| Age | Commit message (Collapse) | Author |
|---|
|
Based on Aras Pranckevičius' patch.
|
|
|
|
|
|
|
|
|
|
|
|
Makes the build happy on non-GCC platforms.
|
|
New shader compiler need libc++ runtime. This works already if
we are using llvm so this just covers the !llvm case.
|
|
Remove --never-interactive because it is already specified in the
source using %option. Use -o instead of --outfile. Some of the
%option commands may also need to be removed for compatibility with
older versions (e.g., 2.5.4) of flex.
This should fix bugzilla #29209.
|
|
|
|
Bison version 2.3 doesn't seem to support %name-prefix in the source.
This should fix bugzilla #29207.
|
|
|
|
Without this, the parser will generate obtuse, useless error
diagnostics when reservered word that are not used by the grammar are
encountered in a shader.
Fixes bugzilla #29519.
|
|
Accidentally having a variable called 'sig' within an if-statement
cause the higher scope 'sig' to always be NULL. As a result a new
function signature was created for a function definition even when one
already existed from a prototype declaration.
Fixes piglit test case glsl-function-prototype (bugzilla #29520).
|
|
This was previously being appended to the output string *after* a copy
of the supposedly final string was made and handed to the caller. So
the diagnostic was never actually visible to the user.
We fix this by moving the check for an unterminated #if from
glcpp_parser_destroy to the calling function, preprocess.
This fixes the test case 083-unterminated-if.c.
|
|
Due to a recent change to glcpp-parse.y.
|
|
This is more clear than the previously-generated diagnostic which was
something confusing like "enexpected newline".
This change makse test 080-if-witout-expression.c now pass.
|
|
Rather than telling the user what to fix, the standard convention is to
describe what the detected problem is. With this change, test
081-elif-without-expression now passes.
|
|
Which are proving to be useful since some of these tests are not yet
acting as desired, (in particular, the unterminated if test is not
generating any diagnostic).
|
|
After a recent change to glcpp-lex.l
|
|
Error messages make more sense this way since the convention is for
the first line of a file to be numbered from 1, rather than 0.
|
|
This allows writing tests that verify diagnostics from the preprocessor.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The glcpp-test script was leaving around bogus *.valgrind-errors files if
a valgrind test was interrupted.
|
|
After a recent change to glcpp-parse.y
|
|
This adds a couple of test cases to expand our coverage of invalid #if and
being skipped, (either by being nested inside an #if/#elif that evaluates to
zero or by being after an #if/#elif that evaluates to non-zero).
|
|
|
|
|
|
We're already using the return-value of cmp to print either PASS or
FAIL and in the case of failure, we're subsequently running and
showing the output of diff. So any warnings/errors from cmp itself are
not actually needed, and can be quite confusing.
|
|
Commit d4a04f315560704bf1103df0b93723e468725df7 caused this test case
to produce an additional blank line, which is otherwise harmless, but
does need to be reflected in the .expected file for the test to pass.
|
|
After making a minor change to the .y file.
|
|
Since we have a custom structure for YYLTYPE locations, we need to use
an %initial-action directive to avoid triggering use of uninitialized
memory when, for example, printing error messages.
We apparently don't yet have a test case that allowed valgrind to find
this bug for us, but valgrind found a similar problem in the other
parser, so we fix this one as well.
|
|
After making a minor change to the .y file.
|
|
Since we have a custom structure for YYLTYPE locations, we need to use
an %initial-action directive to avoid triggering use of uninitialized
memory when, for example, printing error messages.
Thanks to valgrind for noticing this bug.
|
|
It's rather easy to produce two constant multiplies separated by other
multiplies while writing a BRDF shader, and non-obvious enough in the
resulting codegen that I didn't catch it in my demo code until just
recently. Cuts 3 965 instructions from my demo (<1%), and 20 from
glsl-fs-raytrace (1.3%).
|
|
All the current HW backends transform subtract to adding the negation,
so I haven't bothered peepholing it back out in Mesa IR. This allows
some subtract of subtract to get removed in ir_algebraic.
|
|
Whereas constant folding evaluates constant expressions at rvalue
nodes, constant propagation tracks constant components of vectors
across execution to replace (possibly swizzled) variable dereferences
with constant values, triggering possible constant folding or reduced
variable liveness.
|
|
|
|
This lets us track copies across basic block boundaries. The loop
doesn't get a filled out list of available copies into it yet, though.
glsl-fs-raytrace drops from 585 to 535 Mesa IR instructions out of the
compiler, and it appears that Yo Frankie's largest shaders decrease in
Mesa IR count by about 10% as well.
|
|
|
|
Some signatures were being generated with the wrong function name.
|
|
It was being incorrectly added as gl_DepthRangeParameters, which is the
type name, not the variable name.
|
|
|
