[RISCV] The assignment within the loop was not successful.

[shawn-xc]

I wrote a simple C code example and found that the assignment of regOp in the testFunc function within the loop was not successful , and it always retained the value outside the loop. For detailed code information, please refer to：https://godbolt.org/z/vP9oKhe54.

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: None (shawn-xc)

I wrote a simple C code example and found that the assignment of stRegCmd.regOp in the CtlPushNonAipr2AiprCmd function within the loop was not successful, and it always retained the value outside the loop. For detailed code information, please refer to：https://godbolt.org/z/crxxbozzhjoa.

[mikhailramalho]

the link seems to be broken, is it https://godbolt.org/z/rxxozzhjo instead?

[shawn-xc]

the link seems to be broken, is it https://godbolt.org/z/rxxozzhjo instead?

Please see this link https://godbolt.org/z/65ndKod7G.

[topperc]

The store was pushed out of the loop by middle end LICM. It became a store of %2 phi in for.cond.cleanup.

; Loop:                                                                          
for.cond:                                         ; preds = %CopyFunc.exit, %entry
  %2 = phi i8 [ %regOp.promoted, %entry ], [ 8, %CopyFunc.exit ]                 
  %pNo10.0 = phi i32 [ 0, %entry ], [ %inc, %CopyFunc.exit ]                     
  %cmp = icmp ult i32 %pNo10.0, %cnt                                             
  br i1 %cmp, label %CopyFunc.exit, label %for.cond.cleanup                      
                                                                                 
CopyFunc.exit:                                    ; preds = %for.cond            
  %and = shl i32 %pNo10.0, 5                                                     
  %bf.shl = and i32 %and, 992                                                    
  %bf.set = or disjoint i32 %bf.shl, %1                                          
  %unPpn14 = getelementptr inbounds nuw i8, ptr %cmd, i32 8                      
  store i32 %bf.set, ptr %unPpn14, align 4                                       
  %unPpn.sroa.5.0.unPpn14.sroa_idx = getelementptr inbounds nuw i8, ptr %cmd, i32 12
  store i32 %unPpn.coerce.fca.1.extract, ptr %unPpn.sroa.5.0.unPpn14.sroa_idx, align 4, !tbaa !12
  %arrayidx.i = getelementptr inbounds nuw [64 x %struct.qList], ptr inttoptr (i32 285409580 to ptr), i32 0, i32 %id
  %call.i.i = tail call ptr @PoolMemAlloc(ptr noundef nonnull inttoptr (i32 285409280 to ptr), i32 noundef range(i32 8, 1029) 28) #7
  store ptr %arrayidx.i, ptr %call.i.i, align 4, !tbaa !14                       
  %prev.i.i = getelementptr inbounds nuw i8, ptr %arrayidx.i, i32 4              
  store ptr %call.i.i, ptr %prev.i.i, align 4, !tbaa !18                         
  %prev2.i.i = getelementptr inbounds nuw i8, ptr %call.i.i, i32 4               
  store ptr %call.i.i, ptr %prev2.i.i, align 4, !tbaa !18                        
  %3 = load ptr, ptr %prev.i.i, align 4, !tbaa !18                               
  store ptr %call.i.i, ptr %3, align 4, !tbaa !14                                
  %entry3.i = getelementptr inbounds nuw i8, ptr %call.i.i, i32 8                
  call void @llvm.memcpy.p0.p0.i32(ptr noundef nonnull align 4 dereferenceable(56) %entry3.i, ptr noundef nonnull readonly align 4 dereferenceable(56) %cmd, i32 56, i1 false), !tbaa.struct !19
  %inc = add nuw i32 %pNo10.0, 1                                                 
  br label %for.cond, !llvm.loop !20                                             
                                                                                 
; Exit blocks                                                                    
for.cond.cleanup:                                 ; preds = %for.cond            
  %.lcssa = phi i8 [ %2, %for.cond ]                                             
  store i8 %.lcssa, ptr %regOp, align 1                                          
  call void @llvm.lifetime.end.p0(i64 20, ptr nonnull %cmd) #7                   
  ret i32 0

That store was later delete by SROA. I think because cmd object was not used after the store since its a local variable.

