add support for pentium class CPUs which do not have cmov,

PR4841.  Patch by Craig Smith!


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@98496 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 64 insertions, 16 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 528bfaf56b..7af4f4c39e 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -8478,6 +8478,11 @@ X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   case X86::CMOV_V4F32:
   case X86::CMOV_V2F64:
   case X86::CMOV_V2I64:
+  case X86::CMOV_GR16:
+  case X86::CMOV_GR32:
+  case X86::CMOV_RFP32:
+  case X86::CMOV_RFP64:
+  case X86::CMOV_RFP80:
     return EmitLoweredSelect(MI, BB, EM);
 
   case X86::FP32_TO_INT16_IN_MEM:
diff --git a/lib/Target/X86/X86InstrFPStack.td b/lib/Target/X86/X86InstrFPStack.td
index ae24bfbb2d..b7309180bb 100644
--- a/lib/Target/X86/X86InstrFPStack.td
+++ b/lib/Target/X86/X86InstrFPStack.td
@@ -350,20 +350,27 @@ def FBLDm    : FPI<0xDF, MRM4m, (outs), (ins f32mem:$src), "fbld\t$src">;
 def FBSTPm   : FPI<0xDF, MRM6m, (outs f32mem:$dst), (ins), "fbstp\t$dst">;
 
 // Floating point cmovs.
+class FpIf32CMov<dag outs, dag ins, FPFormat fp, list<dag> pattern> :
+  FpI_<outs, ins, fp, pattern>, Requires<[FPStackf32, HasCMov]>;
+class FpIf64CMov<dag outs, dag ins, FPFormat fp, list<dag> pattern> :
+  FpI_<outs, ins, fp, pattern>, Requires<[FPStackf64, HasCMov]>;
+
 multiclass FPCMov<PatLeaf cc> {
-  def _Fp32  : FpIf32<(outs RFP32:$dst), (ins RFP32:$src1, RFP32:$src2),
+  def _Fp32  : FpIf32CMov<(outs RFP32:$dst), (ins RFP32:$src1, RFP32:$src2),
                        CondMovFP,
                      [(set RFP32:$dst, (X86cmov RFP32:$src1, RFP32:$src2,
                                         cc, EFLAGS))]>;
-  def _Fp64  : FpIf64<(outs RFP64:$dst), (ins RFP64:$src1, RFP64:$src2),
+  def _Fp64  : FpIf64CMov<(outs RFP64:$dst), (ins RFP64:$src1, RFP64:$src2),
                        CondMovFP,
                      [(set RFP64:$dst, (X86cmov RFP64:$src1, RFP64:$src2,
                                         cc, EFLAGS))]>;
   def _Fp80  : FpI_<(outs RFP80:$dst), (ins RFP80:$src1, RFP80:$src2),
                      CondMovFP,
                      [(set RFP80:$dst, (X86cmov RFP80:$src1, RFP80:$src2,
-                                        cc, EFLAGS))]>;
+                                        cc, EFLAGS))]>,
+                                        Requires<[HasCMov]>;
 }
+
 let Uses = [EFLAGS], isTwoAddress = 1 in {
 defm CMOVB  : FPCMov<X86_COND_B>;
 defm CMOVBE : FPCMov<X86_COND_BE>;
@@ -375,6 +382,7 @@ defm CMOVNE : FPCMov<X86_COND_NE>;
 defm CMOVNP : FPCMov<X86_COND_NP>;
 }
 
+let Predicates = [HasCMov] in {
 // These are not factored because there's no clean way to pass DA/DB.
 def CMOVB_F  : FPI<0xC0, AddRegFrm, (outs RST:$op), (ins),
                   "fcmovb\t{$op, %st(0)|%ST(0), $op}">, DA;
@@ -392,6 +400,7 @@ def CMOVNE_F : FPI<0xC8, AddRegFrm, (outs RST:$op), (ins),
                   "fcmovne\t{$op, %st(0)|%ST(0), $op}">, DB;
 def CMOVNP_F : FPI<0xD8, AddRegFrm, (outs RST:$op), (ins),
                   "fcmovnu\t{$op, %st(0)|%ST(0), $op}">, DB;
+} // Predicates = [HasCMov]
 
 // Floating point loads & stores.
 let canFoldAsLoad = 1 in {
diff --git a/lib/Target/X86/X86InstrInfo.td b/lib/Target/X86/X86InstrInfo.td
index e6a240d600..be5741645b 100644
--- a/lib/Target/X86/X86InstrInfo.td
+++ b/lib/Target/X86/X86InstrInfo.td
@@ -297,6 +297,8 @@ def tls32addr : ComplexPattern<i32, 4, "SelectTLSADDRAddr",
 
 //===----------------------------------------------------------------------===//
 // X86 Instruction Predicate Definitions.
+def HasCMov      : Predicate<"Subtarget->hasCMov()">;
+def NoCMov       : Predicate<"!Subtarget->hasCMov()">;
 def HasMMX       : Predicate<"Subtarget->hasMMX()">;
 def HasSSE1      : Predicate<"Subtarget->hasSSE1()">;
 def HasSSE2      : Predicate<"Subtarget->hasSSE2()">;
@@ -1213,19 +1215,7 @@ let isTwoAddress = 1 in {
 // Conditional moves
 let Uses = [EFLAGS] in {
 
