[llvm-bugs] [Bug 43481] New: [MIPS] Missing ANDI optimization
via llvm-bugs
llvm-bugs at lists.llvm.org
Fri Sep 27 08:27:54 PDT 2019
https://bugs.llvm.org/show_bug.cgi?id=43481
Bug ID: 43481
Summary: [MIPS] Missing ANDI optimization
Product: libraries
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: enhancement
Priority: P
Component: Backend: MIPS
Assignee: unassignedbugs at nondot.org
Reporter: arichardson.kde at gmail.com
CC: llvm-bugs at lists.llvm.org
We noticed that clang generates unnecessary SLL instructions in one of our
benchmark hot paths (and apparently GCC does not): https://godbolt.org/z/xKcCqT
Consider
```
#include <stdint.h>
uint64_t foo(uint64_t a)
{
uint64_t b = a / 16;
uint64_t c = b & 0x7UL;
return ((uint64_t)1UL << c);
}
```
At present, the MIPS backend produces:
```
foo(unsigned long): # @foo(unsigned long)
sll $1, $4, 0
srl $1, $1, 4
andi $1, $1, 7
daddiu $2, $zero, 1
jr $ra
dsllv $2, $2, $1
```
The `sll` is introduced because the `srl` requires that its input be zero
extended (which, well, seems silly, but so it goes). In any case, because
`andi` has a 16-bit immediate, some arithmetic and lookahead could find that
`0x7 << 4` fits and so make this be
```
andi $1, $4, 0x70
srl $1, $1, 4
daddiu $2, $zero, 1
jr $ra
dsllv $2, $2, $1
```
Alternatively, whatever's concluding that it can use 32-bit values internally
could stop doing that and this could instead just be
```
dsrl $1, $4, 4
andi $1, $1, 7
daddiu $2, $zero, 1
jr $ra
dsllv $2, $2, $1
```
This problem was found by Nathaniel Wesley Filardo and reported as
https://github.com/CTSRD-CHERI/llvm-project/issues/343
--
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20190927/ba6be177/attachment.html>
More information about the llvm-bugs
mailing list