[PATCH] DAGCombiner: Don't unnecessarily swap operands in ReassociateOps

Tom Stellard thomas.stellard at amd.com
Tue Feb 10 12:22:03 PST 2015


Hi resistor,

In the case where op = add, y = base_ptr, and x = offset, this
transform:

(op y, (op x, c1)) -> (op (op x, y), c1)

breaks the canonical form of add by putting the base pointer in the
second operand and the offset in the first.

This fix is important for the R600 target, because for some address
spaces the base pointer and the offset are stored in separate register
classes.  The old pattern caused the ISel code for matching addressing
modes to put the base pointer and offset in the wrong register classes,
which required no-trivial code transformations to fix.

http://reviews.llvm.org/D7540

Files:
  lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  test/CodeGen/R600/dagcombine-reassociate-bug.ll

Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -786,9 +786,9 @@
         return SDValue();
       }
       if (N1.hasOneUse()) {
-        // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
+        // reassoc. (op x, (op y, c1)) -> (op (op x, y), c1) iff x+c1 has one
         // use
-        SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
+        SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0, N1.getOperand(0));
         if (!OpNode.getNode())
           return SDValue();
         AddToWorklist(OpNode.getNode());
Index: test/CodeGen/R600/dagcombine-reassociate-bug.ll
===================================================================
--- /dev/null
+++ test/CodeGen/R600/dagcombine-reassociate-bug.ll
@@ -0,0 +1,33 @@
+; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck %s
+
+; Test for a bug where DAGCombiner::ReassociateOps() was creating adds
+; with offset in the first operand and base pointers in the second.
+
+; CHECK-LABEL: {{^}}store_same_base_ptr:
+; CHECK: buffer_store_dword v{{[0-9]+}}, [[VADDR:v\[[0-9]+:[0-9]+\]]], [[SADDR:s\[[0-9]+:[0-9]+\]]]
+; CHECK: buffer_store_dword v{{[0-9]+}}, [[VADDR]], [[SADDR]]
+; CHECK: buffer_store_dword v{{[0-9]+}}, [[VADDR]], [[SADDR]]
+; CHECK: buffer_store_dword v{{[0-9]+}}, [[VADDR]], [[SADDR]]
+
+define void @store_same_base_ptr(i32 addrspace(1)* %out) {
+entry:
+  %id = call i32 @llvm.r600.read.tidig.x() #0
+  %offset = sext i32 %id to i64
+  %offset0 = add i64 %offset, 1027
+  %ptr0 = getelementptr i32 addrspace(1)* %out, i64 %offset0
+  store i32 3, i32 addrspace(1)* %ptr0
+  %offset1 = add i64 %offset, 1026
+  %ptr1 = getelementptr i32 addrspace(1)* %out, i64 %offset1
+  store i32 2, i32 addrspace(1)* %ptr1
+  %offset2 = add i64 %offset, 1025
+  %ptr2 = getelementptr i32 addrspace(1)* %out, i64 %offset2
+  store i32 1, i32 addrspace(1)* %ptr2
+  %offset3 = add i64 %offset, 1024
+  %ptr3 = getelementptr i32 addrspace(1)* %out, i64 %offset3
+  store i32 0, i32 addrspace(1)* %ptr3
+  ret void
+}
+
+declare i32 @llvm.r600.read.tidig.x() #0
+
+attributes #0 = { readnone }

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D7540.19701.patch
Type: text/x-patch
Size: 2283 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150210/e57ce255/attachment.bin>


More information about the llvm-commits mailing list