wafer-lsp 0.1.0__py3-none-any.whl → 0.1.2__py3-none-any.whl

wafer_lsp/parsers/hip_parser.py

@@ -0,0 +1,688 @@
+"""
+HIP Parser for wafer-lsp.
+
+Extracts GPU constructs from HIP (AMD GPU) source files:
+- __global__ kernels
+- __device__ helper functions  
+- __shared__ memory allocations (LDS)
+- Kernel launch sites (<<<>>> and hipLaunchKernelGGL)
+- Wavefront-sensitive patterns for diagnostics
+"""
+
+import re
+from dataclasses import dataclass, field
+from typing import Any
+
+from .base_parser import BaseParser
+
+
+@dataclass(frozen=True)
+class HIPParameter:
+    """A kernel or device function parameter."""
+    name: str
+    type_str: str
+    is_pointer: bool = False
+    is_const: bool = False
+
+
+@dataclass(frozen=True)
+class HIPKernel:
+    """A __global__ GPU kernel function."""
+    name: str
+    line: int
+    end_line: int
+    parameters: list[str]
+    parameter_info: list[HIPParameter]
+    attributes: list[str]  # e.g., __launch_bounds__(256)
+    docstring: str | None = None
+
+
+@dataclass(frozen=True)
+class HIPDeviceFunction:
+    """A __device__ helper function."""
+    name: str
+    line: int
+    end_line: int
+    parameters: list[str]
+    parameter_info: list[HIPParameter]
+    return_type: str = "void"
+    is_inline: bool = False
+
+
+@dataclass(frozen=True)
+class SharedMemoryAllocation:
+    """A __shared__ memory (LDS) allocation."""
+    name: str
+    type_str: str
+    line: int
+    size_bytes: int | None  # None if dynamic/unknown
+    array_size: str | None  # Array dimension if static
+    is_dynamic: bool = False
+
+
+@dataclass(frozen=True)
+class KernelLaunchSite:
+    """Where a kernel is launched."""
+    kernel_name: str
+    line: int
+    grid_dim: str | None  # Grid dimensions if determinable
+    block_dim: str | None  # Block dimensions if determinable
+    shared_mem_bytes: str | None  # Dynamic shared memory size
+    stream: str | None  # CUDA stream if specified
+    is_hip_launch_kernel_ggl: bool = False  # True if using hipLaunchKernelGGL
+
+
+@dataclass(frozen=True)
+class WavefrontPattern:
+    """A pattern that might indicate wavefront-size assumptions.
+    
+    These patterns often indicate code written for CUDA's 32-thread warps
+    that may behave incorrectly on AMD's 64-thread wavefronts.
+    """
+    pattern_type: str  # "warp_size_32", "ballot_32bit", "shuffle_mask", "lane_calc_32"
+    line: int
+    code_snippet: str
+    problematic_value: str  # The specific value causing concern (e.g., "32", "0xFFFFFFFF")
+    severity: str = "warning"  # "warning", "error", "info"
+
+
+class HIPParser(BaseParser):
+    """Parser for HIP (AMD GPU) source files.
+    
+    Extracts kernels, device functions, shared memory allocations,
+    launch sites, and wavefront-sensitive patterns using regex-based parsing.
