[Polly] [PPCGCodeGeneration] Teach `SubtreeValues` to include possibly required base pointers, not remove all base pointers blindly.

[bollu]

Summary

Currently, we remove all ScopArrayInfo base pointers from SubtreeValues.
We do this on the assumption that if an array is used, it will be passed
by the relevant ScopArrayInfo in
createKernelFunction/createKernelFunctionDecl.

However, there can be operations on a base pointer that we _do not
model_, eg. ptrtoint.

This will lead to incorrect code generation since we blindly copy
instructions using BlockGenerator. When BlockGenerator tries to clone
the ptrtoint instruction, It will create a kernel-to-host reference
which will cause us to fail verify.

[efriedma]

Did you forget to include part of your patch? This doesn't contain any code changes.

[bollu]

@efriedma Thanks, I totally screwed up the rebase :)

[bollu]

• add changes in PPCGCodeGen that were missed due to incorrect rebase.

[bollu]

• [NFC wrt patch] remove debug code

[philip.pfaffe]

Do you have a testcase for the change this patch is introducing?

[bollu]

• update code and test case

[philip.pfaffe]

lib/CodeGen/PPCGCodeGeneration.cpp

1289 ↗

(On Diff #123168)

Is this related?

1498 ↗

(On Diff #123168)

This lambda can perfectly be inlined into it's caller.

[philip.pfaffe]

test/GPGPU/invariant-load-hoisting.ll

20

Please don't hardcode unnamed registers!

test/GPGPU/only-remove-arrays-from-subtree-that-are-used-by-kernel.ll

44 ↗

(On Diff #123168)

Don't hardcode unnamed registers.

test/GPGPU/invariant-load-hoisting-read-in-kernel.ll

 ; RUN: opt %loadPolly -disable-output -polly-acc-dump-kernel-ir \
 ; RUN: -polly-codegen-ppcg -polly-scops \
 ; RUN: -polly-invariant-load-hoisting < %s | FileCheck %s
 
 ; REQUIRES: pollyacc
 
 ; Verify that invariant loads used in a kernel statement are correctly forwarded
 ; as subtree value to the GPU kernel.
 
-; CHECK:  define ptx_kernel void @FUNC_foo_SCOP_0_KERNEL_0({{.*}} float %polly.access.p.load)
+; CHECK:  define ptx_kernel void @FUNC_foo_SCOP_0_KERNEL_0({{.*}} float %polly.access.p.load{{.*}})
 ; CHECK:   store float %polly.access.p.load, float* %indvar2f.phiops
 
 define void @foo(float* %A, float* %p) {
 entry:
   br label %loop
 
 loop:
   %indvar = phi i64 [0, %entry], [%indvar.next, %loop]

test/GPGPU/invariant-load-hoisting.ll

 ; SCOP:       Function: f
 ; SCOP-NEXT:  Region: %entry.split---%for.end26
 ; SCOP-NEXT:  Max Loop Depth:  3
 ; SCOP-NEXT:  Invariant Accesses: {
 ; SCOP-NEXT:          ReadAccess := [Reduction Type: NONE] [Scalar: 0]
 ; SCOP-NEXT:              [n, tmp12] -> { Stmt_for_body6[i0, i1, i2] -> MemRef_invariant[0] };
 ; SCOP-NEXT:          Execution Context: [n, tmp12] -> {  : n > 0 }
 ; SCOP-NEXT:  }
-; HOST-IR:      call void @polly_launchKernel(i8* %209, i32 %215, i32 1, i32 32, i32 1, i32 1, i8* %polly_launch_0_params_i8ptr)
-; HOST-IR-NEXT: call void @polly_freeKernel(i8* %209)
+; HOST-IR:      call void @polly_launchKernel(i8* %211, i32 %217, i32 1, i32 32, i32 1, i32 1, i8* %polly_launch_0_params_i8ptr)
+; HOST-IR-NEXT: call void @polly_freeKernel(i8* %211)
 
-; KERNEL-IR: define ptx_kernel void @FUNC_f_SCOP_0_KERNEL_0(i8 addrspace(1)* %MemRef_B, i8 addrspace(1)* %MemRef_A, i32 %n, i32 %tmp12, i32 %polly.preload.tmp21.merge)
+; KERNEL-IR: define ptx_kernel void @FUNC_f_SCOP_0_KERNEL_0(i8 addrspace(1)* %MemRef_B, i8 addrspace(1)* %MemRef_A, i32 %n, i32 %tmp12, i32* %invariant, i32 %polly.preload.tmp21.merge)
 
 
 ; Check that we generate correct GPU code in case of invariant load hoisting.
 ;
 ;
 ;    static const int N = 3000;
 ;
 ;    void f(int A[N][N], int *invariant, int B[N][N], int n) {