I haven't checked to see if LICM was correct to move the store out of the loop yet.

[shawn-xc]

The store was pushed out of the loop by middle end LICM. It became a store of %2 phi in for.cond.cleanup.

; Loop:                                                                          
for.cond:                                         ; preds = %CopyFunc.exit, %entry
  %2 = phi i8 [ %regOp.promoted, %entry ], [ 8, %CopyFunc.exit ]                 
  %pNo10.0 = phi i32 [ 0, %entry ], [ %inc, %CopyFunc.exit ]                     
  %cmp = icmp ult i32 %pNo10.0, %cnt                                             
  br i1 %cmp, label %CopyFunc.exit, label %for.cond.cleanup                      
                                                                                 
CopyFunc.exit:                                    ; preds = %for.cond            
  %and = shl i32 %pNo10.0, 5                                                     
  %bf.shl = and i32 %and, 992                                                    
  %bf.set = or disjoint i32 %bf.shl, %1                                          
  %unPpn14 = getelementptr inbounds nuw i8, ptr %cmd, i32 8                      
  store i32 %bf.set, ptr %unPpn14, align 4                                       
  %unPpn.sroa.5.0.unPpn14.sroa_idx = getelementptr inbounds nuw i8, ptr %cmd, i32 12
  store i32 %unPpn.coerce.fca.1.extract, ptr %unPpn.sroa.5.0.unPpn14.sroa_idx, align 4, !tbaa !12
  %arrayidx.i = getelementptr inbounds nuw [64 x %struct.qList], ptr inttoptr (i32 285409580 to ptr), i32 0, i32 %id
  %call.i.i = tail call ptr @PoolMemAlloc(ptr noundef nonnull inttoptr (i32 285409280 to ptr), i32 noundef range(i32 8, 1029) 28) #7
  store ptr %arrayidx.i, ptr %call.i.i, align 4, !tbaa !14                       
  %prev.i.i = getelementptr inbounds nuw i8, ptr %arrayidx.i, i32 4              
  store ptr %call.i.i, ptr %prev.i.i, align 4, !tbaa !18                         
  %prev2.i.i = getelementptr inbounds nuw i8, ptr %call.i.i, i32 4               
  store ptr %call.i.i, ptr %prev2.i.i, align 4, !tbaa !18                        
  %3 = load ptr, ptr %prev.i.i, align 4, !tbaa !18                               
  store ptr %call.i.i, ptr %3, align 4, !tbaa !14                                
  %entry3.i = getelementptr inbounds nuw i8, ptr %call.i.i, i32 8                
  call void @llvm.memcpy.p0.p0.i32(ptr noundef nonnull align 4 dereferenceable(56) %entry3.i, ptr noundef nonnull readonly align 4 dereferenceable(56) %cmd, i32 56, i1 false), !tbaa.struct !19
  %inc = add nuw i32 %pNo10.0, 1                                                 
  br label %for.cond, !llvm.loop !20                                             
                                                                                 
; Exit blocks                                                                    
for.cond.cleanup:                                 ; preds = %for.cond            
  %.lcssa = phi i8 [ %2, %for.cond ]                                             
  store i8 %.lcssa, ptr %regOp, align 1                                          
  call void @llvm.lifetime.end.p0(i64 20, ptr nonnull %cmd) #7                   
  ret i32 0

That store was later delete by SROA. I think because cmd object was not used after the store since its a local variable.

I haven't checked to see if LICM was correct to move the store out of the loop yet.

I modified the code *data = *dataSrc in CopyFunc function to the following format.

for (u32 i = 0; i < sizeof(MEM_DATA_56B) / sizeof(u32); i++) {
    data->val[i] = dataSrc->val[i];
}

I found the cmd.regOp assignment is OK.