We just want to mark the whole thing used, not mark from each element
the whole size in use. Fixes undefined URB entry writes on i965,
which blew up with debugging enabled.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
From section 7.1 (Vertex Shader Special Variables) of the GLSL 1.30
spec:
"It is an error for a shader to statically write both
gl_ClipVertex and gl_ClipDistance."
Fixes piglit test mixing-clip-distance-and-clip-vertex-disallowed.c.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The check now applies both when explicitly declaring the size of
gl_TexCoord and when implicitly setting the size of gl_TexCoord by
accessing it using integral constant expressions.
This is prep work for adding similar size checks to gl_ClipDistance.
Fixes piglit tests texcoord/implicit-access-max.{frag,vert}.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
From the GLSL 1.30 spec, section 7.1 (Vertex Shader Special Variables):
The gl_ClipDistance array is predeclared as unsized and must be
sized by the shader either redeclaring it with a size or indexing it
only with integral constant expressions.
Fixes piglit tests clip-distance-implicit-length.vert,
clip-distance-implicit-nonconst-access.vert, and
{vs,fs}-clip-distance-explicitly-sized.shader_test.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The list of numbers in (constant type (<numbers>)) needs to contain
exactly type->components() numbers (16 for a mat4, 3 for a vec3, etc.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Throwing away the extra numbers ought to match the existing behavior.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Each of these vecN constants only provided one component, which is
illegal. The printed IR is meant to contain exactly as many components
as are necessary; the IR reader does not splat single values.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Using multiply and reciprocal for integer division involves potentially
lossy floating point conversions. This is okay for older GPUs that
represent integers as floating point, but undesirable for GPUs with
native integer division instructions.
TGSI, for example, has UDIV/IDIV instructions for integer division,
so it makes sense to handle this directly. Likewise for i965.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Bryan Cain <bryancain3@gmail.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Forgotten in the patch that enabled the extension.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
while debugging texelFetchOffset we kept hitting the assert.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Otherwise we continue and hit the "Illegal formal parameter mode"
assertion.
Fixes negative compile test texelFetchOffset.frag in piglit.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
It's the same as GL_AMD_conservative_depth. The specs have slight
differences in wording, but don't differ in content or behavior.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
One unique aspect of TXS is that it doesn't have a coordinate.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
There is no ir_hierarchical_visitor::visit(ir_if *) method, since ir_if
is not a leaf node. Instead, there are visit_enter and visit_leave
methods. Use visit_enter arbitrarily (either would work fine, though
visit_enter will catch errors sooner).
Found thanks to a warning emitted by Clang.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
This builds the static library libmesa_glsl and executable glsl_compiler
from glsl. glsl_compiler is only installed for engineering build.
Reviewed-by: Chad Versace <chad@chad-versace.us>
This patch fixes a bug when lowering an integer division:
x/y
to a multiplication by a reciprocal:
int(float(x)*reciprocal(float(y)))
If x was a plain int and y was an ivecN, the lowering pass
incorrectly assigned the type of the product to be float, when in fact
it should be vecN. This caused mesa to abort with an IR validation
error.
Fixes piglit tests {fs,vs}-op-div-int-ivec{2,3,4}.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The vs-varying-array-mat2-col-row-wr test writes a mat2[3] constant to
a mat2[3] varying out array, and also statically accesses element 1 of
it on the VS and FS sides. At link time it would get trimmed down to
just 2 elements, and then codegen of the VS would end up generating
assignments to the unallocated last entry of the array. On the new
i965 VS backend, that happened to land on the vertex position.
Some issues remain in this test on softpipe, i965/old-vs and
i965/new-vs on visual inspection, but i965 is passing because only one
green pixel is probed, not the whole split green/red quad.
This patch extends ir_validate.cpp to check the following
characteristics of each ir_call:
- The number of actual parameters must match the number of formal
parameters in the signature.
- The type of each actual parameter must match the type of the
corresponding formal parameter in the signature.
- Each "out" or "inout" actual parameter must be an lvalue.
Reviewed-by: Chad Versace <chad@chad-versace.us>
These functions don't modify the target instruction, so it makes sense
to make them const. This allows these functions to be called from ir
validation code (which uses const to ensure that it doesn't
accidentally modify the IR being validated).