+    
+    Why regex instead of full AST parsing:
+    - HIP/CUDA syntax is C++ which is complex to fully parse
+    - Key constructs (__global__, __device__, __shared__) are lexically distinct
+    - Fast enough for real-time IDE use (<100ms)
+    - Works on incomplete/invalid code during editing
+    """
+
+    # Regex patterns for HIP constructs
+    
+    # Matches __global__ function declarations
+    # Captures: return type, function name
+    # Note: __launch_bounds__ can be on the same line or line before
+    _KERNEL_PATTERN = re.compile(
+        r'(?:template\s*<[^>]*>\s*)?'  # Optional template
+        r'(?:__launch_bounds__\s*\([^)]+\)\s*[\n\r]*)?'  # Optional launch bounds before __global__
+        r'__global__\s+'
+        r'(?:__device__\s+)?'  # Optional __device__ (CUDA allows both)
+        r'(?:__launch_bounds__\s*\([^)]+\)\s*)?'  # Optional launch bounds after __global__
+        r'([\w\s\*&:<>,]+?)\s+'  # Return type (usually void)
+        r'(\w+)\s*'  # Function name (captured)
+        r'\(',  # Start of parameter list
+        re.MULTILINE | re.DOTALL
+    )
+    
+    # Separate pattern to detect __launch_bounds__ near __global__
+    _LAUNCH_BOUNDS_PATTERN = re.compile(
+        r'__launch_bounds__\s*\(([^)]+)\)',
+        re.MULTILINE
+    )
+    
+    # Matches __device__ function declarations (not __global__)
+    _DEVICE_FUNC_PATTERN = re.compile(
+        r'(?:template\s*<[^>]*>\s*)?'  # Optional template
+        r'__device__\s+'
+        r'(?!__global__)'  # Not followed by __global__
+        r'(__forceinline__\s+)?'  # Optional __forceinline__ (captured)
+        r'(?:__host__\s+)?'  # Optional __host__
+        r'(?:inline\s+)?'
+        r'([\w\s\*&:<>,]+?)\s+'  # Return type (captured)
+        r'(\w+)\s*'  # Function name (captured)
+        r'\(',
+        re.MULTILINE
+    )
+    
+    # Matches __shared__ variable declarations (including 2D arrays)
+    _SHARED_MEMORY_PATTERN = re.compile(
+        r'__shared__\s+'
+        r'([\w\s:<>,]+?)\s+'  # Type (captured)
+        r'(\w+)\s*'  # Variable name (captured)
+        r'(\[[^\]]*\](?:\s*\[[^\]]*\])?)?'  # Optional array size(s) including 2D (captured)
+        r'\s*[;=]',
+        re.MULTILINE
+    )
+    
+    # Matches kernel launches with <<<>>> syntax
+    _LAUNCH_SYNTAX_PATTERN = re.compile(
+        r'(\w+)\s*'  # Kernel name (captured)
+        r'<<<\s*'
+        r'([^,>]+)\s*,\s*'  # Grid dim (captured)
+        r'([^,>]+)'  # Block dim (captured)
+        r'(?:\s*,\s*([^,>]+))?'  # Optional shared mem (captured)
+        r'(?:\s*,\s*([^>]+))?'  # Optional stream (captured)
+        r'\s*>>>',
+        re.MULTILINE
+    )
+    
+    # Matches hipLaunchKernelGGL calls
+    _HIP_LAUNCH_PATTERN = re.compile(
+        r'hipLaunchKernelGGL\s*\(\s*'
+        r'(\w+)\s*,\s*'  # Kernel name (captured)
+        r'([^,]+)\s*,\s*'  # Grid dim (captured)
+        r'([^,]+)\s*,\s*'  # Block dim (captured)
+        r'([^,]+)\s*,\s*'  # Shared mem (captured)
+        r'([^,)]+)'  # Stream (captured)
+        r'(?:\s*,)?',  # Optional comma before args
+        re.MULTILINE
+    )
+    
+    # Patterns that suggest incorrect wavefront size assumptions (CUDA's 32 vs AMD's 64)
+    _WARP_SIZE_32_PATTERNS = [
