[ARM] Copy SMAX(lhs, 0) and SMIN(lhs, 0) patterns from AArch64 to ARM

They work on ARM too.

Author: AZero13

llvm/lib/Target/ARM/ARMISelLowering.cpp

@@ -5540,6 +5540,24 @@ SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
                       DAG.getConstant(VT.getSizeInBits() - 1, dl, VT));
       return DAG.getNode(ISD::OR, dl, VT, Shift, DAG.getConstant(1, dl, VT));
     }
+
+    // Check for SMAX(lhs, 0) and SMIN(lhs, 0) patterns.
+    // (SELECT_CC setgt, lhs, 0, lhs, 0) -> (BIC lhs, (SRA lhs, typesize-1))
+    // (SELECT_CC setlt, lhs, 0, lhs, 0) -> (AND lhs, (SRA lhs, typesize-1))
+    // Both require less instructions than compare and conditional select.
+    if ((CC == ISD::SETGT || CC == ISD::SETLT) && LHS == TrueVal && RHSC &&
+        RHSC->isZero() && CFVal && CFVal->isZero() &&
+        LHS.getValueType() == RHS.getValueType()) {
+      EVT VT = LHS.getValueType();
+      SDValue Shift =
+          DAG.getNode(ISD::SRA, dl, VT, LHS,
+                      DAG.getConstant(VT.getSizeInBits() - 1, dl, VT));
+
+      if (CC == ISD::SETGT)
+        Shift = DAG.getNOT(dl, Shift, VT);
+
+      return DAG.getNode(ISD::AND, dl, VT, LHS, Shift);
+    }
   }
 
