From 2c76e6d023f1c4c4a5d4e19ca00f7df2c26c9f2b Mon Sep 17 00:00:00 2001
From: Henrik Rydgard <hrydgard@gmail.com>
Date: Mon, 24 Mar 2014 11:19:11 +0100
Subject: [PATCH] Correctly keep track of "full alpha" in vertices (x86 jit
 only).

---
 GPU/GLES/TransformPipeline.cpp | 13 ++++++++-----
 GPU/GLES/TransformPipeline.h   |  6 ++++++
 GPU/GLES/VertexDecoderX86.cpp  | 22 ++++++++++++++++++++++
 3 files changed, 36 insertions(+), 5 deletions(-)

diff --git a/GPU/GLES/TransformPipeline.cpp b/GPU/GLES/TransformPipeline.cpp
index 0b01d00a3a..263fdfb4b1 100644
--- a/GPU/GLES/TransformPipeline.cpp
+++ b/GPU/GLES/TransformPipeline.cpp
@@ -265,11 +265,6 @@ void TransformDrawEngine::SetupVertexDecoder(u32 vertType) {
 	if (vertTypeID != lastVType_) {
 		dec_ = GetVertexDecoder(vertTypeID);
 		lastVType_ = vertTypeID;
-
-		// TODO: Add functionality to VertexDecoder to scan for non-full alpha in the two other formats,
-		// which are quite common.
-		int colorType = vertTypeID & GE_VTYPE_COL_MASK;
-		gstate_c.vertexFullAlpha = colorType == GE_VTYPE_COL_NONE || colorType == GE_VTYPE_COL_565;
 	}
 }
 
@@ -566,6 +561,8 @@ void TransformDrawEngine::DoFlush() {
 					vai->numVerts = indexGen.VertexCount();
 					vai->prim = indexGen.Prim();
 					vai->maxIndex = indexGen.MaxIndex();
+					vai->flags = gstate_c.vertexFullAlpha ? VAI_FLAG_VERTEXFULLALPHA : 0;
+
 					goto rotateVBO;
 				}
 
@@ -645,6 +642,8 @@ void TransformDrawEngine::DoFlush() {
 					vertexCount = vai->numVerts;
 					maxIndex = vai->maxIndex;
 					prim = static_cast<GEPrimitiveType>(vai->prim);
+
+					gstate_c.vertexFullAlpha = vai->flags & VAI_FLAG_VERTEXFULLALPHA;
 					break;
 				}
 
@@ -665,6 +664,8 @@ void TransformDrawEngine::DoFlush() {
 					vertexCount = vai->numVerts;
 					maxIndex = vai->maxIndex;
 					prim = static_cast<GEPrimitiveType>(vai->prim);
+
+					gstate_c.vertexFullAlpha = vai->flags & VAI_FLAG_VERTEXFULLALPHA;
 					break;
 				}
 
@@ -717,6 +718,7 @@ rotateVBO:
 			glBindBuffer(GL_ARRAY_BUFFER, 0);
 	} else {
 		DecodeVerts();
+
 		LinkedShader *program = shaderManager_->ApplyFragmentShader(vshader, prim, lastVType_);
 		gpuStats.numUncachedVertsDrawn += indexGen.VertexCount();
 		prim = indexGen.Prim();
@@ -737,6 +739,7 @@ rotateVBO:
 	decodeCounter_ = 0;
 	dcid_ = 0;
 	prevPrim_ = GE_PRIM_INVALID;
+	gstate_c.vertexFullAlpha = true;
 
 #ifndef MOBILE_DEVICE
 	host->GPUNotifyDraw();
diff --git a/GPU/GLES/TransformPipeline.h b/GPU/GLES/TransformPipeline.h
index a9c16eaca8..bee7621bd3 100644
--- a/GPU/GLES/TransformPipeline.h
+++ b/GPU/GLES/TransformPipeline.h
@@ -43,6 +43,10 @@ struct DecVtxFormat;
 // DRAWN_ONCE -> death
 // DRAWN_RELIABLE -> death
 
