Fix a miscompile in the DAG combiner. Previously, we would incorrectly

try to reduce the width of this load, and would end up transforming:

  (truncate (lshr (sextload i48 <ptr> as i64), 32) to i32)
to
  (truncate (zextload i32 <ptr+4> as i64) to i32)

We lost the sext attached to the load while building the narrower i32
load, and replaced it with a zext because lshr always zext's the
results. Instead, bail out of this combine when there is a conflict
between a sextload and a zext narrowing. The rest of the DAG combiner
still optimize the code down to the proper single instruction:

  movswl 6(...),%eax

Which is exactly what we wanted. Previously we read past the end *and*
missed the sign extension:

  movl 6(...), %eax

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@169802 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index e774c138d2810f92c3521f09f9e56987508ba2c2..1c28d6dcaf4c76ee84fc66c421f85857ece73f89 100644 (file)
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -5068,11 +5068,15 @@ SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
       // At this point, we must have a load or else we can't do the transform.
       if (!isa<LoadSDNode>(N0)) return SDValue();
 
+      // Because a SRL must be assumed to *need* to zero-extend the high bits
+      // (as opposed to anyext the high bits), we can't combine the zextload
+      // lowering of SRL and an sextload.
+      if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
+        return SDValue();
+
       // If the shift amount is larger than the input type then we're not
       // accessing any of the loaded bytes.  If the load was a zextload/extload
       // then the result of the shift+trunc is zero/undef (handled elsewhere).
-      // If the load was a sextload then the result is a splat of the sign bit
-      // of the extended byte.  This is not worth optimizing for.
       if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
         return SDValue();
     }

diff --git a/test/CodeGen/X86/sext-load.ll b/test/CodeGen/X86/sext-load.ll
index c9b39d3a489e2d2e8dd76e230ef4fb6906278b28..58c93229a2c0a66ed2857f1edbb65132121078cb 100644 (file)
--- a/test/CodeGen/X86/sext-load.ll
+++ b/test/CodeGen/X86/sext-load.ll
@@ -1,9 +1,30 @@
-; RUN: llc < %s -march=x86 | grep movsbl
+; RUN: llc < %s -march=x86 | FileCheck %s
 
-define i32 @foo(i32 %X) nounwind  {
+; When doing sign extension, use the sext-load lowering to take advantage of
+; x86's sign extension during loads.
+;
+; CHECK: test1:
+; CHECK:      movsbl {{.*}}, %eax
+; CHECK-NEXT: ret
+define i32 @test1(i32 %X) nounwind  {
 entry:
        %tmp12 = trunc i32 %X to i8             ; <i8> [#uses=1]
        %tmp123 = sext i8 %tmp12 to i32         ; <i32> [#uses=1]
        ret i32 %tmp123
 }
 
+; When using a sextload representation, ensure that the sign extension is
+; preserved even when removing shifted-out low bits.
+;
+; CHECK: test2:
+; CHECK:      movswl {{.*}}, %eax
+; CHECK-NEXT: ret
+define i32 @test2({i16, [6 x i8]}* %this) {
+entry:
+  %b48 = getelementptr inbounds { i16, [6 x i8] }* %this, i32 0, i32 1
+  %cast = bitcast [6 x i8]* %b48 to i48*
+  %bf.load = load i48* %cast, align 2
+  %bf.ashr = ashr i48 %bf.load, 32
+  %bf.cast = trunc i48 %bf.ashr to i32
+  ret i32 %bf.cast
+}