wangtao
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[BlockGenerator] Also eliminate dead code not originating from BB 

After having generated the code for a ScopStmt, we run a simple dead-code
elimination that drops all instructions that are known to be and remain unused.
Until this change, we only considered instructions for dead-code elimination, if
they have a corresponding instruction in the original BB that belongs to
ScopStmt. However, when generating code we do not only copy code from the BB
belonging to a ScopStmt, but also generate code for operands referenced from BB.
After this change, we now also considers code for dead code elimination, which
does not have a corresponding instruction in BB.

This fixes a bug in Polly-ACC where such dead-code referenced CPU code from
within a GPU kernel, which is possible as we do not guarantee that all variables
that are used in known-dead-code are moved to the GPU.

llvm-svn: 278103
This commit is contained in:
Tobias Grosser 2016-08-09 08:59:05 +00:00
parent c99f72868d
commit c59b3ce044
5 changed files with 127 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
polly/lib/CodeGen/BlockGenerators.cpp
View File

@ -279,23 +279,20 @@ void BlockGenerator::copyInstruction(ScopStmt &Stmt, Instruction *Inst,
}
void BlockGenerator::removeDeadInstructions(BasicBlock *BB, ValueMapT &BBMap) {
for (auto I = BB->rbegin(), E = BB->rend(); I != E; I++) {
Instruction *Inst = &*I;
Value *NewVal = BBMap[Inst];
if (!NewVal)
continue;
Instruction *NewInst = dyn_cast<Instruction>(NewVal);
if (!NewInst)
continue;
auto NewBB = Builder.GetInsertBlock();
for (auto I = NewBB->rbegin(); I != NewBB->rend(); I++) {
Instruction *NewInst = &*I;
if (!isInstructionTriviallyDead(NewInst))
continue;
BBMap.erase(Inst);
for (auto Pair : BBMap)
if (Pair.second == NewInst) {
BBMap.erase(Pair.first);
}
NewInst->eraseFromParent();
I = NewBB->rbegin();
}
}

