[PATCH] D25584: [polly codegen] Handle multi-dimensional invariant load.
Friedman, Eli via llvm-commits
llvm-commits at lists.llvm.org
Mon Oct 17 15:04:54 PDT 2016
On 10/15/2016 1:36 AM, Johannes Doerfert wrote:
> LGTM but we should check why code generation bails and places the
> "false" as a guard.
l1 has the execution context "l2 > 0", therefore loading l1 requires
loading l2. But l1 is a dimension for l2, so loading l2 requires
loading l1. Code generation sees this cycle and abandons the scop.
We can probably add a special case to check for a false dependence on
the dimension.
-Eli
>
> On 10/14, Eli Friedman wrote:
>> efriedma updated the summary for this revision.
>> efriedma updated this revision to Diff 74724.
>> efriedma added a comment.
>>
>> Updated to correctly iterate over all the inputs to the SCEV value.
>>
>>
>> Repository:
>> rL LLVM
>>
>> https://reviews.llvm.org/D25584
>>
>> Files:
>> lib/CodeGen/IslNodeBuilder.cpp
>> test/Isl/CodeGen/invariant-load-dimension.ll
>>
>>
>> Index: test/Isl/CodeGen/invariant-load-dimension.ll
>> ===================================================================
>> --- /dev/null
>> +++ test/Isl/CodeGen/invariant-load-dimension.ll
>> @@ -0,0 +1,54 @@
>> +; RUN: opt < %s -analyze -polly-scops -polly-process-unprofitable -polly-invariant-load-hoisting | FileCheck %s -check-prefix=SCOPS
>> +; RUN: opt -S < %s -polly-codegen -polly-process-unprofitable -polly-invariant-load-hoisting | FileCheck %s -check-prefix=CODEGEN
>> +
>> +target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n8:16:32-S64"
>> +
>> +%S = type { i32, i32, [12 x %L] }
>> +%L = type { i32, i32, double, i32, i32, i32, i32, i32 }
>> +
>> +define void @test(%S* %cpi, i1 %b) {
>> +; SCOPS-LABEL: Region: %if.then14---%exit
>> +; SCOPS: Invariant Accesses: {
>> +; SCOPS-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]
>> +; SCOPS-NEXT: [l2, l1] -> { Stmt_for_body_i[i0] -> MemRef_cpi[0, 0] };
>> +; SCOPS-NEXT: Execution Context: [l2, l1] -> { : }
>> +; SCOPS-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]
>> +; SCOPS-NEXT: [l2, l1] -> { Stmt_for_body_lr_ph_i[] -> MemRef_cpi[0, 1] };
>> +; SCOPS-NEXT: Execution Context: [l2, l1] -> { : l2 > 0 }
>> +; SCOPS-NEXT: }
>> +; SCOPS: Arrays {
>> +; SCOPS-NEXT: i32 MemRef_cpi[*][%l1]; // Element size 4
>> +; SCOPS-NEXT: }
>> +
>> +; FIXME: Figure out how to actually generate code for this loop.
>> +; CODEGEN-LABEL: @test(
>> +; CODEGEN: polly.preload.begin:
>> +; CODEGEN-NEXT: br i1 false
>> +
>> +entry:
>> + %nt = getelementptr inbounds %S, %S* %cpi, i32 0, i32 1
>> + br i1 %b, label %if.then14, label %exit
>> +
>> +if.then14:
>> + %ns = getelementptr inbounds %S, %S* %cpi, i32 0, i32 0
>> + %l0 = load i32, i32* %ns, align 8
>> + %cmp12.i = icmp sgt i32 %l0, 0
>> + br i1 %cmp12.i, label %for.body.lr.ph.i, label %exit
>> +
>> +for.body.lr.ph.i:
>> + %l1 = load i32, i32* %nt, align 4
>> + br label %for.body.i
>> +
>> +for.body.i:
>> + %phi = phi i32 [ 0, %for.body.lr.ph.i ], [ %inc, %for.body.i ]
>> + %mul.i163 = mul nsw i32 %phi, %l1
>> + %cv = getelementptr inbounds %S, %S* %cpi, i32 0, i32 2, i32 %mul.i163, i32 0
>> + store i32 0, i32* %cv, align 8
>> + %inc = add nuw nsw i32 %phi, 1
>> + %l2 = load i32, i32* %ns, align 8
>> + %cmp.i164 = icmp slt i32 %inc, %l2
>> + br i1 %cmp.i164, label %for.body.i, label %exit
>> +
>> +exit:
>> + ret void
>> +}
>> Index: lib/CodeGen/IslNodeBuilder.cpp
>> ===================================================================
>> --- lib/CodeGen/IslNodeBuilder.cpp
>> +++ lib/CodeGen/IslNodeBuilder.cpp
>> @@ -1107,6 +1107,25 @@
>> ExecutionCtx = isl_set_intersect(ExecutionCtx, BaseExecutionCtx);
>> }
>>
>> + // If the size of a dimension is dependent on another class, make sure it is
>> + // preloaded.
>> + for (unsigned i = 1, e = SAI->getNumberOfDimensions(); i < e; ++i) {
>> + const SCEV *Dim = SAI->getDimensionSize(i);
>> + SetVector<Value *> Values;
>> + findValues(Dim, SE, Values);
>> + for (auto *Val : Values) {
>> + if (auto *BaseIAClass = S.lookupInvariantEquivClass(Val)) {
>> + if (!preloadInvariantEquivClass(*BaseIAClass))
>> + return false;
>> +
>> + // After we preloaded the BaseIAClass we adjusted the BaseExecutionCtx
>> + // and we need to refine the ExecutionCtx.
