[ARM] Fold away CMP/CSINC from CMOV

This makes use of the code in D114013 to fold away unnecessary
CMPZ/CSINC starting from a CMOV, in a similar way to how we fold away
CSINV/CSINC/etc

Differential Revision: https://reviews.llvm.org/D115185

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 0d8f265..e38938e 100644 (file)
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -18027,6 +18027,23 @@ ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
   if (!VT.isInteger())
       return SDValue();
 
+  // Fold away an unneccessary CMPZ/CMOV
+  // CMOV A, B, C1, $cpsr, (CMPZ (CMOV 1, 0, C2, D), 0) ->
+  // if C1==EQ -> CMOV A, B, C2, $cpsr, D
+  // if C1==NE -> CMOV A, B, NOT(C2), $cpsr, D
+  if (N->getConstantOperandVal(2) == ARMCC::EQ ||
+      N->getConstantOperandVal(2) == ARMCC::NE) {
+    ARMCC::CondCodes Cond;
+    if (SDValue C = IsCMPZCSINC(N->getOperand(4).getNode(), Cond)) {
+      if (N->getConstantOperandVal(2) == ARMCC::NE)
+        Cond = ARMCC::getOppositeCondition(Cond);
+      return DAG.getNode(N->getOpcode(), SDLoc(N), MVT::i32, N->getOperand(0),
+                         N->getOperand(1),
+                         DAG.getTargetConstant(Cond, SDLoc(N), MVT::i32),
+                         N->getOperand(3), C);
+    }
+  }
+
   // Materialize a boolean comparison for integers so we can avoid branching.
   if (isNullConstant(FalseVal)) {
     if (CC == ARMCC::EQ && isOneConstant(TrueVal)) {

diff --git a/llvm/test/CodeGen/Thumb2/mve-pred-or.ll b/llvm/test/CodeGen/Thumb2/mve-pred-or.ll
index 2310e38..cb3f554 100644 (file)
--- a/llvm/test/CodeGen/Thumb2/mve-pred-or.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-pred-or.ll
@@ -380,24 +380,22 @@ define arm_aapcs_vfpcc <2 x i64> @cmpeqz_v2i1(<2 x i64> %a, <2 x i64> %b) {
 ; CHECK-NEXT:    vmov r0, r1, d0
 ; CHECK-NEXT:    orrs r0, r1
 ; CHECK-NEXT:    vmov r1, r2, d2
-; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r1, r2
 ; CHECK-NEXT:    cset r1, eq
 ; CHECK-NEXT:    cmp r0, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    movne r1, #1
+; CHECK-NEXT:    it eq
+; CHECK-NEXT:    moveq r1, #1
 ; CHECK-NEXT:    rsbs r0, r1, #0
 ; CHECK-NEXT:    movs r1, #0
 ; CHECK-NEXT:    bfi r1, r0, #0, #8
 ; CHECK-NEXT:    vmov r0, r2, d1
 ; CHECK-NEXT:    orrs r0, r2
 ; CHECK-NEXT:    vmov r2, r3, d3
-; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r2, r3
 ; CHECK-NEXT:    cset r2, eq
 ; CHECK-NEXT:    cmp r0, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    movne r2, #1
+; CHECK-NEXT:    it eq
+; CHECK-NEXT:    moveq r2, #1
 ; CHECK-NEXT:    rsbs r0, r2, #0
 ; CHECK-NEXT:    bfi r1, r0, #8, #8
 ; CHECK-NEXT:    vmsr p0, r1
@@ -422,14 +420,12 @@ define arm_aapcs_vfpcc <2 x i64> @cmpeq_v2i1(<2 x i64> %a, <2 x i64> %b, <2 x i6
 ; CHECK-NEXT:    vmov r1, r2, d0
 ; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r1, r2
-; CHECK-NEXT:    vmov r12, r2, d5
-; CHECK-NEXT:    cset r1, eq
-; CHECK-NEXT:    cmp r1, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    movne r0, #1
+; CHECK-NEXT:    it eq
+; CHECK-NEXT:    moveq r0, #1
 ; CHECK-NEXT:    rsbs r0, r0, #0
 ; CHECK-NEXT:    movs r1, #0
 ; CHECK-NEXT:    bfi r1, r0, #0, #8
+; CHECK-NEXT:    vmov r12, r2, d5
 ; CHECK-NEXT:    vmov r3, r0, d3
 ; CHECK-NEXT:    eors r0, r2
 ; CHECK-NEXT:    eor.w r2, r3, r12
@@ -437,10 +433,8 @@ define arm_aapcs_vfpcc <2 x i64> @cmpeq_v2i1(<2 x i64> %a, <2 x i64> %b, <2 x i6
 ; CHECK-NEXT:    vmov r2, r3, d1
 ; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r2, r3
-; CHECK-NEXT:    cset r2, eq
-; CHECK-NEXT:    cmp r2, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    movne r0, #1
+; CHECK-NEXT:    it eq
+; CHECK-NEXT:    moveq r0, #1
 ; CHECK-NEXT:    rsbs r0, r0, #0
 ; CHECK-NEXT:    bfi r1, r0, #8, #8
 ; CHECK-NEXT:    vmsr p0, r1