   if (Subtarget->hasV8_1MMainlineOps() && CFVal && CTVal &&

llvm/test/CodeGen/ARM/min-max-combine.ll

@@ -83,26 +83,20 @@ define i8 @smini8_zero(i8 %a) {
 ; ARM-LABEL: smini8_zero:
 ; ARM:       @ %bb.0:
 ; ARM-NEXT:    sxtb r0, r0
-; ARM-NEXT:    cmp r0, #0
-; ARM-NEXT:    movpl r0, #0
+; ARM-NEXT:    and r0, r0, r0, asr #31
 ; ARM-NEXT:    bx lr
 ;
 ; THUMB-LABEL: smini8_zero:
 ; THUMB:       @ %bb.0:
-; THUMB-NEXT:    sxtb r0, r0
-; THUMB-NEXT:    cmp r0, #0
-; THUMB-NEXT:    bmi .LBB3_2
-; THUMB-NEXT:  @ %bb.1:
-; THUMB-NEXT:    movs r0, #0
-; THUMB-NEXT:  .LBB3_2:
+; THUMB-NEXT:    sxtb r1, r0
+; THUMB-NEXT:    asrs r0, r1, #31
+; THUMB-NEXT:    ands r0, r1
 ; THUMB-NEXT:    bx lr
 ;
 ; THUMB2-LABEL: smini8_zero:
 ; THUMB2:       @ %bb.0:
 ; THUMB2-NEXT:    sxtb r0, r0
-; THUMB2-NEXT:    cmp r0, #0
-; THUMB2-NEXT:    it pl
-; THUMB2-NEXT:    movpl r0, #0
+; THUMB2-NEXT:    and.w r0, r0, r0, asr #31
 ; THUMB2-NEXT:    bx lr
   %c = call i8 @llvm.smin.i8(i8 %a, i8 0)
   ret i8 %c
@@ -114,26 +108,20 @@ define i16 @smini16_zero(i16 %a) {
 ; ARM-LABEL: smini16_zero:
 ; ARM:       @ %bb.0:
 ; ARM-NEXT:    sxth r0, r0
-; ARM-NEXT:    cmp r0, #0
-; ARM-NEXT:    movpl r0, #0
+; ARM-NEXT:    and r0, r0, r0, asr #31
 ; ARM-NEXT:    bx lr
 ;
 ; THUMB-LABEL: smini16_zero:
 ; THUMB:       @ %bb.0:
-; THUMB-NEXT:    sxth r0, r0
-; THUMB-NEXT:    cmp r0, #0
-; THUMB-NEXT:    bmi .LBB4_2
-; THUMB-NEXT:  @ %bb.1:
-; THUMB-NEXT:    movs r0, #0
-; THUMB-NEXT:  .LBB4_2:
+; THUMB-NEXT:    sxth r1, r0
+; THUMB-NEXT:    asrs r0, r1, #31
+; THUMB-NEXT:    ands r0, r1
 ; THUMB-NEXT:    bx lr
 ;
 ; THUMB2-LABEL: smini16_zero:
 ; THUMB2:       @ %bb.0:
 ; THUMB2-NEXT:    sxth r0, r0
-; THUMB2-NEXT:    cmp r0, #0
-; THUMB2-NEXT:    it pl
-; THUMB2-NEXT:    movpl r0, #0
+; THUMB2-NEXT:    and.w r0, r0, r0, asr #31
 ; THUMB2-NEXT:    bx lr
   %c = call i16 @llvm.smin.i16(i16 %a, i16 0)
   ret i16 %c
@@ -144,24 +132,18 @@ declare i32 @llvm.smin.i32(i32 %a, i32 %b) readnone
 define i32 @smini32_zero(i32 %a) {
 ; ARM-LABEL: smini32_zero:
 ; ARM:       @ %bb.0:
-; ARM-NEXT:    cmp r0, #0
-; ARM-NEXT:    movpl r0, #0
+; ARM-NEXT:    and r0, r0, r0, asr #31
 ; ARM-NEXT:    bx lr
 ;
 ; THUMB-LABEL: smini32_zero:
 ; THUMB:       @ %bb.0:
-; THUMB-NEXT:    cmp r0, #0
-; THUMB-NEXT:    bmi .LBB5_2
-; THUMB-NEXT:  @ %bb.1:
-; THUMB-NEXT:    movs r0, #0
-; THUMB-NEXT:  .LBB5_2:
+; THUMB-NEXT:    asrs r1, r0, #31
+; THUMB-NEXT:    ands r0, r1
 ; THUMB-NEXT:    bx lr
 ;
 ; THUMB2-LABEL: smini32_zero:
 ; THUMB2:       @ %bb.0:
-; THUMB2-NEXT:    cmp r0, #0
-; THUMB2-NEXT:    it pl
-; THUMB2-NEXT:    movpl r0, #0
+; THUMB2-NEXT:    and.w r0, r0, r0, asr #31
 ; THUMB2-NEXT:    bx lr
   %c = call i32 @llvm.smin.i32(i32 %a, i32 0)
   ret i32 %c

llvm/test/CodeGen/Thumb2/LowOverheadLoops/predicated-liveout-unknown-lanes.ll

@@ -6,8 +6,7 @@ define arm_aapcs_vfpcc <4 x float> @arm_max_no_idx_f32_mve(ptr %pSrc, i32 %block
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r7, lr}
 ; CHECK-NEXT:    push {r7, lr}
-; CHECK-NEXT:    subs r2, r1, #4
-; CHECK-NEXT:    movw r3, #0
+; CHECK-NEXT:    movs r3, #0
 ; CHECK-NEXT:    movt r3, #65408
 ; CHECK-NEXT:    vdup.32 q0, r3
 ; CHECK-NEXT:    dlstp.32 lr, r1

llvm/test/CodeGen/Thumb2/LowOverheadLoops/unpredload.ll

@@ -5,7 +5,6 @@ define void @arm_cmplx_mag_squared_q15_mve(ptr %pSrc, ptr %pDst, i32 %blockSize)
 ; CHECK-LABEL: arm_cmplx_mag_squared_q15_mve:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    push {r7, lr}
-; CHECK-NEXT:    subs.w r3, r2, #8
 ; CHECK-NEXT:    dlstp.16 lr, r2
 ; CHECK-NEXT:  .LBB0_1: @ %do.body
 ; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1