[X86] Allow legacy vcvtps2ph intrinsics to select EVEX encoded instructions. Rely on EVEX->VEX to convert back.

topperc · topperc · commit c30df5f3d389 · 2017-11-08T04:00:30.000Z
Missed store folding opportunities will be fixed in a subsequent commit. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@317661 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
@@ -7698,11 +7698,11 @@ multiclass f16c_ph2ps<RegisterClass RC, X86MemOperand x86memop> {
              T8PD, VEX, Sched<[WriteCvtF2FLd]>;
 }
 
-multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
+multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop> {
   def rr : Ii8<0x1D, MRMDestReg, (outs VR128:$dst),
                (ins RC:$src1, i32u8imm:$src2),
                "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}",
-               [(set VR128:$dst, (Int RC:$src1, imm:$src2))]>,
+               [(set VR128:$dst, (X86cvtps2ph RC:$src1, imm:$src2))]>,
                TAPD, VEX, Sched<[WriteCvtF2F]>;
   let hasSideEffects = 0, mayStore = 1,
       SchedRW = [WriteCvtF2FLd, WriteRMW] in
@@ -7715,6 +7715,8 @@ multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop, Intrinsic Int> {
 let Predicates = [HasF16C, NoVLX] in {
   defm VCVTPH2PS  : f16c_ph2ps<VR128, f64mem>;
   defm VCVTPH2PSY : f16c_ph2ps<VR256, f128mem>, VEX_L;
+  defm VCVTPS2PH  : f16c_ps2ph<VR128, f64mem>;
+  defm VCVTPS2PHY : f16c_ps2ph<VR256, f128mem>, VEX_L;
 
   // Pattern match vcvtph2ps of a scalar i64 load.
   def : Pat<(v4f32 (X86cvtph2ps (v8i16 (vzmovl_v2i64 addr:$src)))),
@@ -7724,23 +7726,17 @@ let Predicates = [HasF16C, NoVLX] in {
   def : Pat<(v4f32 (X86cvtph2ps (v8i16 (bitconvert
               (v2i64 (scalar_to_vector (loadi64 addr:$src))))))),
             (VCVTPH2PSrm addr:$src)>;
-}
-
-let Predicates = [HasF16C] in {
-  defm VCVTPS2PH  : f16c_ps2ph<VR128, f64mem, int_x86_vcvtps2ph_128>;
-  defm VCVTPS2PHY : f16c_ps2ph<VR256, f128mem, int_x86_vcvtps2ph_256>, VEX_L;
 
-  def : Pat<(store (f64 (extractelt (bc_v2f64 (v8i16
-                  (int_x86_vcvtps2ph_128 VR128:$src1, i32:$src2))), (iPTR 0))),
-                   addr:$dst),
-                   (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
-  def : Pat<(store (i64 (extractelt (bc_v2i64 (v8i16
-                  (int_x86_vcvtps2ph_128 VR128:$src1, i32:$src2))), (iPTR 0))),
-                   addr:$dst),
-                   (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
-  def : Pat<(store (v8i16 (int_x86_vcvtps2ph_256 VR256:$src1, i32:$src2)),
-                   addr:$dst),
-                   (VCVTPS2PHYmr addr:$dst, VR256:$src1, imm:$src2)>;
+  def : Pat<(store (f64 (extractelt
+                         (bc_v2f64 (v8i16 (X86cvtps2ph VR128:$src1, i32:$src2))),
+                         (iPTR 0))), addr:$dst),
+            (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
+  def : Pat<(store (i64 (extractelt
+                         (bc_v2i64 (v8i16 (X86cvtps2ph VR128:$src1, i32:$src2))),
+                         (iPTR 0))), addr:$dst),
+            (VCVTPS2PHmr addr:$dst, VR128:$src1, imm:$src2)>;
+  def : Pat<(store (v8i16 (X86cvtps2ph VR256:$src1, i32:$src2)), addr:$dst),
+            (VCVTPS2PHYmr addr:$dst, VR256:$src1, imm:$src2)>;
 }
 
