X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86InstrShiftRotate.td;h=1185941d344d0f35fcca92dc5123f5e02e45a6c5;hb=7fd324a31fbfd237f43d38d3a780a19fbf909ba3;hp=fe7d0ecf896541290a7e25830f2b38ec0ebcdb72;hpb=6bcdb5b9034a02901175e94e89b4815f28f2f141;p=oota-llvm.git

diff --git a/lib/Target/X86/X86InstrShiftRotate.td b/lib/Target/X86/X86InstrShiftRotate.td
index fe7d0ecf896..1185941d344 100644
--- a/lib/Target/X86/X86InstrShiftRotate.td
+++ b/lib/Target/X86/X86InstrShiftRotate.td
@@ -51,6 +51,7 @@ def SHL64ri  : RIi8<0xC1, MRM4r, (outs GR64:$dst),
 
 // NOTE: We don't include patterns for shifts of a register by one, because
 // 'add reg,reg' is cheaper (and we have a Pat pattern for shift-by-one).
+let hasSideEffects = 0 in {
 def SHL8r1   : I<0xD0, MRM4r, (outs GR8:$dst), (ins GR8:$src1),
                  "shl{b}\t$dst", [], IIC_SR>;
 def SHL16r1  : I<0xD1, MRM4r, (outs GR16:$dst), (ins GR16:$src1),
@@ -59,8 +60,9 @@ def SHL32r1  : I<0xD1, MRM4r, (outs GR32:$dst), (ins GR32:$src1),
                  "shl{l}\t$dst", [], IIC_SR>;
 def SHL64r1  : RI<0xD1, MRM4r, (outs GR64:$dst), (ins GR64:$src1),
                  "shl{q}\t$dst", [], IIC_SR>;
+} // hasSideEffects = 0
 } // isConvertibleToThreeAddress = 1
-} // Constraints = "$src = $dst" 
+} // Constraints = "$src = $dst"
 
 
 // FIXME: Why do we need an explicit "Uses = [CL]" when the instr has a pattern
@@ -333,6 +335,7 @@ def SAR64m1 : RI<0xD1, MRM7m, (outs), (ins i64mem:$dst),
 // Rotate instructions
 //===----------------------------------------------------------------------===//
 
+let hasSideEffects = 0 in {
 let Constraints = "$src1 = $dst" in {
 def RCL8r1 : I<0xD0, MRM2r, (outs GR8:$dst), (ins GR8:$src1),
                "rcl{b}\t$dst", [], IIC_SR>;
@@ -455,6 +458,7 @@ def RCR32mCL : I<0xD3, MRM3m, (outs), (ins i32mem:$dst),
 def RCR64mCL : RI<0xD3, MRM3m, (outs), (ins i64mem:$dst),
                   "rcr{q}\t{%cl, $dst|$dst, CL}", [], IIC_SR>;
 }
+} // hasSideEffects = 0
 
 let Constraints = "$src1 = $dst" in {
 // FIXME: provide shorter instructions when imm8 == 1
@@ -896,4 +900,59 @@ let Predicates = [HasBMI2] in {
             (RORX32mi addr:$src, (ROT32L2R_imm8 imm:$shamt))>;
   def : Pat<(rotl (loadi64 addr:$src), (i8 imm:$shamt)),
             (RORX64mi addr:$src, (ROT64L2R_imm8 imm:$shamt))>;
+
+  // Prefer SARX/SHRX/SHLX over SAR/SHR/SHL with variable shift BUT not
+  // immedidate shift, i.e. the following code is considered better
+  //
+  //  mov %edi, %esi
+  //  shl $imm, %esi
+  //  ... %edi, ...
+  //
+  // than
+  //
+  //  movb $imm, %sil
+  //  shlx %sil, %edi, %esi
+  //  ... %edi, ...
+  //
+  let AddedComplexity = 1 in {
+    def : Pat<(sra GR32:$src1, GR8:$src2),
+              (SARX32rr GR32:$src1,
+                        (INSERT_SUBREG
+                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
+    def : Pat<(sra GR64:$src1, GR8:$src2),
+              (SARX64rr GR64:$src1,
+                        (INSERT_SUBREG
+                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
+
+    def : Pat<(srl GR32:$src1, GR8:$src2),
+              (SHRX32rr GR32:$src1,
+                        (INSERT_SUBREG
+                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
+    def : Pat<(srl GR64:$src1, GR8:$src2),
+              (SHRX64rr GR64:$src1,
+                        (INSERT_SUBREG
+                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
+
+    def : Pat<(shl GR32:$src1, GR8:$src2),
+              (SHLX32rr GR32:$src1,
+                        (INSERT_SUBREG
+                          (i32 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
+    def : Pat<(shl GR64:$src1, GR8:$src2),
+              (SHLX64rr GR64:$src1,
+                        (INSERT_SUBREG
+                          (i64 (IMPLICIT_DEF)), GR8:$src2, sub_8bit))>;
+  }
+
+  // Patterns on SARXrm/SHRXrm/SHLXrm are explicitly omitted to favor
+  //
+  //  mov (%ecx), %esi
+  //  shl $imm, $esi
+  //
+  // over
+  //
+  //  movb $imm %al
+  //  shlx %al, (%ecx), %esi
+  //
+  // As SARXrr/SHRXrr/SHLXrr is favored on variable shift, the peephole
+  // optimization will fold them into SARXrm/SHRXrm/SHLXrm if possible.
 }