[Turbofan] Add ARM NEON instructions for implementing SIMD.

test/cctest/test-assembler-arm.cc

 // Copyright 2012 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 1237 matching lines @@
     uint32_t srcA0;
     uint32_t srcA1;
     uint32_t dstA0;
     uint32_t dstA1;
     uint32_t dstA2;
     uint32_t dstA3;
     uint32_t dstA4;
     uint32_t dstA5;
     uint32_t dstA6;
     uint32_t dstA7;
-    uint32_t vmov_src[4], vmov_dst[4];
-    uint32_t veor_src[4], veor_dst[4];
+    uint32_t vmov_src[4], vmov_dst[4], vmvn[4];
+    int32_t vcvt_s32_f32[4];
+    uint32_t vcvt_u32_f32[4];
+    float vcvt_f32_s32[4], vcvt_f32_u32[4];
+    uint32_t vdup1[4], vdup2[4], vdup3[4], vdup4[4];
+    uint32_t veor[4], vadd[4], vsub[4];
+    uint32_t vtst[4], vceq[4], vbsl[4], vtbl[2], vtbx[2];
+    float vaddf[4], vsubf[4];
   } T;
   T t;
 
   // Create a function that accepts &t, and loads, manipulates, and stores
   // the doubles and floats.
   Assembler assm(isolate, NULL, 0);
 
 
   if (CpuFeatures::IsSupported(NEON)) {
     CpuFeatureScope scope(&assm, NEON);
 
-    __ stm(db_w, sp, r4.bit() | lr.bit());
+    __ stm(db_w, sp, r4.bit() | r5.bit() | lr.bit());
     // Move 32 bytes with neon.
     __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, src0))));
     __ vld1(Neon8, NeonListOperand(d0, 4), NeonMemOperand(r4));
     __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, dst0))));
     __ vst1(Neon8, NeonListOperand(d0, 4), NeonMemOperand(r4));
 
     // Expand 8 bytes into 8 words(16 bits).
     __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, srcA0))));
     __ vld1(Neon8, NeonListOperand(d0), NeonMemOperand(r4));
     __ vmovl(NeonU8, q0, d0);
     __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, dstA0))));
     __ vst1(Neon8, NeonListOperand(d0, 2), NeonMemOperand(r4));
 
     // The same expansion, but with different source and destination registers.
     __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, srcA0))));
     __ vld1(Neon8, NeonListOperand(d1), NeonMemOperand(r4));
     __ vmovl(NeonU8, q1, d1);
     __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, dstA4))));
     __ vst1(Neon8, NeonListOperand(d2, 2), NeonMemOperand(r4));
 
-    // Test vmov for q-registers.
+    // vmov for q-registers.
     __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vmov_src))));
-    __ vld1(Neon8, NeonListOperand(d0, 2), NeonMemOperand(r4));
+    __ vld1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
     __ vmov(q1, q0);
     __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vmov_dst))));
-    __ vst1(Neon8, NeonListOperand(d2, 2), NeonMemOperand(r4));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
 
