[Build] LLVM-18

[hughperkins]

Issue: #

Brief Summary

Based heavily on taichi-dev/taichi#8735 by @johnnynunez

• revised approach to pass upgrade

copilot:summary

Walkthrough

copilot:walkthrough

[Copilot]

Pull Request Overview

This PR upgrades the LLVM version from 15.0.7 to 18.1.8, adapting the codebase to accommodate breaking changes in LLVM's API, particularly the transition from the legacy pass manager to the new pass manager infrastructure and the migration to opaque pointer types.

• Updates LLVM prebuilt binaries to version 18.1.8
• Migrates CUDA and CPU JIT compilation to use the new PassBuilder API with ModulePassManager
• Adapts pointer type handling for LLVM 18's opaque pointers using PointerType::getUnqual()
• Temporarily skips tests for quantized floats with 8 exponent bits

Reviewed Changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 13 comments.

Show a summary per file

File	Description
.github/workflows/scripts/ti_build/llvm.py	Updates LLVM version to 18.1.8 and build version, removes AMDGPU-specific path
CMakeLists.txt	Removes version detection code
gstaichi/runtime/llvm/llvm_context_pass.h	Comments out deprecated PassManagerBuilder include
gstaichi/runtime/cuda/jit_cuda.h	Adds new pass manager headers, removes legacy PassManagerBuilder
gstaichi/runtime/cuda/jit_cuda.cpp	Migrates to new PassBuilder API with ModulePassManager and updates to CodeGenOptLevel enum
gstaichi/runtime/cpu/jit_cpu.cpp	Updates to use toPtr() for ExecutorAddr conversion and Host.h location
gstaichi/codegen/cpu/codegen_cpu.cpp	Migrates to new PassBuilder API, updates to CodeGenOptLevel and CodeGenFileType enums
gstaichi/runtime/amdgpu/jit_amdgpu.cpp	Removes deprecated adjustPassManager call
gstaichi/rhi/cuda/cuda_context.cpp	Updates maximum compute capability from 86 to 90 for newer GPUs
gstaichi/codegen/llvm/struct_llvm.cpp	Replaces typed pointer methods with opaque PointerType::getUnqual()
gstaichi/codegen/llvm/llvm_codegen_utils.cpp	Updates pointer type comparison for opaque pointers
gstaichi/codegen/llvm/codegen_llvm.cpp	Updates pointer type creation and adds TI_NOT_IMPLEMENTED for deprecated path
gstaichi/codegen/cuda/codegen_cuda.cpp	Updates pointer type creation to use PointerType::getUnqual()
tests/python/test_quant_float_shared_exp.py	Adds pytest.skip for exponent_bits == 8 tests

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

With LLVM 18's opaque pointers, the updated logic only compares address spaces, but ignores that the function description (lines 12-23) indicates the function should check if types are renamed versions of each other. This change fundamentally breaks the original purpose of the function - it should be comparing pointer types more comprehensively, not just address spaces. The original recursive check via getPointerElementType() was checking type equivalence, but now two pointers with different pointee types will be considered the same as long as they have the same address space.

Suggested change

	return ptr_a->getAddressSpace() == ptr_b->getAddressSpace();
	if (ptr_a->getAddressSpace() != ptr_b->getAddressSpace()) {
	return false;
	}
	#if LLVM_VERSION_MAJOR >= 18
	// Opaque pointers: cannot compare element types directly
	if (ptr_a->isOpaque() || ptr_b->isOpaque()) {
	// Fallback to type name comparison for opaque pointers
	auto a_name = type_name(a);
	auto b_name = type_name(b);
	return a_name == b_name;
	}
	#endif
	// Non-opaque pointers: compare element types recursively
	return is_same_type(ptr_a->getPointerElementType(), ptr_b->getPointerElementType());

[hughperkins]

Note: I"m going to ignore the things copilot said, unless anyone bumps any of its comments.

[hughperkins]

though copilot makes a good point about amd code actually 🤔 Basically we cant support amd without having ci for amd though I feel.

[johnnynunez]

though copilot makes a good point about amd code actually 🤔 Basically we cant support amd without having ci for amd though I feel.

There are a lot of code optimizations i think.
I am not expert on this, i was only available to run on llvm 20.
But from taichi there was a lot of to do and ideas to do.
Upgrade to that version is a good point to support newer devices