Force Inline all the InterlockedAdd functions in stdlib (#3965)

This change forcibly inlines the InterlockedAdd functions when using byteAddress buffer.

The IR generated when using nonUniformResourceInst on RWByteAddressBuffer:
    buffer[NonUniformResourceIndex(uint(0))].InterlockedAdd(0, 1);

follows the sequence of a call into an index lookup that is wrapped by a nonuniformResourceIndex:
    %ld = nonUniformResourceIndex(0)
    Call RWStructBufferInterlockedAdd(%ld, 0, 1)

This prevents NonUniformResource decoration of the buffer because it is wrapped by the function call to
InterlockedAdd, that further expands to:
    %gep = getElement(%buffer, 0)
    SpirvAsmInst(..., rwStructuredBufferGEP(%gep, 0), ...)

By Force-Inlining the atomic functions, the buffer / resource is made visible to the nonUniformResourceIndex inst,
allowing the decoration.
Identified while debugging tests/spirv/coherent-2.slang

0 files changed, 0 insertions, 0 deletions