Pick a conservative register class when creating a small live range for remat.

The rematerialized instruction may require a more constrained register class
than the register being spilled. In the test case, the spilled register has been
inflated to the DPR register class, but we are rematerializing a load of the
ssub_0 sub-register which only exists for DPR_VFP2 registers.

The register class is reinflated after spilling, so the conservative choice is
only temporary.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@128610 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/InlineSpiller.cpp b/lib/CodeGen/InlineSpiller.cpp
index d0c750777f3e12e8607c86ad888995d5f89bcaf1..86f4cfcc92f3b63a787478423c7c383aee877ac0 100644 (file)
--- a/lib/CodeGen/InlineSpiller.cpp
+++ b/lib/CodeGen/InlineSpiller.cpp
@@ -627,7 +627,7 @@ bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg,
   }
 
   // Alocate a new register for the remat.
-  LiveInterval &NewLI = Edit->createFrom(VirtReg.reg, LIS, VRM);
+  LiveInterval &NewLI = Edit->createFrom(Original, LIS, VRM);
   NewLI.markNotSpillable();
 
   // Finally we can rematerialize OrigMI before MI.

diff --git a/test/CodeGen/ARM/crash-greedy.ll b/test/CodeGen/ARM/crash-greedy.ll
new file mode 100644 (file)
index 0000000..3a94110
--- /dev/null
+++ b/test/CodeGen/ARM/crash-greedy.ll
@@ -0,0 +1,61 @@
+; RUN: llc < %s -regalloc=greedy -mcpu=cortex-a8 -relocation-model=pic -disable-fp-elim | FileCheck %s
+;
+; ARM tests that crash or fail with the greedy register allocator.
+
+target triple = "thumbv7-apple-darwin"
+
+declare double @exp(double)
+
+; CHECK remat_subreg
+define void @remat_subreg(float* nocapture %x, i32* %y, i32 %n, i32 %z, float %c, float %lambda, float* nocapture %ret_f, float* nocapture %ret_df) nounwind {
+entry:
+  %conv16 = fpext float %lambda to double
+  %mul17 = fmul double %conv16, -1.000000e+00
+  br i1 undef, label %cond.end.us, label %cond.end
+
+cond.end.us:                                      ; preds = %entry
+  unreachable
+
+cond.end:                                         ; preds = %cond.end, %entry
+  %mul = fmul double undef, 0.000000e+00
+  %add = fadd double undef, %mul
+  %add46 = fadd double undef, undef
+  %add75 = fadd double 0.000000e+00, undef
+  br i1 undef, label %for.end, label %cond.end
+
+for.end:                                          ; preds = %cond.end
+  %conv78 = sitofp i32 %z to double
+  %conv83 = fpext float %c to double
+  %mul84 = fmul double %mul17, %conv83
+  %call85 = tail call double @exp(double %mul84) nounwind
+  %mul86 = fmul double %conv78, %call85
+  %add88 = fadd double 0.000000e+00, %mul86
+; CHECK: blx _exp
+  %call100 = tail call double @exp(double %mul84) nounwind
+  %mul101 = fmul double undef, %call100
+  %add103 = fadd double %add46, %mul101
+  %mul111 = fmul double undef, %conv83
+  %mul119 = fmul double %mul111, undef
+  %add121 = fadd double undef, %mul119
+  %div = fdiv double 1.000000e+00, %conv16
+  %div126 = fdiv double %add, %add75
+  %sub = fsub double %div, %div126
+  %div129 = fdiv double %add103, %add88
+  %add130 = fadd double %sub, %div129
+  %conv131 = fptrunc double %add130 to float
+  store float %conv131, float* %ret_f, align 4
+  %mul139 = fmul double %div129, %div129
+  %div142 = fdiv double %add121, %add88
+  %sub143 = fsub double %mul139, %div142
+; %lambda is passed on the stack, and the stack slot load is rematerialized.
+; The rematted load of a float constrains the D register used for the mul.
+; CHECK: vldr
+  %mul146 = fmul float %lambda, %lambda
+  %conv147 = fpext float %mul146 to double
+  %div148 = fdiv double 1.000000e+00, %conv147
+  %sub149 = fsub double %sub143, %div148
+  %conv150 = fptrunc double %sub149 to float
+  store float %conv150, float* %ret_df, align 4
+  ret void
+}
+