-    // Test veor for q-registers.
-    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, veor_src))));
-    __ vld1(Neon8, NeonListOperand(d0, 2), NeonMemOperand(r4));
-    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, veor_dst))));
-    __ vld1(Neon8, NeonListOperand(d2, 2), NeonMemOperand(r4));
+    // vmvn.
+    __ mov(r4, Operand(0xFF));
+    __ vdup(q0, r4, Neon16);
+    __ vmvn(q1, q0);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vmvn))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // vcvt for q-registers.
+    __ vmov(s0, -1.5);
+    __ vmov(s1, -1);
+    __ vmov(s2, 1);
+    __ vmov(s3, 1.5);
+    __ vcvt_s32_f32(q1, q0);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vcvt_s32_f32))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+    __ vcvt_u32_f32(q1, q0);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vcvt_u32_f32))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+    __ mov(r4, Operand(-1));
+    __ mov(r5, Operand(1));
+    __ vmov(d0, r4, r5);
+    __ vmov(d1, d0);  // q0 = [-1, 1, -1, 1]
+    __ vcvt_f32_s32(q1, q0);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vcvt_f32_s32))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+    __ vcvt_f32_u32(q1, q0);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vcvt_f32_u32))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // int vdup.
+    __ mov(r4, Operand(0xa));
+    __ vdup(q0, r4, Neon8);
+    __ vdup(q1, r4, Neon16);
+    __ vdup(q2, r4, Neon32);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vdup1))));
+    __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vdup2))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vdup3))));
+    __ vst1(Neon8, NeonListOperand(q2), NeonMemOperand(r4));
+    // float vdup.
+    __ vmov(s0, -1.0);
+    __ vdup(q0, s0);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vdup4))));
+    __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+    // veor.
+    __ mov(r4, Operand(0x00aa));
+    __ vdup(q0, r4, Neon16);
+    __ mov(r4, Operand(0x0055));
+    __ vdup(q1, r4, Neon16);
     __ veor(q1, q1, q0);
-    __ vst1(Neon8, NeonListOperand(d2, 2), NeonMemOperand(r4));
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, veor))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // vadd(integer).
+    __ mov(r4, Operand(0x8001));
+    __ vdup(q0, r4, Neon16);
+    __ mov(r4, Operand(0x8002));
+    __ vdup(q1, r4, Neon16);
+    __ vadd(q1, q1, q0, Neon16);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vadd))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // vadd(float).
+    __ vmov(s4, 1.0);
+    __ vdup(q0, s4);
+    __ vdup(q1, s4);
+    __ vadd(q1, q1, q0);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vaddf))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // vsub(integer).

[Rodolph Perfetta (ARM), 2016/12/08 18:08:28]

It would be nice to have test for each NeonSize for the instruction below.

[bbudge, 2016/12/10 21:33:04]

Done. Here and for similar instructions.

+    __ mov(r4, Operand(0x0001));
+    __ vdup(q0, r4, Neon16);
+    __ mov(r4, Operand(0x0002));
+    __ vdup(q1, r4, Neon16);
+    __ vsub(q1, q0, q1, Neon16);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vsub))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // vsub(float).
+    __ vmov(s4, 2.0);
+    __ vdup(q0, s4);
+    __ vmov(s4, 1.0);
+    __ vdup(q1, s4);
+    __ vsub(q1, q1, q0);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vsubf))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // vceq.
+    __ mov(r4, Operand(0x03));
+    __ vdup(q0, r4, Neon8);
+    __ mov(r4, Operand(0x03));
+    __ vdup(q1, r4, Neon16);
+    __ vceq(q1, q0, q1, Neon8);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vceq))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // vtst.
+    __ mov(r4, Operand(0x03));
+    __ vdup(q0, r4, Neon8);
+    __ mov(r4, Operand(0x02));
+    __ vdup(q1, r4, Neon16);
+    __ vtst(q1, q0, q1, Neon8);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vtst))));
+    __ vst1(Neon8, NeonListOperand(q1), NeonMemOperand(r4));
+
+    // vbsl.
+    __ mov(r4, Operand(0x00ff));
+    __ vdup(q0, r4, Neon16);
+    __ mov(r4, Operand(0x01));
+    __ vdup(q1, r4, Neon8);
+    __ mov(r4, Operand(0x02));
+    __ vdup(q2, r4, Neon8);
+    __ vbsl(q0, q1, q2);
+    __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, vbsl))));
+    __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+    // vtb[l/x].
+    __ mov(r4, Operand(0x06040200));
+    __ mov(r5, Operand(0xff050301));
+    __ vmov(d2, r4, r5);  // d2 = ff05030106040200
+    __ vtbl(d0, NeonListOperand(d2, 1), d2);
+    __ vstr(d0, r0, offsetof(T, vtbl));
+    __ vtbx(d2, NeonListOperand(d2, 1), d2);
+    __ vstr(d2, r0, offsetof(T, vtbx));
 
