From f6866098c5b69a96a2de487ed6fee13c58e46855 Mon Sep 17 00:00:00 2001
From: Ben Vanik <ben.vanik@gmail.com>
Date: Sun, 12 Jan 2014 15:41:18 -0800
Subject: [PATCH] Probably correct vupkh|lsb.

---
 src/alloy/backend/ivm/ivm_intcode.cc       | 20 ++++++++
 src/alloy/frontend/ppc/ppc_emit_altivec.cc | 57 +++++++++++++---------
 src/alloy/hir/opcodes.h                    |  2 +
 3 files changed, 56 insertions(+), 23 deletions(-)

diff --git a/src/alloy/backend/ivm/ivm_intcode.cc b/src/alloy/backend/ivm/ivm_intcode.cc
index 5a5c98a5e..4aed60d15 100644
--- a/src/alloy/backend/ivm/ivm_intcode.cc
+++ b/src/alloy/backend/ivm/ivm_intcode.cc
@@ -3790,6 +3790,8 @@ int Translate_PACK(TranslationContext& ctx, Instr* i) {
     IntCode_PACK_FLOAT16_2,
     IntCode_PACK_FLOAT16_4,
     IntCode_PACK_SHORT_2,
+    IntCode_INVALID_TYPE,
+    IntCode_INVALID_TYPE,
   };
   return DispatchToC(ctx, i, fns[i->flags]);
 }
@@ -3844,12 +3846,30 @@ uint32_t IntCode_UNPACK_SHORT_2(IntCodeState& ics, const IntCode* i) {
   dest.f4[3] = 1.0f; // 3?
   return IA_NEXT;
 }
+uint32_t IntCode_UNPACK_S8_IN_16_LO(IntCodeState& ics, const IntCode* i) {
+  const vec128_t& src1 = ics.rf[i->src1_reg].v128;
+  vec128_t& dest = ics.rf[i->dest_reg].v128;
+  for (int n = 0; n < 8; n++) {
+    dest.s8[n] = (int16_t)(int8_t)src1.b16[8 + n];
+  }
+  return IA_NEXT;
+}
+uint32_t IntCode_UNPACK_S8_IN_16_HI(IntCodeState& ics, const IntCode* i) {
+  const vec128_t& src1 = ics.rf[i->src1_reg].v128;
+  vec128_t& dest = ics.rf[i->dest_reg].v128;
+  for (int n = 0; n < 8; n++) {
+    dest.s8[n] = (int16_t)(int8_t)src1.b16[n];
+  }
+  return IA_NEXT;
+}
 int Translate_UNPACK(TranslationContext& ctx, Instr* i) {
   static IntCodeFn fns[] = {
     IntCode_UNPACK_D3DCOLOR,
     IntCode_UNPACK_FLOAT16_2,
     IntCode_UNPACK_FLOAT16_4,
     IntCode_UNPACK_SHORT_2,
+    IntCode_UNPACK_S8_IN_16_LO,
+    IntCode_UNPACK_S8_IN_16_HI,
   };
   return DispatchToC(ctx, i, fns[i->flags]);
 }
diff --git a/src/alloy/frontend/ppc/ppc_emit_altivec.cc b/src/alloy/frontend/ppc/ppc_emit_altivec.cc
index 416005aef..9c6c9efa8 100644
--- a/src/alloy/frontend/ppc/ppc_emit_altivec.cc
+++ b/src/alloy/frontend/ppc/ppc_emit_altivec.cc
@@ -1691,44 +1691,55 @@ XEEMITTER(vpkuwus128,     VX128(5, 960),    VX128  )(PPCHIRBuilder& f, InstrData
   return 1;
 }
 
-
 XEEMITTER(vupkhpx,        0x1000034E, VX  )(PPCHIRBuilder& f, InstrData& i) {
   XEINSTRNOTIMPLEMENTED();
   return 1;
 }
 
-XEEMITTER(vupkhsb,        0x1000020E, VX  )(PPCHIRBuilder& f, InstrData& i) {
-  XEINSTRNOTIMPLEMENTED();
-  return 1;
-}
-XEEMITTER(vupkhsb128,     VX128(6, 896),    VX128  )(PPCHIRBuilder& f, InstrData& i) {
-  XEINSTRNOTIMPLEMENTED();
-  return 1;
-}
-
-XEEMITTER(vupkhsh,        0x1000024E, VX  )(PPCHIRBuilder& f, InstrData& i) {
-  XEINSTRNOTIMPLEMENTED();
-  return 1;
-}
-
 XEEMITTER(vupklpx,        0x100003CE, VX  )(PPCHIRBuilder& f, InstrData& i) {
   XEINSTRNOTIMPLEMENTED();
   return 1;
 }
 