 // Patterns for  matching conversions from float to half-float and vice versa.
diff --git a/lib/Target/X86/X86IntrinsicsInfo.h b/lib/Target/X86/X86IntrinsicsInfo.h
@@ -1588,6 +1588,8 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(ssse3_pshuf_b_128, INTR_TYPE_2OP, X86ISD::PSHUFB, 0),
   X86_INTRINSIC_DATA(vcvtph2ps_128,     INTR_TYPE_1OP, X86ISD::CVTPH2PS, 0),
   X86_INTRINSIC_DATA(vcvtph2ps_256,     INTR_TYPE_1OP, X86ISD::CVTPH2PS, 0),
+  X86_INTRINSIC_DATA(vcvtps2ph_128,     INTR_TYPE_2OP, X86ISD::CVTPS2PH, 0),
+  X86_INTRINSIC_DATA(vcvtps2ph_256,     INTR_TYPE_2OP, X86ISD::CVTPS2PH, 0),
   X86_INTRINSIC_DATA(xop_vpcomb,        INTR_TYPE_3OP, X86ISD::VPCOM, 0),
   X86_INTRINSIC_DATA(xop_vpcomd,        INTR_TYPE_3OP, X86ISD::VPCOM, 0),
   X86_INTRINSIC_DATA(xop_vpcomq,        INTR_TYPE_3OP, X86ISD::VPCOM, 0),
diff --git a/test/CodeGen/X86/f16c-intrinsics.ll b/test/CodeGen/X86/f16c-intrinsics.ll
@@ -121,12 +121,12 @@ define <8 x i16> @test_x86_vcvtps2ph_128(<4 x float> %a0) {
 ;
 ; X32-AVX512VL-LABEL: test_x86_vcvtps2ph_128:
 ; X32-AVX512VL:       # BB#0:
-; X32-AVX512VL-NEXT:    vcvtps2ph $0, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x00]
+; X32-AVX512VL-NEXT:    vcvtps2ph $0, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x00]
 ; X32-AVX512VL-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_vcvtps2ph_128:
 ; X64-AVX512VL:       # BB#0:
-; X64-AVX512VL-NEXT:    vcvtps2ph $0, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x00]
+; X64-AVX512VL-NEXT:    vcvtps2ph $0, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x00]
 ; X64-AVX512VL-NEXT:    retq # encoding: [0xc3]
   %res = call <8 x i16> @llvm.x86.vcvtps2ph.128(<4 x float> %a0, i32 0) ; <<8 x i16>> [#uses=1]
   ret <8 x i16> %res
@@ -148,13 +148,13 @@ define <8 x i16> @test_x86_vcvtps2ph_256(<8 x float> %a0) {
 ;
 ; X32-AVX512VL-LABEL: test_x86_vcvtps2ph_256:
 ; X32-AVX512VL:       # BB#0:
-; X32-AVX512VL-NEXT:    vcvtps2ph $0, %ymm0, %xmm0 # encoding: [0xc4,0xe3,0x7d,0x1d,0xc0,0x00]
+; X32-AVX512VL-NEXT:    vcvtps2ph $0, %ymm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x1d,0xc0,0x00]
 ; X32-AVX512VL-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
 ; X32-AVX512VL-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_vcvtps2ph_256:
 ; X64-AVX512VL:       # BB#0:
-; X64-AVX512VL-NEXT:    vcvtps2ph $0, %ymm0, %xmm0 # encoding: [0xc4,0xe3,0x7d,0x1d,0xc0,0x00]
+; X64-AVX512VL-NEXT:    vcvtps2ph $0, %ymm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x1d,0xc0,0x00]
 ; X64-AVX512VL-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
 ; X64-AVX512VL-NEXT:    retq # encoding: [0xc3]
   %res = call <8 x i16> @llvm.x86.vcvtps2ph.256(<8 x float> %a0, i32 0) ; <<8 x i16>> [#uses=1]
@@ -238,13 +238,13 @@ define void @test_x86_vcvtps2ph_256_m(<8 x i16>* nocapture %d, <8 x float> %a) n
 ; X32-AVX512VL-LABEL: test_x86_vcvtps2ph_256_m:
 ; X32-AVX512VL:       # BB#0: # %entry
 ; X32-AVX512VL-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
-; X32-AVX512VL-NEXT:    vcvtps2ph $3, %ymm0, (%eax) # encoding: [0xc4,0xe3,0x7d,0x1d,0x00,0x03]
+; X32-AVX512VL-NEXT:    vcvtps2ph $3, %ymm0, (%eax) # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x1d,0x00,0x03]
 ; X32-AVX512VL-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
 ; X32-AVX512VL-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_vcvtps2ph_256_m:
 ; X64-AVX512VL:       # BB#0: # %entry
-; X64-AVX512VL-NEXT:    vcvtps2ph $3, %ymm0, (%rdi) # encoding: [0xc4,0xe3,0x7d,0x1d,0x07,0x03]
+; X64-AVX512VL-NEXT:    vcvtps2ph $3, %ymm0, (%rdi) # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x1d,0x07,0x03]
