From 5436cde0fc857e4d4c1b58432f48ab8c9ae33bbd Mon Sep 17 00:00:00 2001
From: Ben Vanik <ben.vanik@gmail.com>
Date: Tue, 27 May 2014 14:27:07 -0700
Subject: [PATCH] Saturating unsigned VECTOR_ADD.

---
 src/alloy/backend/x64/x64_emitter.cc   |  1 +
 src/alloy/backend/x64/x64_emitter.h    |  1 +
 src/alloy/backend/x64/x64_sequences.cc | 36 +++++++++++++++++++++++++-
 3 files changed, 37 insertions(+), 1 deletion(-)

diff --git a/src/alloy/backend/x64/x64_emitter.cc b/src/alloy/backend/x64/x64_emitter.cc
index 6e94a660a..684fcaa86 100644
--- a/src/alloy/backend/x64/x64_emitter.cc
+++ b/src/alloy/backend/x64/x64_emitter.cc
@@ -444,6 +444,7 @@ Address X64Emitter::GetXmmConstPtr(XmmConst id) {
     /* XMMOneOver255          */ vec128f(1.0f / 255.0f, 1.0f / 255.0f, 1.0f / 255.0f, 1.0f / 255.0f),
     /* XMMShiftMaskPS         */ vec128i(0x0000001Fu, 0x0000001Fu, 0x0000001Fu, 0x0000001Fu),
     /* XMMShiftByteMask       */ vec128i(0x000000FFu, 0x000000FFu, 0x000000FFu, 0x000000FFu),
+    /* XMMUnsignedDwordMax    */ vec128i(0xFFFFFFFFu, 0x00000000u, 0xFFFFFFFFu, 0x00000000u),
   };
   // TODO(benvanik): cache base pointer somewhere? stack? It'd be nice to
   // prevent this move.
diff --git a/src/alloy/backend/x64/x64_emitter.h b/src/alloy/backend/x64/x64_emitter.h
index d67348e18..3ac92be3f 100644
--- a/src/alloy/backend/x64/x64_emitter.h
+++ b/src/alloy/backend/x64/x64_emitter.h
@@ -52,6 +52,7 @@ enum XmmConst {
   XMMOneOver255         = 13,
   XMMShiftMaskPS        = 14,
   XMMShiftByteMask      = 15,
+  XMMUnsignedDwordMax   = 16,
 };
 
 // Unfortunately due to the design of xbyak we have to pass this to the ctor.
diff --git a/src/alloy/backend/x64/x64_sequences.cc b/src/alloy/backend/x64/x64_sequences.cc
index 6e6d85da9..9ece83e67 100644
--- a/src/alloy/backend/x64/x64_sequences.cc
+++ b/src/alloy/backend/x64/x64_sequences.cc
@@ -2566,7 +2566,41 @@ EMITTER(VECTOR_ADD, MATCH(I<OPCODE_VECTOR_ADD, V128<>, V128<>, V128<>>)) {
             }
             break;
           case INT32_TYPE:
-            XEASSERTALWAYS();
+            if (saturate) {
+              if (is_unsigned) {
+                // We reuse all these temps...
+                XEASSERT(src1 != e.xmm0 && src1 != e.xmm1 && src1 != e.xmm2);
+                XEASSERT(src2 != e.xmm0 && src2 != e.xmm1 && src2 != e.xmm2);
+                // Clamp to 0xFFFFFFFF.
+                // Wish there was a vpaddusd...
+                // | A | B | C | D |
+                // |     B |     D |
+                e.db(0xCC);
+                e.vpsllq(e.xmm0, src1, 32);
+                e.vpsllq(e.xmm1, src2, 32);
+                e.vpsrlq(e.xmm0, 32);
+                e.vpsrlq(e.xmm1, 32);
+                e.vpaddq(e.xmm0, e.xmm1);
+                e.vpcmpgtq(e.xmm0, e.GetXmmConstPtr(XMMUnsignedDwordMax));
+                e.vpsllq(e.xmm0, 32);
+                e.vpsrlq(e.xmm0, 32);
+                // |     A |     C |
+                e.vpsrlq(e.xmm1, src1, 32);
+                e.vpsrlq(e.xmm2, src2, 32);
+                e.vpaddq(e.xmm1, e.xmm2);
+                e.vpcmpgtq(e.xmm1, e.GetXmmConstPtr(XMMUnsignedDwordMax));
+                e.vpsllq(e.xmm1, 32);
+                // xmm0 = mask for with saturated dwords == 111...
+                e.vpor(e.xmm0, e.xmm1);
+                e.vpaddd(dest, src1, src2);
+                // dest.f[n] = xmm1.f[n] ? xmm1.f[n] : dest.f[n];
+                e.vblendvps(dest, dest, e.xmm1, e.xmm1);
+              } else {
+                XEASSERTALWAYS();
+              }
+            } else {
+              e.vpaddd(dest, src1, src2);
+            }
             break;
           case FLOAT32_TYPE:
             e.vaddps(dest, src1, src2);