| Issue |
165647
|
| Summary |
TypeBasedAliasAnalysis induces incorrect optimization in SLP-VECTORIZER
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
ParkHanbum
|
C Code sample
https://godbolt.org/z/1avvGh7sx
```
void src_assign_to_ptr_trunc(unsigned short *s, unsigned short* t)
{
s[0] = t[0];
s[1] = t[1];
s[2] = t[2];
s[3] = t[3];
}
void src_assign_to_ptr_ext(unsigned int *s, unsigned short* t)
{
s[0] = t[0];
s[1] = t[1];
s[2] = t[2];
s[3] = t[3];
}
```
The example code above is currently optimized as vector operations in the clang trunk.
```
define dso_local void @tgt_assign_to_ptr_trunc(unsigned short*, unsigned int*)(ptr noundef writeonly captures(none) initializes((0, 8)) %0, ptr noundef readonly captures(none) %1) local_unnamed_addr #1 {
%3 = load <4 x i32>, ptr %1, align 4
%4 = trunc <4 x i32> %3 to <4 x i16>
store <4 x i16> %4, ptr %0, align 2
ret void
}
define dso_local void @tgt_assign_to_ptr_ext(unsigned int*, unsigned short*)(ptr noundef writeonly captures(none) initializes((0, 16)) %0, ptr noundef readonly captures(none) %1) local_unnamed_addr #1 {
%3 = load <4 x i16>, ptr %1, align 2
%4 = zext <4 x i16> %3 to <4 x i32>
store <4 x i32> %4, ptr %0, align 4
ret void
}
```
Alive link : https://alive2.llvm.org/ce/z/SYn2VG
However, through Alive2, we can see that such optimization is not correct.
On the other hand, if the target is a pointer of the same type rather than a different type, it does not optimize using vector operations.
```
void src_assign_to_ptr_same(unsigned int *s, unsigned int* t)
{
s[0] = t[0];
s[1] = t[1];
s[2] = t[2];
s[3] = t[3];
}
```
```
define dso_local void @src_assign_to_ptr_same(unsigned int*, unsigned int*)(ptr noundef writeonly captures(none) initializes((0, 16)) %0, ptr noundef readonly captures(none) %1) local_unnamed_addr #1 {
%3 = load i32, ptr %1, align 4
store i32 %3, ptr %0, align 4
%4 = getelementptr inbounds nuw i8, ptr %1, i64 4
%5 = load i32, ptr %4, align 4
%6 = getelementptr inbounds nuw i8, ptr %0, i64 4
store i32 %5, ptr %6, align 4
%7 = getelementptr inbounds nuw i8, ptr %1, i64 8
%8 = load i32, ptr %7, align 4
%9 = getelementptr inbounds nuw i8, ptr %0, i64 8
store i32 %8, ptr %9, align 4
%10 = getelementptr inbounds nuw i8, ptr %1, i64 12
%11 = load i32, ptr %10, align 4
%12 = getelementptr inbounds nuw i8, ptr %0, i64 12
store i32 %11, ptr %12, align 4
ret void
}
```
Here's my analysis of why this situation has arisen.
`SLP-VECTORIZER` checks dependencies between memory locations.
During this process, the internally called `AAResults::alias` function determines alias relationships
between two given memorylocations through registered `AliasAnalysis` instances.
TypeBasedAliasAnalysis is also called during this process to determine whether memory locations are aliases.
However, since TypeBaseAlias only checks the types of the two memory locations,
it determines that truncation/extension occurs and thus deems it a NoAlias case.
Consequently, the SLP-VECTORIZER determines there is no aliasing between the two memory locations
and proceeds with optimization.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