41
polly/test/GPGPU/host-control-flow.ll
View File

@ -73,46 +73,45 @@
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A = add nsw i64 %polly.access.mul.MemRef_A, %8
; KERNEL-IR-NEXT: %polly.access.MemRef_A = getelementptr float, float* %polly.access.cast.MemRef_A, i64 %polly.access.add.MemRef_A
; KERNEL-IR-NEXT: %tmp_p_scalar_ = load float, float* %polly.access.MemRef_A, align 4
; KERNEL-IR-NEXT: %9 = add i64 %6, 1
; KERNEL-IR-NEXT: %polly.access.cast.MemRef_A1 = bitcast i8* %MemRef_A to float*
; KERNEL-IR-NEXT: %pexp.pdiv_r2 = urem i64 %c0, 2
; KERNEL-IR-NEXT: %polly.access.mul.MemRef_A3 = mul nsw i64 %pexp.pdiv_r2, 100
; KERNEL-IR-NEXT: %10 = mul nsw i64 32, %b0
; KERNEL-IR-NEXT: %11 = add nsw i64 %10, %t0
; KERNEL-IR-NEXT: %12 = add nsw i64 %11, 1
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A4 = add nsw i64 %polly.access.mul.MemRef_A3, %12
; KERNEL-IR-NEXT: %9 = mul nsw i64 32, %b0
; KERNEL-IR-NEXT: %10 = add nsw i64 %9, %t0
; KERNEL-IR-NEXT: %11 = add nsw i64 %10, 1
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A4 = add nsw i64 %polly.access.mul.MemRef_A3, %11
; KERNEL-IR-NEXT: %polly.access.MemRef_A5 = getelementptr float, float* %polly.access.cast.MemRef_A1, i64 %polly.access.add.MemRef_A4
; KERNEL-IR-NEXT: %tmp2_p_scalar_ = load float, float* %polly.access.MemRef_A5, align 4
; KERNEL-IR-NEXT: %p_add = fadd float %tmp_p_scalar_, %tmp2_p_scalar_
; KERNEL-IR-NEXT: %polly.access.cast.MemRef_A6 = bitcast i8* %MemRef_A to float*
; KERNEL-IR-NEXT: %pexp.pdiv_r7 = urem i64 %c0, 2
; KERNEL-IR-NEXT: %polly.access.mul.MemRef_A8 = mul nsw i64 %pexp.pdiv_r7, 100
; KERNEL-IR-NEXT: %13 = mul nsw i64 32, %b0
; KERNEL-IR-NEXT: %14 = add nsw i64 %13, %t0
; KERNEL-IR-NEXT: %15 = add nsw i64 %14, 2
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A9 = add nsw i64 %polly.access.mul.MemRef_A8, %15
; KERNEL-IR-NEXT: %12 = mul nsw i64 32, %b0
; KERNEL-IR-NEXT: %13 = add nsw i64 %12, %t0
; KERNEL-IR-NEXT: %14 = add nsw i64 %13, 2
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A9 = add nsw i64 %polly.access.mul.MemRef_A8, %14
; KERNEL-IR-NEXT: %polly.access.MemRef_A10 = getelementptr float, float* %polly.access.cast.MemRef_A6, i64 %polly.access.add.MemRef_A9
; KERNEL-IR-NEXT: %tmp3_p_scalar_ = load float, float* %polly.access.MemRef_A10, align 4
; KERNEL-IR-NEXT: %p_add12 = fadd float %p_add, %tmp3_p_scalar_
; KERNEL-IR-NEXT: %polly.access.cast.MemRef_A11 = bitcast i8* %MemRef_A to float*
; KERNEL-IR-NEXT: %16 = add nsw i64 %c0, 1
; KERNEL-IR-NEXT: %pexp.pdiv_r12 = urem i64 %16, 2
; KERNEL-IR-NEXT: %15 = add nsw i64 %c0, 1
; KERNEL-IR-NEXT: %pexp.pdiv_r12 = urem i64 %15, 2
; KERNEL-IR-NEXT: %polly.access.mul.MemRef_A13 = mul nsw i64 %pexp.pdiv_r12, 100
; KERNEL-IR-NEXT: %17 = mul nsw i64 32, %b0
; KERNEL-IR-NEXT: %18 = add nsw i64 %17, %t0
; KERNEL-IR-NEXT: %19 = add nsw i64 %18, 1
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A14 = add nsw i64 %polly.access.mul.MemRef_A13, %19
; KERNEL-IR-NEXT: %16 = mul nsw i64 32, %b0
; KERNEL-IR-NEXT: %17 = add nsw i64 %16, %t0
; KERNEL-IR-NEXT: %18 = add nsw i64 %17, 1
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A14 = add nsw i64 %polly.access.mul.MemRef_A13, %18
; KERNEL-IR-NEXT: %polly.access.MemRef_A15 = getelementptr float, float* %polly.access.cast.MemRef_A11, i64 %polly.access.add.MemRef_A14
; KERNEL-IR-NEXT: %tmp4_p_scalar_ = load float, float* %polly.access.MemRef_A15, align 4
; KERNEL-IR-NEXT: %p_add17 = fadd float %tmp4_p_scalar_, %p_add12
; KERNEL-IR-NEXT: %polly.access.cast.MemRef_A16 = bitcast i8* %MemRef_A to float*
; KERNEL-IR-NEXT: %20 = add nsw i64 %c0, 1
; KERNEL-IR-NEXT: %pexp.pdiv_r17 = urem i64 %20, 2
; KERNEL-IR-NEXT: %19 = add nsw i64 %c0, 1
; KERNEL-IR-NEXT: %pexp.pdiv_r17 = urem i64 %19, 2
; KERNEL-IR-NEXT: %polly.access.mul.MemRef_A18 = mul nsw i64 %pexp.pdiv_r17, 100
; KERNEL-IR-NEXT: %21 = mul nsw i64 32, %b0
; KERNEL-IR-NEXT: %22 = add nsw i64 %21, %t0
; KERNEL-IR-NEXT: %23 = add nsw i64 %22, 1
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A19 = add nsw i64 %polly.access.mul.MemRef_A18, %23
; KERNEL-IR-NEXT: %20 = mul nsw i64 32, %b0
; KERNEL-IR-NEXT: %21 = add nsw i64 %20, %t0
; KERNEL-IR-NEXT: %22 = add nsw i64 %21, 1
; KERNEL-IR-NEXT: %polly.access.add.MemRef_A19 = add nsw i64 %polly.access.mul.MemRef_A18, %22
; KERNEL-IR-NEXT: %polly.access.MemRef_A20 = getelementptr float, float* %polly.access.cast.MemRef_A16, i64 %polly.access.add.MemRef_A19
; KERNEL-IR-NEXT: store float %p_add17, float* %polly.access.MemRef_A20, align 4
; KERNEL-IR-NEXT: br label %polly.merge

