scalar.fptrunc
← scalar dialect
Float precision truncation using round-to-nearest, ties-to-even. The result format must be strictly narrower than the input format. Special values follow the destination format: f8E4M3 saturates finite overflow and infinities to its signed maximum finite value while preserving NaNs; IEEE formats preserve infinities and NaNs. Without flags, the conversion preserves subnormal inputs and results. daz permits replacing a subnormal input with its signed zero before conversion; ftz permits replacing a subnormal result with its signed zero.
Operation contract
| Property |
Value |
| Semantic phase |
executable |
| Traits |
Pure, DistributionTransfer, SafeToSpeculate |
Signature
| Kind |
Name |
Type |
Cardinality |
Description |
| Operand |
input |
float |
required |
— |
| Result |
result |
float |
required |
— |
| Attribute |
subnormal |
flags FloatConversionFlags |
optional |
— |
Verification constraints
ElementWidthLessThan(result, input)
Examples
%result = scalar.fptrunc %input : f32 to f16