[AOT] Fix too long filename generation (#8738)

This change addresses the issue that when there is a complex kernel with
very long signature list, the generated filename possibly would get too
long to be recognized by python pathlib.

The out filename is constructed as
`kernel_name`.`8-char_sig_hash`_`suffix`.`ext`. `suffix` represents
signature hints in the form `param_id+c/d`, it holds all indexes to map
to the signature list.
This change tries to skip constant params in the signature list to
shorten the `suffix`, since they do not carry hints or extra
information.

This example show how this change work on a kernel with 71 parameters in
which there are 40 constants.
Before
```
kernel_out_name.f84b6f9a_0d1d2d3d456d7d8910111213d14d15d1617d18d19d2021d22d23d2425d26d27d2829d30d31323334353637383940414243444546474849505152535455565758596061626364656667686970.h
```
After
```
kernel_out_name.f84b6f9a_0d1d2d3d6d7d13d14d15d17d18d19d21d22d23d25d26d27d29d30d.h
```

Author: leonling-ll

python/triton/tools/compile.py

@@ -163,11 +163,10 @@ def constexpr(s):
     # dump C stub code
     suffix = ''
     for i, ty in enumerate(signature.values()):
-        suffix += str(i)
         if hints.get((i, ), None) == 1:
-            suffix += 'c'
+            suffix += f'{i}c'
         if hints.get((i, ), None) == 16:
-            suffix += 'd'
+            suffix += f'{i}d'
     func_name = '_'.join([out_name, sig_hash, suffix])
     asm = ccinfo.asm[backend.binary_ext]  # store binary data once
 

python/triton/tools/link.py

@@ -34,7 +34,7 @@ def __init__(self) -> None:
         # [kernel_name, c signature]
         self.linker_directives = re.compile("//[\\s]*tt-linker:[\\s]*([\\w]+):(.+):(.+)")
         # [name, hash, suffix]
-        self.kernel_name = re.compile("^([\\w]+)_([\\w]+)_([\\w]+)$")
+        self.kernel_name = re.compile("^([\\w]+)_([\\w]+)_([\\w]*)$")
         # [(type, name)]
         self.c_sig = re.compile("[\\s]*(\\w+)\\s(\\w+)[,]?")
         # [d|c]
@@ -88,22 +88,24 @@ def _match_suffix(self, suffix: str, c_sig: str):
         s2i = {"c": 1, "d": 16}
         num_specs = 0
         sizes = []
-        # scan through suffix, first find the index,
-        # then see if it is followed by d or c
+        # scan through suffix, suffix only includes indexes followed by d or c.
         for i in range(len(args)):
-            pos = suffix.find(str(i))
-            if pos == -1:
-                raise LinkerError(f"{suffix} is not a valid kernel suffix")
+            pos = 0
+            idx_matched = suffix.startswith(str(i))
+            if not idx_matched:
+                continue
             pos += len(str(i))
             if self.arg_suffix.match(suffix, pos):
                 num_specs += 1
                 sizes.extend([None] * (i - len(sizes)))
                 sizes.append(s2i[suffix[pos]])
                 pos += 1
-            if i < len(args) - 1:
-                suffix = suffix[pos:]
-            else:
-                sizes.extend([None] * (len(args) - len(sizes)))
+            suffix = suffix[pos:]
+
+        if len(suffix) > 0:
+            raise Exception(f"Has invalid extra suffix: {suffix}")
+        sizes.extend([None] * (len(args) - len(sizes)))
+
         return num_specs, sizes
 
     def _add_kernel(self, name: str, ker: KernelLinkerMeta):