llvm dev - Dec 2016 - llvm.trap() from -O1 optimization

Hello,

This is likely to be some small issue, but I am unable to figure out
the solution. The following code, when run through `opt' with the -O1
flag, replaces the call to foo() with a call to llvm.trap() followed
by an `unreachable' instruction.

====================target datalayout =
"e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
target triple = "armv7-unknown-linux-gnu"

define arm_aapcs_vfpcc void @foo() {
entry:
    ret void
}

define arm_aapcs_vfpcc void @bar() {
entry:
    call void @foo()
    ret void
}
====================
If I drop the "arm_aapcs_vfpcc" for foo(), then the `unreachable'
instruction and the call to trap() are no longer introduced in the
code.

I am running `opt' on a 64-bit x86 machine. I have synced llvm, clang,
and compiler-rt to the latest commits, but that did not help.

$ clang --version
clang version 4.0.0 (https://github.com/llvm-mirror/clang.git
631f1e728a6903f320a40e7b5035a63acd0acdc7)
(https://github.com/llvm-mirror/llvm.git
23156f19246677a896ffbda72eadbcbbc1c47964)

Any ideas on how to prevent the unreachable instruction? Thanks!

Ashay

> On Dec 13, 2016, at 9:31 PM, Ashay Rane via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> Hello,
> 
> This is likely to be some small issue, but I am unable to figure out
> the solution. The following code, when run through `opt' with the -O1
> flag, replaces the call to foo() with a call to llvm.trap() followed
> by an `unreachable' instruction.
> 
> ====================> target datalayout =
"e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
> target triple = "armv7-unknown-linux-gnu"
> 
> define arm_aapcs_vfpcc void @foo() {
> entry:
>    ret void
> }
> 
> define arm_aapcs_vfpcc void @bar() {
> entry:
>    call void @foo()
You’re not using the right calling convention here: you need call
arm_aapcs_vfpcc void @foo().

Instcombine detects the mismatch, knows it is incorrect, and thus optimize away
the call with “undefined behavior”-like code.

— 
Mehdi

>    ret void
> }
> ====================> 
> If I drop the "arm_aapcs_vfpcc" for foo(), then the
`unreachable'
> instruction and the call to trap() are no longer introduced in the
> code.
> 
> I am running `opt' on a 64-bit x86 machine. I have synced llvm, clang,
> and compiler-rt to the latest commits, but that did not help.
> 
> $ clang --version
> clang version 4.0.0 (https://github.com/llvm-mirror/clang.git
> 631f1e728a6903f320a40e7b5035a63acd0acdc7)
> (https://github.com/llvm-mirror/llvm.git
> 23156f19246677a896ffbda72eadbcbbc1c47964)
> 
> Any ideas on how to prevent the unreachable instruction? Thanks!
> 
> Ashay
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

On 13 December 2016 at 21:31, Ashay Rane via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> target datalayout =
"e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
> target triple = "armv7-unknown-linux-gnu"
You probably also want the triple "armv7-unknown-linux-gnueabihf".
Straight away that would make your arm_aapcs_vfpcc annotations
redundant because that's the default.

Also, plain "gnu" is a really weird old ABI that LLVM doesn't
actually
support so you could easily get strange behaviour from the backend not
dealing with it properly. And even if the backend gets it right
somehow, your surrounding libraries are almost certainly not "gnu".

Cheers.

Tim.