| Issue |
76559
|
| Summary |
Missed optimization for `fdiv` followed by `fmul`
|
| Labels |
missed-optimization,
llvm
|
| Assignees |
|
| Reporter |
rikhuijzer
|
I wonder whether rewriting `@src` to `@trg` would be a valid optimization?
```ll
define float @src(float %0) {
%2 = fdiv float %0, 6.000000e+00
%3 = fmul float %2, 3.000000e+00
ret float %3
}
define float @trg(float %0) {
%2 = fdiv float %0, 2.000000e+00
ret float %2
}
```
since $\frac{x}{6.0} * 3.0 = \frac{x}{2.0}$.
I've tried to verify this via `alive-tv`, but it times out due to the floating point numbers. I've tried locally with `--smt-to=3600000` (1 hour timeout).
Intuitively, I would say that the optimization is valid because it should make the results **more** accurate.
For example, the result for 7664.71 is more accurate via `f2` (with the suggested optimization) than `f1` (without the suggested optimization) in Julia:
```julia
julia> f1(x) = x / 6.0 * 3.0;
julia> f1(0.2)
0.1
julia> f2(x) = x / 2.0;
julia> f2(0.2)
0.1
julia> @code_llvm dump_module=false debuginfo=:none f1(0.2)
define double @julia_f1_1438(double %0) #0 {
top:
%1 = fdiv double %0, 6.000000e+00
%2 = fmul double %1, 3.000000e+00
ret double %2
}
julia> @code_llvm dump_module=false debuginfo=:none f2(0.2)
define double @julia_f2_1440(double %0) #0 {
top:
%1 = fmul double %0, 5.000000e-01
ret double %1
}
julia> f1(7664.71) * 2
7664.709999999999
julia> f2(7664.71) * 2
7664.71
```
So would an optimization for $\frac{x}{a} * b$ be valid for certain `a` and `b`?
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