     // Restore and return.
-    __ ldm(ia_w, sp, r4.bit() | pc.bit());
+    __ ldm(ia_w, sp, r4.bit() | r5.bit() | pc.bit());
 
     CodeDesc desc;
     assm.GetCode(&desc);
     Handle<Code> code = isolate->factory()->NewCode(
         desc, Code::ComputeFlags(Code::STUB), Handle<Code>());
 #ifdef DEBUG
     OFStream os(stdout);
     code->Print(os);
 #endif
     F3 f = FUNCTION_CAST<F3>(code->entry());
@@ skipping 18 matching lines @@
     t.dstA0 = 0;
     t.dstA1 = 0;
     t.dstA2 = 0;
     t.dstA3 = 0;
     t.dstA4 = 0;
     t.dstA5 = 0;
     t.dstA6 = 0;
     t.dstA7 = 0;
     t.vmov_src[0] = t.vmov_src[1] = t.vmov_src[2] = t.vmov_src[3] = 1;
     t.vmov_dst[0] = t.vmov_dst[1] = t.vmov_dst[2] = t.vmov_dst[3] = 0;
-    t.veor_src[0] = t.veor_src[1] = t.veor_src[2] = t.veor_src[3] = 0xAA;
-    t.veor_dst[0] = t.veor_dst[1] = t.veor_dst[2] = t.veor_dst[3] = 0x55;
     Object* dummy = CALL_GENERATED_CODE(isolate, f, &t, 0, 0, 0, 0);
     USE(dummy);
     CHECK_EQ(0x01020304u, t.dst0);
     CHECK_EQ(0x11121314u, t.dst1);
     CHECK_EQ(0x21222324u, t.dst2);
     CHECK_EQ(0x31323334u, t.dst3);
     CHECK_EQ(0x41424344u, t.dst4);
     CHECK_EQ(0x51525354u, t.dst5);
     CHECK_EQ(0x61626364u, t.dst6);
     CHECK_EQ(0x71727374u, t.dst7);
     CHECK_EQ(0x00430044u, t.dstA0);
     CHECK_EQ(0x00410042u, t.dstA1);
     CHECK_EQ(0x00830084u, t.dstA2);
     CHECK_EQ(0x00810082u, t.dstA3);
     CHECK_EQ(0x00430044u, t.dstA4);
     CHECK_EQ(0x00410042u, t.dstA5);
     CHECK_EQ(0x00830084u, t.dstA6);
     CHECK_EQ(0x00810082u, t.dstA7);
     CHECK_EQ(1u, t.vmov_dst[0]);
     CHECK_EQ(1u, t.vmov_dst[1]);
     CHECK_EQ(1u, t.vmov_dst[2]);
     CHECK_EQ(1u, t.vmov_dst[3]);
-    CHECK_EQ(0xFFu, t.veor_dst[0]);
-    CHECK_EQ(0xFFu, t.veor_dst[1]);
-    CHECK_EQ(0xFFu, t.veor_dst[2]);
-    CHECK_EQ(0xFFu, t.veor_dst[3]);
+    CHECK_EQ(-1, t.vcvt_s32_f32[0]);
+    CHECK_EQ(-1, t.vcvt_s32_f32[1]);
+    CHECK_EQ(1, t.vcvt_s32_f32[2]);
+    CHECK_EQ(1, t.vcvt_s32_f32[3]);
+    CHECK_EQ(0xFFFFFFFFu, t.vcvt_u32_f32[0]);
+    CHECK_EQ(0xFFFFFFFFu, t.vcvt_u32_f32[1]);
+    CHECK_EQ(1u, t.vcvt_u32_f32[2]);
+    CHECK_EQ(1u, t.vcvt_u32_f32[3]);
+    CHECK_EQ(-1.f, t.vcvt_f32_s32[0]);
+    CHECK_EQ(1.f, t.vcvt_f32_s32[1]);
+    CHECK_EQ(-1.f, t.vcvt_f32_s32[2]);
+    CHECK_EQ(1.f, t.vcvt_f32_s32[3]);
+    CHECK_EQ(static_cast<float>(0xFFFFFFFFu), t.vcvt_f32_u32[0]);
+    CHECK_EQ(1.f, t.vcvt_f32_u32[1]);
+    CHECK_EQ(static_cast<float>(0xFFFFFFFFu), t.vcvt_f32_u32[2]);
+    CHECK_EQ(1.f, t.vcvt_f32_u32[3]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0xFF00FF00, t.vmvn[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0x0a0a0a0au, t.vdup1[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0x000a000au, t.vdup2[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0x0000000au, t.vdup3[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0xbf800000u, t.vdup4[i]);  // -1.0f
+    for (int i = 0; i < 4; i++) CHECK_EQ(0x00ff00ffu, t.veor[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(2.0, t.vaddf[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0x00030003u, t.vadd[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(-1.0, t.vsubf[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0xffffffffu, t.vsub[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0x00ff00ffu, t.vceq[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0x00ff00ffu, t.vtst[i]);
+    for (int i = 0; i < 4; i++) CHECK_EQ(0x02010201u, t.vbsl[i]);
+    CHECK_EQ(0x05010400u, t.vtbl[0]);
+    CHECK_EQ(0x00030602u, t.vtbl[1]);
+    CHECK_EQ(0x05010400u, t.vtbx[0]);
+    CHECK_EQ(0xff030602u, t.vtbx[1]);
   }
 }
 
