Avoid unecessary opsize byte in segment move to memory

Segment moves to memory are always 16-bit. Remove invalid 32 and 64
bit variants.

Recommiting with missing clang inline assembly test change.

Fixes PR34478.

Reviewers: rnk, craig.topper

Subscribers: llvm-commits, hiraditya

Differential Revision: https://reviews.llvm.org/D39847

llvm-svn: 318797

5 files changed, 16 insertions, 20 deletions

diff --git a/llvm/lib/Target/X86/MCTargetDesc/X86BaseInfo.h b/llvm/lib/Target/X86/MCTargetDesc/X86BaseInfo.h
index 7c6444ba58a..f65ba1b6005 100644
--- a/llvm/lib/Target/X86/MCTargetDesc/X86BaseInfo.h
+++ b/llvm/lib/Target/X86/MCTargetDesc/X86BaseInfo.h
@@ -366,13 +366,15 @@ namespace X86II {
     // OpSize - OpSizeFixed implies instruction never needs a 0x66 prefix.
     // OpSize16 means this is a 16-bit instruction and needs 0x66 prefix in
     // 32-bit mode. OpSize32 means this is a 32-bit instruction needs a 0x66
-    // prefix in 16-bit mode.
+    // prefix in 16-bit mode. OpSizeIgnore means that the instruction may
+    // take a optional 0x66 byte but should not emit with one.
     OpSizeShift = 7,
     OpSizeMask = 0x3 << OpSizeShift,
 
-    OpSizeFixed = 0 << OpSizeShift,
-    OpSize16    = 1 << OpSizeShift,
-    OpSize32    = 2 << OpSizeShift,
+    OpSizeFixed  = 0 << OpSizeShift,
+    OpSize16     = 1 << OpSizeShift,
+    OpSize32     = 2 << OpSizeShift,
+    OpSizeIgnore = 3 << OpSizeShift,
 
     // AsSize - AdSizeX implies this instruction determines its need of 0x67
     // prefix from a normal ModRM memory operand. The other types indicate that
diff --git a/llvm/lib/Target/X86/X86InstrFormats.td b/llvm/lib/Target/X86/X86InstrFormats.td
index f9572dab05c..371b5046ff3 100644
--- a/llvm/lib/Target/X86/X86InstrFormats.td
+++ b/llvm/lib/Target/X86/X86InstrFormats.td
@@ -157,9 +157,10 @@ def EncEVEX   : Encoding<3>;
 class OperandSize<bits<2> val> {
   bits<2> Value = val;
 }
-def OpSizeFixed : OperandSize<0>; // Never needs a 0x66 prefix.
-def OpSize16    : OperandSize<1>; // Needs 0x66 prefix in 32-bit mode.
-def OpSize32    : OperandSize<2>; // Needs 0x66 prefix in 16-bit mode.
+def OpSizeFixed  : OperandSize<0>; // Never needs a 0x66 prefix.
+def OpSize16     : OperandSize<1>; // Needs 0x66 prefix in 32-bit mode.
+def OpSize32     : OperandSize<2>; // Needs 0x66 prefix in 16-bit mode.
+def OpSizeIgnore : OperandSize<3>; // Takes 0x66 prefix, never emits.
 
 // Address size for encodings that change based on mode.
 class AddressSize<bits<2> val> {
@@ -174,6 +175,7 @@ def AdSize64 : AddressSize<3>; // Encodes a 64-bit address.
 // emitter that various prefix bytes are required.
 class OpSize16 { OperandSize OpSize = OpSize16; }
 class OpSize32 { OperandSize OpSize = OpSize32; }
+class OpSizeIgnore { OperandSize OpSize = OpSizeIgnore; }
 class AdSize16 { AddressSize AdSize = AdSize16; }
 class AdSize32 { AddressSize AdSize = AdSize32; }
 class AdSize64 { AddressSize AdSize = AdSize64; }
diff --git a/llvm/lib/Target/X86/X86InstrInfo.td b/llvm/lib/Target/X86/X86InstrInfo.td
index 37f1d97faac..926080347e7 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.td
+++ b/llvm/lib/Target/X86/X86InstrInfo.td
@@ -3165,8 +3165,8 @@ def : InstAlias<"jmpl\t$seg, $off",  (FARJMP32i  i32imm:$off, i16imm:$seg)>, Req
 // Force mov without a suffix with a segment and mem to prefer the 'l' form of
 // the move.  All segment/mem forms are equivalent, this has the shortest
 // encoding.
-def : InstAlias<"mov\t{$mem, $seg|$seg, $mem}", (MOV32sm SEGMENT_REG:$seg, i32mem:$mem), 0>;
-def : InstAlias<"mov\t{$seg, $mem|$mem, $seg}", (MOV32ms i32mem:$mem, SEGMENT_REG:$seg), 0>;
+def : InstAlias<"mov\t{$mem, $seg|$seg, $mem}", (MOV16sm SEGMENT_REG:$seg, i16mem:$mem), 0>;
+def : InstAlias<"mov\t{$seg, $mem|$mem, $seg}", (MOV16ms i16mem:$mem, SEGMENT_REG:$seg), 0>;
 
