[X86] Generate testl instruction through truncates.

[deadalnix]

Summary

This was introduced in D42646 but ended up being reverted because the original implementation was buggy.

Depends on D42646

[niravd]

This looks good modulo test cases.

[deadalnix]

@niravd I added a test case in rL323868 . The problem was by trying to extract the lower 32 bits of a 16 bit register, which is obviously not possible.

lib/Target/X86/X86ISelDAGToDAG.cpp

3049

The problem doesn't happen with i8 because it is checked here.

3074

We check more i16 here to avoid running into the same problem we did before.

[niravd]

LGTM modulo comment nit.

lib/Target/X86/X86ISelDAGToDAG.cpp

3074

The i16 comparision took me a moment to understand. Can you leave a comment here explaining that N0 must be of size at least i32?

[deadalnix]

Add a comment to explain the check for MVT::i16 .

lib/Target/X86/X86ISelDAGToDAG.cpp

if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
       N0 = N0.getOperand(0);
 
     // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
     // use a smaller encoding.
     // Look past the truncate if CMP is the only use of it.
     if ((N0.getOpcode() == ISD::AND ||
          (N0.getResNo() == 0 && N0.getOpcode() == X86ISD::AND)) &&
         N0.getNode()->hasOneUse() &&
         N0.getValueType() != MVT::i8 &&
         X86::isZeroNode(N1)) {
       ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
       if (!C) break;
       uint64_t Mask = C->getZExtValue();
 
-      // For example, convert "testl %eax, $8" to "testb %al, $8"
+      MVT VT;
+      int SubRegOp;
+      unsigned Op;
+
       if (isUInt<8>(Mask) &&
           (!(Mask & 0x80) || hasNoSignedComparisonUses(Node))) {
-        SDValue Imm = CurDAG->getTargetConstant(Mask, dl, MVT::i8);
-        SDValue Reg = N0.getOperand(0);
-
-        // Extract the l-register.
-        SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl,
-                                                        MVT::i8, Reg);
-
-        // Emit a testb.
-        SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
-                                                 Subreg, Imm);
-        // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
-        // one, do not call ReplaceAllUsesWith.
-        ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
-                    SDValue(NewNode, 0));
-        CurDAG->RemoveDeadNode(Node);
-        return;
-      }
-
+        // For example, convert "testl %eax, $8" to "testb %al, $8"
+        VT = MVT::i8;
+        SubRegOp = X86::sub_8bit;
+        Op = X86::TEST8ri;
+      } else if (OptForMinSize && isUInt<16>(Mask) &&
+                 (!(Mask & 0x8000) || hasNoSignedComparisonUses(Node))) {
         // For example, "testl %eax, $32776" to "testw %ax, $32776".
         // NOTE: We only want to form TESTW instructions if optimizing for
         // min size. Otherwise we only save one byte and possibly get a length
         // changing prefix penalty in the decoders.
-      if (OptForMinSize && isUInt<16>(Mask) && N0.getValueType() != MVT::i16 &&
-          (!(Mask & 0x8000) || hasNoSignedComparisonUses(Node))) {
-        SDValue Imm = CurDAG->getTargetConstant(Mask, dl, MVT::i16);
-        SDValue Reg = N0.getOperand(0);
-
-        // Extract the 16-bit subregister.
-        SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_16bit, dl,
-                                                        MVT::i16, Reg);
-
-        // Emit a testw.
-        SDNode *NewNode = CurDAG->getMachineNode(X86::TEST16ri, dl, MVT::i32,
-                                                 Subreg, Imm);
-        // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
-        // one, do not call ReplaceAllUsesWith.
-        ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
-                    SDValue(NewNode, 0));
-        CurDAG->RemoveDeadNode(Node);
-        return;
+        VT = MVT::i16;
+        SubRegOp = X86::sub_16bit;
+        Op = X86::TEST16ri;
+      } else if (isUInt<32>(Mask) && N0.getValueType() != MVT::i16 &&
+                 (!(Mask & 0x80000000) || hasNoSignedComparisonUses(Node))) {
+        // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
+        // NOTE: We only want to run that transform if N0 is 32 or 64 bits.
+        // Otherwize, we find ourselves in a position where we have to do
+        // promotion. If previous passes did not promote the and, we assume
+        // they had a good reason not to and do not promote here.
+        VT = MVT::i32;
+        SubRegOp = X86::sub_32bit;
+        Op = X86::TEST32ri;
+      } else {
+        // No eligible transformation was found.
+        break;
       }
 
-      // For example, "testq %rax, $268468232" to "testl %eax, $268468232".
-      if (isUInt<32>(Mask) && N0.getValueType() == MVT::i64 &&
-          (!(Mask & 0x80000000) || hasNoSignedComparisonUses(Node))) {
-        SDValue Imm = CurDAG->getTargetConstant(Mask, dl, MVT::i32);
-        SDValue Reg = N0.getOperand(0);
+      SDValue Imm = CurDAG->getTargetConstant(Mask, dl, VT);
 