+        # threadIdx.x < 32, threadIdx.x % 32, threadIdx.x & 31
+        (re.compile(r'threadIdx\s*\.\s*[xyz]\s*[<%&]\s*32\b'), "warp_size_32", "32"),
+        (re.compile(r'threadIdx\s*\.\s*[xyz]\s*&\s*31\b'), "warp_size_32", "31"),
+        (re.compile(r'threadIdx\s*\.\s*[xyz]\s*/\s*32\b'), "warp_size_32", "32"),
+        
+        # Lane/warp index calculations with hard-coded 32
+        (re.compile(r'\blaneId\s*=\s*[^;]*%\s*32\b'), "lane_calc_32", "32"),
+        (re.compile(r'\bwarpId\s*=\s*[^;]*/\s*32\b'), "lane_calc_32", "32"),
+        (re.compile(r'\blane\s*=\s*[^;]*&\s*31\b'), "lane_calc_32", "31"),
+        
+        # Ballot result compared to 32-bit mask
+        (re.compile(r'__ballot\s*\([^)]*\)\s*==\s*0x[Ff]{8}\b'), "ballot_32bit", "0xFFFFFFFF"),
+        (re.compile(r'__ballot\s*\([^)]*\)\s*!=\s*0x[Ff]{8}\b'), "ballot_32bit", "0xFFFFFFFF"),
+        (re.compile(r'__ballot\s*\([^)]*\)\s*&\s*0x[Ff]{8}\b'), "ballot_32bit", "0xFFFFFFFF"),
+        
+        # Shuffle operations with mask suggesting 32-thread warp
+        (re.compile(r'__shfl(?:_sync)?\s*\([^)]*,\s*0x[Ff]{8}\s*\)'), "shuffle_mask", "0xFFFFFFFF"),
+        # Match __shfl_down(val, offset, 32) or similar with explicit width=32
+        (re.compile(r'__shfl_down(?:_sync)?\s*\([^,]+,\s*[^,]+,\s*32\s*\)'), "shuffle_mask", "32"),
+        (re.compile(r'__shfl_up(?:_sync)?\s*\([^,]+,\s*[^,]+,\s*32\s*\)'), "shuffle_mask", "32"),
+        (re.compile(r'__shfl_xor(?:_sync)?\s*\([^,]+,\s*[^,]+,\s*32\s*\)'), "shuffle_mask", "32"),
+        (re.compile(r'__shfl(?:_sync)?\s*\([^,]+,\s*[^,]+,\s*32\s*\)'), "shuffle_mask", "32"),
+        
+        # activemask() compared to 32-bit value
+        (re.compile(r'__activemask\s*\(\s*\)\s*==\s*0x[Ff]{8}\b'), "ballot_32bit", "0xFFFFFFFF"),
+        
+        # Hard-coded warp size
+        (re.compile(r'#define\s+WARP_SIZE\s+32\b'), "warp_size_32", "32"),
+        (re.compile(r'const(?:expr)?\s+\w+\s+(?:warp|WARP)_?(?:size|SIZE)\s*=\s*32\b'), "warp_size_32", "32"),
+    ]
+
+    def parse_file(self, content: str) -> dict[str, Any]:
+        """Parse a HIP source file and extract GPU constructs.
+        
+        Args:
+            content: The source file content as a string.
+            
+        Returns:
+            Dictionary containing:
+            - kernels: List of HIPKernel
+            - device_functions: List of HIPDeviceFunction
+            - shared_memory: List of SharedMemoryAllocation
+            - launch_sites: List of KernelLaunchSite
+            - wavefront_patterns: List of WavefrontPattern
+        """
+        kernels = self._extract_kernels(content)
+        device_functions = self._extract_device_functions(content)
+        shared_memory = self._extract_shared_memory(content)
+        launch_sites = self._extract_launch_sites(content)
+        wavefront_patterns = self._extract_wavefront_patterns(content)
+        
+        return {
+            "kernels": kernels,
+            "device_functions": device_functions,
+            "shared_memory": shared_memory,
+            "launch_sites": launch_sites,
+            "wavefront_patterns": wavefront_patterns,
+        }
+
+    def _extract_kernels(self, content: str) -> list[HIPKernel]:
+        """Extract all __global__ kernel functions."""
