[llvm-bugs] [Bug 32001] New: Regression with r279460 getelementptr argument goes missing

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=32001

            Bug ID: 32001
           Summary: Regression with r279460 getelementptr argument goes
                    missing
           Product: tools
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: opt
          Assignee: unassignedbugs at nondot.org
          Reporter: tarceri at itsqueeze.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 18005
  --> https://bugs.llvm.org/attachment.cgi?id=18005&action=edit
Unoptimised shader

commit f991e38d156c4c10c609ca8425a7c31b951ecbed
Author: James Molloy <james.molloy at arm.com>
Date:   Thu Sep 1 10:44:35 2016 +0000

    [SimplifyCFG] Change the algorithm in SinkThenElseCodeToEnd

    r279460 rewrote this function to be able to handle more than two incoming
edges and took pains to ensure this didn't regress anything.


On AMGGPU at least this caused a regression (possibly indirectly). I've
included a before and after bellow, you can see that the !amdgpu.uniform !0
goes missing.

I've attached the unoptimised version. I tried to debug this with 'lcc
-march=amdgcn -mcpu=polaris10 llvm_broken_preopt.ll' but it didn't seem to hit
the SimplifyCFG path when doing this.

BEFORE:

  br i1 %27, label %else5, label %if1

if1:                                              ; preds = %main_body
  %30 = getelementptr [32 x <8 x i32>], [32 x <8 x i32>] addrspace(2)* %2, i64
0, i64 0, !amdgpu.uniform !0
  %31 = load <8 x i32>, <8 x i32> addrspace(2)* %30, align 32, !invariant.load
!0
  %32 = bitcast [32 x <8 x i32>] addrspace(2)* %2 to [0 x <4 x i32>]
addrspace(2)*
  %33 = getelementptr [0 x <4 x i32>], [0 x <4 x i32>] addrspace(2)* %32, i64
0, i64 3, !amdgpu.uniform !0
  %34 = load <4 x i32>, <4 x i32> addrspace(2)* %33, align 16, !invariant.load
!0
  %35 = bitcast float %28 to i32
  %36 = bitcast float %29 to i32
  %37 = insertelement <2 x i32> undef, i32 %35, i32 0
  %38 = insertelement <2 x i32> %37, i32 %36, i32 1
  %39 = call <4 x float> @llvm.SI.image.sample.v2i32(<2 x i32> %38, <8 x i32>
%31, <4 x i32> %34, i32 15, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)
  br label %endif9

else5:                                            ; preds = %main_body
  %40 = getelementptr [32 x <8 x i32>], [32 x <8 x i32>] addrspace(2)* %2, i64
0, i64 2, !amdgpu.uniform !0
  %41 = load <8 x i32>, <8 x i32> addrspace(2)* %40, align 32, !invariant.load
!0
  %42 = bitcast [32 x <8 x i32>] addrspace(2)* %2 to [0 x <4 x i32>]
addrspace(2)*
  %43 = getelementptr [0 x <4 x i32>], [0 x <4 x i32>] addrspace(2)* %42, i64
0, i64 7, !amdgpu.uniform !0
  %44 = load <4 x i32>, <4 x i32> addrspace(2)* %43, align 16, !invariant.load
!0
  %45 = bitcast float %28 to i32
  %46 = bitcast float %29 to i32
  %47 = insertelement <2 x i32> undef, i32 %45, i32 0
  %48 = insertelement <2 x i32> %47, i32 %46, i32 1
  %49 = call <4 x float> @llvm.SI.image.sample.v2i32(<2 x i32> %48, <8 x i32>
%41, <4 x i32> %44, i32 15, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)
  br label %endif9

endif9:

AFTER:

%30 = select i1 %27, i64 0, i64 2
  %31 = getelementptr [32 x <8 x i32>], [32 x <8 x i32>] addrspace(2)* %2, i64
0, i64 %30
  %32 = load <8 x i32>, <8 x i32> addrspace(2)* %31, align 32, !invariant.load
!0
  %33 = bitcast [32 x <8 x i32>] addrspace(2)* %2 to [0 x <4 x i32>]
addrspace(2)*
  %34 = select i1 %27, i64 3, i64 7
  %35 = getelementptr [0 x <4 x i32>], [0 x <4 x i32>] addrspace(2)* %33, i64
0, i64 %34
  %36 = load <4 x i32>, <4 x i32> addrspace(2)* %35, align 16, !invariant.load
!0
  %37 = bitcast float %28 to i32
  %38 = bitcast float %29 to i32
  %39 = insertelement <2 x i32> undef, i32 %37, i32 0
  %40 = insertelement <2 x i32> %39, i32 %38, i32 1
  %41 = call <4 x float> @llvm.SI.image.sample.v2i32(<2 x i32> %40, <8 x i32>
%32, <4 x i32> %36, i32 15, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20170218/dc48901a/attachment.html>