Reviewed-by: Chad Versace <chad@chad-versace.us>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
When an out parameter undergoes an implicit type conversion, we need
to store it in a temporary, and then after the call completes, convert
the resulting value. In other words, we convert code like the
following:
void f(out int x);
float value;
f(value);
Into IR that's equivalent to this:
void f(out int x);
float value;
int out_parameter_conversion;
f(out_parameter_conversion);
value = float(out_parameter_conversion);
This transformation needs to happen during ast-to-IR convertion (as
opposed to, say, a lowering pass), because it is invalid IR for formal
and actual parameters to have types that don't match.
Fixes piglit tests
spec/glsl-1.20/compiler/qualifiers/out-conversion-int-to-float.vert and
spec/glsl-1.20/execution/qualifiers/vs-out-conversion-*.shader_test,
and bug 39651.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=39651
Reviewed-by: Chad Versace <chad@chad-versace.us>
Previously if-statements were lowered from inner-most to outer-most
(i.e., bottom-up). All assignments within an if-statement would have
the condition of the if-statement appended to its existing condition.
As a result the assignments from a deeply nested if-statement would
have a very long and complex condition.
Several shaders in the OpenGL ES2 conformance test suite contain
non-constant array indexing that has been lowered by the shader
writer. These tests usually look something like:
if (i == 0) {
value = array[0];
} else if (i == 1) {
value = array[1];
} else ...
The IR for the last assignment ends up as:
(assign (expression bool && (expression bool ! (var_ref if_to_cond_assign_condition) ) (expression bool && (expression bool ! (var_ref if_to_cond_assign_condition@20) ) (expression bool && (expression bool ! (var_ref if_to_cond_assign_condition@22) ) (expression bool && (expression bool ! (var_ref if_to_cond_assign_condition@24) ) (var_ref if_to_cond_assign_condition@26) ) ) ) ) (x) (var_ref value) (array_ref (var_ref array) (constant int (5)))
The Mesa IR that is generated from this is just as awesome as you
might expect.
Three changes are made to the way if-statements are lowered.
1. Two condition variables, if_to_cond_assign_then and
if_to_cond_assign_else, are created for each if-then-else structure.
The former contains the "positive" condition, and the later contains
the "negative" condtion. This change was implemented in the previous
patch.
2. Each condition variable is added to a hash-table when it is created.
3. When lowering an if-statement, assignments to existing condtion
variables get the current condition anded. This ensures that nested
condition variables are only set to true when the condition variable
for all outer if-statements is also true.
Changes #1 and #3 combine to ensure the correctness of the resulting
code.
4. When a condition assignment is encountered with a condition that is
a dereference of a previously added condition variable, the condition
is not modified.
Change #4 prevents the continuous accumulation of conditions on
assignments.
If the original if-statements were:
if (x) {
if (a && b && c && d && e) {
...
} else {
...
}
} else {
if (g && h && i && j && k) {
...
} else {
...
}
}
The lowered code will be
if_to_cond_assign_then@1 = x;
if_to_cond_assign_then@2 = a && b && c && d && e
&& if_to_cond_assign_then@1;
...
if_to_cond_assign_else@2 = !if_to_cond_assign_then
&& if_to_cond_assign_then@1;
...
if_to_cond_assign_else@1 = !if_to_cond_assign_then@1;
if_to_cond_assign_then@3 = g && h && i && j;
&& if_to_cond_assign_else@1;
...
if_to_cond_assign_else@3 = !if_to_cond_assign_then
&& if_to_cond_assign_else@1;
...
Depending on how instructions are emitted, there may be an extra
instruction due to the duplication of the '&&
if_to_cond_assign_{then,else}@1' on the nested else conditions. In
addition, this may cause some unnecessary register pressure since in
the simple case (where the nested conditions are not complex) the
nested then-condition variables are live longer than strictly
necessary.
Before this change, one of the shaders in the OpenGL ES2 conformance
test suite's acos_float_frag_xvary generated 348 Mesa IR instructions.
After this change it only generates 124. Many, but not all, of these
instructions would have also been eliminated by CSE.
