[AArch64] Add SUBS(CSEL) fold from brcond. (#142103)

This folds away subs(csel(1, 0, cc)) from brcond, that can be produced
in certain places from compares that are not already subs (like adc/sbc
generated from i128 add_with_overflow intrinsics).

Author: davemgreen

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -24668,6 +24668,19 @@ static SDValue performBRCONDCombine(SDNode *N,
   if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
     return SDValue();
 
+  // Fold away brcond(NE, cmp(csel(1, 0, CC, Cmp), 1)) -> brcond(~CC, Cmp)
+  if (isCMP(Cmp) && CC == AArch64CC::NE && isOneConstant(Cmp.getOperand(1))) {
+    SDValue CSel = Cmp.getOperand(0);
+    auto CSelCC = getCSETCondCode(CSel);
+    if (CSelCC) {
+      SDLoc DL(N);
+      return DAG.getNode(
+          N->getOpcode(), DL, N->getVTList(), Chain, Dest,
+          DAG.getConstant(getInvertedCondCode(*CSelCC), DL, MVT::i32),
+          CSel.getOperand(3));
+    }
+  }
+
   unsigned CmpOpc = Cmp.getOpcode();
   if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
     return SDValue();

llvm/test/CodeGen/AArch64/Atomics/aarch64-atomicrmw-lsfe.ll

@@ -451,11 +451,10 @@ define dso_local bfloat @atomicrmw_fsub_bfloat_aligned_monotonic(ptr %ptr, bfloa
 ; -O0-LABEL: atomicrmw_fsub_bfloat_aligned_monotonic:
 ; -O0:    add w8, w8, w9
 ; -O0:    add w8, w8, w9
-; -O0:    ldaxrh w9, [x11]
-; -O0:    cmp w9, w8, uxth
+; -O0:    ldaxrh w8, [x11]
+; -O0:    cmp w8, w9, uxth
 ; -O0:    stlxrh w10, w12, [x11]
-; -O0:    subs w8, w9, w8, uxth
-; -O0:    subs w8, w8, #1
+; -O0:    subs w9, w8, w9, uxth
 ;
 ; -O1-LABEL: atomicrmw_fsub_bfloat_aligned_monotonic:
 ; -O1:    ldxrh w9, [x0]
@@ -470,11 +469,10 @@ define dso_local bfloat @atomicrmw_fsub_bfloat_aligned_acquire(ptr %ptr, bfloat
 ; -O0-LABEL: atomicrmw_fsub_bfloat_aligned_acquire:
 ; -O0:    add w8, w8, w9
 ; -O0:    add w8, w8, w9
-; -O0:    ldaxrh w9, [x11]
-; -O0:    cmp w9, w8, uxth
+; -O0:    ldaxrh w8, [x11]
+; -O0:    cmp w8, w9, uxth
 ; -O0:    stlxrh w10, w12, [x11]
-; -O0:    subs w8, w9, w8, uxth
-; -O0:    subs w8, w8, #1
+; -O0:    subs w9, w8, w9, uxth
 ;
 ; -O1-LABEL: atomicrmw_fsub_bfloat_aligned_acquire:
 ; -O1:    ldaxrh w9, [x0]
@@ -489,11 +487,10 @@ define dso_local bfloat @atomicrmw_fsub_bfloat_aligned_release(ptr %ptr, bfloat
 ; -O0-LABEL: atomicrmw_fsub_bfloat_aligned_release:
 ; -O0:    add w8, w8, w9
 ; -O0:    add w8, w8, w9
-; -O0:    ldaxrh w9, [x11]
-; -O0:    cmp w9, w8, uxth
+; -O0:    ldaxrh w8, [x11]
+; -O0:    cmp w8, w9, uxth
 ; -O0:    stlxrh w10, w12, [x11]
-; -O0:    subs w8, w9, w8, uxth
-; -O0:    subs w8, w8, #1
+; -O0:    subs w9, w8, w9, uxth
 ;
 ; -O1-LABEL: atomicrmw_fsub_bfloat_aligned_release:
 ; -O1:    ldxrh w9, [x0]
@@ -508,11 +505,10 @@ define dso_local bfloat @atomicrmw_fsub_bfloat_aligned_acq_rel(ptr %ptr, bfloat
 ; -O0-LABEL: atomicrmw_fsub_bfloat_aligned_acq_rel:
 ; -O0:    add w8, w8, w9
 ; -O0:    add w8, w8, w9
-; -O0:    ldaxrh w9, [x11]
-; -O0:    cmp w9, w8, uxth
+; -O0:    ldaxrh w8, [x11]
+; -O0:    cmp w8, w9, uxth
 ; -O0:    stlxrh w10, w12, [x11]
-; -O0:    subs w8, w9, w8, uxth
-; -O0:    subs w8, w8, #1
+; -O0:    subs w9, w8, w9, uxth
 ;
 ; -O1-LABEL: atomicrmw_fsub_bfloat_aligned_acq_rel:
 ; -O1:    ldaxrh w9, [x0]
@@ -527,11 +523,10 @@ define dso_local bfloat @atomicrmw_fsub_bfloat_aligned_seq_cst(ptr %ptr, bfloat
 ; -O0-LABEL: atomicrmw_fsub_bfloat_aligned_seq_cst:
 ; -O0:    add w8, w8, w9
 ; -O0:    add w8, w8, w9
-; -O0:    ldaxrh w9, [x11]
-; -O0:    cmp w9, w8, uxth
+; -O0:    ldaxrh w8, [x11]
+; -O0:    cmp w8, w9, uxth
 ; -O0:    stlxrh w10, w12, [x11]
-; -O0:    subs w8, w9, w8, uxth
-; -O0:    subs w8, w8, #1
+; -O0:    subs w9, w8, w9, uxth
 ;
 ; -O1-LABEL: atomicrmw_fsub_bfloat_aligned_seq_cst:
 ; -O1:    ldaxrh w9, [x0]