spirv: OpenCL printf support

[ivanstepanovftw]

I want to have std.debug.printf() implemented for GPGPU, because it lowers entry point for newcomers, and this feature allows to catch bugs faster (like I had with signed_integer - 0).

To test this PR you need to clone zig-spirv-test-executor in root dir and run:

./build/stage3/bin/zig build -p stage4 -Denable-llvm -Dno-lib
./stage4/bin/zig build --build-file ./zig-spirv-test-executor/build.zig
export RUSTICL_ENABLE=llvmpipe
./stage4/bin/zig test ./test/behavior/asm.zig \
    --test-runner ./zig-spirv-test-executor/src/test_runner.zig \
    -target spirv64-opencl-none \
    -mcpu generic+int64+float64+float16 \
    -fno-llvm \
    --test-cmd ./zig-spirv-test-executor/zig-out/bin/zig-spirv-test-executor \
    --test-cmd --platform \
    --test-cmd rusticl \
    --test-cmd-bin

[ivanstepanovftw]

Currently, PoCL works, rusticl does not (nor llvmpipe, nor radeonsi), probably because of extra OpPtrCastToGeneric (if so, then it is a Mesa bug).

---

Currently, std.fmt.bufPrint is commented because of cannot call function pointers and TODO (SPIR-V): switch with ranges errors.

Details

lib/std/io/Writer.zig:12:5: error: cannot call function pointers
pub fn write(self: Self, bytes: []const u8) anyerror!usize {
~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
lib/std/unicode.zig:26:5: error: TODO (SPIR-V): switch with ranges
pub fn utf8ByteSequenceLength(first_byte: u8) !u3 {
~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
lib/std/fmt.zig:1783:5: error: TODO (SPIR-V): switch with ranges
pub fn digitToChar(digit: u8, case: Case) u8 {
~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
lib/std/unicode.zig:1617:5: error: TODO (SPIR-V): switch with ranges
pub fn isSurrogateCodepoint(c: u21) bool {
~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

---

I am not sure about the design (Zig formatting and printf(%s, after_format)). I fear this approach might not be worth the effort.
Since it is just printf-debugging function, which is not implemented, this may be worth to have it like so, with Zig formatting and buffer.
(... if it will work given error "cannot call function pointers", with writergate as a possible solution)
@alichraghi, what do you think about this design, is this acceptable (at least for the first time), or it should look like comptime mapping Zig's fmt to expected C-like printf?

[alichraghi]

I suggest you to not integrate it with std.debug.print and instead have std.gpu.printf which doesn't try to integrate with any zig formatting and is just a simple call to OpenCL's printf.

[ivanstepanovftw]

Rebased.

[ivanstepanovftw]

#compiler > spirv printf discussing on how to replace this switch expression with comptime generated asm expression to allow as many arguments as possible

[ivanstepanovftw]

Some edge cases like this one with constant mixed with a variable do not compile, i.e.:

var a: i32 = 10;
_ = &a;
_ = printf("%d %d\n", .{ 10, a });

[alichraghi]

I have a few nitpicks that i can fix myself but otherwise LGTM. Hopefully we can switch to an std.format version later

[ivanstepanovftw]

Tell me when you want to merge this. Expect no further changes from me.