[SelectOpt] Optimise big select groups in the latch of a non-inner loop to branches (llvm#119728)

Loop latches often have a loop-carried dependency, and if they have
several SelectLike instructions in one select group, it is usually
profitable to convert it to branches rather than keep selects.

Author: igogo-x86

llvm/lib/CodeGen/SelectOptimize.cpp

@@ -1044,6 +1044,18 @@ bool SelectOptimizeImpl::isConvertToBranchProfitableBase(
     return true;
   }
 
+  // If latch has a select group with several elements, it is usually profitable
+  // to convert it to branches. We let `optimizeSelectsInnerLoops` decide if
+  // conversion is profitable for innermost loops.
+  auto *BB = SI.getI()->getParent();
+  auto *L = LI->getLoopFor(BB);
+  if (L && !L->isInnermost() && L->getLoopLatch() == BB &&
+      ASI.Selects.size() >= 3) {
+    OR << "Converted to branch because select group in the latch block is big.";
+    EmitAndPrintRemark(ORE, OR);
+    return true;
+  }
+
   ORmiss << "Not profitable to convert to branch (base heuristic).";
   EmitAndPrintRemark(ORE, ORmiss);
   return false;

llvm/test/CodeGen/AArch64/selectopt.ll

@@ -875,3 +875,124 @@ if.end:
   %exitcond.not = icmp eq i64 %indvars.iv.next, %wide.trip.count
   br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
 }
+
+declare i64 @payload(i64, ptr, ptr, i64)
+
+define void @outer_latch_heuristic(ptr %dst, ptr %src, i64 %p, i64 %dim) {
+; CHECKOO-LABEL: @outer_latch_heuristic(
+; CHECKOO-NEXT:  entry:
+; CHECKOO-NEXT:    br label [[OUTER_LOOP:%.*]]
+; CHECKOO:       outer.loop:
+; CHECKOO-NEXT:    [[K_020_US:%.*]] = phi i64 [ [[INC7_US:%.*]], [[SELECT_END:%.*]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECKOO-NEXT:    [[J:%.*]] = phi i64 [ [[J_NEXT:%.*]], [[SELECT_END]] ], [ 0, [[ENTRY]] ]
+; CHECKOO-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], [[SELECT_END]] ], [ 0, [[ENTRY]] ]
+; CHECKOO-NEXT:    [[ARRAYIDX_US:%.*]] = getelementptr inbounds ptr, ptr [[SRC:%.*]], i64 [[I]]
+; CHECKOO-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARRAYIDX_US]], align 8
+; CHECKOO-NEXT:    [[ARRAYIDX1_US:%.*]] = getelementptr inbounds ptr, ptr [[SRC]], i64 [[J]]
+; CHECKOO-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[ARRAYIDX1_US]], align 8
+; CHECKOO-NEXT:    br label [[INNER_LOOP:%.*]]
+; CHECKOO:       inner.loop:
+; CHECKOO-NEXT:    [[LSR_IV:%.*]] = phi i64 [ [[DIM:%.*]], [[OUTER_LOOP]] ], [ [[LSR_IV_NEXT:%.*]], [[INNER_LOOP]] ]
+; CHECKOO-NEXT:    [[DIFF_04_I_US:%.*]] = phi i64 [ [[CALL_I_US:%.*]], [[INNER_LOOP]] ], [ 0, [[OUTER_LOOP]] ]
+; CHECKOO-NEXT:    [[CALL_I_US]] = tail call i64 @payload(i64 [[DIFF_04_I_US]], ptr [[TMP0]], ptr [[TMP1]], i64 [[P:%.*]])
+; CHECKOO-NEXT:    [[LSR_IV_NEXT]] = add i64 [[LSR_IV]], -1
+; CHECKOO-NEXT:    [[EXITCOND_NOT_I_US:%.*]] = icmp eq i64 [[LSR_IV_NEXT]], 0