 // Match 'movq <largeimm>, <reg>' as an alias for movabsq.
 def : InstAlias<"mov{q}\t{$imm, $reg|$reg, $imm}", (MOV64ri GR64:$reg, i64imm:$imm), 0>;
diff --git a/llvm/lib/Target/X86/X86InstrSystem.td b/llvm/lib/Target/X86/X86InstrSystem.td
index 86b3f21018f..abe20a2dd3e 100644
--- a/llvm/lib/Target/X86/X86InstrSystem.td
+++ b/llvm/lib/Target/X86/X86InstrSystem.td
@@ -175,11 +175,7 @@ def MOV64rs : RI<0x8C, MRMDestReg, (outs GR64:$dst), (ins SEGMENT_REG:$src),
                  "mov{q}\t{$src, $dst|$dst, $src}", [], IIC_MOV_REG_SR>;
 let mayStore = 1 in {
 def MOV16ms : I<0x8C, MRMDestMem, (outs), (ins i16mem:$dst, SEGMENT_REG:$src),
-                "mov{w}\t{$src, $dst|$dst, $src}", [], IIC_MOV_MEM_SR>, OpSize16;
-def MOV32ms : I<0x8C, MRMDestMem, (outs), (ins i32mem:$dst, SEGMENT_REG:$src),
-                "mov{l}\t{$src, $dst|$dst, $src}", [], IIC_MOV_MEM_SR>, OpSize32;
-def MOV64ms : RI<0x8C, MRMDestMem, (outs), (ins i64mem:$dst, SEGMENT_REG:$src),
-                 "mov{q}\t{$src, $dst|$dst, $src}", [], IIC_MOV_MEM_SR>;
+                "mov{w}\t{$src, $dst|$dst, $src}", [], IIC_MOV_MEM_SR>, OpSizeIgnore;
 }
 def MOV16sr : I<0x8E, MRMSrcReg, (outs SEGMENT_REG:$dst), (ins GR16:$src),
                 "mov{w}\t{$src, $dst|$dst, $src}", [], IIC_MOV_SR_REG>, OpSize16;
@@ -189,11 +185,7 @@ def MOV64sr : RI<0x8E, MRMSrcReg, (outs SEGMENT_REG:$dst), (ins GR64:$src),
                  "mov{q}\t{$src, $dst|$dst, $src}", [], IIC_MOV_SR_REG>;
 let mayLoad = 1 in {
 def MOV16sm : I<0x8E, MRMSrcMem, (outs SEGMENT_REG:$dst), (ins i16mem:$src),
-                "mov{w}\t{$src, $dst|$dst, $src}", [], IIC_MOV_SR_MEM>, OpSize16;
-def MOV32sm : I<0x8E, MRMSrcMem, (outs SEGMENT_REG:$dst), (ins i32mem:$src),
-                "mov{l}\t{$src, $dst|$dst, $src}", [], IIC_MOV_SR_MEM>, OpSize32;
-def MOV64sm : RI<0x8E, MRMSrcMem, (outs SEGMENT_REG:$dst), (ins i64mem:$src),
-                 "mov{q}\t{$src, $dst|$dst, $src}", [], IIC_MOV_SR_MEM>;
+                "mov{w}\t{$src, $dst|$dst, $src}", [], IIC_MOV_SR_MEM>, OpSizeIgnore;
 }
 } // SchedRW
 
diff --git a/llvm/lib/Target/X86/X86SchedSandyBridge.td b/llvm/lib/Target/X86/X86SchedSandyBridge.td
index 8f8ea9d8feb..c86c48ce06b 100644
--- a/llvm/lib/Target/X86/X86SchedSandyBridge.td
+++ b/llvm/lib/Target/X86/X86SchedSandyBridge.td
@@ -1550,7 +1550,7 @@ def SBWriteResGroup49 : SchedWriteRes<[SBPort5,SBPort23]> {
   let ResourceCycles = [1,1];
 }
 def: InstRW<[SBWriteResGroup49], (instregex "JMP(16|32|64)m")>;
-def: InstRW<[SBWriteResGroup49], (instregex "MOV64sm")>;
+def: InstRW<[SBWriteResGroup49], (instregex "MOV16sm")>;
 
 def SBWriteResGroup50 : SchedWriteRes<[SBPort23,SBPort05]> {
   let Latency = 6;