MIPS[64]: Support for some SIMD operations (4)

src/compiler/mips/code-generator-mips.cc

Index: src/compiler/mips/code-generator-mips.cc
diff --git a/src/compiler/mips/code-generator-mips.cc b/src/compiler/mips/code-generator-mips.cc
index c2b444aee09f9f7912d4aa6ffbdca3b20aab800c..2a46c3d8c1796eb67b0076abc2c79babfba601e1 100644
--- a/src/compiler/mips/code-generator-mips.cc
+++ b/src/compiler/mips/code-generator-mips.cc
@@ -1602,6 +1602,113 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
     case kAtomicCompareExchangeWord32:
       UNREACHABLE();
       break;
+    case kMipsFloat32x4Abs: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ bclri_w(i.OutputSimd128Register(), i.InputSimd128Register(0), 31);
+      break;
+    }
+    case kMipsFloat32x4Neg: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ bnegi_w(i.OutputSimd128Register(), i.InputSimd128Register(0), 31);
+      break;
+    }
+    case kMipsFloat32x4RecipApprox: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ frcp_w(i.OutputSimd128Register(), i.InputSimd128Register(0));
+      break;
+    }
+    case kMipsFloat32x4RecipRefine: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      Simd128Register dst = i.OutputSimd128Register();
+      // Emulate with 2.0f - a * b
+      __ ldi_w(kSimd128ScratchReg, 2);
+      __ ffint_u_w(kSimd128ScratchReg, kSimd128ScratchReg);
+      __ fmul_w(dst, i.InputSimd128Register(0), i.InputSimd128Register(1));
+      __ fsub_w(dst, kSimd128ScratchReg, dst);

[bbudge, 2017/03/28 17:21:40]

I didn't realize until now that only ARM has refinement instructions, and they
have to be emulated on other platforms as you do here. That makes me think we
should remove these instructions, since these sequences are better left to the
programmer, who can schedule them to reduce stalling.

This is fine for now, just a heads up. What do you think about eliminating the
refine instructions?

[dusan.simicic, 2017/03/29 12:19:33]

From MIPS perspective, it is better not to implement those refinement
operations, as we don't have corresponding MSA instructions. Also, I didn't find
anything about Float32x4RecipRefine and Float32x4RecipSqrtRefine requirements on
the following pages: 

https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Obje...

tc39.github.io/ecmascript_simd/

https://github.com/stoklund/portable-simd/blob/master/webassembly-mapping.md

Are you familiar that refinement operations are mandatory by some other
document?
 
Maybe it's better to open an issue on WASM/design github page and involve other
people in discussion.

+      break;
+    }
+    case kMipsFloat32x4RecipSqrtApprox: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ frsqrt_w(i.OutputSimd128Register(), i.InputSimd128Register(0));
+      break;
+    }
+    case kMipsFloat32x4RecipSqrtRefine: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      Simd128Register dst = i.OutputSimd128Register();
+      // Emulate with (3.0f - a * b) * 0.5f;
+      __ ldi_w(kSimd128ScratchReg, 3);
+      __ ffint_u_w(kSimd128ScratchReg, kSimd128ScratchReg);
+      __ fmul_w(dst, i.InputSimd128Register(0), i.InputSimd128Register(1));
+      __ fsub_w(dst, kSimd128ScratchReg, dst);
+      __ ldi_w(kSimd128ScratchReg, 0x3f);
+      __ slli_w(kSimd128ScratchReg, kSimd128ScratchReg, 24);
+      __ fmul_w(dst, dst, kSimd128ScratchReg);

[bbudge, 2017/03/28 17:21:40]

Same comment here, and for MIPS64.

+      break;
+    }
+    case kMipsFloat32x4Add: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fadd_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsFloat32x4Sub: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fsub_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsFloat32x4Mul: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fmul_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsFloat32x4Max: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fmax_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsFloat32x4Min: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fmin_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsFloat32xEqual: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fceq_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsFloat32x4NotEqual: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fcne_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsFloat32xLessThan: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fclt_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsFloat32xLessThanOrEqual: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ fcle_w(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                i.InputSimd128Register(1));
+      break;
+    }
+    case kMipsInt32x4FromFloat32x4: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ ftrunc_s_w(i.OutputSimd128Register(), i.InputSimd128Register(0));
+      break;
+    }
+    case kMipsUint32x4FromFloat32x4: {
+      CpuFeatureScope msa_scope(masm(), MIPS_SIMD);
+      __ ftrunc_u_w(i.OutputSimd128Register(), i.InputSimd128Register(0));
+      break;
+    }
   }
   return kSuccess;
 }  // NOLINT(readability/fn_size)