Be less aggressive about hinting in RAFast.

In particular, don't spill dirty registers only to satisfy a hint. It is
not worth it.

The attached test case provides an example where the fast allocator
would spill a register when other registers are available.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@132900 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/RegAllocFast.cpp b/lib/CodeGen/RegAllocFast.cpp
index 97652036f988ab99e770103af5d9c055bb4e261c..65ebdf827df3541038fe994ac5ab13f779d5f23c 100644 (file)
--- a/lib/CodeGen/RegAllocFast.cpp
+++ b/lib/CodeGen/RegAllocFast.cpp
@@ -487,14 +487,12 @@ void RAFast::allocVirtReg(MachineInstr *MI, LiveRegEntry &LRE, unsigned Hint) {
 
   // Take hint when possible.
   if (Hint) {
-    switch(calcSpillCost(Hint)) {
-    default:
-      definePhysReg(MI, Hint, regFree);
-      // Fall through.
-    case 0:
+    // Ignore the hint if we would have to spill a dirty register.
+    unsigned Cost = calcSpillCost(Hint);
+    if (Cost < spillDirty) {
+      if (Cost)
+        definePhysReg(MI, Hint, regFree);
       return assignVirtToPhysReg(LRE, Hint);
-    case spillImpossible:
-      break;
     }
   }
 

diff --git a/test/CodeGen/ARM/fast-isel-static.ll b/test/CodeGen/ARM/fast-isel-static.ll
index 2d79674028caa2220e1b262673153839dcbb0364..a86e3251f715b32aca44544fce8ce871385a190b 100644 (file)
--- a/test/CodeGen/ARM/fast-isel-static.ll
+++ b/test/CodeGen/ARM/fast-isel-static.ll
@@ -23,7 +23,7 @@ entry:
   %z = alloca float, align 4
   store float 0.000000e+00, float* %ztot, align 4
   store float 1.000000e+00, float* %z, align 4
-; CHECK-LONG: blx     r2
+; CHECK-LONG: blx     r
 ; CHECK-NORM: bl      _myadd
   call void @myadd(float* %ztot, float* %z)
   ret i32 0

diff --git a/test/CodeGen/X86/2011-06-12-FastAllocSpill.ll b/test/CodeGen/X86/2011-06-12-FastAllocSpill.ll
new file mode 100644 (file)
index 0000000..a51dad0
--- /dev/null
+++ b/test/CodeGen/X86/2011-06-12-FastAllocSpill.ll
@@ -0,0 +1,52 @@
+; RUN: llc < %s -O0 -disable-fp-elim -relocation-model=pic -stats |& FileCheck %s
+;
+; This test should not cause any spilling with RAFast.
+;
+; CHECK: Number of copies coalesced
+; CHECK-NOT: Number of stores added
+;
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-apple-darwin10.0.0"
+
+%0 = type { i64, i64, i8*, i8* }
+%1 = type opaque
+%2 = type opaque
+%3 = type <{ i8*, i32, i32, void (%4*)*, i8*, i64 }>
+%4 = type { i8**, i32, i32, i8**, %5*, i64 }
+%5 = type { i64, i64 }
+%6 = type { i8*, i32, i32, i8*, %5* }
+
+@0 = external hidden constant %0
+
+define hidden void @f() ssp {
+bb:
+  %tmp5 = alloca i64, align 8
+  %tmp6 = alloca void ()*, align 8
+  %tmp7 = alloca %3, align 8
+  store i64 0, i64* %tmp5, align 8
+  br label %bb8
+
+bb8:                                              ; preds = %bb23, %bb
+  %tmp15 = getelementptr inbounds %3* %tmp7, i32 0, i32 4
+  store i8* bitcast (%0* @0 to i8*), i8** %tmp15
+  %tmp16 = bitcast %3* %tmp7 to void ()*
+  store void ()* %tmp16, void ()** %tmp6, align 8
+  %tmp17 = load void ()** %tmp6, align 8
+  %tmp18 = bitcast void ()* %tmp17 to %6*
+  %tmp19 = getelementptr inbounds %6* %tmp18, i32 0, i32 3
+  %tmp20 = bitcast %6* %tmp18 to i8*
+  %tmp21 = load i8** %tmp19
+  %tmp22 = bitcast i8* %tmp21 to void (i8*)*
+  call void %tmp22(i8* %tmp20)
+  br label %bb23
+
+bb23:                                             ; preds = %bb8
+  %tmp24 = load i64* %tmp5, align 8
+  %tmp25 = add i64 %tmp24, 1
+  store i64 %tmp25, i64* %tmp5, align 8
+  %tmp26 = icmp ult i64 %tmp25, 10
+  br i1 %tmp26, label %bb8, label %bb27
+
+bb27:                                             ; preds = %bb23
+  ret void
+}

diff --git a/test/CodeGen/X86/fast-isel-gep.ll b/test/CodeGen/X86/fast-isel-gep.ll
index 48abfd0f26e57312db849644c08c137c875f4047..1a2e34ec7f5c4b137752058c31dcf85f08d5dbfd 100644 (file)
--- a/test/CodeGen/X86/fast-isel-gep.ll
+++ b/test/CodeGen/X86/fast-isel-gep.ll
@@ -24,7 +24,7 @@ define i32 @test2(i64 %t3, i32* %t1) nounwind {
        %t15 = load i32* %t9            ; <i32> [#uses=1]
        ret i32 %t15
 ; X32: test2:
-; X32:         movl    (%edx,%ecx,4), %eax
+; X32:         movl    (%edx,%ecx,4), %e
 ; X32:         ret
 
 ; X64: test2: