[clang] [llvm] [InstCombine] Canonicalize `(sitofp x)` -> `(uitofp x)` if `x >= 0` (PR #82404)
via cfe-commits
cfe-commits at lists.llvm.org
Wed Mar 20 10:56:57 PDT 2024
asmok-g wrote:
Well, I'm not sure how proper that wold be as a reproducer,
I extracted the mentioned test to a program:
```
#include <cmath>
#include "third_party/swiftshader/src/Reactor/Coroutine.hpp"
#include "third_party/swiftshader/src/Reactor/Reactor.hpp"
using float4 = float[4];
using int4 = int[4];
static std::string testName() { return "test name"; }
using namespace rr;
int main() {
Function<Void(Pointer<Int4>, Pointer<Int4>)> function;
{
Pointer<Int4> x = function.Arg<0>();
Pointer<Int4> y = function.Arg<1>();
*y = Abs(*x);
}
auto routine = function(testName().c_str());
auto callable = (void (*)(int4 *, int4 *))routine->getEntry();
int input[] = {1, -1, 0, (int)0x80000000};
for (int x : input) {
int4 v_in = {x, x, x, x};
int4 v_out = {0};
callable(&v_in, &v_out);
float expected = abs(x);
if (v_out[0] != expected) {
return 10;
}
}
}
```
Before the change program exits with 0, after it exists with 10.
More precisely: `v_out[0]` is: -2147483648, `expected` is: 2.14748365e+09
[out_at.txt](https://github.com/llvm/llvm-project/files/14670227/out_at.txt)
[out_before.txt](https://github.com/llvm/llvm-project/files/14670228/out_before.txt)
I attached the generated IR before and after the change. It's only one line diff:
```
call void @llvm.memset.p0.i64(ptr noundef nonnull align 16 dereferenceable(16) %9, i8 0, i64 16, i1 false)
call void %49(ptr noundef nonnull %8, ptr noundef nonnull %9) #16
%61 = call i32 @llvm.abs.i32(i32 %60, i1 true)
- %62 = sitofp i32 %61 to float
+ %62 = uitofp i32 %61 to float
%63 = load i32, ptr %9, align 16
%64 = sitofp i32 %63 to float
%65 = fcmp une float %64, %62
```
generation command is `clang -O1 -fsized-deallocation '-std=gnu++20' -c pre.ii -emit-llvm -S -o /tmp/ir_before.ll`
I'm trying to reduce the preprocessed file, tho not sure i'll keep the semantics of the failure.
https://github.com/llvm/llvm-project/pull/82404
More information about the cfe-commits
mailing list