[llvm] r301044 - InferAddressSpaces: Infer for just GEPs

[Matt Arsenault via llvm-commits]

Author: arsenm
Date: Fri Apr 21 16:35:04 2017
New Revision: 301044

URL: http://llvm.org/viewvc/llvm-project?rev=301044&view=rev
Log:
InferAddressSpaces: Infer for just GEPs

Fixes leaving intermediate flat addressing computations
where a GEP instruction's source is a constant expression.

Still leaves behind a trivial addrspacecast + gep pair that
instcombine is able to handle, which ideally could be folded
here directly.

Added:
    llvm/trunk/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll
Modified:
    llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp

Modified: llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp?rev=301044&r1=301043&r2=301044&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/InferAddressSpaces.cpp Fri Apr 21 16:35:04 2017
@@ -297,10 +297,14 @@ InferAddressSpaces::collectFlatAddressEx
                                                  &Visited);
   };
 
-  // We only explore address expressions that are reachable from loads and
-  // stores for now because we aim at generating faster loads and stores.
+  // Look at operations that may be interesting accelerate by moving to a known
+  // address space. We aim at generating after loads and stores, but pure
+  // addressing calculations may also be faster.
   for (Instruction &I : instructions(F)) {
-    if (auto *LI = dyn_cast<LoadInst>(&I))
+    if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
+      if (!GEP->getType()->isVectorTy())
+        PushPtrOperand(GEP->getPointerOperand());
+    } else if (auto *LI = dyn_cast<LoadInst>(&I))
       PushPtrOperand(LI->getPointerOperand());
     else if (auto *SI = dyn_cast<StoreInst>(&I))
       PushPtrOperand(SI->getPointerOperand());

Added: llvm/trunk/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll?rev=301044&view=auto
==============================================================================
--- llvm/trunk/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll (added)
+++ llvm/trunk/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll Fri Apr 21 16:35:04 2017
@@ -0,0 +1,48 @@
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -infer-address-spaces %s | FileCheck %s
+
+; Test that pure GetElementPtr instructions not directly connected to
+; a memory operation are inferred.
+
+ at lds = internal unnamed_addr addrspace(3) global [648 x double] undef, align 8
+
+; CHECK-LABEL: @constexpr_gep_addrspacecast(
+; CHECK: %gep0 = getelementptr inbounds double, double addrspace(4)* addrspacecast (double addrspace(3)* getelementptr inbounds ([648 x double], [648 x double] addrspace(3)* @lds, i64 0, i64 384) to double addrspace(4)*), i64 %idx0
+; CHECK-NEXT: %asc = addrspacecast double addrspace(4)* %gep0 to double addrspace(3)*
+; CHECK-NEXT: store double 1.000000e+00, double addrspace(3)* %asc
+define void @constexpr_gep_addrspacecast(i64 %idx0, i64 %idx1) {
+  %gep0 = getelementptr inbounds double, double addrspace(4)* getelementptr ([648 x double], [648 x double] addrspace(4)* addrspacecast ([648 x double] addrspace(3)* @lds to [648 x double] addrspace(4)*), i64 0, i64 384), i64 %idx0
+  %asc = addrspacecast double addrspace(4)* %gep0 to double addrspace(3)*
+  store double 1.0, double addrspace(3)* %asc, align 8
+  ret void
+}
+
+; CHECK-LABEL: @constexpr_gep_gep_addrspacecast(
+; CHECK: %gep0 = getelementptr inbounds double, double addrspace(3)* getelementptr inbounds ([648 x double], [648 x double] addrspace(3)* @lds, i64 0, i64 384), i64 %idx0
+; CHECK-NEXT: %1 = addrspacecast double addrspace(3)* %gep0 to double addrspace(4)*
+; CHECK-NEXT: %gep1 = getelementptr inbounds double, double addrspace(4)* %1, i64 %idx1
+; CHECK-NEXT: %asc = addrspacecast double addrspace(4)* %gep1 to double addrspace(3)*
+; CHECK-NEXT: store double 1.000000e+00, double addrspace(3)* %asc, align 8
+define void @constexpr_gep_gep_addrspacecast(i64 %idx0, i64 %idx1) {
+  %gep0 = getelementptr inbounds double, double addrspace(4)* getelementptr ([648 x double], [648 x double] addrspace(4)* addrspacecast ([648 x double] addrspace(3)* @lds to [648 x double] addrspace(4)*), i64 0, i64 384), i64 %idx0
+  %gep1 = getelementptr inbounds double, double addrspace(4)* %gep0, i64 %idx1
+  %asc = addrspacecast double addrspace(4)* %gep1 to double addrspace(3)*
+  store double 1.0, double addrspace(3)* %asc, align 8
+  ret void
+}
+
+; Don't crash
+; CHECK-LABEL: @vector_gep(
+; CHECK: %cast = addrspacecast <4 x [1024 x i32] addrspace(3)*> %array to <4 x [1024 x i32] addrspace(4)*>
+define amdgpu_kernel void @vector_gep(<4 x [1024 x i32] addrspace(3)*> %array) nounwind {
+  %cast = addrspacecast <4 x [1024 x i32] addrspace(3)*> %array to <4 x [1024 x i32] addrspace(4)*>
+  %p = getelementptr [1024 x i32], <4 x [1024 x i32] addrspace(4)*> %cast, <4 x i16> zeroinitializer, <4 x i16> <i16 16, i16 16, i16 16, i16 16>
+  %p0 = extractelement <4 x i32 addrspace(4)*> %p, i32 0
+  %p1 = extractelement <4 x i32 addrspace(4)*> %p, i32 1
+  %p2 = extractelement <4 x i32 addrspace(4)*> %p, i32 2
+  %p3 = extractelement <4 x i32 addrspace(4)*> %p, i32 3
+  store i32 99, i32 addrspace(4)* %p0
+  store i32 99, i32 addrspace(4)* %p1
+  store i32 99, i32 addrspace(4)* %p2
+  store i32 99, i32 addrspace(4)* %p3
+  ret void
+}