In spite of all this, note
that comparisons expression results are 1 bit wide.
Continue cleaning up shadowed variables, flagged by turning on -Wshadow.
No intended change in functionality. Patch looks right, and is tested
to compile and run on my machine. No regressions in test suite.
The condition expression needs its width tested, even if the width
is not used. Also clean up some handling of widths/types for other
expression types.
The right-operand of shift expressions is self-determined, but we still
need to run a test_width to get the PExpr decorated with types and
expression widths.
The l-value doesn't really constrain the size of unsized expressions
because there are possible sub-expressions that may pull high bits
down to the low bits. (Divide, for example.)
If either of the operands of a ternary expression are unsigned, then
both are treated as unsigned. It works just like a binary expression
in that regard.
The results of comparisons are unsigned, but the arguments may be signed
and the calculation performed may be signed. Handle the padding properly
for comparisons so that the math is properly signed.
Later passes need the intermediate results for width and size so that
some special cases, were self-determined arguments occur, can be
processed properly during elaboration. This can be especially tricky
and interesting for ternary expressions.
This patch fixes the expression width calculation for a multiply
operation with an unsized operation. The expression width needs
to be at least the minimum of the maximum multiply result width
and the width of an integer.
This patch removes all the checks for constant expressions performed
during the parsing phase, as these checks are (mostly) repeated during
elaboration. It adds the missing check in the elaboration phase (the
RHS of a register initialisation), and improves the error reporting
and error recovery in other checks.
This patch fixes pr2132552, which was caused by a fault in the parser
constant expression checking.
Multiply of any expression with constant 0 will always return zero.
We can handle this early, during elaboration, and save a lot of code
downstream the trouble.
Also, while we are at it, fix up test_width to re-test the left
expression width if the right expression width is unsized. This allows
for the left expression code to adapt to the unsized-ness of the
expression context.
There are cases where the r-value doesn't pad itself to the width
that is requested by the call to elaborate_expr. This impacts the
elaboration of PGAssign. Pad/sign extend as appropreate.
The arguments to bitwise operators are padded if *either* of the operands
is unsigned. This is according to the sign/unsigned expression rules
of Verilog, and also matches the behavior of the "Big-3."
The addition of UINT_MAX in netlist.cc requires #include <climits> when compiling with gcc-4.3.2.
I also noticed that commit 8704e3e used c style includes in c++ sources contrary to the style used in the rest of the c++ code and fixed those.
Simple error setting the expression width of the FALSE clause of a
ternary expression when the expression is short cuirted to false.
Fix a simple error where the type of a system function is not returned
through the test_width function.
The arguments to user defined functions are self-determined. And if
the result is real valued, we can call them lossless self-determined.
Treat these arguments like r-value elaboration for assignments.
Also clean up the binary divide elaboration a little bit.
Also, floating point literals are unsized with width==1.
Widths of real values are always 1. When paired with vectorable types
in expressions, the vectorable type is processed as losslessly self-
determined. ("unsized" in the test_width methods.)
Don't force the expression with to be different from the l-value if the
l-value is the immediate destination. This saves the effort of handling
overly wide constant values in simple cases.
Also, in determined contexts, signed unary minus does not need to pad the
width of the expression.
the PGAssign elaborate method used the test_width to get the width
of the r-value expression. This should be completely sufficient to
get the width of the expression, so always use a defined width to
elaborate the expression.
When ternary expressions are self-determined, use the test_width
method to get a proper reading of the expression width. Also improve
the test_width method to handle unsized operands.
The true and false expression clauses must have compatible types,
which are not necesarily identical. In particular, VT_BOOL and
VT_LOGIC are compatible for the purposes of ternary arguments. The
test in NetETernary::synthesize was incorrect.
In the process, fix the type handling of NetConst objects to allow
for IVL_VT_BOOL constants. This is information that the downstream
may find useful, so should be handled correctly.
Nothing to do with tab width! Eliminates useless
trailing spaces and tabs, and nearly all <space><tab>
pairings. No change to derived files (e.g., .vvp),
non-master files (e.g., lxt2_write.c) or the new tgt-vhdl
directory.
Low priority, simple entropy reduction. Please apply
unless it deletes some steganographic content you want
to keep.
In continuous assignment, the width of the expression needs to come
from the expression itself, and not just from the width of the l-value.
Use the PExpr::test_width method to get the width of the expression
to pass to the elaborate.
When the condition expression of a ternary is constant 1 or 0, we can
short-circuit the elaboration by only processing the clause (true or
false) that we need. This saves compile time and execution time.
It is possible for the left operand of a left shift to be not known
early during elaborate expression. In that case, make a punt expression
and expect it to be resolved later.
The expr:::synthesize methods need not deal with saturating left or
right shifts if they are dealt with early, in elaborate_expr methods.
So the elaborate_expr for shift takes on much more responsibility.
In signed contextx, the right shift and unary minus expressions need
to be sign extended before they operate, otherwise there may be bad
results in the high bits in the greater context.
The NetBranch object is connected, but not like an object, so the
NetPins object base class is factored out from NetObj to handle the
connectivity, and the NetBranch class uses the NetPins to connect a
branch.
Also, account for the fact that nets with a discipline are by default
real-valued.
Detect function call expressions that turn out to be calls to the
access function of a nature. Elaborate the access function and stub
the emit code. This gets the access function just short of the code
generator.
Contribution statements have an l-value and r-value. Parse those
expressions into pform so that elaboration has something to work with.
In this process, this patch also changes the PECallFunction class to
use the vector template instead of the svector template. The latter
doesn't add anything over the STL vector template, so this is a start
of working the svector out.
Part select of an entire signal returns just the NetESignal itself,
but since part selects are always unsigned, even if the selected
signal is signed, we need to cast the NetESignal to unsigned first.
