Add separate intrinsics for MMX / SSE shifts with i32 integer operands. This allow...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50601 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/IntrinsicsX86.td b/include/llvm/IntrinsicsX86.td
index 5bcfa7922ba0e4e32db5d704518fad4f35974186..0484926d07698594e9c70ba5213198ebe09db572 100644 (file)
--- a/include/llvm/IntrinsicsX86.td
+++ b/include/llvm/IntrinsicsX86.td
@@ -324,9 +324,6 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_psll_q : GCCBuiltin<"__builtin_ia32_psllq128">,
               Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
                          llvm_v2i64_ty], [IntrNoMem]>;
-  def int_x86_sse2_psll_dq : GCCBuiltin<"__builtin_ia32_pslldqi128">,
-              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
-                         llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psrl_w : GCCBuiltin<"__builtin_ia32_psrlw128">,
               Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
@@ -336,15 +333,44 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_sse2_psrl_q : GCCBuiltin<"__builtin_ia32_psrlq128">,
               Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
                          llvm_v2i64_ty], [IntrNoMem]>;
-  def int_x86_sse2_psrl_dq : GCCBuiltin<"__builtin_ia32_psrldqi128">,
-              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
-                         llvm_i32_ty], [IntrNoMem]>;
   def int_x86_sse2_psra_w : GCCBuiltin<"__builtin_ia32_psraw128">,
               Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
                          llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_sse2_psra_d : GCCBuiltin<"__builtin_ia32_psrad128">,
               Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty,
                          llvm_v4i32_ty], [IntrNoMem]>;
+
+  def int_x86_sse2_pslli_w : GCCBuiltin<"__builtin_ia32_psllwi128">,
+              Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_sse2_pslli_d : GCCBuiltin<"__builtin_ia32_pslldi128">,
+              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_sse2_pslli_q : GCCBuiltin<"__builtin_ia32_psllqi128">,
+              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_sse2_psrli_w : GCCBuiltin<"__builtin_ia32_psrlwi128">,
+              Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_sse2_psrli_d : GCCBuiltin<"__builtin_ia32_psrldi128">,
+              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_sse2_psrli_q : GCCBuiltin<"__builtin_ia32_psrlqi128">,
+              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_sse2_psrai_w : GCCBuiltin<"__builtin_ia32_psrawi128">,
+              Intrinsic<[llvm_v8i16_ty, llvm_v8i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_sse2_psrai_d : GCCBuiltin<"__builtin_ia32_psradi128">,
+              Intrinsic<[llvm_v4i32_ty, llvm_v4i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+
+  def int_x86_sse2_psll_dq : GCCBuiltin<"__builtin_ia32_pslldqi128">,
+              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_sse2_psrl_dq : GCCBuiltin<"__builtin_ia32_psrldqi128">,
+              Intrinsic<[llvm_v2i64_ty, llvm_v2i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
 }
 
 // Integer comparison ops
@@ -939,6 +965,33 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_mmx_psra_d : GCCBuiltin<"__builtin_ia32_psrad">,
               Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
                          llvm_v1i64_ty], [IntrNoMem]>;
+
+  def int_x86_mmx_pslli_w : GCCBuiltin<"__builtin_ia32_psllwi">,
+              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_mmx_pslli_d : GCCBuiltin<"__builtin_ia32_pslldi">,
+              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_mmx_pslli_q : GCCBuiltin<"__builtin_ia32_psllqi">,
+              Intrinsic<[llvm_v1i64_ty, llvm_v1i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+
+  def int_x86_mmx_psrli_w : GCCBuiltin<"__builtin_ia32_psrlwi">,
+              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_mmx_psrli_d : GCCBuiltin<"__builtin_ia32_psrldi">,
+              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_mmx_psrli_q : GCCBuiltin<"__builtin_ia32_psrlqi">,
+              Intrinsic<[llvm_v1i64_ty,   llvm_v1i64_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+
+  def int_x86_mmx_psrai_w : GCCBuiltin<"__builtin_ia32_psrawi">,
+              Intrinsic<[llvm_v4i16_ty, llvm_v4i16_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_mmx_psrai_d : GCCBuiltin<"__builtin_ia32_psradi">,
+              Intrinsic<[llvm_v2i32_ty, llvm_v2i32_ty,
+                         llvm_i32_ty], [IntrNoMem]>;
 }
 