+enum {
+	VAI_FLAG_VERTEXFULLALPHA = 1,
+};
+
 // Try to keep this POD.
 class VertexArrayInfo {
 public:
@@ -57,6 +61,7 @@ public:
 		lastFrame = gpuStats.numFlips;
 		numVerts = 0;
 		drawsUntilNextFullHash = 0;
+		flags = 0;
 	}
 	~VertexArrayInfo();
 
@@ -85,6 +90,7 @@ public:
 	int numFrames;
 	int lastFrame;  // So that we can forget.
 	u16 drawsUntilNextFullHash;
+	u8 flags;
 };
 
 // Handles transform, lighting and drawing.
diff --git a/GPU/GLES/VertexDecoderX86.cpp b/GPU/GLES/VertexDecoderX86.cpp
index 32223cf42e..0c770903da 100644
--- a/GPU/GLES/VertexDecoderX86.cpp
+++ b/GPU/GLES/VertexDecoderX86.cpp
@@ -556,6 +556,11 @@ void VertexDecoderJitCache::Jit_TcFloatThrough() {
 void VertexDecoderJitCache::Jit_Color8888() {
 	MOV(32, R(tempReg1), MDisp(srcReg, dec_->coloff));
 	MOV(32, MDisp(dstReg, dec_->decFmt.c0off), R(tempReg1));
+
+	CMP(32, R(tempReg1), Imm32(0xFF000000));
+	FixupBranch skip = J_CC(CC_GE, false);
+	MOV(8, M(&gstate_c.textureFullAlpha), Imm8(0));
+	SetJumpTarget(skip);
 }
 
 static const u32 MEMORY_ALIGNED16(nibbles[4]) = { 0x0f0f0f0f, 0x0f0f0f0f, 0x0f0f0f0f, 0x0f0f0f0f, };
@@ -625,6 +630,11 @@ void VertexDecoderJitCache::Jit_Color4444() {
 	OR(32, R(tempReg2), R(tempReg3));
 
 	MOV(32, MDisp(dstReg, dec_->decFmt.c0off), R(tempReg2));
+
+	CMP(32, R(tempReg2), Imm32(0xFF000000));
+	FixupBranch skip = J_CC(CC_AE, false);
+	MOV(8, M(&gstate_c.textureFullAlpha), Imm8(0));
+	SetJumpTarget(skip);
 }
 
 void VertexDecoderJitCache::Jit_Color565() {
@@ -696,6 +706,11 @@ void VertexDecoderJitCache::Jit_Color5551() {
 	OR(32, R(tempReg2), R(tempReg1));
 
 	MOV(32, MDisp(dstReg, dec_->decFmt.c0off), R(tempReg2));
+
+	CMP(32, R(tempReg2), Imm32(0xFF000000));
+	FixupBranch skip = J_CC(CC_AE, false);
+	MOV(8, M(&gstate_c.textureFullAlpha), Imm8(0));
+	SetJumpTarget(skip);
 }
 
 void VertexDecoderJitCache::Jit_Color8888Morph() {
@@ -890,6 +905,13 @@ void VertexDecoderJitCache::Jit_WriteMorphColor(int outOff) {
 	PACKSSDW(fpScratchReg, R(fpScratchReg));
 	PACKUSWB(fpScratchReg, R(fpScratchReg));
 	MOVD_xmm(MDisp(dstReg, outOff), fpScratchReg);
+
+	// TODO: May be a faster way to do this without the MOVD.
+	MOVD_xmm(R(tempReg1), fpScratchReg);
+	CMP(32, R(tempReg1), Imm32(0xFF000000));
+	FixupBranch skip = J_CC(CC_AE, false);
+	MOV(8, M(&gstate_c.textureFullAlpha), Imm8(0));
+	SetJumpTarget(skip);
 }
 
 // Copy 3 bytes and then a zero. Might as well copy four.