-XEEMITTER(vupklsb,        0x1000028E, VX  )(PPCHIRBuilder& f, InstrData& i) {
-  XEINSTRNOTIMPLEMENTED();
-  return 1;
+int InstrEmit_vupkhsb_(PPCHIRBuilder& f, uint32_t vd, uint32_t vb) {
+  // bytes 0-7 expanded to halfwords 0-8 and sign extended
+  Value* v = f.Unpack(f.LoadVR(vb), PACK_TYPE_S8_IN_16_HI);
+  f.StoreVR(vd, v);
+  return 0;
 }
-XEEMITTER(vupklsb128,     VX128(6, 960),    VX128  )(PPCHIRBuilder& f, InstrData& i) {
-  XEINSTRNOTIMPLEMENTED();
-  return 1;
+XEEMITTER(vupkhsb,        0x1000020E, VX  )(PPCHIRBuilder& f, InstrData& i) {
+  return InstrEmit_vupkhsb_(f, i.VX.VD, i.VX.VB);
+}
+XEEMITTER(vupkhsb128,     VX128(6, 896),    VX128  )(PPCHIRBuilder& f, InstrData& i) {
+  return InstrEmit_vupkhsb_(f, VX128_VD128, VX128_VB128);
 }
 
-XEEMITTER(vupklsh,        0x100002CE, VX  )(PPCHIRBuilder& f, InstrData& i) {
-  XEINSTRNOTIMPLEMENTED();
+int InstrEmit_vupkhsh_(PPCHIRBuilder& f, uint32_t vd, uint32_t va, uint32_t vb) {
   return 1;
 }
+XEEMITTER(vupkhsh,        0x1000024E, VX  )(PPCHIRBuilder& f, InstrData& i) {
+  return InstrEmit_vupkhsh_(f, i.VX.VD, i.VX.VA, i.VX.VB);
+}
+
+int InstrEmit_vupklsb_(PPCHIRBuilder& f, uint32_t vd, uint32_t vb) {
+  // bytes 8-15 expanded to halfwords 0-8 and sign extended
+  Value* v = f.Unpack(f.LoadVR(vb), PACK_TYPE_S8_IN_16_LO);
+  f.StoreVR(vd, v);
+  return 0;
+}
+XEEMITTER(vupklsb,        0x1000028E, VX  )(PPCHIRBuilder& f, InstrData& i) {
+  return InstrEmit_vupklsb_(f, i.VX.VD, i.VX.VB);
+}
+XEEMITTER(vupklsb128,     VX128(6, 960),    VX128  )(PPCHIRBuilder& f, InstrData& i) {
+  return InstrEmit_vupklsb_(f, VX128_VD128, VX128_VB128);
+}
+
+int InstrEmit_vupklsh_(PPCHIRBuilder& f, uint32_t vd, uint32_t va, uint32_t vb) {
+  return 1;
+}
+XEEMITTER(vupklsh,        0x100002CE, VX  )(PPCHIRBuilder& f, InstrData& i) {
+  return InstrEmit_vupklsh_(f, i.VX.VD, i.VX.VA, i.VX.VB);
+}
 
 XEEMITTER(vpkd3d128,      VX128_4(6, 1552), VX128_4)(PPCHIRBuilder& f, InstrData& i) {
   const uint32_t vd = i.VX128_4.VD128l | (i.VX128_4.VD128h << 5);
diff --git a/src/alloy/hir/opcodes.h b/src/alloy/hir/opcodes.h
index 282a9600a..af3c36b72 100644
--- a/src/alloy/hir/opcodes.h
+++ b/src/alloy/hir/opcodes.h
@@ -71,6 +71,8 @@ enum PackType {
   PACK_TYPE_FLOAT16_2 = 1,
   PACK_TYPE_FLOAT16_4 = 2,
   PACK_TYPE_SHORT_2 = 3,
+  PACK_TYPE_S8_IN_16_LO = 4,
+  PACK_TYPE_S8_IN_16_HI = 5,
 };