The part select of a vector is converted by the compiler during
elaboration to a 0-based canonical address. But since it is legal
to address bits below the LSB, the canonical address can be negative.
So make the part select base for selecting from signals work with
signed arithmetic and make the code generator generate negative
indices when needed.
Part select of parameter names is fixed up to be structurally similar
to part select of signals, and also to behave similarly. (Though not
identically, for reason.)
Part selects to signals are allowed to be off the ends of the signal
itself. The bits that are beyond the vector return X. This may mean
creating constant X bits on one or both ends of the result.
A recent change to sb_to_idx() made it return a signed value
where it previously returned an unsigned value. This patch adds
explicit casts to remove the two signed vs unsigned comparison
warning messages.
Fix
elab_expr.cc:1561: warning: comparison between signed and unsigned integer expressions
caused by revised prototype of NetNet::sb_to_idx() in
commit dfb7bf5211
IEEE1364 has specific names for the various generations of Verilog that
are supported. Icarus Verilog should stick to those names for selection
the language feature set.
In the process, the extensions that were tied to the 2x generations
are pulled out out and given their own enable flags. The makes all the
feature control more regular and understandable.
Implement in behavioral the abs/min/max operators for real values.
The parser treats these builtin functions as operators, unary or
binary, and elaborates them appropriately.
Also add enough code generator support to handle real valued expressions
in thread context.
1364-2001 does not allow a replication count of zero, but
1364-2005 allows them when the replication is enclosed in
a concatenation with other valid bits. This patch adds
code to perform these checks for procedural expressions.
Because of the NetNet object must have a width greater
than zero a replication used in a continuous assignment
can never have a count of zero.
Logical and/or take as inputs condition expressions, which are scalar
expressions. Be sure to reduce vectors using proper logic to get the
right condition value.
When left-shifting unsized constants in self-determined contexts, the
constant value is normally pared down to its minimum required width.
But this practically guarantees loss of bits. Instead, detect this
special case and give the unsized constant a width of an integer.
Still allow for the super-special case that the shifted value and the
shift amount are constants. In that case, the result width (and value)
can be calculated precisely and there is no need to resort to default
widths.
Binary bit based operators were not padding arguments if they had
different widths. The reduction operators and ! were not correctly
evaluating their arguments in a self-determined context.
This patch reworks much of the ternary code to short circuit when
possible and supports real values better. It adds a blend operator
for real values that returns 0.0 when the values differ and the value
when they match. This deviates slightly from the standard which
specifies that the value for reals is always 0.0 when the conditional
is 'bx. There are also a couple bug fixes.
These fixes have not been ported to continuous assignments yet.
Ternary operators used at compile time and in procedural assignments
should be complete (short circuit and support real values).
This patch adds bit based power support to normal expressions.
It also pushes the constant unsigned bit based calculation to
the runtime until the bit based method can be copied to the
compiler. Continuous assignments also need to use this type
of calculation.
This patch adds a check and prints a warning message when the power
operator is used with unsigned bit based values. It also fixes a couple
of typos and adds an asserts if the above power operator happens to
get to the tgt-vvp back end.
The compiler attempts to precalculate the results of <= comparisons.
Fix a few corner cases where the arguments are signed. Also fix the
important test of constant against non-constant where we try to test
if the non-constant value can possibly make the test fail.
Rework the handling of file names to use a perm_string heap to hold
the file names, instead of the custom file name heap in the lexor.
Also rename the get_line to get_fileline to reflect its real duties.
This latter chage touched a lot of files.
Comparison operands are signed only of both operands are signed.
If one is unsigned, then they both are unsigned. This does not
affect the signedness of the comparison itself, which is unsigned.
Signed-off-by: Stephen Williams <steve@icarus.com>
Fixed some signed/unsigned warnings, including one that caused an
actual runtime error if not dealt with safely.
Signed-off-by: Stephen Williams <steve@icarus.com>
The argument to $unsigned is self-determined no matter what the
context the $unsigned itself is in. This is important only where
the result can be negative but the result width is context-determined.
Do ocntext fitting manually to prevent the context fitting it and
affecting the expression argument.
Signed-off-by: Stephen Williams <steve@icarus.com>
Fix the handling of expressions that have unsized integers and are
in self determined context. Unsized integers are generally assumed
to have at least 32bits.
Signed-off-by: Stephen Williams <steve@icarus.com>
In expressions, user defined functions have a clearly defined
width and the test_width expression needs to express that. Note
that the $signed/$unsigned functions are special and magical.
Signed-off-by: Stephen Williams <steve@icarus.com>
distinguish between self-determined and value-preserving when
evaluating constants that are arguments to concatenations.
Signed-off-by: Stephen Williams <steve@icarus.com>
This fixes r-value type references to memory words for memories
that have a non-zero base address. Elaborate the expression
needed to get the canonical word index.
Real value are vector width of 1, fix real literal to reflect this.
fix leaking real registers in code generation for function arguments.
Load of signal should handle conversion from real to vector. Function
arguments, type vector passed a real value, are an example where this
comes up.
Signed-off-by: Stephen Williams <steve@icarus.com>
more general concept of arrays. The NetMemory and NetEMemory
classes are removed from the ivl core program, and the IVL_LPM_RAM
lpm type is removed from the ivl_target API.
scaled to the local units. Change the internal
implementation of vpiSystemTime the $time functions
to properly account for this. Also add $simtime
to get the simulation time.
I now generate the scopes and notice the parameters
in a separate pass over the pform. Once the scopes
are generated, I can process overrides and evalutate
paremeters before elaboration begins.
Martin Whitaker
Larry Doolittle
Cary R
Stephen Williams
Patrick Doyle
Martin Whitaker
Jared Casper
Martin Whitaker
steve