+; CHECKOO-NEXT:    br i1 [[EXITCOND_NOT_I_US]], label [[LATCH:%.*]], label [[INNER_LOOP]]
+; CHECKOO:       latch:
+; CHECKOO-NEXT:    [[CMP2_US:%.*]] = icmp sgt i64 [[CALL_I_US]], -1
+; CHECKOO-NEXT:    [[DIFF_0_LCSSA_I_LOBIT_US:%.*]] = lshr i64 [[CALL_I_US]], 63
+; CHECKOO-NEXT:    [[CMP2_US_FROZEN:%.*]] = freeze i1 [[CMP2_US]]
+; CHECKOO-NEXT:    br i1 [[CMP2_US_FROZEN]], label [[SELECT_TRUE_SINK:%.*]], label [[SELECT_FALSE_SINK:%.*]]
+; CHECKOO:       select.true.sink:
+; CHECKOO-NEXT:    [[TMP2:%.*]] = add nsw i64 [[J]], 1
+; CHECKOO-NEXT:    br label [[SELECT_END]]
+; CHECKOO:       select.false.sink:
+; CHECKOO-NEXT:    [[TMP3:%.*]] = add nsw i64 1, [[I]]
+; CHECKOO-NEXT:    br label [[SELECT_END]]
+; CHECKOO:       select.end:
+; CHECKOO-NEXT:    [[I_NEXT]] = phi i64 [ [[I]], [[SELECT_TRUE_SINK]] ], [ [[TMP3]], [[SELECT_FALSE_SINK]] ]
+; CHECKOO-NEXT:    [[J_NEXT]] = phi i64 [ [[TMP2]], [[SELECT_TRUE_SINK]] ], [ [[J]], [[SELECT_FALSE_SINK]] ]
+; CHECKOO-NEXT:    [[COND_IN_US:%.*]] = phi ptr [ [[ARRAYIDX1_US]], [[SELECT_TRUE_SINK]] ], [ [[ARRAYIDX_US]], [[SELECT_FALSE_SINK]] ]
+; CHECKOO-NEXT:    [[INC4_US:%.*]] = zext i1 [[CMP2_US]] to i64
+; CHECKOO-NEXT:    [[COND_US:%.*]] = load ptr, ptr [[COND_IN_US]], align 8
+; CHECKOO-NEXT:    [[ARRAYIDX6_US:%.*]] = getelementptr inbounds ptr, ptr [[DST:%.*]], i64 [[K_020_US]]
+; CHECKOO-NEXT:    store ptr [[COND_US]], ptr [[ARRAYIDX6_US]], align 8
+; CHECKOO-NEXT:    [[INC7_US]] = add i64 [[K_020_US]], 1
+; CHECKOO-NEXT:    [[EXITCOND23_NOT:%.*]] = icmp eq i64 [[K_020_US]], 1000
+; CHECKOO-NEXT:    br i1 [[EXITCOND23_NOT]], label [[EXIT:%.*]], label [[OUTER_LOOP]]
+; CHECKOO:       exit:
+; CHECKOO-NEXT:    ret void
+;
+; CHECKII-LABEL: @outer_latch_heuristic(
+; CHECKII-NEXT:  entry:
+; CHECKII-NEXT:    br label [[OUTER_LOOP:%.*]]
+; CHECKII:       outer.loop:
+; CHECKII-NEXT:    [[K_020_US:%.*]] = phi i64 [ [[INC7_US:%.*]], [[LATCH:%.*]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECKII-NEXT:    [[J:%.*]] = phi i64 [ [[J_NEXT:%.*]], [[LATCH]] ], [ 0, [[ENTRY]] ]
+; CHECKII-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], [[LATCH]] ], [ 0, [[ENTRY]] ]
+; CHECKII-NEXT:    [[ARRAYIDX_US:%.*]] = getelementptr inbounds ptr, ptr [[SRC:%.*]], i64 [[I]]
+; CHECKII-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARRAYIDX_US]], align 8
+; CHECKII-NEXT:    [[ARRAYIDX1_US:%.*]] = getelementptr inbounds ptr, ptr [[SRC]], i64 [[J]]
+; CHECKII-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[ARRAYIDX1_US]], align 8
+; CHECKII-NEXT:    br label [[INNER_LOOP:%.*]]
+; CHECKII:       inner.loop:
+; CHECKII-NEXT:    [[LSR_IV:%.*]] = phi i64 [ [[DIM:%.*]], [[OUTER_LOOP]] ], [ [[LSR_IV_NEXT:%.*]], [[INNER_LOOP]] ]
+; CHECKII-NEXT:    [[DIFF_04_I_US:%.*]] = phi i64 [ [[CALL_I_US:%.*]], [[INNER_LOOP]] ], [ 0, [[OUTER_LOOP]] ]