 ; X64-AVX512VL-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
 ; X64-AVX512VL-NEXT:    retq # encoding: [0xc3]
 entry:
@@ -268,15 +268,15 @@ define void @test_x86_vcvtps2ph_128_m(<4 x i16>* nocapture %d, <4 x float> %a) n
 ; X32-AVX512VL-LABEL: test_x86_vcvtps2ph_128_m:
 ; X32-AVX512VL:       # BB#0: # %entry
 ; X32-AVX512VL-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
-; X32-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x03]
+; X32-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x03]
 ; X32-AVX512VL-NEXT:    vpmovzxwd %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x33,0xc0]
 ; X32-AVX512VL-NEXT:    # xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
 ; X32-AVX512VL-NEXT:    vpmovdw %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x33,0x00]
 ; X32-AVX512VL-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_vcvtps2ph_128_m:
 ; X64-AVX512VL:       # BB#0: # %entry
-; X64-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x03]
+; X64-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x03]
 ; X64-AVX512VL-NEXT:    vpmovzxwd %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x33,0xc0]
 ; X64-AVX512VL-NEXT:    # xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
 ; X64-AVX512VL-NEXT:    vpmovdw %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x33,0x07]
@@ -303,12 +303,14 @@ define void @test_x86_vcvtps2ph_128_m2(double* nocapture %hf4x16, <4 x float> %f
 ; X32-AVX512VL-LABEL: test_x86_vcvtps2ph_128_m2:
 ; X32-AVX512VL:       # BB#0: # %entry
 ; X32-AVX512VL-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
-; X32-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, (%eax) # encoding: [0xc4,0xe3,0x79,0x1d,0x00,0x03]
+; X32-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x03]
+; X32-AVX512VL-NEXT:    vmovlps %xmm0, (%eax) # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x13,0x00]
 ; X32-AVX512VL-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_vcvtps2ph_128_m2:
 ; X64-AVX512VL:       # BB#0: # %entry
-; X64-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, (%rdi) # encoding: [0xc4,0xe3,0x79,0x1d,0x07,0x03]
+; X64-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x03]
+; X64-AVX512VL-NEXT:    vmovlps %xmm0, (%rdi) # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x13,0x07]
 ; X64-AVX512VL-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = tail call <8 x i16> @llvm.x86.vcvtps2ph.128(<4 x float> %f4x32, i32 3)
@@ -333,12 +335,14 @@ define void @test_x86_vcvtps2ph_128_m3(i64* nocapture %hf4x16, <4 x float> %f4x3
 ; X32-AVX512VL-LABEL: test_x86_vcvtps2ph_128_m3:
 ; X32-AVX512VL:       # BB#0: # %entry
 ; X32-AVX512VL-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
-; X32-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, (%eax) # encoding: [0xc4,0xe3,0x79,0x1d,0x00,0x03]
+; X32-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x03]
+; X32-AVX512VL-NEXT:    vmovlps %xmm0, (%eax) # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x13,0x00]
 ; X32-AVX512VL-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-AVX512VL-LABEL: test_x86_vcvtps2ph_128_m3:
 ; X64-AVX512VL:       # BB#0: # %entry
-; X64-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, (%rdi) # encoding: [0xc4,0xe3,0x79,0x1d,0x07,0x03]
+; X64-AVX512VL-NEXT:    vcvtps2ph $3, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc0,0x03]
+; X64-AVX512VL-NEXT:    vmovlps %xmm0, (%rdi) # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x13,0x07]
 ; X64-AVX512VL-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = tail call <8 x i16> @llvm.x86.vcvtps2ph.128(<4 x float> %f4x32, i32 3)
