xref: /freebsd-10.2-release/contrib/llvm/patches/patch-r262261-llvm-r199187-sparc.diff

Pull in r199187 from upstream llvm trunk (by Jakob Stoklund Olesen):

  Always let value types influence register classes.

  When creating a virtual register for a def, the value type should be
  used to pick the register class. If we only use the register class
  constraint on the instruction, we might pick a too large register class.

  Some registers can store values of different sizes. For example, the x86
  xmm registers can hold f32, f64, and 128-bit vectors. The three
  different value sizes are represented by register classes with identical
  register sets: FR32, FR64, and VR128. These register classes have
  different spill slot sizes, so it is important to use the right one.

  The register class constraint on an instruction doesn't necessarily care
  about the size of the value its defining. The value type determines
  that.

  This fixes a problem where InstrEmitter was picking 32-bit register
  classes for 64-bit values on SPARC.

Introduced here: http://svnweb.freebsd.org/changeset/base/262261

Index: test/CodeGen/SPARC/spillsize.ll
===================================================================
--- test/CodeGen/SPARC/spillsize.ll
+++ test/CodeGen/SPARC/spillsize.ll
@@ -0,0 +1,25 @@
+; RUN: llc < %s -verify-machineinstrs | FileCheck %s
+target datalayout = "E-m:e-i64:64-n32:64-S128"
+target triple = "sparcv9"
+
+; CHECK-LABEL: spill4
+; This function spills two values: %p and the materialized large constant.
+; Both must use 8-byte spill and fill instructions.
+; CHECK: stx %{{..}}, [%fp+
+; CHECK: stx %{{..}}, [%fp+
+; CHECK: ldx [%fp+
+; CHECK: ldx [%fp+
+define void @spill4(i64* nocapture %p) {
+entry:
+  %val0 = load i64* %p
+  %cmp0 = icmp ult i64 %val0, 385672958347594845
+  %cm80 = zext i1 %cmp0 to i64
+  store i64 %cm80, i64* %p, align 8
+  tail call void asm sideeffect "", "~{i0},~{i1},~{i2},~{i3},~{i4},~{i5},~{g2},~{g3},~{g4},~{g5},~{l0},~{l1},~{l2},~{l3},~{l4},~{l5},~{l6},~{l7},~{o0},~{o1},~{o2},~{o3},~{o4},~{o5},~{o7}"()
+  %arrayidx1 = getelementptr inbounds i64* %p, i64 1
+  %val = load i64* %arrayidx1
+  %cmp = icmp ult i64 %val, 385672958347594845
+  %cm8 = select i1 %cmp, i64 10, i64 20
+  store i64 %cm8, i64* %arrayidx1, align 8
+  ret void
+}
Index: lib/CodeGen/SelectionDAG/InstrEmitter.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/InstrEmitter.cpp
+++ lib/CodeGen/SelectionDAG/InstrEmitter.cpp
@@ -220,10 +220,19 @@ void InstrEmitter::CreateVirtualRegisters(SDNode *
     unsigned VRBase = 0;
     const TargetRegisterClass *RC =
       TRI->getAllocatableClass(TII->getRegClass(II, i, TRI, *MF));
-    // If the register class is unknown for the given definition, then try to
-    // infer one from the value type.
-    if (!RC && i < NumResults)
-      RC = TLI->getRegClassFor(Node->getSimpleValueType(i));
+    // Always let the value type influence the used register class. The
+    // constraints on the instruction may be too lax to represent the value
+    // type correctly. For example, a 64-bit float (X86::FR64) can't live in
+    // the 32-bit float super-class (X86::FR32).
+    if (i < NumResults && TLI->isTypeLegal(Node->getSimpleValueType(i))) {
+      const TargetRegisterClass *VTRC =
+        TLI->getRegClassFor(Node->getSimpleValueType(i));
+      if (RC)
+        VTRC = TRI->getCommonSubClass(RC, VTRC);
+      if (VTRC)
+        RC = VTRC;
+    }
+
     if (II.OpInfo[i].isOptionalDef()) {
       // Optional def must be a physical register.
       unsigned NumResults = CountResults(Node);