Reviewed-by: Eric Anholt <eric@anholt.net>
Now the condition (for the then-clause) and the inverse condition (for
the else-clause) get written to separate temporary variables. In the
presence of complex conditions, this shouldn't result in more code
being generated. If the original if-statement was
if (a && b && c && d && e) {
...
} else {
...
}
The lowered code will be
if_to_cond_assign_then = a && b && c && d && e;
...
if_to_cond_assign_else = !if_to_cond_assign_then;
...
Reviewed-by: Eric Anholt <eric@anholt.net>
At least one of the invariants verified by IR validation concerns the
relative ordering of toplevel constructs in the IR: references to
global variables must come after the declarations of those global
variables.
Since linking affects the ordering of toplevel constructs in the IR,
it's possible that a bug in the linker will cause invalid IR to be
generated, even if all the pre-linked shaders are valid. (In fact,
such a bug was fixed by the previous commit.)
Bugs like this are easily masked by further optimization passes,
particularly inlining. So to make them easier to track down, this
patch addes an IR validation step right after linking, and before
final optimization occurs. The validation only occurs on debug
builds.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
When link_functions.cpp adds a new function to the final linked
program, it needs to add it after any global variable declarations
that the function refers to, otherwise the IR will be invalid (because
variable declarations must occur before variable accesses). The
easiest way to do that is to have the linker emit functions to the
tail of the final linked program.
The linker used to emit functions to the head of the final linked
program, in an effort to keep callees sorted before their callers.
However, this was not reliable: it didn't work for functions declared
or defined in the same compilation unit as main, for diamond-shaped
patterns in the call graph, or for some obscure cases involving
overloaded functions. And no code currently relies on this sort
order.
No Piglit regressions with i965 Ironlake.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
process_array_type() contains an assertion to verify that no IR
instructions are generated while processing the expression that
specifies the size of the array. This assertion needs to happen
_after_ checking whether the expression is constant. Otherwise we may
crash on an illegal shader rather than reporting an error.
Fixes piglit tests array-size-non-builtin-function.vert and
array-size-with-side-effect.vert.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Rearranged the logic for converting the ast for a function call to
hir, so that we constant fold before emitting any IR. Previously we
would emit some IR, and then only later detect whether we could
constant fold. The unnecessary IR would usually get cleaned up by a
later optimization step, however in the case of a builtin function
being used to compute an array size, it was causing an assertion.
Fixes Piglit test array-size-constant-relational.vert.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38625
The ast-to-hir conversion needs to emit function signatures in two
circumstances: when a function declaration (or definition) is
encountered, and when a built-in function is encountered.
To avoid emitting a function signature in an illegal place (such as
inside a function), emit_function() checked whether we were inside a
function definition, and if so, emitted the signature before the
function definition.
However, this didn't cover the case of emitting function signatures
for built-in functions when those built-in functions are called from
inside the constant integer expression that specifies the length of a
global array. This failed because when processing an array length, we
are emitting IR into a dummy exec_list (see process_array_type() in
ast_to_hir.cpp). process_array_type() later checks (via an assertion)
that no instructions were emitted to the dummy exec_list, based on the
reasonable assumption that we shouldn't need to emit instructions to
calculate the value of a constant.
This patch changes emit_function() so that it emits function
signatures at toplevel in all cases.
This partially fixes bug 38625
(https://bugs.freedesktop.org/show_bug.cgi?id=38625). The remainder
of the fix is in the patch that follows.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
opt_dead_functions contained a shortcut to skip processing the first
function's body, based on the assumption that IR functions are
topologically sorted, with callees always coming before their callers
(therefore the first function cannot contain any calls).
This assumption turns out not to be true in general. For example, the
following code snippet gets translated to IR that violates this
assumption:
void f();
void g();
void f() { g(); }
void g() { ... }
In practice, the shortcut didn't cause bugs because of a coincidence
of the circumstances in which opt_dead_functions is called:
(a) we do inlining right before dead function elimination, and
inlining (when successful) eliminates all calls.
(b) for user-defined functions, inlining is always successful, because
previous optimization passes (during compilation) have reduced
them to a form that is eligible for inlining.