+        kernels: list[HIPKernel] = []
+        
+        for match in self._KERNEL_PATTERN.finditer(content):
+            line = content[:match.start()].count('\n')
+            
+            attributes = []
+            
+            # Look for __launch_bounds__ in the matched region or nearby
+            # Search from a bit before the match to where __global__ ends
+            search_start = max(0, match.start() - 50)
+            search_end = match.end()
+            search_region = content[search_start:search_end]
+            
+            lb_match = self._LAUNCH_BOUNDS_PATTERN.search(search_region)
+            if lb_match:
+                attributes.append(f"__launch_bounds__({lb_match.group(1)})")
+            
+            kernel_name = match.group(2)
+            
+            # Extract parameters
+            params, param_info = self._extract_parameters(content, match.end() - 1)
+            
+            # Find end of function (approximate by finding matching brace)
+            end_line = self._find_function_end(content, match.end())
+            
+            # Extract docstring (comment immediately before the kernel)
+            docstring = self._extract_docstring(content, match.start())
+            
+            kernels.append(HIPKernel(
+                name=kernel_name,
+                line=line,
+                end_line=end_line,
+                parameters=params,
+                parameter_info=param_info,
+                attributes=attributes,
+                docstring=docstring,
+            ))
+        
+        return kernels
+
+    def _extract_device_functions(self, content: str) -> list[HIPDeviceFunction]:
+        """Extract all __device__ helper functions."""
+        device_funcs: list[HIPDeviceFunction] = []
+        
+        for match in self._DEVICE_FUNC_PATTERN.finditer(content):
+            # Skip if this is part of a __global__ __device__ combination
+            prefix = content[max(0, match.start() - 50):match.start()]
+            if '__global__' in prefix:
+                continue
+            
+            line = content[:match.start()].count('\n')
+            
+            # Group 1 is __forceinline__ (optional), group 2 is return type, group 3 is func name
+            forceinline_match = match.group(1)
+            return_type = match.group(2).strip()
+            func_name = match.group(3)
+            
+            # Check if inline (either captured in pattern or in prefix)
+            is_inline = bool(forceinline_match) or \
+                        '__forceinline__' in content[max(0, match.start() - 30):match.start()] or \
+                        'inline' in content[max(0, match.start() - 20):match.start()]
+            
+            # Extract parameters
+            params, param_info = self._extract_parameters(content, match.end() - 1)
+            
+            end_line = self._find_function_end(content, match.end())
+            
+            device_funcs.append(HIPDeviceFunction(
+                name=func_name,
+                line=line,
+                end_line=end_line,
+                parameters=params,
+                parameter_info=param_info,
+                return_type=return_type,
+                is_inline=is_inline,
+            ))
+        
+        return device_funcs
+
+    def _extract_shared_memory(self, content: str) -> list[SharedMemoryAllocation]:
+        """Extract all __shared__ memory declarations."""
+        shared_mem: list[SharedMemoryAllocation] = []
+        
+        for match in self._SHARED_MEMORY_PATTERN.finditer(content):
+            line = content[:match.start()].count('\n')
+            
+            type_str = match.group(1).strip()
+            var_name = match.group(2)
+            array_dims = match.group(3)  # Could be [n] or [n][m] for 2D
+            
+            # Clean up the array dimension string
+            array_size = None
+            if array_dims:
+                # Remove brackets and combine dimensions
+                array_size = array_dims.strip()
+            
+            # Try to compute size in bytes
+            size_bytes = self._estimate_shared_mem_size(type_str, array_size)
+            
+            # Check if it's dynamic (extern __shared__)
+            is_dynamic = 'extern' in content[max(0, match.start() - 20):match.start()]
+            
+            shared_mem.append(SharedMemoryAllocation(
+                name=var_name,
+                type_str=type_str,
+                line=line,
+                size_bytes=size_bytes,
+                array_size=array_size,
+                is_dynamic=is_dynamic,
+            ))
+        
+        return shared_mem
+
+    def _extract_launch_sites(self, content: str) -> list[KernelLaunchSite]:
+        """Extract all kernel launch sites."""