37
polly/test/GPGPU/remove-dead-instructions-in-stmt-2.ll Normal file
View File

@ -0,0 +1,37 @@
; RUN: opt %loadPolly -polly-codegen-ppcg -polly-acc-dump-kernel-ir \
; RUN: -disable-output < %s | \
; RUN: FileCheck %s -check-prefix=KERNEL-IR
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
; KERNEL-IR: store i32 0, i32* %polly.access.MemRef_sum_c, align 4
; KERNEL-IR-NEXT: br label %polly.merge
define void @kernel_dynprog([50 x [50 x i32]]* %sum_c) {
entry:
br label %for.cond1.preheader
for.cond1.preheader: ; preds = %entry
br label %for.body3
for.cond1.loopexit: ; preds = %for.end
%indvars.iv.next49 = add nuw nsw i64 %indvars.iv48, 1
%exitcond57 = icmp ne i64 %indvars.iv.next56, 49
br i1 %exitcond57, label %for.body3, label %for.inc55
for.body3: ; preds = %for.cond1.loopexit, %for.cond1.preheader
%indvars.iv55 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next56, %for.cond1.loopexit ]
%indvars.iv48 = phi i64 [ 1, %for.cond1.preheader ], [ %indvars.iv.next49, %for.cond1.loopexit ]
%indvars.iv.next56 = add nuw nsw i64 %indvars.iv55, 1
%arrayidx10 = getelementptr inbounds [50 x [50 x i32]], [50 x [50 x i32]]* %sum_c, i64 %indvars.iv55, i64 %indvars.iv48, i64 %indvars.iv55
store i32 0, i32* %arrayidx10, align 4
%cmp1334 = icmp slt i64 %indvars.iv.next56, %indvars.iv48
br label %for.end
for.end: ; preds = %for.body3
br label %for.cond1.loopexit
for.inc55: ; preds = %for.cond1.loopexit
unreachable
}

60
polly/test/GPGPU/remove-dead-instructions-in-stmt.ll Normal file
View File

@ -0,0 +1,60 @@
; RUN: opt %loadPolly -polly-codegen-ppcg -polly-acc-dump-kernel-ir \
; RUN: -disable-output < %s | \
; RUN: FileCheck %s -check-prefix=KERNEL-IR
; Ensure that no dead instructions are emitted between the store and the
; branch instruction of the ScopStmt. At some point, our dead-code-elimination
; did not remove code that was inserted to compute the old (unused) branch
; condition. This code referred to CPU registers and consequently resulted
; in invalid bitcode.
; KERNEL-IR: store i32 0, i32* %polly.access.MemRef_sum_c, align 4
; KERNEL-IR-NEXT: br label %polly.merge
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
define void @kernel_dynprog([50 x [50 x i32]]* %sum_c) {
entry:
br label %for.cond1.preheader
for.cond1.preheader: ; preds = %entry
br label %for.body3
for.cond4.for.cond1.loopexit_crit_edge: ; preds = %for.end
br label %for.cond1.loopexit
for.cond1.loopexit: ; preds = %for.cond4.for.cond1.loopexit_crit_edge
br i1 undef, label %for.body3, label %for.inc55
for.body3: ; preds = %for.cond1.loopexit, %for.cond1.preheader
%indvars.iv55 = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next56, %for.cond1.loopexit ]
%indvars.iv.next56 = add nuw nsw i64 %indvars.iv55, 1
br label %for.body6
for.body6: ; preds = %for.end, %for.body3
%indvars.iv50 = phi i64 [ 0, %for.body3 ], [ %indvars.iv.next51, %for.end ]
%arrayidx10 = getelementptr inbounds [50 x [50 x i32]], [50 x [50 x i32]]* %sum_c, i64 %indvars.iv55, i64 %indvars.iv50, i64 %indvars.iv55
store i32 0, i32* %arrayidx10, align 4
%cmp1334 = icmp slt i64 %indvars.iv.next56, %indvars.iv50
br i1 %cmp1334, label %for.body14.lr.ph, label %for.end
for.body14.lr.ph: ; preds = %for.body6
br label %for.body14
for.body14: ; preds = %for.body14, %for.body14.lr.ph
%arrayidx32 = getelementptr inbounds [50 x [50 x i32]], [50 x [50 x i32]]* %sum_c, i64 %indvars.iv55, i64 %indvars.iv50, i64 0
br i1 false, label %for.body14, label %for.cond12.for.end_crit_edge
for.cond12.for.end_crit_edge: ; preds = %for.body14
br label %for.end
for.end: ; preds = %for.cond12.for.end_crit_edge, %for.body6
%indvars.iv.next51 = add nuw nsw i64 %indvars.iv50, 1
%lftr.wideiv53 = trunc i64 %indvars.iv.next51 to i32
%exitcond54 = icmp ne i32 %lftr.wideiv53, 50
br i1 %exitcond54, label %for.body6, label %for.cond4.for.cond1.loopexit_crit_edge
for.inc55: ; preds = %for.cond1.loopexit
unreachable
}

1
polly/test/ScopInfo/int2ptr_ptr2int_2.ll
View File

@ -28,6 +28,7 @@
;
; IR: polly.loop_preheader:
; IR-NEXT: %scevgep = getelementptr i64, i64* %ptr, i64 1
; IR-NEXT: %26 = add i64 %val, 1
; IR-NEXT: br label %polly.loop_header
;
;