>> + isl_set *BaseExecutionCtx = isl_set_copy(BaseIAClass->ExecutionContext);
>> + ExecutionCtx = isl_set_intersect(ExecutionCtx, BaseExecutionCtx);
>> + }
>> + }
>> + }
>> +
>> Instruction *AccInst = MA->getAccessInstruction();
>> Type *AccInstTy = AccInst->getType();
>>
>>
>>
>> --
>> You received this message because you are subscribed to the Google Groups "Polly Development" group.
>> To unsubscribe from this group and stop receiving emails from it, send an email to polly-dev+unsubscribe at googlegroups.com.
>> For more options, visit https://groups.google.com/d/optout.
>> Index: test/Isl/CodeGen/invariant-load-dimension.ll
>> ===================================================================
>> --- /dev/null
>> +++ test/Isl/CodeGen/invariant-load-dimension.ll
>> @@ -0,0 +1,54 @@
>> +; RUN: opt < %s -analyze -polly-scops -polly-process-unprofitable -polly-invariant-load-hoisting | FileCheck %s -check-prefix=SCOPS
>> +; RUN: opt -S < %s -polly-codegen -polly-process-unprofitable -polly-invariant-load-hoisting | FileCheck %s -check-prefix=CODEGEN
>> +
>> +target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n8:16:32-S64"
>> +
>> +%S = type { i32, i32, [12 x %L] }
>> +%L = type { i32, i32, double, i32, i32, i32, i32, i32 }
>> +
>> +define void @test(%S* %cpi, i1 %b) {
>> +; SCOPS-LABEL: Region: %if.then14---%exit
>> +; SCOPS: Invariant Accesses: {
>> +; SCOPS-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]
>> +; SCOPS-NEXT: [l2, l1] -> { Stmt_for_body_i[i0] -> MemRef_cpi[0, 0] };
>> +; SCOPS-NEXT: Execution Context: [l2, l1] -> { : }
>> +; SCOPS-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]
>> +; SCOPS-NEXT: [l2, l1] -> { Stmt_for_body_lr_ph_i[] -> MemRef_cpi[0, 1] };
>> +; SCOPS-NEXT: Execution Context: [l2, l1] -> { : l2 > 0 }
>> +; SCOPS-NEXT: }
>> +; SCOPS: Arrays {
>> +; SCOPS-NEXT: i32 MemRef_cpi[*][%l1]; // Element size 4
>> +; SCOPS-NEXT: }
>> +
>> +; FIXME: Figure out how to actually generate code for this loop.
>> +; CODEGEN-LABEL: @test(
>> +; CODEGEN: polly.preload.begin:
>> +; CODEGEN-NEXT: br i1 false
>> +
>> +entry:
>> + %nt = getelementptr inbounds %S, %S* %cpi, i32 0, i32 1
>> + br i1 %b, label %if.then14, label %exit
>> +
>> +if.then14:
>> + %ns = getelementptr inbounds %S, %S* %cpi, i32 0, i32 0
>> + %l0 = load i32, i32* %ns, align 8
>> + %cmp12.i = icmp sgt i32 %l0, 0
>> + br i1 %cmp12.i, label %for.body.lr.ph.i, label %exit
>> +
>> +for.body.lr.ph.i:
>> + %l1 = load i32, i32* %nt, align 4
>> + br label %for.body.i
>> +
>> +for.body.i:
>> + %phi = phi i32 [ 0, %for.body.lr.ph.i ], [ %inc, %for.body.i ]
>> + %mul.i163 = mul nsw i32 %phi, %l1
>> + %cv = getelementptr inbounds %S, %S* %cpi, i32 0, i32 2, i32 %mul.i163, i32 0
>> + store i32 0, i32* %cv, align 8
>> + %inc = add nuw nsw i32 %phi, 1
>> + %l2 = load i32, i32* %ns, align 8
>> + %cmp.i164 = icmp slt i32 %inc, %l2
>> + br i1 %cmp.i164, label %for.body.i, label %exit
>> +
>> +exit:
>> + ret void
>> +}
>> Index: lib/CodeGen/IslNodeBuilder.cpp
>> ===================================================================
>> --- lib/CodeGen/IslNodeBuilder.cpp
>> +++ lib/CodeGen/IslNodeBuilder.cpp
>> @@ -1107,6 +1107,25 @@
>> ExecutionCtx = isl_set_intersect(ExecutionCtx, BaseExecutionCtx);
>> }
>>
>> + // If the size of a dimension is dependent on another class, make sure it is
>> + // preloaded.
>> + for (unsigned i = 1, e = SAI->getNumberOfDimensions(); i < e; ++i) {
>> + const SCEV *Dim = SAI->getDimensionSize(i);
>> + SetVector<Value *> Values;
>> + findValues(Dim, SE, Values);
>> + for (auto *Val : Values) {
>> + if (auto *BaseIAClass = S.lookupInvariantEquivClass(Val)) {
>> + if (!preloadInvariantEquivClass(*BaseIAClass))
>> + return false;
>> +
>> + // After we preloaded the BaseIAClass we adjusted the BaseExecutionCtx
>> + // and we need to refine the ExecutionCtx.
>> + isl_set *BaseExecutionCtx = isl_set_copy(BaseIAClass->ExecutionContext);
>> + ExecutionCtx = isl_set_intersect(ExecutionCtx, BaseExecutionCtx);
>> + }
>> + }
>> + }
>> +
>> Instruction *AccInst = MA->getAccessInstruction();
>> Type *AccInstTy = AccInst->getType();
>>
>
--
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
More information about the llvm-commits
mailing list