Missing optimization (fold of ashr and sext) #3010
Description
| Bugzilla Link | 2638 |
| Resolution | FIXED |
| Resolved on | Nov 07, 2018 00:21 |
| Version | unspecified |
| OS | All |
| Attachments | Input .c file |
| Reporter | LLVM Bugzilla Contributor |
| CC | @lattner |
Extended Description
LLVM misses some optimization of the attached test case:
ddunbar@ddunbar2:CodeGen$ llvm-gcc -c --emit-llvm -o - ~/rt/gaz.c | opt -std-compile-opts | llvm-dis
; ModuleID = ''
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
target triple = "i386-apple-darwin9"
define i32 @foo(i32 %i) nounwind {
entry:
%tmp12 = trunc i32 %i to i8 ; [#uses=1]
%tmp16 = shl i8 %tmp12, 6 ; [#uses=1]
%sextr23 = ashr i8 %tmp16, 6 ; [#uses=1]
%sextr2324 = sext i8 %sextr23 to i32 ; [#uses=1]
ret i32 %sextr2324
}
For reference, here is the bytecode that comes from the IR generated by clang, after optimization:
-- clang --
; ModuleID = ''
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
target triple = "i686-apple-darwin9.4.0"
define i32 @foo(i32 %i) nounwind {
entry:
%tmp9 = shl i32 %i, 30 ; [#uses=1]
%tmp10 = ashr i32 %tmp9, 30 ; [#uses=1]
ret i32 %tmp10
}
The difference comes from the way the bitfield load is generated....
No clue if this is worthwhile.