| Issue |
63883
|
| Summary |
Suboptimal code loading aligned odd size _BitInt
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
VoxSciurorum
|
A `_BitInt(17)` known to be aligned to 4 bytes can be loaded with a single memory reference. Code generation generates two, apparently because it treats the value as fundamentally 24 bits. I observe this on aarch64 (code below) and amd64. I expect it happens on any machine without a native 17 or 24 bit type.
IR example:
```
; ModuleID = 'extract17.c'
source_filename = "extract17.c"
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
target triple = "aarch64-unknown-freebsd"
define dso_local i32 @f(ptr nocapture noundef readonly %0) local_unnamed_addr {
%2 = load i17, ptr %0, align 4
%3 = sext i17 %2 to i32
ret i32 %3
}
```
Code looks like
```
ldrb w8, [x0, #2]
ldrh w9, [x0]
bfi w9, w8, #16, #16
sbfx w0, w9, #0, #17
ret
```
but should look like the corresponding code with a bit field of width 17,
```
ldr w8, [x0]
sbfx w0, w8, #0, #17
ret
```
C code:
```
struct Field { int value : 17; };
struct BitInt { _BitInt(17) value; int aligner; };
int f1(struct Field *sp) { return sp->value; }
int f2(struct BitInt *sp) { return sp->value; }
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