diff --git a/llvm/test/CodeGen/Thumb2/mve-pred-vselect.ll b/llvm/test/CodeGen/Thumb2/mve-pred-vselect.ll
index b53b1d9..747021e 100644 (file)
--- a/llvm/test/CodeGen/Thumb2/mve-pred-vselect.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-pred-vselect.ll
@@ -73,28 +73,26 @@ define arm_aapcs_vfpcc <2 x i64> @cmpeqz_v2i1(<2 x i64> %a, <2 x i64> %b, <2 x i
 ; CHECK-NEXT:    vmov r0, r1, d4
 ; CHECK-NEXT:    orrs r0, r1
 ; CHECK-NEXT:    vmov r1, r2, d0
-; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r1, r2
 ; CHECK-NEXT:    vmov r2, r3, d2
 ; CHECK-NEXT:    cset r1, eq
 ; CHECK-NEXT:    orrs r2, r3
 ; CHECK-NEXT:    cset r2, eq
 ; CHECK-NEXT:    cmp r0, #0
-; CHECK-NEXT:    csel r0, r1, r2, ne
+; CHECK-NEXT:    csel r0, r1, r2, eq
 ; CHECK-NEXT:    movs r1, #0
 ; CHECK-NEXT:    rsbs r0, r0, #0
 ; CHECK-NEXT:    bfi r1, r0, #0, #8
 ; CHECK-NEXT:    vmov r0, r2, d5
-; CHECK-NEXT:    orrs r0, r2
+; CHECK-NEXT:    orr.w r12, r0, r2
 ; CHECK-NEXT:    vmov r2, r3, d1
-; CHECK-NEXT:    cset r12, eq
 ; CHECK-NEXT:    orrs r2, r3
 ; CHECK-NEXT:    vmov r3, r0, d3
 ; CHECK-NEXT:    cset r2, eq
 ; CHECK-NEXT:    orrs r0, r3
 ; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    cmp.w r12, #0
-; CHECK-NEXT:    csel r0, r2, r0, ne
+; CHECK-NEXT:    csel r0, r2, r0, eq
 ; CHECK-NEXT:    rsbs r0, r0, #0
 ; CHECK-NEXT:    bfi r1, r0, #8, #8
 ; CHECK-NEXT:    vmsr p0, r1

