[PATCH] D22103: [ARM] Skip inline asm memory operands in DAGToDAGISel

[Diana Picus via llvm-commits]

rovka created this revision.
rovka added reviewers: jmolloy, bogner, echristo.
rovka added subscribers: llvm-commits, rengolin.
Herald added subscribers: mehdi_amini, aemerson.

The current logic for handling inline asm operands in DAGToDAGISel interprets
the operands by looking for constants, which should represent the flags
describing the kind of operand we're dealing with (immediate, memory, register
def etc). The operands representing actual data are skipped only if they are
non-const, with the exception of immediate operands which are skipped explicitly
when a flag describing an immediate is found.
   
The oversight is that memory operands may be const too (e.g. for device drivers
reading a fixed address), so we should explicitly skip the operand following a
flag describing a memory operand. If we don't, we risk interpreting that
constant as a flag, which is definitely not intended.

http://reviews.llvm.org/D22103

Files:
  lib/Target/ARM/ARMISelDAGToDAG.cpp
  test/CodeGen/ARM/inlineasm3.ll

Index: test/CodeGen/ARM/inlineasm3.ll
===================================================================
--- test/CodeGen/ARM/inlineasm3.ll
+++ test/CodeGen/ARM/inlineasm3.ll
@@ -121,3 +121,11 @@
   %0 = tail call <4 x i32> asm "vld1.s32 {${0:e}[], ${0:f}[]}, [$1]", "=w,r"(i32* %p) nounwind
   ret <4 x i32> %0
 }
+
+; Bugzilla PR26038
+
+define i32 @fn1() local_unnamed_addr nounwind {
+entry:
+  %0 = tail call i32 asm "ldrh  $0, $1", "=r,*Q"(i8* inttoptr (i32 5 to i8*)) nounwind
+  ret i32 %0
+}
Index: lib/Target/ARM/ARMISelDAGToDAG.cpp
===================================================================
--- lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -4244,6 +4244,17 @@
     if (Changed && InlineAsm::isUseOperandTiedToDef(Flag, DefIdx))
       IsTiedToChangedOp = OpChanged[DefIdx];
 
+    // Memory operands to inline asm in the SelectionDAG are modeled with two
+    // operands: a constant of value InlineAsm::Kind_Mem followed by the input
+    // operand. If we get here and we have a Kind_Mem, skip the next operand (so
+    // it doesn't get misinterpreted), and continue. We do this here because
+    // it's important to update the OpChanged array correctly before moving on.
+    if (Kind == InlineAsm::Kind_Mem) {
+      SDValue op = N->getOperand(++i);
+      AsmNodeOperands.push_back(op);
+      continue;
+    }
+
     if (Kind != InlineAsm::Kind_RegUse && Kind != InlineAsm::Kind_RegDef
         && Kind != InlineAsm::Kind_RegDefEarlyClobber)
       continue;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D22103.63102.patch
Type: text/x-patch
Size: 1520 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160707/f83d7d68/attachment.bin>