[PATCH] D51066: [ARM] Handle all-ones mask explicitly in targetShrinkDemandedConstant.

[Eli Friedman via Phabricator via llvm-commits]

efriedma created this revision.
efriedma added reviewers: samparker, SjoerdMeijer, t.p.northover.
Herald added a reviewer: javed.absar.
Herald added a reviewer: bollu.
Herald added subscribers: chrib, kristof.beyls.

This avoids a potential infinite loop setting and unsetting bits in the mask.

Testcase reduced from a failure on the polly-aosp bot.


Repository:
  rL LLVM

https://reviews.llvm.org/D51066

Files:
  lib/Target/ARM/ARMISelLowering.cpp
  test/CodeGen/ARM/demanded-bits-and.ll


Index: test/CodeGen/ARM/demanded-bits-and.ll
===================================================================
--- /dev/null
+++ test/CodeGen/ARM/demanded-bits-and.ll
@@ -0,0 +1,35 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=arm-eabi < %s | FileCheck %s
+
+; Make sure this doesn't hang, and there are no unnecessary
+; "and" instructions.
+
+define dso_local void @f(i16* %p) {
+; CHECK-LABEL: f:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:  .LBB0_1: @ %bb
+; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    ldrh r1, [r0]
+; CHECK-NEXT:    and r2, r1, #255
+; CHECK-NEXT:    add r3, r2, r1, lsr #8
+; CHECK-NEXT:    add r2, r3, r2
+; CHECK-NEXT:    add r1, r2, r1, lsr #8
+; CHECK-NEXT:    add r1, r1, #2
+; CHECK-NEXT:    lsr r1, r1, #2
+; CHECK-NEXT:    strh r1, [r0]
+; CHECK-NEXT:    b .LBB0_1
+entry:
+  br label %bb
+
+bb:
+  %_p_scalar_ = load i16, i16* %p, align 2
+  %p_and = and i16 %_p_scalar_, 255
+  %p_ = lshr i16 %_p_scalar_, 8
+  %p_add = add nuw nsw i16 %p_, 2
+  %p_add14 = add nuw nsw i16 %p_add, %p_and
+  %p_add18 = add nuw nsw i16 %p_add14, %p_and
+  %p_add19 = add nuw nsw i16 %p_add18, %p_
+  %p_200 = lshr i16 %p_add19, 2
+  store i16 %p_200, i16* %p, align 2
+  br label %bb
+}
Index: lib/Target/ARM/ARMISelLowering.cpp
===================================================================
--- lib/Target/ARM/ARMISelLowering.cpp
+++ lib/Target/ARM/ARMISelLowering.cpp
@@ -13630,14 +13630,21 @@
 
   unsigned Mask = C->getZExtValue();
 
-  // If mask is zero, nothing to do.
-  if (!Mask)
-    return false;
-
   unsigned Demanded = DemandedAPInt.getZExtValue();
   unsigned ShrunkMask = Mask & Demanded;
   unsigned ExpandedMask = Mask | ~Demanded;
 
+  // If the mask is all zeros, let the target-independent code replace the
+  // result with zero.
+  if (ShrunkMask == 0)
+    return false;
+
+  // If the mask is all ones, erase the AND. (Currently, the target-independent
+  // code won't do this, so we have to do it explicitly to avoid an infinite
+  // loop in obscure cases.)
+  if (ExpandedMask == ~0U)
+    return TLO.CombineTo(Op, Op.getOperand(0));
+
   auto IsLegalMask = [ShrunkMask, ExpandedMask](unsigned Mask) -> bool {
     return (ShrunkMask & Mask) == ShrunkMask && (~ExpandedMask & Mask) == 0;
   };


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D51066.161831.patch
Type: text/x-patch
Size: 2334 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180821/e6984106/attachment.bin>