Long addressing was not being used to access the values, which could lead to mis-evaluation of comparisons against values in global structs, unions, or arrays, depending on the memory layout.
This could sometimes affect the c99desinit.c test, when run with large memory model and at least intermediate code peephole optimization. It could also affect this simpler test (depending on memory layout):
#pragma memorymodel 1
#pragma optimize 1
struct S {
void *p;
} s = {&s};
int main(void) {
return s.p != &s; /* should be 0 */
}
Previously, the assembly-level optimizations applied to code in asm statements. In many cases, this was fine (and could even do useful optimizations), but occasionally the optimizations could be invalid. This was especially the case if the assembly involved tricky things like self-modifying code.
To avoid these problems, this patch makes the assembly optimizers ignore code from asm statements, so it is always emitted as-is, without any changes.
This fixes#34.
This differs from the usual ORCA/C behavior of treating all floating-point parameters as extended. With the option enabled, they will still be passed in the extended format, but will be converted to their declared type at the start of the function. This is needed for strict standards conformance, because you should be able to take the address of a parameter and get a usable pointer to its declared type. The difference in types can also affect the behavior of _Generic expressions.
The implementation of this is based on ORCA/Pascal, which already did the same thing (unconditionally) with real/double/comp parameters.
They now use a jmp (addr,X) instruction, rather than a more complicated code sequence using rts. This is an improvement that was suggested in an old Genie message from Todd Whitesel.
This converts comparisons like x > N (with constant N) to instead be evaluated as x >= N+1, since >= comparisons generate better code. This is possible as long as N is not the maximum value in the type, but in that case the comparison is always false. There are also a few other tweaks to the generated code in some cases.
The TAY instruction would set the flags, but that is unnecessary because pc_cnv is a "NeedsCondition" operation (and some other conversions also do not reliably set the flags).
The code is also changed to preserve the Y register, possibly facilitating register optimizations.
This was a problem introduced in commit 393b7304a05. It could cause a compiler error for unoptimized array indexing code, e.g.:
int a[100];
int main(void) {
return a[5];
}
This affects 16-bit subtractions where where only the left operand is "complex" (i.e. most things other than constants and simple loads). They were using an unnecessarily complicated code path suitable for the case where both operands are complex.
These could occur because the code for certain operations was assumed to set the z flag based on the result value, but did not actually do so. The affected operations were shifts, loads or stores of bit-fields, and ? : expressions.
Here is an example showing the problem with a shift:
#pragma optimize 1
int main(void) {
int i = 1, j = 0;
return (i >> j) ? 1 : 0;
}
Here is an example showing the problem with a bit-field load:
struct {
signed int i : 16;
} s = {1};
int main(void) {
return (s.i) ? 1 : 0;
}
Here is an example showing the problem with a bit-field store:
#pragma optimize 1
struct {
signed int i : 16;
} s;
int main(void) {
return (s.i = 1) ? 1 : 0;
}
Here is an example showing the problem with a ? : expression:
#pragma optimize 1
int main(void) {
int a = 5;
return (a ? (a<<a) : 0) ? 0 : 1;
}
This optimizes most multiplications by a power of 2 or the sum of two powers of 2, converting them to equivalent operations using shifts which should be faster than the general-purpose multiplication routine.
This changes unsigned 16-bit multiplies to use the new ~CUMul2 routine in ORCALib, rather than ~UMul2 in SysLib. They differ in that ~CUMul2 gives the low-order 16 bits of the true result in case of overflow. The C standards require this behavior for arithmetic on unsigned types.
The desired location for the quad result was not saved, so it could be overwritten when generating code for the left operand. This could result in incorrect code that might trash the stack.
Here is an example affected by this:
#pragma optimize 1
int main(void) {
long long a, b=2;
char c = (a=1,b);
}
Previously, one-byte loads were typically done by reading a 16-bit value and then masking off the upper 8 bits. This is a problem when accessing softswitches or slot IO locations, because reading the subsequent byte may have some undesired effect. Now, ORCA/C will do an 8-bit read for such cases, if the volatile qualifier is used.
There were also a couple optimizations that could occasionally result in not all the bytes of a larger value actually being read. These are now disabled for volatile loads that may access softswitches or IO.
These changes should make ORCA/C more suitable for writing low-level software like device drivers.
This affects functions whose body spans multiple files due to includes, or is treated as doing so due to #line directives. ORCA/C will now generate a COP 6 instruction to record each source file change, allowing debuggers to properly track the flow of execution across files.
This does not seem to be necessary for any of the debuggers (at least in their latest versions), and it obviously causes problems with case-sensitive filesystems.
It could wind up storing garbage in the upper 8 bits of the destination, because it was not doing a proper 8-bit to 16-bit conversion.
This is an old bug, but the change in commit 95f518244212c caused it to be triggered in more cases, e.g. in the C7.5.1.1.CC test case.
Here is a case that could exhibit the bug even before that:
#pragma optimize 1
#include <stdio.h>
int main(void) {
int k[1];
int i = 0;
unsigned char uch = 'm';
k[i] = uch;
printf("%i\n", k[0]);
}
The C standards generally allow floating-point operations to be done with extra range and precision, but they require that explicit casts convert to the actual type specified. ORCA/C was not previously doing that.
This patch relies on some new library routines (currently in ORCALib) to do this precision reduction.
This fixes#64.
This was already done by the optimizer, but it is simple enough to just do it all the time. This avoids most performance regressions from the previous commit, and also generates more efficient code for long long stores (in the common cases where the value of an assignment expression is not used in any larger expression).
The value of an assignment expression should be exactly what gets written to the destination, without any extra range or precision. Since floating-point expressions generally do have extra precision, we need to load the actual stored value to get rid of it.
This means that floating-point constants can now have the range and precision of the extended type (aka long double), and floating-point constant expressions evaluated within the compiler also have that same range and precision (matching expressions evaluated at run time). This new behavior is intended to match the behavior specified in the C99 and later standards for FLT_EVAL_METHOD 2.
This fixes the previous problem where long double constants and constant expressions of type long double were not represented and evaluated with the full range and precision that they should be. It also gives extra range and precision to constants and constant expressions of type double or float. This may have pluses and minuses, but at any rate it is consistent with the existing behavior for expressions evaluated at run time, and with one of the possible models of floating point evaluation specified in the C standards.
This works when both operands are simple loads, such that they can be broken up into operations on their subwords in a standard format.
Currently, this is implemented for bitwise binary ops, but it can also be expanded to arithmetic, etc.
This optimization works when the return value is stored directly to a local variable and not used otherwise (typically only recognized when using intermediate code peephole optimization).
This allows them to bypass the intermediate step of loading the value onto the stack. Currently, this only works for simple cases where a value is loaded and immediately stored.
For the moment, this does not really do anything, but it lays the groundwork for not always having to load quadword values to the stack before operating on or storing them.
