Re: [RFC PATCH v3 11/12] powerpc: Remove unreachable() from WARN_ON() (gcc issue ?)

[Christophe Leroy <christophe.leroy@csgroup.eu>]

Hi Segher, your help might be welcome,

Le 01/07/2022 à 08:56, Sathvika Vasireddy a écrit :
> Hi Chen,
> 
> Thanks for pitching in and providing your inputs :-)
> 
> On 01/07/22 07:43, Chen Zhongjin wrote:
>> Hi everyone,
>>
>> Hope I'm not too late for this discussion.
>>
>> I'm not familiar with ppc so it spent me some time to reproduce this. 
>> But at last I didn't make it.
>>
>> What I did:
>>
>> 1 checkout to tip/objtool/core
>>
>> 2 apply this patch
>>
>> 3 recover the unreachable() after WARN_ENTRY(..
>>
>> 4 compile on defconfig/allyesconfig
>>
>> If anyone can point out which file will generate this problem it will 
>> be perfect.
> 
> To reproduce this problem, you may want to apply this patch series on 
> top of objtool/core branch of the tip tree, and compile on 
> ppc64le_defconfig.
> 
> There are a couple of C files that are generating these warnings. One 
> such file is: arch/powerpc/kvm/../../../virt/kvm/kvm_main.o which gives
> *arch/powerpc/kvm/../../../virt/kvm/kvm_main.o: warning: objtool: 
> kvm_mmu_notifier_release+0x6c: unannotated intra-function call* warning.
> 
> With unreachable() in __WARN_FLAGS(), we get unannotated intra-function 
> call warnings, but without unreachable() like in patch 11/12 or with 
> just the builtin variant of unreachable (__builtin_unreachable()) 
> instead of unreachable(), we do not get those warnings.
> 


I made a simple test aside our issue to see if I get something similar 
to ARM. I don't if it is the same at the end, but it looks odd anyway:

int test(int x)
{
	switch(x) {
	case 0 : return 3;
	case 1 : return 5;
	default : unreachable();
	}
}

0000000000001c80 <test>:
     1c80:       a6 02 08 7c     mflr    r0
     1c84:       01 00 00 48     bl      1c84 <test+0x4>
                         1c84: R_PPC64_REL24     _mcount
     1c88:       00 00 23 2c     cmpdi   r3,0
     1c8c:       14 00 82 41     beq     1ca0 <test+0x20>
     1c90:       01 00 03 2c     cmpwi   r3,1
     1c94:       05 00 60 38     li      r3,5
     1c98:       20 00 82 4d     beqlr
     1c9c:       00 00 42 60     ori     r2,r2,0
     1ca0:       03 00 60 38     li      r3,3
     1ca4:       20 00 80 4e     blr

So it looks like it takes no real benefit from the unreachable marking.

Now with __builtin_unreachable() instead of unreachable() :

int test(int x)
{
	switch(x) {
	case 0 : return 3;
	case 1 : return 5;
	default : __builtin_unreachable();
	}
}

0000000000001c80 <test>:
     1c80:       a6 02 08 7c     mflr    r0
     1c84:       01 00 00 48     bl      1c84 <test+0x4>
                         1c84: R_PPC64_REL24     _mcount
     1c88:       00 00 63 20     subfic  r3,r3,0
     1c8c:       10 19 63 7c     subfe   r3,r3,r3
     1c90:       bc 07 63 54     rlwinm  r3,r3,0,30,30
     1c94:       03 00 63 38     addi    r3,r3,3
     1c98:       20 00 80 4e     blr


Here the generated code takes advantage of the unreachable marking and 
results in a branchless code.


Christophe