c4c679c232
At the end of a region statement, the PHINode must be generated while the current IRBuilder's block is the region's exit node. For obvious reasons: The PHINode references the region's exiting block. A partial write would insert new control flow, i.e. insert new basic blocks between the exiting blocks and the current block. We fix this by generating the PHI nodes (region exit values) before generating any MemoryAccess's stores. This should fix the AOSP buildbot. Reported-by: Eli Friedman <efriedma@quicinc.com> llvm-svn: 361204
45 lines
1.3 KiB
LLVM
45 lines
1.3 KiB
LLVM
; RUN: opt %loadPolly -polly-import-jscop -polly-import-jscop-postfix=transformed -polly-codegen -S < %s | FileCheck %s
|
|
;
|
|
; This text case has a partial write of PHI in a region-statement. It
|
|
; requires that the new PHINode from the region's exiting block is
|
|
; generated before before the partial memory write.
|
|
;
|
|
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
|
|
|
|
define void @region_multiexit_partialwrite(i32* %arg, i64 %arg1, i32* %arg2) {
|
|
bb:
|
|
br label %bb3
|
|
|
|
bb3:
|
|
%tmp = phi i64 [ %tmp17, %bb10 ], [ 1, %bb ]
|
|
%tmp4 = getelementptr inbounds i32, i32* %arg, i64 %tmp
|
|
%tmp5 = load i32, i32* %tmp4, align 4
|
|
%tmp6 = icmp slt i32 %tmp5, 0
|
|
br i1 %tmp6, label %bb7, label %bb9
|
|
|
|
bb7:
|
|
%tmp8 = select i1 undef, i32 -2147483648, i32 undef
|
|
br label %bb10
|
|
|
|
bb9:
|
|
br label %bb10
|
|
|
|
bb10:
|
|
%tmp11 = phi i32 [ %tmp8, %bb7 ], [ undef, %bb9 ]
|
|
%tmp16 = getelementptr inbounds i32, i32* %arg2, i64 %tmp
|
|
store i32 %tmp11, i32* %tmp16, align 4
|
|
%tmp17 = add nuw i64 %tmp, 1
|
|
%tmp18 = icmp eq i64 %tmp17, %arg1
|
|
br i1 %tmp18, label %bb19, label %bb3
|
|
|
|
bb19:
|
|
ret void
|
|
}
|
|
|
|
|
|
; CHECK: polly.stmt.bb10.exit:
|
|
; CHECK-NEXT: %polly.tmp11 = phi i32 [ %p_tmp8, %polly.stmt.bb7 ], [ undef, %polly.stmt.bb9 ]
|
|
|
|
; CHECK: polly.stmt.bb10.exit.Stmt_bb3__TO__bb10_Write1.partial:
|
|
; CHECK: store i32 %polly.tmp11
|