+        launch_sites: list[KernelLaunchSite] = []
+        
+        # <<<>>> syntax
+        for match in self._LAUNCH_SYNTAX_PATTERN.finditer(content):
+            line = content[:match.start()].count('\n')
+            
+            launch_sites.append(KernelLaunchSite(
+                kernel_name=match.group(1),
+                line=line,
+                grid_dim=match.group(2).strip() if match.group(2) else None,
+                block_dim=match.group(3).strip() if match.group(3) else None,
+                shared_mem_bytes=match.group(4).strip() if match.group(4) else None,
+                stream=match.group(5).strip() if match.group(5) else None,
+                is_hip_launch_kernel_ggl=False,
+            ))
+        
+        # hipLaunchKernelGGL
+        for match in self._HIP_LAUNCH_PATTERN.finditer(content):
+            line = content[:match.start()].count('\n')
+            
+            launch_sites.append(KernelLaunchSite(
+                kernel_name=match.group(1),
+                line=line,
+                grid_dim=match.group(2).strip() if match.group(2) else None,
+                block_dim=match.group(3).strip() if match.group(3) else None,
+                shared_mem_bytes=match.group(4).strip() if match.group(4) else None,
+                stream=match.group(5).strip() if match.group(5) else None,
+                is_hip_launch_kernel_ggl=True,
+            ))
+        
+        return launch_sites
+
+    def _extract_wavefront_patterns(self, content: str) -> list[WavefrontPattern]:
+        """Extract patterns that might indicate incorrect wavefront size assumptions.
+        
+        AMD GPUs use 64-thread wavefronts (CDNA) or configurable 32/64 (RDNA),
+        while CUDA uses 32-thread warps. Code written for CUDA may have
+        hard-coded assumptions about warp size that break on AMD.
+        """
+        patterns: list[WavefrontPattern] = []
+        lines = content.split('\n')
+        
+        for i, line in enumerate(lines):
+            # Skip comments
+            stripped = line.strip()
+            if stripped.startswith('//') or stripped.startswith('/*'):
+                continue
+            
+            # Skip lines that use warpSize (correct portable code)
+            if 'warpSize' in line or '__AMDGCN_WAVEFRONT_SIZE__' in line:
+                continue
+            
+            for pattern_re, pattern_type, problematic_value in self._WARP_SIZE_32_PATTERNS:
+                match = pattern_re.search(line)
+                if match:
+                    patterns.append(WavefrontPattern(
+                        pattern_type=pattern_type,
+                        line=i,
+                        code_snippet=line.strip(),
+                        problematic_value=problematic_value,
+                    ))
+        
+        return patterns
+
+    def _extract_parameters(self, content: str, paren_start: int) -> tuple[list[str], list[HIPParameter]]:
+        """Extract function parameters starting from the opening parenthesis.
+        
+        Returns:
+            Tuple of (parameter names list, detailed parameter info list)
+        """
+        if paren_start >= len(content) or content[paren_start] != '(':
+            return [], []
+        
+        # Find matching closing paren
+        depth = 0
+        param_end = paren_start
+        
+        for i in range(paren_start, len(content)):
+            char = content[i]
+            if char == '(':
+                depth += 1
+            elif char == ')':
+                depth -= 1
+                if depth == 0:
+                    param_end = i
+                    break
+        
+        if param_end == paren_start:
+            return [], []
+        
+        param_str = content[paren_start + 1:param_end]
+        
+        # Parse parameters handling templates and nested parens
+        params: list[str] = []
+        param_info: list[HIPParameter] = []
+        current_param = ""
+        template_depth = 0
+        paren_depth = 0
+        
+        for char in param_str:
+            if char == '<':
+                template_depth += 1
+                current_param += char
+            elif char == '>':
+                template_depth -= 1
+                current_param += char
+            elif char == '(':
+                paren_depth += 1
+                current_param += char
+            elif char == ')':
+                paren_depth -= 1
+                current_param += char
+            elif char == ',' and template_depth == 0 and paren_depth == 0:
+                param = self._parse_single_parameter(current_param.strip())
+                if param:
+                    params.append(param.name)
+                    param_info.append(param)
+                current_param = ""
+            else:
+                current_param += char
+        
+        # Handle last parameter
+        if current_param.strip():
+            param = self._parse_single_parameter(current_param.strip())
+            if param:
+                params.append(param.name)
+                param_info.append(param)
+        
+        return params, param_info
+
+    def _parse_single_parameter(self, param_str: str) -> HIPParameter | None:
+        """Parse a single parameter string into HIPParameter.
