[PowerPC] Fix failure with STBRX when store is narrower than the bswap

[nemanjai]

Summary

Revision https://reviews.llvm.org/rL296811 causes https://bugs.llvm.org/show_bug.cgi?id=32140.
However, the code previously generated for the test case in that bug was wrong, so that revision is certainly needed. This patch just adds the truncate that was missing in that revision.

[hfinkel]

LGTM

llvm/trunk/lib/Target/PowerPC/PPCISelLowering.cpp

if (cast<StoreSDNode>(N)->isUnindexed() &&
       // Do an any-extend to 32-bits if this is a half-word input.
       if (BSwapOp.getValueType() == MVT::i16)
         BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
 
       // If the type of BSWAP operand is wider than stored memory width
       // it need to be shifted to the right side before STBRX.
       EVT mVT = cast<StoreSDNode>(N)->getMemoryVT();
       if (Op1VT.bitsGT(mVT)) {
-        int shift = Op1VT.getSizeInBits() - mVT.getSizeInBits();
+        int Shift = Op1VT.getSizeInBits() - mVT.getSizeInBits();
         BSwapOp = DAG.getNode(ISD::SRL, dl, Op1VT, BSwapOp,
-                              DAG.getConstant(shift, dl, MVT::i32));
+                              DAG.getConstant(Shift, dl, MVT::i32));
+        // Need to truncate if this is a bswap of i64 stored as i32/i16.
+        if (Op1VT == MVT::i64)
+          BSwapOp = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BSwapOp);
       }
 
       SDValue Ops[] = {
         N->getOperand(0), BSwapOp, N->getOperand(2), DAG.getValueType(mVT)
       };
       return
         DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
                                 Ops, cast<StoreSDNode>(N)->getMemoryVT(),

llvm/trunk/test/CodeGen/PowerPC/pr32140.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=powerpc64le-linux-gnu -mcpu=pwr8 < %s | FileCheck %s
+; RUN: llc -mtriple=powerpc64-linux-gnu -mcpu=pwr8 < %s | FileCheck %s
+
+@as = common local_unnamed_addr global i16 0, align 2
+@bs = common local_unnamed_addr global i16 0, align 2
+@ai = common local_unnamed_addr global i32 0, align 4
+@bi = common local_unnamed_addr global i32 0, align 4
+
+define void @bswapStorei64Toi32() {
+; CHECK-LABEL: bswapStorei64Toi32:
+; CHECK:       # BB#0: # %entry
+; CHECK:         lwa 3, 0(3)
+; CHECK-NEXT:    rldicl 3, 3, 32, 32
+; CHECK-NEXT:    stwbrx 3, 0, 4
+; CHECK-NEXT:    blr
+entry:
+  %0 = load i32, i32* @ai, align 4
+  %conv.i = sext i32 %0 to i64
+  %or26.i = tail call i64 @llvm.bswap.i64(i64 %conv.i)
+  %conv = trunc i64 %or26.i to i32
+  store i32 %conv, i32* @bi, align 4
+  ret void
+}
+
+define void @bswapStorei32Toi16() {
+; CHECK-LABEL: bswapStorei32Toi16:
+; CHECK:       # BB#0: # %entry
+; CHECK:         lha 3, 0(3)
+; CHECK-NEXT:    srwi 3, 3, 16
+; CHECK-NEXT:    sthbrx 3, 0, 4
+; CHECK-NEXT:    blr
+entry:
+  %0 = load i16, i16* @as, align 2
+  %conv.i = sext i16 %0 to i32
+  %or26.i = tail call i32 @llvm.bswap.i32(i32 %conv.i)
+  %conv = trunc i32 %or26.i to i16
+  store i16 %conv, i16* @bs, align 2
+  ret void
+}
+
+define void @bswapStorei64Toi16() {
+; CHECK-LABEL: bswapStorei64Toi16:
+; CHECK:       # BB#0: # %entry
+; CHECK:         lha 3, 0(3)
+; CHECK-NEXT:    rldicl 3, 3, 16, 48
+; CHECK-NEXT:    sthbrx 3, 0, 4
+; CHECK-NEXT:    blr
+entry:
+  %0 = load i16, i16* @as, align 2
+  %conv.i = sext i16 %0 to i64
+  %or26.i = tail call i64 @llvm.bswap.i64(i64 %conv.i)
+  %conv = trunc i64 %or26.i to i16
+  store i16 %conv, i16* @bs, align 2
+  ret void
+}
+
+declare i32 @llvm.bswap.i32(i32)
+declare i64 @llvm.bswap.i64(i64)