[PowerPC] Relax the checking on AND/AND8 in isSignOrZeroExtended

[sfertile]

Summary

Separate the handling of AND/AND8 out from PHI/OR/ISEL checking. The reasoning is the others need all their operands to be sign/zero extended for their output to also be sign/zero extended. This is true for AND and sign-extension, but for zero-extension we only need at least one of the input operands to be sign extended for the result to also be zero extended.

[nemanjai]

LGTM.
I think this code may be off by default at this time due to issues on some external benchmarks. If that's the case, it should be fine to commit this, but you'll probably need the flag in the test case to enable the transformation that uses this.

lib/Target/PowerPC/PPCInstrInfo.cpp

2336 ↗

(On Diff #119540)

Please turn this into an assert. If we have non-register input operands to a reg+reg instruction, it's an error and we shouldn't just continue.

[inouehrs]

I am submitting a patch that may conflict with this change. I expect that it is trivial to resolve the conflict.
https://reviews.llvm.org/D40554

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

case PPC::XORIS8: {
     const MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
     if (SrcMI != NULL)
       return isSignOrZeroExtended(*SrcMI, SignExt, Depth);
 
     return false;
   }
 
   // If all incoming values are sign-/zero-extended,
-  // the output of AND, OR, ISEL or PHI is also sign-/zero-extended.
-  case PPC::AND:
-  case PPC::AND8:
+  // the output of OR, ISEL or PHI is also sign-/zero-extended.
   case PPC::OR:
   case PPC::OR8:
   case PPC::ISEL:
   case PPC::PHI: {
     if (Depth >= MAX_DEPTH)
       return false;
 
     // The input registers for PHI are operand 1, 3, ...
for (unsigned I = 1; I != E; I += D) {
           return false;
       }
       else
         return false;
     }
     return true;
   }
 
+  // If at least one of the incoming values of an AND is zero extended
+  // then the output is also zero-extended. If both of the incoming values
+  // are sign-extended then the output is also sign extended.
+  case PPC::AND:
+  case PPC::AND8: {
+    if (Depth >= MAX_DEPTH)
+       return false;
+
+    assert(MI.getOperand(1).isReg() && MI.getOperand(2).isReg());
+
+    unsigned SrcReg1 = MI.getOperand(1).getReg();
+    unsigned SrcReg2 = MI.getOperand(2).getReg();
+
+    if (!TargetRegisterInfo::isVirtualRegister(SrcReg1) ||
+        !TargetRegisterInfo::isVirtualRegister(SrcReg2))
+       return false;
+
+    const MachineInstr *MISrc1 = MRI->getVRegDef(SrcReg1);
+    const MachineInstr *MISrc2 = MRI->getVRegDef(SrcReg2);
+    if (!MISrc1 || !MISrc2)
+        return false;
+
+    if(SignExt)
+        return isSignOrZeroExtended(*MISrc1, SignExt, Depth+1) &&
+               isSignOrZeroExtended(*MISrc2, SignExt, Depth+1);
+    else
+        return isSignOrZeroExtended(*MISrc1, SignExt, Depth+1) ||
+               isSignOrZeroExtended(*MISrc2, SignExt, Depth+1);
+  }
+
   default:
     break;
   }
   return false;
 }

llvm/trunk/test/CodeGen/PowerPC/zext-and-cmp.ll

+; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu  < %s | FileCheck %s
+
+; Test that we recognize that an 'and' instruction that feeds a comparison
+; to zero can be simplifed by using the record form when one of its operands
+; is known to be zero extended.
+
+@k = common local_unnamed_addr global i32 0, align 4
+
+; Function Attrs: norecurse nounwind
+define signext i32 @cmplwi(i32* nocapture readonly %p, i32* nocapture readonly %q, i32 signext %j, i32 signext %r10) {
+entry:
+  %0 = load i32, i32* %q, align 4
+  %shl = shl i32 %0, %j
+  %1 = load i32, i32* %p, align 4
+  %and = and i32 %shl, %r10
+  %and1 = and i32 %and, %1
+  %tobool = icmp eq i32 %and1, 0
+  br i1 %tobool, label %cleanup, label %if.then
+
+if.then:
+  store i32 %j, i32* @k, align 4
+  br label %cleanup
+
+cleanup:
+  %retval.0 = phi i32 [ 0, %if.then ], [ 1, %entry ]
+  ret i32 %retval.0
+}
+
+; CHECK-LABEL: cmplwi:
+; CHECK:      lwz [[T1:[0-9]+]], 0(3)
+; CHECK:      and. {{[0-9]+}}, {{[0-9]+}}, [[T1]]
+; CHECK-NOT:  cmplwi
+; CHECK-NEXT: beq      0,