 
 TEST(16) {
   // Test the pkh, uxtb, uxtab and uxtb16 instructions.
   CcTest::InitializeVM();
   Isolate* isolate = CcTest::i_isolate();
   HandleScope scope(isolate);
 
@@ skipping 1571 matching lines @@
   const uint32_t test_2 = 0x89abcdef;
   __ mov(r4, Operand(test_1));
   __ mov(r5, Operand(test_2));
   // TODO(bbudge) replace with vdup when implemented.
   __ vmov(d8, r4, r4);
   __ vmov(d9, r4, r4);  // q4 = [1.0, 1.0]
   __ vmov(d10, r5, r5);
   __ vmov(d11, r5, r5);  // q5 = [-1.0, -1.0]
   __ vswp(q4, q5);
   __ add(r6, r0, Operand(static_cast<int32_t>(offsetof(T, vswp_q4))));
-  __ vst1(Neon8, NeonListOperand(d8, 2), NeonMemOperand(r6));
+  __ vst1(Neon8, NeonListOperand(q4), NeonMemOperand(r6));
   __ add(r6, r0, Operand(static_cast<int32_t>(offsetof(T, vswp_q5))));
-  __ vst1(Neon8, NeonListOperand(d10, 2), NeonMemOperand(r6));
+  __ vst1(Neon8, NeonListOperand(q5), NeonMemOperand(r6));
 
   __ ldm(ia_w, sp, r4.bit() | r5.bit() | r6.bit() | r7.bit() | pc.bit());
   __ bx(lr);
 
   CodeDesc desc;
   assm.GetCode(&desc);
   Handle<Code> code = isolate->factory()->NewCode(
       desc, Code::ComputeFlags(Code::STUB), Handle<Code>());
 #ifdef DEBUG
   OFStream os(stdout);
@@ skipping 69 matching lines @@
   HandleScope scope(isolate);
 
   Assembler assm(isolate, NULL, 0);
   __ mov(r0, Operand(isolate->factory()->infinity_value()));
   __ BlockConstPoolFor(1019);
   for (int i = 0; i < 1019; ++i) __ nop();
   __ vldr(d0, MemOperand(r0, 0));
 }
 
 #undef __