| Issue |
150973
|
| Summary |
[Machine uniformity Analysis] uniformity of a value inside and outside a loop
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
jgu222
|
There are cases that a value inside a loop is uniform, but outside a loop, it is divergent.
For example,
```
;
; opt -mtriple amdgcn-unknown-amdhsa -passes='print<uniformity>' -disable-output phi_loop_issue.ll
;
define amdgpu_kernel void @test_ctrl_divergence() {
B0:
%tid = call i32 @llvm.amdgcn.workitem.id.x()
br label %B1
B1:
%i = phi i32 [0, %B0], [%i_n, %B1]
%div.cond = icmp slt i32 %tid, %i
%i_n = add i32 %i, 1
br i1 %div.cond, label %B1, label %B2
B2:
%a = add i32 %i, 10
ret void
}
declare i32 @llvm.amdgcn.workitem.id.x() #0
attributes #0 = {nounwind readnone }
```
The result from uniformity analysis:
```
UniformityInfo for function 'test_ctrl_divergence':
CYCLES WITH DIVERGENT EXIT:
depth=1: entries(B1)
TEMPORAL DIVERGENCE LIST:
Value : %i = phi i32 [ 0, %B0 ], [ %i_n, %B1 ]
Used by : %a = add i32 %i, 10
Outside cycle :depth=1: entries(B1)
BLOCK B0
DEFINITIONS
DIVERGENT: %tid = call i32 @llvm.amdgcn.workitem.id.x()
TERMINATORS
br label %B1
END BLOCK
BLOCK B1
DEFINITIONS
%i = phi i32 [ 0, %B0 ], [ %i_n, %B1 ]
DIVERGENT: %div.cond = icmp slt i32 %tid, %i
%i_n = add i32 %i, 1
TERMINATORS
DIVERGENT: br i1 %div.cond, label %B1, label %B2
END BLOCK
BLOCK B2
DEFINITIONS
DIVERGENT: %a = add i32 %i, 10
TERMINATORS
ret void
END BLOCK
```
As the loop (B1) has a divergent condition, the %i would be divergent outside the loop as each lane would exit the loop with different %i. But within the loop, %i is uniform. As %i is the same both inside the loop and outside the loop, there is no way to represent uniformity for %i.
Traditionally, this is resolved by using LCSSA to separate a value inside a loop and one outside the loop. However, there is no Machine function pass LCSSA so far.
The purpose of the issue is to invite the input to see what is a better solution to this issue, would it be better to port Function pass LCSSA into MachineFunction pass ?
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
