[alexreinking]

constant_integer_bounds had no case for Select, so a select fell through to the full bounds of its type. A select always evaluates to one of its two values, so this PR bounds it by the union of the two values' bounds (the condition is ignored). The three-argument Call::if_then_else gets the same treatment; the two-argument form, which has no else value, is unchanged.

This matters for code that decodes small integer codes through selects and then needs constant_integer_bounds to prove that a later integer sum can't overflow. The GGML app (#9511) does this for ternary (tq1_0) weights, and currently works around the gap by replacing each select with a variable bound in a Scope to the union of its values' bounds.

correctness_constant_interval gains test_constant_integer_bounds_of_select(), which fails before this change and passes after it. It covers:

• a scalar select: select(x > 0, int(u8), -3) → [-3, 255];
• a nested select;
• a vector (int16x8) select;
• a select with one unbounded value (full range);
• a three-argument if_then_else.

A comment in correctness_struct_type that said bounds_helper had no Select case is updated.

Stacked on #9599.

Checklist

• Tests added or updated (not required for docs, CI config, or typo fixes)
• Documentation updated (if public API changed)
• Python bindings updated (if public API changed)
• Benchmarks are included here if the change is intended to affect performance.
• Commits include AI attribution where applicable (see Code of Conduct)

Tested on macOS arm64 with LLVM 22:

• ctest -R 'correctness_(bound|bounds|constant_interval|struct_type|bits_known|lossless_cast|intrinsics|interval|many_inlined_selects|multi_way_select|tuple_select|popc_clz_ctz|overflow|simd_op_check_(arm|x86|wasm|hvx))|fuzz_lossless_cast': 39/39 passed.
• ctest -L correctness, excluding GPU and Hexagon tests: 618 of 621 passed.
  • correctness_cross_compilation failed because the local LLVM build has no Hexagon target.
  • The two auto_schedule_no_* tests failed because the autoscheduler plugin wasn't built; both pass once it is.

[codecov]

Codecov Report

❌ Patch coverage is 83.33333% with 1 line in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (alexreinking/prefetch-fixes@d645519). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
src/ConstantBounds.cpp	83.33%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@                      Coverage Diff                       @@
##             alexreinking/prefetch-fixes    #9604   +/-   ##
==============================================================
  Coverage                               ?   72.95%           
==============================================================
  Files                                  ?      263           
  Lines                                  ?    82494           
  Branches                               ?    20179           
==============================================================
  Hits                                   ?    60184           
  Misses                                 ?    15914           
  Partials                               ?     6396           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.