+        
+        Returns None if param_str cannot be parsed into a valid parameter.
+        """
+        if not param_str:
+            return None
+        
+        # Remove default value if present
+        if '=' in param_str:
+            param_str = param_str.split('=')[0].strip()
+        
+        is_const = 'const' in param_str
+        is_pointer = '*' in param_str or '&' in param_str
+        
+        # Extract name (last token after removing type qualifiers)
+        parts = param_str.replace('*', ' * ').replace('&', ' & ').split()
+        
+        if not parts:
+            return None
+        
+        # Find the parameter name (last identifier)
+        name = parts[-1].strip('*&')
+        
+        # If we only have one part, we can't distinguish type from name
+        # This could be a parameter like `void` (no name) or incomplete code
+        if len(parts) == 1:
+            # Assume it's just a name with unknown type
+            type_str = ""
+        else:
+            # Reconstruct type from all parts except the last (name)
+            type_parts = parts[:-1]
+            type_str = ' '.join(type_parts).replace(' * ', '*').replace(' & ', '&')
+        
+        return HIPParameter(
+            name=name,
+            type_str=type_str,
+            is_pointer=is_pointer,
+            is_const=is_const,
+        )
+
+    def _find_function_end(self, content: str, start_pos: int) -> int:
+        """Find the line number where a function ends (closing brace).
+        
+        Returns:
+            Line number of closing brace, or start line if function body not found.
+            
+        Note: Returns start line (not an approximation) when closing brace cannot be found.
+        This happens for incomplete code during editing - the caller should handle this case.
+        """
+        start_line = content[:start_pos].count('\n')
+        
+        # Find opening brace
+        brace_pos = content.find('{', start_pos)
+        if brace_pos == -1:
+            # No function body found (e.g., forward declaration or incomplete code)
+            return start_line
+        
+        # Find matching closing brace
+        depth = 1
+        for i in range(brace_pos + 1, len(content)):
+            if content[i] == '{':
+                depth += 1
+            elif content[i] == '}':
+                depth -= 1
+                if depth == 0:
+                    return content[:i].count('\n')
+        
+        # Unbalanced braces - incomplete code during editing
+        # Return start line to indicate we couldn't determine the end
+        return start_line
+
+    def _extract_docstring(self, content: str, func_start: int) -> str | None:
+        """Extract documentation comment before a function."""
+        # Look backwards for a comment block
+        search_start = max(0, func_start - 500)
+        prefix = content[search_start:func_start].rstrip()
+        
+        # Check for /// or /** style comments
+        lines = prefix.split('\n')
+        doc_lines: list[str] = []
+        
+        for line in reversed(lines):
+            stripped = line.strip()
+            if stripped.startswith('///'):
+                doc_lines.insert(0, stripped[3:].strip())
+            elif stripped.startswith('*') and not stripped.startswith('*/'):
+                doc_lines.insert(0, stripped[1:].strip())
+            elif stripped.startswith('/**'):
+                doc_lines.insert(0, stripped[3:].strip())
+                break
+            elif stripped.endswith('*/'):
+                doc_lines.insert(0, stripped[:-2].strip())
+            elif stripped == '' or stripped.startswith('//'):
+                continue
+            else:
+                # Not a doc comment, stop
+                if doc_lines:
+                    break
+        
+        if doc_lines:
+            return '\n'.join(doc_lines)
+        return None
+
+    def _estimate_shared_mem_size(self, type_str: str, array_size: str | None) -> int | None:
+        """Estimate shared memory size in bytes.
