[llvm-commits] [llvm] r103946 - in /llvm/trunk: lib/CodeGen/RegAllocFast.cpp test/CodeGen/ARM/2010-05-17-FastAllocCrash.ll

Jakob Stoklund Olesen stoklund at 2pi.dk
Mon May 17 10:18:59 PDT 2010


Author: stoklund
Date: Mon May 17 12:18:59 2010
New Revision: 103946

URL: http://llvm.org/viewvc/llvm-project?rev=103946&view=rev
Log:
Avoid allocating the same physreg to multiple virtregs in one instruction.

While that approach works wonders for register pressure, it tends to break
everything.

This should unbreak the arm-linux builder and fix a number of miscompilations.

Added:
    llvm/trunk/test/CodeGen/ARM/2010-05-17-FastAllocCrash.ll
Modified:
    llvm/trunk/lib/CodeGen/RegAllocFast.cpp

Modified: llvm/trunk/lib/CodeGen/RegAllocFast.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/RegAllocFast.cpp?rev=103946&r1=103945&r2=103946&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/RegAllocFast.cpp (original)
+++ llvm/trunk/lib/CodeGen/RegAllocFast.cpp Mon May 17 12:18:59 2010
@@ -471,6 +471,7 @@
 
   unsigned BestReg = 0, BestCost = spillImpossible;
   for (TargetRegisterClass::iterator I = AOB; I != AOE; ++I) {
+    if (UsedInInstr.test(*I)) continue;
     unsigned Cost = calcSpillCost(*I);
     // Cost is 0 when all aliases are already disabled.
     if (Cost == 0)

Added: llvm/trunk/test/CodeGen/ARM/2010-05-17-FastAllocCrash.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/2010-05-17-FastAllocCrash.ll?rev=103946&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/ARM/2010-05-17-FastAllocCrash.ll (added)
+++ llvm/trunk/test/CodeGen/ARM/2010-05-17-FastAllocCrash.ll Mon May 17 12:18:59 2010
@@ -0,0 +1,105 @@
+; RUN: llc < %s -regalloc=fast -verify-machineinstrs
+target triple = "arm-pc-linux-gnu"
+
+; This test case would accidentally use the same physreg for two virtregs
+; because allocVirtReg forgot to check if registers were already used in the
+; instruction.
+; This caused the RegScavenger to complain, but -verify-machineinstrs also
+; catches it.
+
+%struct.CHESS_POSITION = type { i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i32, i8, i8, [64 x i8], i8, i8, i8, i8, i8 }
+
+ at search = external global %struct.CHESS_POSITION  ; <%struct.CHESS_POSITION*> [#uses=1]
+ at bishop_mobility_rr45 = external global [64 x [256 x i32]] ; <[64 x [256 x i32]]*> [#uses=1]
+
+declare fastcc i32 @FirstOne()
+
+define fastcc void @Evaluate() {
+entry:
+  br i1 false, label %cond_false186, label %cond_true
+
+cond_true:                                        ; preds = %entry
+  ret void
+
+cond_false186:                                    ; preds = %entry
+  br i1 false, label %cond_true293, label %bb203
+
+bb203:                                            ; preds = %cond_false186
+  ret void
+
+cond_true293:                                     ; preds = %cond_false186
+  br i1 false, label %cond_true298, label %cond_next317
+
+cond_true298:                                     ; preds = %cond_true293
+  br i1 false, label %cond_next518, label %cond_true397.preheader
+
+cond_next317:                                     ; preds = %cond_true293
+  ret void
+
+cond_true397.preheader:                           ; preds = %cond_true298
+  ret void
+
+cond_next518:                                     ; preds = %cond_true298
+  br i1 false, label %bb1069, label %cond_true522
+
+cond_true522:                                     ; preds = %cond_next518
+  ret void
+
+bb1069:                                           ; preds = %cond_next518
+  br i1 false, label %cond_next1131, label %bb1096
+
+bb1096:                                           ; preds = %bb1069
+  ret void
+
+cond_next1131:                                    ; preds = %bb1069
+  br i1 false, label %cond_next1207, label %cond_true1150
+
+cond_true1150:                                    ; preds = %cond_next1131
+  ret void
+
+cond_next1207:                                    ; preds = %cond_next1131
+  br i1 false, label %cond_next1219, label %cond_true1211
+
+cond_true1211:                                    ; preds = %cond_next1207
+  ret void
+
+cond_next1219:                                    ; preds = %cond_next1207
+  br i1 false, label %cond_true1223, label %cond_next1283
+
+cond_true1223:                                    ; preds = %cond_next1219
+  br i1 false, label %cond_true1254, label %cond_true1264
+
+cond_true1254:                                    ; preds = %cond_true1223
+  br i1 false, label %bb1567, label %cond_true1369.preheader
+
+cond_true1264:                                    ; preds = %cond_true1223
+  ret void
+
+cond_next1283:                                    ; preds = %cond_next1219
+  ret void
+
+cond_true1369.preheader:                          ; preds = %cond_true1254
+  ret void
+
+bb1567:                                           ; preds = %cond_true1254
+  %tmp1591 = load i64* getelementptr inbounds (%struct.CHESS_POSITION* @search, i32 0, i32 4) ; <i64> [#uses=1]
+  %tmp1572 = tail call fastcc i32 @FirstOne()     ; <i32> [#uses=1]
+  %tmp1594 = load i32* undef                      ; <i32> [#uses=1]
+  %tmp1594.upgrd.5 = trunc i32 %tmp1594 to i8     ; <i8> [#uses=1]
+  %shift.upgrd.6 = zext i8 %tmp1594.upgrd.5 to i64 ; <i64> [#uses=1]
+  %tmp1595 = lshr i64 %tmp1591, %shift.upgrd.6    ; <i64> [#uses=1]
+  %tmp1595.upgrd.7 = trunc i64 %tmp1595 to i32    ; <i32> [#uses=1]
+  %tmp1596 = and i32 %tmp1595.upgrd.7, 255        ; <i32> [#uses=1]
+  %gep.upgrd.8 = zext i32 %tmp1596 to i64         ; <i64> [#uses=1]
+  %tmp1598 = getelementptr [64 x [256 x i32]]* @bishop_mobility_rr45, i32 0, i32 %tmp1572, i64 %gep.upgrd.8 ; <i32*> [#uses=1]
+  %tmp1599 = load i32* %tmp1598                   ; <i32> [#uses=1]
+  %tmp1602 = sub i32 0, %tmp1599                  ; <i32> [#uses=1]
+  br i1 undef, label %cond_next1637, label %cond_true1607
+
+cond_true1607:                                    ; preds = %bb1567
+  ret void
+
+cond_next1637:                                    ; preds = %bb1567
+  %tmp1662 = sub i32 %tmp1602, 0                  ; <i32> [#uses=0]
+  ret void
+}





More information about the llvm-commits mailing list