 // Pack ops.

diff --git a/lib/Target/X86/X86InstrMMX.td b/lib/Target/X86/X86InstrMMX.td
index d4413e9fbccb6f9466fb0e9ed48779ef1a16a0ed..d484695f60ea2214e70dd6e154cd57e2cdcb8f7b 100644 (file)
--- a/lib/Target/X86/X86InstrMMX.td
+++ b/lib/Target/X86/X86InstrMMX.td
@@ -118,7 +118,8 @@ let isTwoAddress = 1 in {
   }
 
   multiclass MMXI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
-                                string OpcodeStr, Intrinsic IntId> {
+                                string OpcodeStr, Intrinsic IntId,
+                                Intrinsic IntId2> {
     def rr : MMXI<opc, MRMSrcReg, (outs VR64:$dst),
                                   (ins VR64:$src1, VR64:$src2),
                   !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
@@ -131,11 +132,7 @@ let isTwoAddress = 1 in {
     def ri : MMXIi8<opc2, ImmForm, (outs VR64:$dst),
                                    (ins VR64:$src1, i32i8imm:$src2),
                     !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
-           [(set VR64:$dst, (IntId VR64:$src1,
-                             (v1i64 (bitconvert
-                                     (v2i32 (vector_shuffle immAllZerosV,
-                                     (v2i32 (scalar_to_vector (i32 imm:$src2))),
-                                             MMX_MOVL_shuffle_mask))))))]>;
+           [(set VR64:$dst, (IntId2 VR64:$src1, (i32 imm:$src2)))]>;
   }
 }
 
@@ -283,23 +280,23 @@ let isTwoAddress = 1 in {
 
 // Shift Instructions
 defm MMX_PSRLW : MMXI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
-                                    int_x86_mmx_psrl_w>;
+                                    int_x86_mmx_psrl_w, int_x86_mmx_psrli_w>;
 defm MMX_PSRLD : MMXI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
-                                    int_x86_mmx_psrl_d>;
+                                    int_x86_mmx_psrl_d, int_x86_mmx_psrli_d>;
 defm MMX_PSRLQ : MMXI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
-                                    int_x86_mmx_psrl_q>;
+                                    int_x86_mmx_psrl_q, int_x86_mmx_psrli_q>;
 
 defm MMX_PSLLW : MMXI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
-                                    int_x86_mmx_psll_w>;
+                                    int_x86_mmx_psll_w, int_x86_mmx_pslli_w>;
 defm MMX_PSLLD : MMXI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
-                                    int_x86_mmx_psll_d>;
+                                    int_x86_mmx_psll_d, int_x86_mmx_pslli_d>;
 defm MMX_PSLLQ : MMXI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
-                                    int_x86_mmx_psll_q>;
+                                    int_x86_mmx_psll_q, int_x86_mmx_pslli_q>;
 
 defm MMX_PSRAW : MMXI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
-                                    int_x86_mmx_psra_w>;
+                                    int_x86_mmx_psra_w, int_x86_mmx_psrai_w>;
 defm MMX_PSRAD : MMXI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
-                                    int_x86_mmx_psra_d>;
+                                    int_x86_mmx_psra_d, int_x86_mmx_psrai_d>;
 
 // Comparison Instructions
 defm MMX_PCMPEQB : MMXI_binop_rm_int<0x74, "pcmpeqb", int_x86_mmx_pcmpeq_b>;

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index e8ee9dc1c4a11539c8d5f24d6aec54aa79dcff19..9c07de6fd0303d4a8147efe6a4c0ecd44a2738e4 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -1780,6 +1780,21 @@ multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
                                         (bitconvert (memopv2i64 addr:$src2))))]>;
 }
 
+multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
+                             string OpcodeStr,
+                             Intrinsic IntId, Intrinsic IntId2> {
+  def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
+               !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
+               [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
+  def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
+               !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
+               [(set VR128:$dst, (IntId VR128:$src1,
+                                        (bitconvert (memopv2i64 addr:$src2))))]>;
+  def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
+               !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
+               [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
+}
+
 /// PDI_binop_rm - Simple SSE2 binary operator.
 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
                         ValueType OpVT, bit Commutable = 0> {
@@ -1854,64 +1869,24 @@ defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
 defm PSADBW : PDI_binop_rm_int<0xE0, "psadbw", int_x86_sse2_psad_bw, 1>;
 
 
-defm PSLLW : PDI_binop_rm_int<0xF1, "psllw", int_x86_sse2_psll_w>;
-defm PSLLD : PDI_binop_rm_int<0xF2, "pslld", int_x86_sse2_psll_d>;
-defm PSLLQ : PDI_binop_rm_int<0xF3, "psllq", int_x86_sse2_psll_q>;
-
-defm PSRLW : PDI_binop_rm_int<0xD1, "psrlw", int_x86_sse2_psrl_w>;
-defm PSRLD : PDI_binop_rm_int<0xD2, "psrld", int_x86_sse2_psrl_d>;
-defm PSRLQ : PDI_binop_rm_int<0xD3, "psrlq", int_x86_sse2_psrl_q>;
-
-defm PSRAW : PDI_binop_rm_int<0xE1, "psraw", int_x86_sse2_psra_w>;
-defm PSRAD : PDI_binop_rm_int<0xE2, "psrad", int_x86_sse2_psra_d>;
-
-// Some immediate variants need to match a bit_convert.
-let Constraints = "$src1 = $dst" in {
-def PSLLWri : PDIi8<0x71, MRM6r, (outs VR128:$dst),
-                                 (ins VR128:$src1, i32i8imm:$src2),
-                    "psllw\t{$src2, $dst|$dst, $src2}",
-                    [(set VR128:$dst, (int_x86_sse2_psll_w VR128:$src1,
-                      (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
-def PSLLDri : PDIi8<0x72, MRM6r, (outs VR128:$dst),
-                                 (ins VR128:$src1, i32i8imm:$src2),
-                    "pslld\t{$src2, $dst|$dst, $src2}",
-                    [(set VR128:$dst, (int_x86_sse2_psll_d VR128:$src1,
-                          (scalar_to_vector (i32 imm:$src2))))]>;
-def PSLLQri : PDIi8<0x73, MRM6r, (outs VR128:$dst),
-                                 (ins VR128:$src1, i32i8imm:$src2),
-                    "psllq\t{$src2, $dst|$dst, $src2}",
-                    [(set VR128:$dst, (int_x86_sse2_psll_q VR128:$src1,
-                      (bc_v2i64 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
-
-def PSRLWri : PDIi8<0x71, MRM2r, (outs VR128:$dst),
-                                 (ins VR128:$src1, i32i8imm:$src2),
-                    "psrlw\t{$src2, $dst|$dst, $src2}",
-                    [(set VR128:$dst, (int_x86_sse2_psrl_w VR128:$src1,
-                      (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
-def PSRLDri : PDIi8<0x72, MRM2r, (outs VR128:$dst),
-                                 (ins VR128:$src1, i32i8imm:$src2),
-                    "psrld\t{$src2, $dst|$dst, $src2}",
-                    [(set VR128:$dst, (int_x86_sse2_psrl_d VR128:$src1,
-                          (scalar_to_vector (i32 imm:$src2))))]>;
-def PSRLQri : PDIi8<0x73, MRM2r, (outs VR128:$dst),
-                                 (ins VR128:$src1, i32i8imm:$src2),
-                    "psrlq\t{$src2, $dst|$dst, $src2}",
-                    [(set VR128:$dst, (int_x86_sse2_psrl_q VR128:$src1,
-                      (bc_v2i64 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
-
-def PSRAWri : PDIi8<0x71, MRM4r, (outs VR128:$dst),
-                                 (ins VR128:$src1, i32i8imm:$src2),
-                    "psraw\t{$src2, $dst|$dst, $src2}",
-                    [(set VR128:$dst, (int_x86_sse2_psra_w VR128:$src1,
-                      (bc_v8i16 (v4i32 (scalar_to_vector (i32 imm:$src2))))))]>;
-def PSRADri : PDIi8<0x72, MRM4r, (outs VR128:$dst),
-                                 (ins VR128:$src1, i32i8imm:$src2),
-                    "psrad\t{$src2, $dst|$dst, $src2}",
-                    [(set VR128:$dst, (int_x86_sse2_psra_d VR128:$src1,
-                          (scalar_to_vector (i32 imm:$src2))))]>;
-}
-
-// PSRAQ doesn't exist in SSE[1-3].
+defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
+                               int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
+defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
+                               int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
+defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
+                               int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
+
+defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
+                               int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
+defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
+                               int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
+defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x72, MRM2r, "psrlq",
+                               int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
+
+defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
+                               int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
+defm PSRAD : PDI_binop_rmi_int<0xE2, 0x71, MRM4r, "psrad",
+                               int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
 
