% of two integers of any width is an integer remainder - #852

What

% of two integers of any width is now an integer remainder with an integer result. Since #571 that was true only when one operand was 64-bit; two narrower integers were cast to number.

Why

Found in the default library: Map and Set compute <uint>hashCode % <uint>buckets.length on every lookup. That compiled to two uitofp, an frem (an fmod call) and an fptosi. Modulo a length of 0 it was NaN converted to an index, which is undefined (ASDAlexander77/TypeScriptCompilerDefaultLib#100 removed the zero length). It is now:

%8 = urem i32 %1, %7

The type of the result

• An integer literal that fits takes the other operand's type, on either side (u % 7 is a u32).
• Otherwise the wider operand's type; of two of the same width and signedness, the left one.
• A signed and an unsigned operand that an s64 holds both of are done in s64, so neither changes its value (s32 % u32, s16 % u16).
• 64-bit operands are as they were after TSLANG_ISSUES 16: % of 64-bit integers is an integer remainder #571.
• A number on either side is still a number's remainder.

Breaking

• Result type: int % int is an integer, not a number. own/own_interface_view.ts put i % 10 into an object literal's field seen through an interface declaring number, which no longer compiled; it now declares the constant : number. That was the only test affected.
• Zero divisor: undefined, as integer division by zero is in C (a fault on x86), where it was NaN. The same holds for the least value of a signed type % -1. This is what TSLANG_ISSUES 16: % of 64-bit integers is an integer remainder #571 already did for 64-bit operands.

Change

• tslang/lib/TypeScript/MLIRGenImpl.h: the % case of the cast step in adjustTypesForBinaryOp. The lowering is unchanged: it already picks remsi/remui by type.
• tslang/test/tester/tests/00int_remainder.ts (new, registered for compile and JIT): s8 to u32, literals on either side, mixed signedness, mixed widths, %=, what >>> gives, and the number path.
• 00int64_remainder.ts: one comment that said narrower integers keep the number path.

Testing

• Windows x64, release: the whole ctest suite passes, 7244 of 7244.
• The default library rebuilt with this compiler for gc, rc, none and own, and its suite passes on all four, compiled and under the JIT. It needs no source change.
• Speed: a loop of Map.get over 200,000 keys went from about 18 ms to about 17 ms. That is small and close to noise.
• Not run locally: debug, Linux, x86, wasm.

🤖 Generated with Claude Code