Update live-in lists when splitting critical edges.

Fixes PR10814. Patch by Jan Sjödin!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@141960 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/MachineBasicBlock.cpp b/lib/CodeGen/MachineBasicBlock.cpp
index 8f0fb46879ac165de8b766fe2175044bc0292d51..4c5fe4c480a68b90b136d8ff10ad2ac907ebef83 100644 (file)
--- a/lib/CodeGen/MachineBasicBlock.cpp
+++ b/lib/CodeGen/MachineBasicBlock.cpp
@@ -571,6 +571,11 @@ MachineBasicBlock::SplitCriticalEdge(MachineBasicBlock *Succ, Pass *P) {
       if (i->getOperand(ni+1).getMBB() == this)
         i->getOperand(ni+1).setMBB(NMBB);
 
+  // Inherit live-ins from the successor
+  for (MachineBasicBlock::livein_iterator I = Succ->livein_begin(),
+        E = Succ->livein_end(); I != E; ++I)
+    NMBB->addLiveIn(*I);
+
   // Update LiveVariables.
   if (LV) {
     // Restore kills of virtual registers that were killed by the terminators.

diff --git a/test/CodeGen/X86/MachineSink-eflags.ll b/test/CodeGen/X86/MachineSink-eflags.ll
new file mode 100644 (file)
index 0000000..5b8c7b2
--- /dev/null
+++ b/test/CodeGen/X86/MachineSink-eflags.ll
@@ -0,0 +1,74 @@
+; RUN: llc < %s | FileCheck %s
+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-pc-linux"
+
+
+%0 = type <{ i64, i64, %1, %1, [21 x %2] }>
+%1 = type <{ i64, i64, i64 }>
+%2 = type <{ i32, i32, i8 addrspace(2)* }>
+%3 = type { i8*, i8*, i8*, i8*, i32 }
+%4 = type <{ %5*, i8*, i32, i32, [4 x i64], [4 x i64], [4 x i64], [4 x i64], [4 x i64] }>
+%5 = type <{ void (i32)*, i8*, i32 (i8*, ...)* }>
+
+define void @foo(i8* nocapture %_stubArgs) nounwind {
+entry:
+ %i0 = alloca i8*, align 8
+ %i2 = alloca i8*, align 8
+ %b.i = alloca [16 x <2 x double>], align 16
+ %conv = bitcast i8* %_stubArgs to i32*
+ %tmp1 = load i32* %conv, align 4
+ %ptr8 = getelementptr i8* %_stubArgs, i64 16
+ %i4 = bitcast i8* %ptr8 to <2 x double>*
+ %ptr20 = getelementptr i8* %_stubArgs, i64 48
+ %i7 = bitcast i8* %ptr20 to <2 x double> addrspace(1)**
+ %tmp21 = load <2 x double> addrspace(1)** %i7, align 8
+ %ptr28 = getelementptr i8* %_stubArgs, i64 64
+ %i9 = bitcast i8* %ptr28 to i32*
+ %tmp29 = load i32* %i9, align 4
+ %ptr32 = getelementptr i8* %_stubArgs, i64 68
+ %i10 = bitcast i8* %ptr32 to i32*
+ %tmp33 = load i32* %i10, align 4
+ %tmp17.i = mul i32 10, 20
+ %tmp19.i = add i32 %tmp17.i, %tmp33
+ %conv21.i = zext i32 %tmp19.i to i64
+ %tmp6.i = and i32 42, -32
+ %tmp42.i = add i32 %tmp6.i, 17
+ %tmp44.i = insertelement <2 x i32> undef, i32 %tmp42.i, i32 1
+ %tmp96676677.i = or i32 17, -4
+ %ptr4438.i = getelementptr inbounds [16 x <2 x double>]* %b.i, i64 0, i64 0
+ %arrayidx4506.i = getelementptr [16 x <2 x double>]* %b.i, i64 0, i64 4
+ %tmp52.i = insertelement <2 x i32> %tmp44.i, i32 0, i32 0
+ %tmp78.i = extractelement <2 x i32> %tmp44.i, i32 1
+ %tmp97.i = add i32 %tmp78.i, %tmp96676677.i
+ %tmp99.i = insertelement <2 x i32> %tmp52.i, i32 %tmp97.i, i32 1
+ %tmp154.i = extractelement <2 x i32> %tmp99.i, i32 1
+ %tmp156.i = extractelement <2 x i32> %tmp52.i, i32 0
+ %tmp158.i = urem i32 %tmp156.i, %tmp1
+ %i38 = mul i32 %tmp154.i, %tmp29
+ %i39 = add i32 %tmp158.i, %i38
+ %conv160.i = zext i32 %i39 to i64
+ %tmp22.sum652.i = add i64 %conv160.i, %conv21.i
+ %arrayidx161.i = getelementptr <2 x double> addrspace(1)* %tmp21, i64 %tmp22.sum652.i
+ %tmp162.i = load <2 x double> addrspace(1)* %arrayidx161.i, align 16
+ %tmp222.i = add i32 %tmp154.i, 1
+ %i43 = mul i32 %tmp222.i, %tmp29
+ %i44 = add i32 %tmp158.i, %i43
+ %conv228.i = zext i32 %i44 to i64
+ %tmp22.sum656.i = add i64 %conv228.i, %conv21.i
+ %arrayidx229.i = getelementptr <2 x double> addrspace(1)* %tmp21, i64 %tmp22.sum656.i
+ %tmp230.i = load <2 x double> addrspace(1)* %arrayidx229.i, align 16
+ %cmp432.i = icmp ult i32 %tmp156.i, %tmp1
+
+; %shl.i should not be sinked below the compare.
+; CHECK: cmpl
+; CHECK-NOT: shlq
+
+ %cond.i = select i1 %cmp432.i, <2 x double> %tmp162.i, <2 x double> zeroinitializer
+ store <2 x double> %cond.i, <2 x double>* %ptr4438.i, align 16
+ %cond448.i = select i1 %cmp432.i, <2 x double> %tmp230.i, <2 x double> zeroinitializer
+ store <2 x double> %cond448.i, <2 x double>* %arrayidx4506.i, align 16
+  ret void
+}
+
+
+

