mi-sched: improve the generic register pressure comparison.

Only compare pressure within the same set. When multiple sets are
affected, we prioritize the most constrained set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189641 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/CodeGen/RegisterPressure.h

@@ -112,14 +112,13 @@ public:
     assert(isValid() && "invalid PressureChange");
     return PSetID - 1;
   }
+  // If PSetID is invalid, return UINT16_MAX to give it lowest priority.
+  unsigned getPSetOrMax() const { return (PSetID - 1) & UINT16_MAX; }
 
   int getUnitInc() const { return UnitInc; }
 
   void setUnitInc(int Inc) { UnitInc = Inc; }
 
-  // If PSetID is invalid, convert to INT_MAX to give it lowest priority.
-  int getRank() const { return (PSetID - 1) & INT_MAX; }
-
   bool operator==(const PressureChange &RHS) const {
     return PSetID == RHS.PSetID && UnitInc == RHS.UnitInc;
   }

lib/CodeGen/MachineScheduler.cpp

@@ -2182,21 +2182,19 @@ static bool tryPressure(const PressureChange &TryP,
                         ConvergingScheduler::SchedCandidate &TryCand,
                         ConvergingScheduler::SchedCandidate &Cand,
                         ConvergingScheduler::CandReason Reason) {
-  if (TryP.isValid() && CandP.isValid()) {
+  int TryRank = TryP.getPSetOrMax();
-    // If both candidates affect the same set, go with the smallest increase.
+  int CandRank = CandP.getPSetOrMax();
-    if (TryP.getPSet() == CandP.getPSet()) {
+  // If both candidates affect the same set, go with the smallest increase.
-      return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
+  if (TryRank == CandRank) {
-                     Reason);
+    return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
-    }
+                   Reason);
-    // If one candidate decreases and the other increases, go with it.
+  }
-    if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
+  // If one candidate decreases and the other increases, go with it.
-                Reason)) {
+  // Invalid candidates have UnitInc==0.
-      return true;
+  if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
-    }
+              Reason)) {
+    return true;
   }
-  // If TryP has lower Rank, it has a higher priority.
-  int TryRank = TryP.getRank();
-  int CandRank = CandP.getRank();
   // If the candidates are decreasing pressure, reverse priority.
   if (TryP.getUnitInc() < 0)
     std::swap(TryRank, CandRank);

test/CodeGen/X86/misched-balance.ll

@@ -1,5 +1,4 @@
-; RUN-disabled: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched -verify-machineinstrs | FileCheck %s
-; RUN: true
 ;
 ; Verify that misched resource/latency balancy heuristics are sane.
 
@@ -16,7 +15,7 @@ entry:
 ; Since mmult1 IR is already in good order, this effectively ensure
 ; the scheduler maintains source order.
 ;
-; CHECK: %for.body
+; CHECK-LABEL: %for.body
 ; CHECK-NOT: %rsp
 ; CHECK: imull 4
 ; CHECK-NOT: {{imull|rsp}}
@@ -46,7 +45,7 @@ entry:
 ; CHECK-NOT: {{imull|rsp}}
 ; CHECK: addl
 ; CHECK-NOT: {{imull|rsp}}
-; CHECK: %end
+; CHECK-LABEL: %end
 for.body:
   %indvars.iv42.i = phi i64 [ %indvars.iv.next43.i, %for.body ], [ 0, %entry ]
   %tmp57 = load i32* %tmp56, align 4
@@ -121,7 +120,7 @@ end:
 ; Unlike the above loop, this IR starts out bad and must be
 ; rescheduled.
 ;
-; CHECK: %for.body
+; CHECK-LABEL: %for.body
 ; CHECK-NOT: %rsp
 ; CHECK: imull 4
 ; CHECK-NOT: {{imull|rsp}}
@@ -151,7 +150,7 @@ end:
 ; CHECK-NOT: {{imull|rsp}}
 ; CHECK: addl
 ; CHECK-NOT: {{imull|rsp}}
-; CHECK: %end
+; CHECK-LABEL: %end
 define void @unrolled_mmult2(i32* %tmp55, i32* %tmp56, i32* %pre, i32* %pre94,
   i32* %pre95, i32* %pre96, i32* %pre97, i32* %pre98, i32* %pre99,
   i32* %pre100, i32* %pre101, i32* %pre102, i32* %pre103, i32* %pre104)
@@ -233,8 +232,8 @@ end:
 ; balanced heuristics are interesting here because we have resource,
 ; latency, and register limits all at once. For now, simply check that
 ; we don't use any callee-saves.
-; CHECK: @encpc1
+; CHECK-LABEL: @encpc1
-; CHECK: %entry
+; CHECK-LABEL: %entry
 ; CHECK-NOT: push
 ; CHECK-NOT: pop
 ; CHECK: ret