 // 128-bit logical shifts.
 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {

diff --git a/lib/VMCore/AutoUpgrade.cpp b/lib/VMCore/AutoUpgrade.cpp
index 2fb0e8079ae32f2ffee44c50fe9293accf49bb0b..7faff565096c494ab904055368e2c3e9691e5198 100644 (file)
--- a/lib/VMCore/AutoUpgrade.cpp
+++ b/lib/VMCore/AutoUpgrade.cpp
@@ -122,7 +122,7 @@ static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
     if (Name.compare(5,10,"x86.mmx.ps",10) == 0 &&
         (Name.compare(13,4,"psll", 4) == 0 ||
          Name.compare(13,4,"psra", 4) == 0 ||
-         Name.compare(13,4,"psrl", 4) == 0)) {
+         Name.compare(13,4,"psrl", 4) == 0) && Name[17] != 'i') {
       
       const llvm::Type *VT = VectorType::get(IntegerType::get(64), 1);
       

diff --git a/test/CodeGen/X86/mmx-shift.ll b/test/CodeGen/X86/mmx-shift.ll
index d68af2d1d334a971119738d91553d31acd7a2950..82eeafd07527e3cfcf46eb180f0179bd7c58f28f 100644 (file)
--- a/test/CodeGen/X86/mmx-shift.ll
+++ b/test/CodeGen/X86/mmx-shift.ll
@@ -1,14 +1,15 @@
 ; RUN: llvm-as < %s | llc -march=x86 -mattr=+mmx | grep psllq | grep 32
+; RUN: llvm-as < %s | llc -march=x86-64 -mattr=+mmx | grep psllq | grep 32
 ; RUN: llvm-as < %s | llc -march=x86 -mattr=+mmx | grep psrad
 
 define i64 @t1(<1 x i64> %mm1) nounwind  {
 entry:
-       %tmp6 = tail call <1 x i64> @llvm.x86.mmx.psll.q( <1 x i64> %mm1, <1 x i64> <i64 32> )          ; <<1 x i64>> [#uses=1]
+       %tmp6 = tail call <1 x i64> @llvm.x86.mmx.pslli.q( <1 x i64> %mm1, i32 32 )             ; <<1 x i64>> [#uses=1]
        %retval1112 = bitcast <1 x i64> %tmp6 to i64            ; <i64> [#uses=1]
        ret i64 %retval1112
 }
 
-declare <1 x i64> @llvm.x86.mmx.psll.q(<1 x i64>, <1 x i64>) nounwind readnone 
+declare <1 x i64> @llvm.x86.mmx.pslli.q(<1 x i64>, i32) nounwind readnone 
 
 define i64 @t2(<2 x i32> %mm1, <2 x i32> %mm2) nounwind  {
 entry: