miscompilation of i512 #3821
Description
| Bugzilla Link | 3449 |
| Resolution | DUPLICATE |
| Resolved on | Jan 31, 2009 02:34 |
| Version | 1.0 |
| OS | All |
| CC | @sunfishcode |
Extended Description
sumarrray-dbl now fails because SROA is promoting a struct to an i512 and codegen is shifting by the wrong amount. Here's a testcase:
define void @test(<8 x double> %P, i64 %Q) nounwind {
%A = load <8 x double>* %P
%B = bitcast <8 x double> %A to i512 ; [#uses=2]
%C = lshr i512 %B, 448 ; [#uses=1]
%D = trunc i512 %C to i64 ; [#uses=1]
volatile store i64 %D, i64* %Q
ret void
}
I get:
$ llvm-as < ~/t.ll | llc -march=x86-64
_test:
movq 24(%rdi), %rax
movq %rax, (%rsi)
ret
bit 448 = byte 56 not byte 24. It looks like it is off by 32 or maybe modulusing by 32 or something weird like that?