| Issue |
114727
|
| Summary |
Missing fold of subtract and comparison by using sign flag on ARM64 and x86-64
|
| Labels |
|
| Assignees |
|
| Reporter |
chandlerc
|
Full example:
https://cpp.compiler-explorer.com/z/GaPEch7dj
For the code in the example, LLVM generates the following ARM64 assembly:
```
make_id(T*, long):
ldr w9, [x0, x1, lsl #2]
mov w8, #-2097152
asr w9, w9, #9
sub w8, w8, w9
cmn w9, #512, lsl #12
csel w0, w9, w8, gt
ret
```
While I would expect, and GCC generates, something that reuses the `sub`:
```
make_id(T*, long):
ldr w1, [x0, x1, lsl 2]
mov w2, -2097152
asr w1, w1, 9
subs w0, w2, w1
csel w0, w0, w1, pl
ret
```
The same failure also occurs on x86-64, with LLVM generating:
```
make_id(T*, long):
mov ecx, dword ptr [rdi + 4*rsi]
sar ecx, 9
mov eax, -2097152
sub eax, ecx
cmp ecx, -2097151
cmovge eax, ecx
ret
```
While I would expect, and GCC generates:
```
make_id(T*, long):
mov edx, DWORD PTR [rdi+rsi*4]
mov eax, -2097152
sar edx, 9
sub eax, edx
cmovs eax, edx
ret
```
The LLVM IR for x86-64 (should be the same for ARM64):
```llvm
%struct.T = type { i32 }
define dso_local noundef range(i32 -2097151, 4194304) i32 @make_id(T*, long)(ptr nocapture noundef readonly %data, i64 noundef %i) local_unnamed_addr {
entry:
%arrayidx = getelementptr inbounds %struct.T, ptr %data, i64 %i
%bf.load = load i32, ptr %arrayidx, align 4
%shr = ashr i32 %bf.load, 9
%cmp = icmp sgt i32 %shr, -2097152
%sub = sub nuw nsw i32 -2097152, %shr
%retval.0 = select i1 %cmp, i32 %shr, i32 %sub
ret i32 %retval.0
}
```
Note that the `nuw` and `nsw` flags are inferred here due to the `ashr`.
The original C++ code in question:
```cpp
#include <assert.h>
#include <stdint.h>
#include <sys/types.h>
enum E : uint8_t {
V1,
V2,
};
struct T {
E e : sizeof(E) * 8;
bool b : 1;
unsigned payload : 23;
};
constexpr int32_t InvalidId = 0b1111'1111'1110'0000'0000'0000'0000'0000;
int32_t make_id(T* data, ssize_t i) {
uint32_t payload = data[i].payload;
constexpr int shift = 32 - 23;
int32_t ext_id = static_cast<int32_t>(payload << shift) >> shift;
int32_t index = static_cast<uint32_t>(InvalidId) - static_cast<uint32_t>(ext_id);
if (index < 0) {
assert(ext_id > InvalidId);
return ext_id;
}
return index;
}
```
---
An attempt at simplifying the test case produces somewhat bizarre results. Not sure what to make of these.
https://cpp.compiler-explorer.com/z/PGYb7vsGs
Here the C++ code is:
```cpp
#include <assert.h>
#include <stdint.h>
#include <sys/types.h>
constexpr int32_t InvalidId = 0b1111'1111'1110'0000'0000'0000'0000'0000;
int32_t make_id(int32_t* data, ssize_t i) {
int32_t ext_id = data[i];
#ifdef __clang__
__builtin_assume(ext_id <= 0b0000'0000'0011'1111'1111'1111'1111'1111);
#endif
int32_t index = static_cast<uint32_t>(InvalidId) - static_cast<uint32_t>(ext_id);
if (index < 0) {
assert(ext_id > InvalidId);
return ext_id;
}
return index;
}
```
And the LLVM IR (for x86-64) is:
```llvm
define dso_local noundef range(i32 -2097151, 2145386497) i32 @make_id(int*, long)(ptr nocapture noundef readonly %data, i64 noundef %i) local_unnamed_addr {
entry:
%arrayidx = getelementptr inbounds i32, ptr %data, i64 %i
%0 = load i32, ptr %arrayidx, align 4
%cmp = icmp slt i32 %0, 4194304
tail call void @llvm.assume(i1 %cmp)
%cmp1 = icmp sgt i32 %0, -2097152
%sub = sub nuw nsw i32 -2097152, %0
%retval.0 = select i1 %cmp1, i32 %0, i32 %sub
ret i32 %retval.0
}
declare void @llvm.assume(i1 noundef) #1
```
Without the `assume`, LLVM thinks that the `assert` above can fire... Which I think is correct? But weirdly, GCC deletes the assert...
But even setting that aside, I'm a bit surprised at the assume being enough for LLVM to infer both `nuw` and `nsw` on the `sub` here. Because `INT_MIN` seems like it isn't precluded by the assume and yet I feel like the subtract somewhat has to wrap if `%0` is `INT_MIN`. So that seems like it could be a miscompile unless I've misremembered how one of the no-wrap flags works.
And after all of that, LLVM still generates the redundant comparison on bath x86-64 and ARM64:
```
make_id(int*, long):
ldr w9, [x0, x1, lsl #2]
mov w8, #-2097152
sub w8, w8, w9
cmn w9, #512, lsl #12
csel w0, w9, w8, gt
ret
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