diff --git a/test/CodeGen/X86/uint64-to-float.ll b/test/CodeGen/X86/uint64-to-float.ll
index d9f753c7a88e7d97995027516657d13267f651cf..1dbbdcf892797de823308647b49e98c5d1b9968e 100644 (file)
--- a/test/CodeGen/X86/uint64-to-float.ll
+++ b/test/CodeGen/X86/uint64-to-float.ll
@@ -6,12 +6,37 @@
 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"
 
-; CHECK: testq %rdi, %rdi
-; CHECK-NEXT: jns LBB0_2
+; FIXME: This test could generate this code:
+;
+; ## BB#0:                                ## %entry
+;      testq   %rdi, %rdi
+;      jns     LBB0_2
+; ## BB#1:
+;      movq    %rdi, %rax
+;      shrq    %rax
+;      andq    $1, %rdi
+;      orq     %rax, %rdi
+;      cvtsi2ssq       %rdi, %xmm0
+;      addss   %xmm0, %xmm0
+;      ret
+; LBB0_2:                                 ## %entry
+;      cvtsi2ssq       %rdi, %xmm0
+;      ret
+;
+; The blocks come from lowering:
+;
+;   %vreg7<def> = CMOV_FR32 %vreg6<kill>, %vreg5<kill>, 15, %EFLAGS<imp-use>; FR32:%vreg7,%vreg6,%vreg5
+;
+; If the instruction had an EFLAGS<kill> flag, it wouldn't need to mark EFLAGS
+; as live-in on the new blocks, and machine sinking would be able to sink
+; everything below the test.
+
 ; CHECK: shrq
-; CHECK-NEXT: andq
+; CHECK: andq
 ; CHECK-NEXT: orq
-; CHECK-NEXT: cvtsi2ss
+; CHECK: testq %rdi, %rdi
+; CHECK-NEXT: jns LBB0_2
+; CHECK: cvtsi2ss
 ; CHECK: LBB0_2
 ; CHECK-NEXT: cvtsi2ss
 define float @test(i64 %a) {