[dgryski]

A range loop over a slice has the condition (index + 1) < len(slice) with a signed compare, while the bounds check on the element is unsigned. LLVM cannot merge the two without knowing that the length is not negative. It keeps the check and adds a smax to the loop count.

The index starts at -1 and goes up by one while it is below the length, and a slice length is never negative. An unsigned compare gives the same result, so this change uses it for that loop condition. Then it matches the bounds check and LLVM removes the check. Other loops, loops over arrays and loops over strings are not changed.

Results

In the standard library, building the test binaries of the packages in make/test.mk and comparing the surviving checks with the ones gc keeps, this removes the check from 291 of the 292 range loop sites where gc removes it and LLVM did not.

At -opt=z a range loop with an early return goes from 21 to 15 instructions and loses its panic call. At -opt=2 LLVM already removed it in that case.

On a Seeeduino XIAO (SAMD21, Cortex-M0+ at 48 MHz, -opt=s) I measured cycles with SysTick. A range loop with a continue statement takes 11.1% fewer cycles, from 18,368 to 16,320. Three other range loops are unchanged and so are two control loops. The cycle counts are exactly repeatable and do not depend on code layout.

On a desktop I did not find a consistent speedup. Over 8 code layouts, some benchmarks in compress/flate and strings.Replacer are 2 to 4% faster at -opt=z, but unchanged code varies by more than that. The compress/lzw compression benchmark is 4 to 6% slower on wasip1 and at -opt=z, in every layout, and I do not know why. The machine code of the affected function is smaller.

Code size changes by -0.01% to -0.33% at -opt=z, -0.05% to -0.16% at -opt=s and -0.21% to +0.50% at -opt=2 on the programs I tried.

Testing

• Added a compiler test for the loop condition. It passes with LLVM 20, 21 and 23.
• Added runtime tests in testdata/slice.go for range loops over nil, empty and resliced slices, changing the slice or the index in the loop body, break and continue. The expected output comes from the reference Go toolchain. They pass on the host and on wasip1 at -opt=0, 1, 2, s and z.
• The compiler tests that already fail on dev with these LLVM versions are unchanged.
• I did not run the Cortex-M size tests locally. builder/testdata/binary-size.txt may need an update from the CI output.