(c) the function that appears first in the IR can't possibly call a
built-in function, because built-in functions are always emitted
before the function that calls them.
It seems unnecessarily fragile to have opt_dead_functions depend on
these coincidences. And the next patch in this series will break (c).
So I'm reverting the shortcut. The consequence will be a slight
increase in link time for complex shaders.
This reverts commit c75427f4c8.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Unlike C++, empty declarations such as
float;
should be valid. The spec is not explicit about this actually.
Some apps that generate their shader sources may rely on this. This was
noted when porting one of them to Linux from Windows.
Reviewed-by: Chad Versace <chad@chad-versace.us>
Note: this is a candidate for the 7.11 branch.
This prevents assertion failures in ralloc_strcat. The ralloc_free in
_mesa_free_shader_program_data can be omitted because freeing the
gl_shader_program in _mesa_delete_shader_program will take care of
this automatically.
A bunch of this code could use a refactor to use ralloc a bit more
effectively. A bunch of the things that are allocated with malloc and
owned by the gl_shader_program should be allocated with ralloc (using
the gl_shader_program as the context).
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
linker_warning is a new function. It's identical to linker_error
except that it doesn't set LinkStatus=false and it prepends "warning: "
on messages instead of "error: ".
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Remove the other places that set LinkStatus to false since they all
immediately follow a call to linker_error. The function linker_error
was previously known as linker_error_printf. The name was changed
because it may seem surprising that a printf function will set an
error flag.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
The previous formula for atan(x,y) returned a value of +/- pi whenever
|x|<0.0001, and used a formula based on atan(y/x) otherwise. This
broke in cases where both x and y were small (e.g. atan(1e-5, 1e-5)).
This patch modifies the formula so that it returns a value of +/- pi
whenever |x|<1e-8*|y|, and uses the formula based on atan(y/x)
otherwise.
The previous formula for asin(x) was algebraically equivalent to:
sign(x)*(pi/2 - sqrt(1-|x|)*(A + B|x| + C|x|^2))
where A, B, and C were arbitrary constants determined by a curve fit.
This formula had a worst case absolute error of 0.00448, an unbounded
worst case relative error, and a discontinuity near x=0.
Changed the formula to:
sign(x)*(pi/2 - sqrt(1-|x|)*(pi/2 + (pi/4-1)|x| + A|x|^2 + B|x|^3))
where A and B are arbitrary constants determined by a curve fit. This
has a worst case absolute error of 0.00039, a worst case relative
error of 0.000405, and no discontinuities.
I don't expect a significant performance degradation, since the extra
multiply-accumulate should be fast compared to the sqrt() computation.
Fixes piglit tests {vs,fs}-asin-float and {vs,fs}-atan-*
The function used a variable named 'score', which was an outright lie.
A signature matches or it doesn't; there is no fuzzy scoring.
Change the return type of parameter_lists_match() to an enum, and
let ir_function::matching_sigature() switch on that enum.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Chad Versace <chad@chad-versace.us>
Array constructors obey narrower conversion rules than other constructors
[1] --- they use the implicit conversion rules [2] instead of the scalar
constructor conversions [3]. But process_array_constructor() was
incorrectly applying the broader rules.
[1] GLSL 1.50 spec, Section 5.4.4 Array Constructors, page 52 (58 of pdf)
[2] GLSL 1.50 spec, Section 4.1.10 Implicit Conversions, page 25 (31 of pdf)
[3] GLSL 1.50 spec, Section 5.4.1 Conversion, page 48 (54 of pdf)
To fix this, first check (with glsl_type::can_be_implicitly_converted_to)
if an implicit conversion is legal before performing the conversion.
Fixes:
piglit:spec/glsl-1.20/compiler/structure-and-array-operations/array-ctor-implicit-conversion-bool-float.vert
piglit:spec/glsl-1.20/compiler/structure-and-array-operations/array-ctor-implicit-conversion-bvec*-vec*.vert
Note: This is a candidate for the 7.10 and 7.11 branches.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Chad Versace <chad@chad-versace.us>
The function is no longer used and has been replaced by
glsl_type::can_implicitly_convert_to().
Note: This is a candidate for the 7.10 and 7.11 branches.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Chad Versace <chad@chad-versace.us>
