Re-enable elimination of unnecessary SUBREG_TO_REG instructions in

LowerSubregs, and fix an x86-64 isel bug that this exposed.

SUBREG_TO_REG for x86-64 implicit zero extension is only safe for
isel to generate when the source is known to always have zeros in
the high 32 bits. The EXTRACT_SUBREG instruction does not clear
the high 32 bits.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54444 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 46 insertions, 12 deletions

diff --git a/lib/CodeGen/LowerSubregs.cpp b/lib/CodeGen/LowerSubregs.cpp
index 1cdd34b77f..30894ad6ad 100644
--- a/lib/CodeGen/LowerSubregs.cpp
+++ b/lib/CodeGen/LowerSubregs.cpp
@@ -108,15 +108,20 @@ bool LowerSubregsInstructionPass::LowerSubregToReg(MachineInstr *MI) {
 
   DOUT << "subreg: CONVERTING: " << *MI;
 
-  // Insert sub-register copy
-  const TargetRegisterClass *TRC0= TRI.getPhysicalRegisterRegClass(DstSubReg);
-  const TargetRegisterClass *TRC1= TRI.getPhysicalRegisterRegClass(InsReg);
-  TII.copyRegToReg(*MBB, MI, DstSubReg, InsReg, TRC0, TRC1);
+  if (DstSubReg == InsReg) {
+    // No need to insert an identify copy instruction.
+    DOUT << "subreg: eliminated!";
+  } else {
+    // Insert sub-register copy
+    const TargetRegisterClass *TRC0= TRI.getPhysicalRegisterRegClass(DstSubReg);
+    const TargetRegisterClass *TRC1= TRI.getPhysicalRegisterRegClass(InsReg);
+    TII.copyRegToReg(*MBB, MI, DstSubReg, InsReg, TRC0, TRC1);
 
 #ifndef NDEBUG
   MachineBasicBlock::iterator dMI = MI;
   DOUT << "subreg: " << *(--dMI);
 #endif
+  }
 
   DOUT << "\n";
   MBB->erase(MI);
diff --git a/lib/Target/X86/X86Instr64bit.td b/lib/Target/X86/X86Instr64bit.td
index a187f86a10..f891a9e818 100644
--- a/lib/Target/X86/X86Instr64bit.td
+++ b/lib/Target/X86/X86Instr64bit.td
@@ -258,6 +258,18 @@ def MOVZX64rm16: I<0xB7, MRMSrcMem, (outs GR64:$dst), (ins i16mem:$src),
                    "movz{wl|x}\t{$src, ${dst:subreg32}|${dst:subreg32}, $src}",
                    [(set GR64:$dst, (zextloadi64i16 addr:$src))]>, TB;
 
+// There's no movzlq instruction, but movl can be used for this purpose, using
+// implicit zero-extension. We need this because the seeming alternative for
+// implementing zext from 32 to 64, an EXTRACT_SUBREG/SUBREG_TO_REG pair, isn't
+// safe because both instructions could be optimized away in the
+// register-to-register case, leaving nothing behind to do the zero extension.
+def MOVZX64rr32 : I<0x89, MRMDestReg, (outs GR64:$dst), (ins GR32:$src),
+                    "mov{l}\t{$src, ${dst:subreg32}|${dst:subreg32}, $src}",
+                    [(set GR64:$dst, (zext GR32:$src))]>;
+def MOVZX64rm32 : I<0x8B, MRMSrcMem, (outs GR64:$dst), (ins i32mem:$src),
+                    "mov{l}\t{$src, ${dst:subreg32}|${dst:subreg32}, $src}",
+                    [(set GR64:$dst, (zextloadi64i32 addr:$src))]>;
+
 let neverHasSideEffects = 1 in {
   let Defs = [RAX], Uses = [EAX] in
   def CDQE : RI<0x98, RawFrm, (outs), (ins),
@@ -1213,9 +1225,6 @@ def : Pat<(i64 (zext GR32:$src)),
 // zextload bool -> zextload byte
 def : Pat<(zextloadi64i1 addr:$src), (MOVZX64rm8 addr:$src)>;
 
-def : Pat<(zextloadi64i32 addr:$src), 
-          (SUBREG_TO_REG (i64 0), (MOV32rm addr:$src), x86_subreg_32bit)>;
-
 // extload
 def : Pat<(extloadi64i1 addr:$src),  (MOVZX64rm8  addr:$src)>;
 def : Pat<(extloadi64i8 addr:$src),  (MOVZX64rm8  addr:$src)>;
@@ -1240,11 +1249,9 @@ def : Pat<(i64 (anyext (loadi32 addr:$src))),
 // Some peepholes
 //===----------------------------------------------------------------------===//
 
-// r & (2^32-1) ==> mov32 + implicit zext
-def : Pat<(and GR64:$src, i64immFFFFFFFF), 
-          (SUBREG_TO_REG (i64 0), 
-            (i32 (EXTRACT_SUBREG GR64:$src, x86_subreg_32bit)),
-            x86_subreg_32bit)>;
+// r & (2^32-1) ==> movz
+def : Pat<(and GR64:$src, i64immFFFFFFFF),
+          (MOVZX64rr32 (i32 (EXTRACT_SUBREG GR64:$src, x86_subreg_32bit)))>;
 // r & (2^16-1) ==> movz
 def : Pat<(and GR64:$src, 0xffff),
           (MOVZX64rr16 (i16 (EXTRACT_SUBREG GR64:$src, x86_subreg_16bit)))>;
diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index 61dcfa95b3..a871fa8dc4 100644
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -396,6 +396,7 @@ X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
     { X86::MOVZX32rr16,     X86::MOVZX32rm16 },
     { X86::MOVZX32rr8,      X86::MOVZX32rm8 },
     { X86::MOVZX64rr16,     X86::MOVZX64rm16 },
+    { X86::MOVZX64rr32,     X86::MOVZX64rm32 },
     { X86::MOVZX64rr8,      X86::MOVZX64rm8 },
     { X86::PSHUFDri,        X86::PSHUFDmi },
     { X86::PSHUFHWri,       X86::PSHUFHWmi },
diff --git a/test/CodeGen/X86/subreg-to-reg-0.ll b/test/CodeGen/X86/subreg-to-reg-0.ll
new file mode 100644
index 0000000000..6b60f65265
--- /dev/null
+++ b/test/CodeGen/X86/subreg-to-reg-0.ll
@@ -0,0 +1,11 @@
+; RUN: llvm-as < %s | llc -march=x86-64 | grep mov | count 1
+
+; Do eliminate the zero-extension instruction and rely on
+; x86-64's implicit zero-extension!
+
+define i64 @foo(i32* %p) nounwind {
+  %t = load i32* %p
+  %n = add i32 %t, 1
+  %z = zext i32 %n to i64
+  ret i64 %z
+}
diff --git a/test/CodeGen/X86/subreg-to-reg-1.ll b/test/CodeGen/X86/subreg-to-reg-1.ll
new file mode 100644
index 0000000000..4e487e1847
--- /dev/null
+++ b/test/CodeGen/X86/subreg-to-reg-1.ll
@@ -0,0 +1,10 @@
+; RUN: llvm-as < %s | llc -march=x86-64 | grep {movl	%e.\*, %e.\*} | count 1
+
+; Don't eliminate or coalesce away the explicit zero-extension!
+
+define i64 @foo(i64 %a) {
+  %b = add i64 %a, 4294967295
+  %c = and i64 %b, 4294967295
+  %d = add i64 %c, 1
+  ret i64 %d
+}