-        // Extract the 32-bit subregister.
-        SDValue Subreg = CurDAG->getTargetExtractSubreg(X86::sub_32bit, dl,
-                                                        MVT::i32, Reg);
-        // Emit a testl.
-        SDNode *NewNode = CurDAG->getMachineNode(X86::TEST32ri, dl, MVT::i32,
-                                                 Subreg, Imm);
-        // Replace SUB|CMP with TEST, since SUB has two outputs while TEST has
-        // one, do not call ReplaceAllUsesWith.
-        ReplaceUses(SDValue(Node, (Opcode == X86ISD::SUB ? 1 : 0)),
-                    SDValue(NewNode, 0));
-        CurDAG->RemoveDeadNode(Node);
-        return;
-      }
-    }
+      // Extract the subregister if necessary.
+      if (N0.getValueType() != VT)
+        Reg = CurDAG->getTargetExtractSubreg(SubRegOp, dl, VT, Reg);
+
+      // Emit a testl or testw.
+      SDNode *NewNode = CurDAG->getMachineNode(Op, dl, MVT::i32, Reg, Imm);
     break;
   }
   case ISD::STORE:
     if (foldLoadStoreIntoMemOperand(Node))
       return;
     break;
   }
 

test/CodeGen/X86/test-shrink-bug.ll

 define void @fail(i16 %a, <2 x i8> %b) {
 ; CHECK-X86-LABEL: fail:
 ; CHECK-X86:       ## %bb.0:
 ; CHECK-X86-NEXT:    subl $12, %esp
 ; CHECK-X86-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-X86-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
 ; CHECK-X86-NEXT:    cmpb $123, {{[0-9]+}}(%esp)
 ; CHECK-X86-NEXT:    sete %al
-; CHECK-X86-NEXT:    andl $263, %ecx ## imm = 0x107
-; CHECK-X86-NEXT:    testw %cx, %cx
+; CHECK-X86-NEXT:    testl $263, %ecx ## imm = 0x107
 ; CHECK-X86-NEXT:    je LBB1_2
 ; CHECK-X86-NEXT:  ## %bb.1:
 ; CHECK-X86-NEXT:    testb %al, %al
 ; CHECK-X86-NEXT:    jne LBB1_2
 ; CHECK-X86-NEXT:  ## %bb.3: ## %no
 ; CHECK-X86-NEXT:    calll _bar
 ; CHECK-X86-NEXT:    addl $12, %esp
 ; CHECK-X86-NEXT:    retl

test/CodeGen/X86/test-shrink.ll

yes:
   ret void
 no:
   ret void
 }
 
 define void @truncand32(i16 inreg %x) nounwind {
 ; CHECK-LINUX64-LABEL: truncand32:
 ; CHECK-LINUX64:       # %bb.0:
-; CHECK-LINUX64-NEXT:    andl $2049, %edi # imm = 0x801
-; CHECK-LINUX64-NEXT:    testw %di, %di
+; CHECK-LINUX64-NEXT:    testl $2049, %edi # imm = 0x801
 ; CHECK-LINUX64-NEXT:    je .LBB11_1
 ; CHECK-LINUX64-NEXT:  # %bb.2: # %no
 ; CHECK-LINUX64-NEXT:    retq
 ; CHECK-LINUX64-NEXT:  .LBB11_1: # %yes
 ; CHECK-LINUX64-NEXT:    pushq %rax
 ; CHECK-LINUX64-NEXT:    callq bar
 ; CHECK-LINUX64-NEXT:    popq %rax
 ; CHECK-LINUX64-NEXT:    retq
 ;
 ; CHECK-WIN32-64-LABEL: truncand32:
 ; CHECK-WIN32-64:       # %bb.0:
 ; CHECK-WIN32-64-NEXT:    subq $40, %rsp
-; CHECK-WIN32-64-NEXT:    andl $2049, %ecx # imm = 0x801
-; CHECK-WIN32-64-NEXT:    testw %cx, %cx
+; CHECK-WIN32-64-NEXT:    testl $2049, %ecx # imm = 0x801
 ; CHECK-WIN32-64-NEXT:    je .LBB11_1
 ; CHECK-WIN32-64-NEXT:  # %bb.2: # %no
 ; CHECK-WIN32-64-NEXT:    addq $40, %rsp
 ; CHECK-WIN32-64-NEXT:    retq
 ; CHECK-WIN32-64-NEXT:  .LBB11_1: # %yes
 ; CHECK-WIN32-64-NEXT:    callq bar
 ; CHECK-WIN32-64-NEXT:    addq $40, %rsp
 ; CHECK-WIN32-64-NEXT:    retq
 ;
 ; CHECK-X86-LABEL: truncand32:
 ; CHECK-X86:       # %bb.0:
-; CHECK-X86-NEXT:    andl $2049, %eax # imm = 0x801
-; CHECK-X86-NEXT:    testw %ax, %ax
+; CHECK-X86-NEXT:    testl $2049, %eax # imm = 0x801
 ; CHECK-X86-NEXT:    je .LBB11_1
 ; CHECK-X86-NEXT:  # %bb.2: # %no
 ; CHECK-X86-NEXT:    retl
 ; CHECK-X86-NEXT:  .LBB11_1: # %yes
 ; CHECK-X86-NEXT:    calll bar
 ; CHECK-X86-NEXT:    retl
   %t = and i16 %x, 2049
   %s = icmp eq i16 %t, 0