-// X86 doesn't have 8-bit conditional moves. Use a customInserter to
-// emit control flow. An alternative to this is to mark i8 SELECT as Promote,
-// however that requires promoting the operands, and can induce additional
-// i8 register pressure. Note that CMOV_GR8 is conservatively considered to
-// clobber EFLAGS, because if one of the operands is zero, the expansion
-// could involve an xor.
-let usesCustomInserter = 1, isTwoAddress = 0, Defs = [EFLAGS] in
-def CMOV_GR8 : I<0, Pseudo,
-                 (outs GR8:$dst), (ins GR8:$src1, GR8:$src2, i8imm:$cond),
-                 "#CMOV_GR8 PSEUDO!",
-                 [(set GR8:$dst, (X86cmov GR8:$src1, GR8:$src2,
-                                          imm:$cond, EFLAGS))]>;
-
+let Predicates = [HasCMov] in {
 let isCommutable = 1 in {
 def CMOVB16rr : I<0x42, MRMSrcReg,       // if <u, GR16 = GR16
                   (outs GR16:$dst), (ins GR16:$src1, GR16:$src2),
@@ -1613,6 +1603,49 @@ def CMOVNO32rm : I<0x41, MRMSrcMem,       // if !overflow, GR32 = [mem32]
                    [(set GR32:$dst, (X86cmov GR32:$src1, (loadi32 addr:$src2),
                                     X86_COND_NO, EFLAGS))]>,
                   TB;
+} // Predicates = [HasCMov]
+
+// X86 doesn't have 8-bit conditional moves. Use a customInserter to
+// emit control flow. An alternative to this is to mark i8 SELECT as Promote,
+// however that requires promoting the operands, and can induce additional
+// i8 register pressure. Note that CMOV_GR8 is conservatively considered to
+// clobber EFLAGS, because if one of the operands is zero, the expansion
+// could involve an xor.
+let usesCustomInserter = 1, isTwoAddress = 0, Defs = [EFLAGS] in {
+def CMOV_GR8 : I<0, Pseudo,
+                 (outs GR8:$dst), (ins GR8:$src1, GR8:$src2, i8imm:$cond),
+                 "#CMOV_GR8 PSEUDO!",
+                 [(set GR8:$dst, (X86cmov GR8:$src1, GR8:$src2,
+                                          imm:$cond, EFLAGS))]>;
+
+let Predicates = [NoCMov] in {
+def CMOV_GR32 : I<0, Pseudo,
+                    (outs GR32:$dst), (ins GR32:$src1, GR32:$src2, i8imm:$cond),
+                    "#CMOV_GR32* PSEUDO!",
+                    [(set GR32:$dst,
+                      (X86cmov GR32:$src1, GR32:$src2, imm:$cond, EFLAGS))]>;
+def CMOV_GR16 : I<0, Pseudo,
+                    (outs GR16:$dst), (ins GR16:$src1, GR16:$src2, i8imm:$cond),
+                    "#CMOV_GR16* PSEUDO!",
+                    [(set GR16:$dst,
+                      (X86cmov GR16:$src1, GR16:$src2, imm:$cond, EFLAGS))]>;
+def CMOV_RFP32 : I<0, Pseudo,
+                    (outs RFP32:$dst), (ins RFP32:$src1, RFP32:$src2, i8imm:$cond),
+                    "#CMOV_RFP32 PSEUDO!",
+                    [(set RFP32:$dst, (X86cmov RFP32:$src1, RFP32:$src2, imm:$cond,
+                                                  EFLAGS))]>;
+def CMOV_RFP64 : I<0, Pseudo,
+                    (outs RFP64:$dst), (ins RFP64:$src1, RFP64:$src2, i8imm:$cond),
+                    "#CMOV_RFP64 PSEUDO!",
+                    [(set RFP64:$dst, (X86cmov RFP64:$src1, RFP64:$src2, imm:$cond,
+                                                  EFLAGS))]>;
+def CMOV_RFP80 : I<0, Pseudo,
+                    (outs RFP80:$dst), (ins RFP80:$src1, RFP80:$src2, i8imm:$cond),
+                    "#CMOV_RFP80 PSEUDO!",
+                    [(set RFP80:$dst, (X86cmov RFP80:$src1, RFP80:$src2, imm:$cond,
+                                                  EFLAGS))]>;
+} // Predicates = [NoCMov]
+} // UsesCustomInserter = 1, isTwoAddress = 0, Defs = [EFLAGS] 
 } // Uses = [EFLAGS]
 
 
diff --git a/lib/Target/X86/X86Subtarget.h b/lib/Target/X86/X86Subtarget.h
index 594a470009..50338d33ac 100644
--- a/lib/Target/X86/X86Subtarget.h
+++ b/lib/Target/X86/X86Subtarget.h
@@ -133,6 +133,7 @@ public:
   PICStyles::Style getPICStyle() const { return PICStyle; }
   void setPICStyle(PICStyles::Style Style)  { PICStyle = Style; }
 
+  bool hasCMov() const { return HasCMov; }
   bool hasMMX() const { return X86SSELevel >= MMX; }
   bool hasSSE1() const { return X86SSELevel >= SSE1; }
   bool hasSSE2() const { return X86SSELevel >= SSE2; }