test/CodeGen/X86/misched-matmul.ll

@@ -1,6 +1,5 @@
 ; REQUIRES: asserts
-; RUN-disabled: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched -stats 2>&1 | FileCheck %s
+; RUN: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched -stats 2>&1 | FileCheck %s
-; RUN: true
 ;
 ; Verify that register pressure heuristics are working in MachineScheduler.
 ;

test/CodeGen/X86/misched-matrix.ll

@@ -1,6 +1,6 @@
 ; RUN: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched \
 ; RUN:          -misched-topdown -verify-machineinstrs \
-; RUN:     | FileCheck %s -check-prefix=TOPDOWN-disabled
+; RUN:     | FileCheck %s -check-prefix=TOPDOWN
 ; RUN: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched \
 ; RUN:          -misched=ilpmin -verify-machineinstrs \
 ; RUN:     | FileCheck %s -check-prefix=ILPMIN
@@ -15,19 +15,19 @@
 ; been reordered with the stores. This tests the scheduler's cheap
 ; alias analysis ability (that doesn't require any AliasAnalysis pass).
 ;
-; TOPDOWN-disabled: %for.body
+; TOPDOWN-LABEL: %for.body
 ; TOPDOWN: movl %{{.*}}, (
 ; TOPDOWN: imull {{[0-9]*}}(
 ; TOPDOWN: movl %{{.*}}, 4(
 ; TOPDOWN: imull {{[0-9]*}}(
 ; TOPDOWN: movl %{{.*}}, 8(
 ; TOPDOWN: movl %{{.*}}, 12(
-; TOPDOWN: %for.end
+; TOPDOWN-LABEL: %for.end
 ;
 ; For -misched=ilpmin, verify that each expression subtree is
 ; scheduled independently, and that the imull/adds are interleaved.
 ;
-; ILPMIN: %for.body
+; ILPMIN-LABEL: %for.body
 ; ILPMIN: movl %{{.*}}, (
 ; ILPMIN: imull
 ; ILPMIN: imull
@@ -53,12 +53,12 @@
 ; ILPMIN: imull
 ; ILPMIN: addl
 ; ILPMIN: movl %{{.*}}, 12(
-; ILPMIN: %for.end
+; ILPMIN-LABEL: %for.end
 ;
 ; For -misched=ilpmax, verify that each expression subtree is
 ; scheduled independently, and that the imull/adds are clustered.
 ;
-; ILPMAX: %for.body
+; ILPMAX-LABEL: %for.body
 ; ILPMAX: movl %{{.*}}, (
 ; ILPMAX: imull
 ; ILPMAX: imull
@@ -84,7 +84,7 @@
 ; ILPMAX: addl
 ; ILPMAX: addl
 ; ILPMAX: movl %{{.*}}, 12(
-; ILPMAX: %for.end
+; ILPMAX-LABEL: %for.end
 
 define void @mmult([4 x i32]* noalias nocapture %m1, [4 x i32]* noalias nocapture %m2,
 [4 x i32]* noalias nocapture %m3) nounwind uwtable ssp {