Fix address-offset folding for plain addi

When folding an addi into a memory access that can take an immediate offset, we
were implicitly assuming that the existing offset was zero. This was incorrect.
If we're dealing with an addi with a plain constant, we can add it to the
existing offset (assuming that doesn't overflow the immediate, etc.), but if we
have anything else (i.e. something that will become a relocation expression),
we'll go back to requiring the existing immediate offset to be zero (because we
don't know what the requirements on that relocation expression might be - e.g.
maybe it is paired with some addis in some relevant way).

On the other hand, when dealing with a plain addi with a regular constant
immediate, the alignment restrictions (from the TOC base pointer, etc.) are
irrelevant.

I've added the test case from PR30280, which demonstrated the bug, but also
demonstrates a missed optimization opportunity (i.e. we don't need the memory
accesses at all).

Fixes PR30280.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@280789 91177308-0d34-0410-b5e6-96231b3b80d8

Author: Hal Finkel

lib/Target/PowerPC/PPCISelDAGToDAG.cpp

@@ -4371,25 +4371,48 @@ void PPCDAGToDAGISel::PeepholePPC64() {
     SDValue HBase = Base.getOperand(0);
 
     int Offset = N->getConstantOperandVal(FirstOp);
-    if (Offset < 0 || Offset > MaxDisplacement) {
-      // If we have a addi(toc@l)/addis(toc@ha) pair, and the addis has only
-      // one use, then we can do this for any offset, we just need to also
-      // update the offset (i.e. the symbol addend) on the addis also.
-      if (Base.getMachineOpcode() != PPC::ADDItocL)
-        continue;
+    if (ReplaceFlags) {
+      if (Offset < 0 || Offset > MaxDisplacement) {
+        // If we have a addi(toc@l)/addis(toc@ha) pair, and the addis has only
+        // one use, then we can do this for any offset, we just need to also
+        // update the offset (i.e. the symbol addend) on the addis also.
+        if (Base.getMachineOpcode() != PPC::ADDItocL)
+          continue;
 
-      if (!HBase.isMachineOpcode() ||
-          HBase.getMachineOpcode() != PPC::ADDIStocHA)
-        continue;
+        if (!HBase.isMachineOpcode() ||
+            HBase.getMachineOpcode() != PPC::ADDIStocHA)
+          continue;
 
-      if (!Base.hasOneUse() || !HBase.hasOneUse())
-        continue;
+        if (!Base.hasOneUse() || !HBase.hasOneUse())
+          continue;
 
-      SDValue HImmOpnd = HBase.getOperand(1);
-      if (HImmOpnd != ImmOpnd)
-        continue;
+        SDValue HImmOpnd = HBase.getOperand(1);
+        if (HImmOpnd != ImmOpnd)
+          continue;
 
-      UpdateHBase = true;
+        UpdateHBase = true;
+      }
+    } else {
+      // If we're directly folding the addend from an addi instruction, then:
+      //  1. In general, the offset on the memory access must be zero.
+      //  2. If the addend is a constant, then it can be combined with a
+      //     non-zero offset, but only if the result meets the encoding
+      //     requirements.
+      if (auto *C = dyn_cast<ConstantSDNode>(ImmOpnd)) {
+        Offset += C->getSExtValue();
+
+        if ((StorageOpcode == PPC::LWA || StorageOpcode == PPC::LD ||
+             StorageOpcode == PPC::STD) && (Offset % 4) != 0)
+          continue;
+
+        if (!isInt<16>(Offset))
+          continue;
+
+        ImmOpnd = CurDAG->getTargetConstant(Offset, SDLoc(ImmOpnd),
+                                            ImmOpnd.getValueType());
+      } else if (Offset != 0) {
+        continue;
+      }
     }
 
     // We found an opportunity.  Reverse the operands from the add

test/CodeGen/PowerPC/addi-offset-fold.ll

@@ -0,0 +1,40 @@
+; RUN: llc < %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-n32:64"
+target triple = "powerpc64le-unknown-linux-gnu"
+
+%struct.S0 = type <{ i32, [5 x i8] }>
+
+; Function Attrs: norecurse nounwind readnone
+define signext i32 @foo([2 x i64] %a.coerce) local_unnamed_addr #0 {
+entry:
+  %a = alloca %struct.S0, align 8
+  %a.coerce.fca.0.extract = extractvalue [2 x i64] %a.coerce, 0
+  %a.coerce.fca.1.extract = extractvalue [2 x i64] %a.coerce, 1
+  %a.0.a.0..sroa_cast = bitcast %struct.S0* %a to i64*
+  store i64 %a.coerce.fca.0.extract, i64* %a.0.a.0..sroa_cast, align 8
+  %tmp.sroa.2.0.extract.trunc = trunc i64 %a.coerce.fca.1.extract to i8
+  %a.8.a.8..sroa_idx = getelementptr inbounds %struct.S0, %struct.S0* %a, i64 0, i32 1, i64 4
+  store i8 %tmp.sroa.2.0.extract.trunc, i8* %a.8.a.8..sroa_idx, align 8
+  %a.4.a.4..sroa_idx = getelementptr inbounds %struct.S0, %struct.S0* %a, i64 0, i32 1
+  %a.4.a.4..sroa_cast = bitcast [5 x i8]* %a.4.a.4..sroa_idx to i40*
+  %a.4.a.4.bf.load = load i40, i40* %a.4.a.4..sroa_cast, align 4
+  %bf.lshr = lshr i40 %a.4.a.4.bf.load, 31
+  %bf.lshr.tr = trunc i40 %bf.lshr to i32
+  %bf.cast = and i32 %bf.lshr.tr, 127
+  ret i32 %bf.cast
+
+; CHECK-LABEL: @foo
+; FIXME: We don't need to do these stores/loads at all.
+; CHECK-DAG: std 3, -24(1)
+; CHECK-DAG: stb 4, -16(1)
+; CHECK: ori 2, 2, 0
+; CHECK-DAG: lbz [[REG1:[0-9]+]], -16(1)
+; CHECK-DAG: lwz [[REG2:[0-9]+]], -20(1)
+; CHECK-DAG: sldi [[REG3:[0-9]+]], [[REG1]], 32
+; CHECK-DAG: or [[REG4:[0-9]+]], [[REG2]], [[REG3]]
+; CHECK: rldicl 3, [[REG4]], 33, 57
+; CHECK: blr
+}
+
+attributes #0 = { nounwind "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "target-cpu"="ppc64le" }
+