+        
+        Handles both 1D (e.g., [256]) and 2D (e.g., [16][16]) arrays.
+        Returns None if size cannot be determined (e.g., non-literal dimension).
+        
+        Why no eval(): We only handle literal integer dimensions and simple multiplication.
+        Variable-based dimensions (e.g., [BLOCK_SIZE]) return None - that's correct behavior
+        since we can't know the value at parse time.
+        """
+        # Size mapping for common types
+        type_sizes = {
+            'char': 1, 'int8_t': 1, 'uint8_t': 1,
+            'short': 2, 'int16_t': 2, 'uint16_t': 2, 'half': 2, '__half': 2,
+            'int': 4, 'int32_t': 4, 'uint32_t': 4, 'float': 4, 'unsigned': 4,
+            'long': 8, 'int64_t': 8, 'uint64_t': 8, 'double': 8, 'long long': 8,
+            'float4': 16, 'float2': 8, 'int4': 16, 'int2': 8,
+            'double2': 16, 'double4': 32,
+        }
+        
+        # Find base type
+        base_type = type_str.strip()
+        type_size: int | None = None
+        for known_type, size in type_sizes.items():
+            if known_type in base_type:
+                type_size = size
+                break
+        
+        if type_size is None:
+            return None
+        
+        if not array_size:
+            return type_size
+        
+        # Parse array dimensions - could be [n] or [n][m]
+        # Extract all bracketed dimensions
+        dims = re.findall(r'\[([^\]]+)\]', array_size)
+        if not dims:
+            return type_size
+        
+        total_elements = 1
+        for dim in dims:
+            dim_value = self._parse_dimension_expression(dim.strip())
+            if dim_value is None:
+                # Non-literal dimension (e.g., variable) - cannot determine size
+                return None
+            total_elements *= dim_value
+        
+        return type_size * total_elements
+
+    def _parse_dimension_expression(self, expr: str) -> int | None:
+        """Parse a dimension expression safely without eval().
+        
+        Handles:
+        - Simple integers: "256"
+        - Simple multiplication: "16 * 16", "BLOCK_SIZE * 4" (only if all literal)
+        
+        Returns None for anything we can't safely parse (variables, complex expressions).
+        This is correct behavior - we don't know variable values at parse time.
+        """
+        expr = expr.strip()
+        
+        # Simple integer
+        if expr.isdigit():
+            return int(expr)
+        
+        # Handle expressions with * (multiplication only)
+        if '*' in expr:
+            parts = expr.split('*')
+            result = 1
+            for part in parts:
+                part = part.strip()
+                if not part.isdigit():
+                    # Contains a variable or non-integer - cannot evaluate
+                    return None
+                result *= int(part)
+            return result
+        
+        # Not a simple literal - could be a variable like BLOCK_SIZE
+        return None
+
+
+def is_hip_file(content: str) -> bool:
+    """Check if content appears to be a HIP file based on content markers."""
+    hip_markers = [
+        '#include <hip/hip_runtime.h>',
+        '#include "hip/hip_runtime.h"',
+        '#include <hip/hip_runtime_api.h>',
+        'hipMalloc',
+        'hipMemcpy',
+        'hipFree',
+        'hipLaunchKernelGGL',
+        'hipDeviceSynchronize',
+        '__HIP_PLATFORM_AMD__',
+        '__HIP_PLATFORM_HCC__',
+        'HIP_KERNEL_NAME',
+    ]
+    
+    content_lower = content.lower()
+    for marker in hip_markers:
+        if marker.lower() in content_lower:
+            return True
+    
+    return False