+; CHECKII-NEXT:    [[CALL_I_US]] = tail call i64 @payload(i64 [[DIFF_04_I_US]], ptr [[TMP0]], ptr [[TMP1]], i64 [[P:%.*]])
+; CHECKII-NEXT:    [[LSR_IV_NEXT]] = add i64 [[LSR_IV]], -1
+; CHECKII-NEXT:    [[EXITCOND_NOT_I_US:%.*]] = icmp eq i64 [[LSR_IV_NEXT]], 0
+; CHECKII-NEXT:    br i1 [[EXITCOND_NOT_I_US]], label [[LATCH]], label [[INNER_LOOP]]
+; CHECKII:       latch:
+; CHECKII-NEXT:    [[CMP2_US:%.*]] = icmp sgt i64 [[CALL_I_US]], -1
+; CHECKII-NEXT:    [[DIFF_0_LCSSA_I_LOBIT_US:%.*]] = lshr i64 [[CALL_I_US]], 63
+; CHECKII-NEXT:    [[I_NEXT]] = add nsw i64 [[DIFF_0_LCSSA_I_LOBIT_US]], [[I]]
+; CHECKII-NEXT:    [[INC4_US:%.*]] = zext i1 [[CMP2_US]] to i64
+; CHECKII-NEXT:    [[J_NEXT]] = add nsw i64 [[J]], [[INC4_US]]
+; CHECKII-NEXT:    [[COND_IN_US:%.*]] = select i1 [[CMP2_US]], ptr [[ARRAYIDX1_US]], ptr [[ARRAYIDX_US]]
+; CHECKII-NEXT:    [[COND_US:%.*]] = load ptr, ptr [[COND_IN_US]], align 8
+; CHECKII-NEXT:    [[ARRAYIDX6_US:%.*]] = getelementptr inbounds ptr, ptr [[DST:%.*]], i64 [[K_020_US]]
+; CHECKII-NEXT:    store ptr [[COND_US]], ptr [[ARRAYIDX6_US]], align 8
+; CHECKII-NEXT:    [[INC7_US]] = add i64 [[K_020_US]], 1
+; CHECKII-NEXT:    [[EXITCOND23_NOT:%.*]] = icmp eq i64 [[K_020_US]], 1000
+; CHECKII-NEXT:    br i1 [[EXITCOND23_NOT]], label [[EXIT:%.*]], label [[OUTER_LOOP]]
+; CHECKII:       exit:
+; CHECKII-NEXT:    ret void
+;
+entry:
+  br label %outer.loop
+
+outer.loop:
+  %k.020.us = phi i64 [ %inc7.us, %latch ], [ 0, %entry ]
+  %j = phi i64 [ %j.next, %latch ], [ 0, %entry ]
+  %i = phi i64 [ %i.next, %latch ], [ 0, %entry ]
+  %arrayidx.us = getelementptr inbounds ptr, ptr %src, i64 %i
+  %4 = load ptr, ptr %arrayidx.us, align 8
+  %arrayidx1.us = getelementptr inbounds ptr, ptr %src, i64 %j
+  %5 = load ptr, ptr %arrayidx1.us, align 8
+  br label %inner.loop
+
+inner.loop:
+  %lsr.iv = phi i64 [ %dim, %outer.loop ], [ %lsr.iv.next, %inner.loop ]
+  %diff.04.i.us = phi i64 [ %call.i.us, %inner.loop ], [ 0, %outer.loop ]
+  %call.i.us = tail call i64 @payload(i64 %diff.04.i.us, ptr %4, ptr %5, i64 %p)
+  %lsr.iv.next = add i64 %lsr.iv, -1
+  %exitcond.not.i.us = icmp eq i64 %lsr.iv.next, 0
+  br i1 %exitcond.not.i.us, label %latch, label %inner.loop
+
+latch:
+  %cmp2.us = icmp sgt i64 %call.i.us, -1
+  %diff.0.lcssa.i.lobit.us = lshr i64 %call.i.us, 63
+  %i.next = add nsw i64 %diff.0.lcssa.i.lobit.us, %i
+  %inc4.us = zext i1 %cmp2.us to i64
+  %j.next = add nsw i64 %j, %inc4.us
+  %cond.in.us = select i1 %cmp2.us, ptr %arrayidx1.us, ptr %arrayidx.us
+  %cond.us = load ptr, ptr %cond.in.us, align 8
+  %arrayidx6.us = getelementptr inbounds ptr, ptr %dst, i64 %k.020.us
+  store ptr %cond.us, ptr %arrayidx6.us, align 8
+  %inc7.us = add i64 %k.020.us, 1
+  %exitcond23.not = icmp eq i64 %k.020.us, 1000
+  br i1 %exitcond23.not, label %exit, label %outer.loop
+
+exit:
+  ret void
+}