diff --git a/llvm/test/CodeGen/Thumb2/mve-pred-xor.ll b/llvm/test/CodeGen/Thumb2/mve-pred-xor.ll
index 8741b2b..e5fef33 100644 (file)
--- a/llvm/test/CodeGen/Thumb2/mve-pred-xor.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-pred-xor.ll
@@ -460,24 +460,22 @@ define arm_aapcs_vfpcc <2 x i64> @cmpeqz_v2i1(<2 x i64> %a, <2 x i64> %b) {
 ; CHECK-NEXT:    vmov r0, r1, d0
 ; CHECK-NEXT:    orrs r0, r1
 ; CHECK-NEXT:    vmov r1, r2, d2
-; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r1, r2
 ; CHECK-NEXT:    cset r1, eq
 ; CHECK-NEXT:    cmp r0, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    eorne r1, r1, #1
+; CHECK-NEXT:    it eq
+; CHECK-NEXT:    eoreq r1, r1, #1
 ; CHECK-NEXT:    rsbs r0, r1, #0
 ; CHECK-NEXT:    movs r1, #0
 ; CHECK-NEXT:    bfi r1, r0, #0, #8
 ; CHECK-NEXT:    vmov r0, r2, d1
 ; CHECK-NEXT:    orrs r0, r2
 ; CHECK-NEXT:    vmov r2, r3, d3
-; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r2, r3
 ; CHECK-NEXT:    cset r2, eq
 ; CHECK-NEXT:    cmp r0, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    eorne r2, r2, #1
+; CHECK-NEXT:    it eq
+; CHECK-NEXT:    eoreq r2, r2, #1
 ; CHECK-NEXT:    rsbs r0, r2, #0
 ; CHECK-NEXT:    bfi r1, r0, #8, #8
 ; CHECK-NEXT:    vmsr p0, r1
@@ -502,14 +500,12 @@ define arm_aapcs_vfpcc <2 x i64> @cmpeq_v2i1(<2 x i64> %a, <2 x i64> %b, <2 x i6
 ; CHECK-NEXT:    vmov r1, r2, d0
 ; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r1, r2
-; CHECK-NEXT:    vmov r12, r2, d5
-; CHECK-NEXT:    cset r1, eq
-; CHECK-NEXT:    cmp r1, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    eorne r0, r0, #1
+; CHECK-NEXT:    it eq
+; CHECK-NEXT:    eoreq r0, r0, #1
 ; CHECK-NEXT:    rsbs r0, r0, #0
 ; CHECK-NEXT:    movs r1, #0
 ; CHECK-NEXT:    bfi r1, r0, #0, #8
+; CHECK-NEXT:    vmov r12, r2, d5
 ; CHECK-NEXT:    vmov r3, r0, d3
 ; CHECK-NEXT:    eors r0, r2
 ; CHECK-NEXT:    eor.w r2, r3, r12
@@ -517,10 +513,8 @@ define arm_aapcs_vfpcc <2 x i64> @cmpeq_v2i1(<2 x i64> %a, <2 x i64> %b, <2 x i6
 ; CHECK-NEXT:    vmov r2, r3, d1
 ; CHECK-NEXT:    cset r0, eq
 ; CHECK-NEXT:    orrs r2, r3
-; CHECK-NEXT:    cset r2, eq
-; CHECK-NEXT:    cmp r2, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    eorne r0, r0, #1
+; CHECK-NEXT:    it eq
+; CHECK-NEXT:    eoreq r0, r0, #1
 ; CHECK-NEXT:    rsbs r0, r0, #0
 ; CHECK-NEXT:    bfi r1, r0, #8, #8
 ; CHECK-NEXT:    vmsr p0, r1

diff --git a/llvm/test/CodeGen/Thumb2/mve-saturating-arith.ll b/llvm/test/CodeGen/Thumb2/mve-saturating-arith.ll
index 5dd8d2d..10ecdc0 100644 (file)
--- a/llvm/test/CodeGen/Thumb2/mve-saturating-arith.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-saturating-arith.ll
@@ -47,12 +47,10 @@ define arm_aapcs_vfpcc <2 x i64> @sadd_int64_t(<2 x i64> %src1, <2 x i64> %src2)
 ; CHECK-NEXT:    cmp r1, #0
 ; CHECK-NEXT:    cset r1, ne
 ; CHECK-NEXT:    cmp r2, #0
-; CHECK-NEXT:    cset r2, mi
-; CHECK-NEXT:    cmp r2, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    eorne r1, r1, #1
-; CHECK-NEXT:    rsbs r1, r1, #0
+; CHECK-NEXT:    it mi
+; CHECK-NEXT:    eormi r1, r1, #1
 ; CHECK-NEXT:    movs r2, #0
+; CHECK-NEXT:    rsbs r1, r1, #0
 ; CHECK-NEXT:    bfi r2, r1, #0, #8
 ; CHECK-NEXT:    vmov r1, r3, d3
 ; CHECK-NEXT:    adds r1, r1, r0
@@ -66,10 +64,8 @@ define arm_aapcs_vfpcc <2 x i64> @sadd_int64_t(<2 x i64> %src1, <2 x i64> %src2)
 ; CHECK-NEXT:    vmov q0[3], q0[1], lr, r5
 ; CHECK-NEXT:    cset r0, ne
 ; CHECK-NEXT:    cmp r3, #0
-; CHECK-NEXT:    cset r3, mi
-; CHECK-NEXT:    cmp r3, #0
-; CHECK-NEXT:    it ne
-; CHECK-NEXT:    eorne r0, r0, #1
+; CHECK-NEXT:    it mi
+; CHECK-NEXT:    eormi r0, r0, #1
 ; CHECK-NEXT:    rsbs r0, r0, #0
 ; CHECK-NEXT:    bfi r2, r0, #8, #8
 ; CHECK-NEXT:    asrs r0, r5, #31