agc: support one-word NOP reservations and share packet sizing

src/graphics/guest_gpu/graphicsRun.cpp

@@ -689,7 +689,7 @@ void CommandProcessor::ProcessPm4(Pm4Execution& execution) {
 		const auto        opcode        = (packet_header >> 8u) & 0xffu;
 		EXIT_NOT_IMPLEMENTED(remaining_dw > total_dw);
 
-		if (packet_header == 0x80000000u) {
+		if (packet_header == 0x80000000u || Pm4::PacketSizeDw(packet_header) == 1u) {
 			cursor.offset_dw++;
 			execution.m_made_progress = true;
 			continue;

src/graphics/guest_gpu/pm4.cpp

@@ -120,15 +120,15 @@ void DumpPm4PacketStream(Common::File* file, const uint32_t* cmd_buffer, uint32_
 		uint32_t len = 0;
 
 		const auto packet_type = static_cast<PacketType>(cmd_id >> 30u);
-		// Type-2 packets are header-only padding; every other packet type requires a body.
+		EXIT_NOT_IMPLEMENTED(dw < 2 && packet_type != PacketType::Type2 &&
-		EXIT_NOT_IMPLEMENTED(dw < 2 && packet_type != PacketType::Type2);
+		                     PacketSizeDw(cmd_id) != 1u);
 
 		switch (packet_type) {
 			case PacketType::Type3: {
 				const bool    sh_gx = (cmd_id & 0x2u) == 0;
 				const uint8_t op    = ((cmd_id >> 8u) & 0xffu);
 				const auto    r     = KYTY_PM4_R(cmd_id);
-				len                 = ((cmd_id >> 16u) & 0x3fffu) + 1;
+				len                 = PacketSizeDw(cmd_id) - 1u;
 
 				EXIT_NOT_IMPLEMENTED(len >= dw);
 

src/graphics/guest_gpu/pm4.h

@@ -19,6 +19,10 @@ class File;
 
 namespace Libs::Graphics::Pm4 {
 
+constexpr uint32_t PacketSizeDw(uint32_t header) {
+	return (header & 0x3fffff00u) == 0x3fff1000u ? 1u : KYTY_PM4_LEN(header);
+}
+
 constexpr uint32_t IT_NOP                       = 0x10;
 constexpr uint32_t IT_SET_BASE                  = 0x11;
 constexpr uint32_t IT_CLEAR_STATE               = 0x12;

src/libs/agc.cpp

@@ -1730,7 +1730,7 @@ int KYTY_SYSV_ABI AgcDriverRegisterWorkloadStream(uint32_t stream_id, const void
 }
 
 uint32_t* KYTY_SYSV_ABI AgcCbNop(CommandBuffer* buf, uint32_t size_in_dwords) {
-	if (buf == nullptr || size_in_dwords < 2) {
+	if (buf == nullptr || size_in_dwords == 0) {
 		return nullptr;
 	}
 
@@ -1741,9 +1741,6 @@ uint32_t* KYTY_SYSV_ABI AgcCbNop(CommandBuffer* buf, uint32_t size_in_dwords) {
 	}
 
 	cmd[0] = KYTY_PM4(size_in_dwords, Pm4::IT_NOP, Pm4::R_ZERO);
-	if (size_in_dwords > 1) {
-		memset(cmd + 1, 0, static_cast<size_t>(size_in_dwords - 1) * 4);
-	}
 
 	return cmd;
 }
@@ -3678,12 +3675,7 @@ int KYTY_SYSV_ABI AgcDmaDataPatchSetSrcAddressOrOffsetOrImmediate(
 }
 
 uint32_t KYTY_SYSV_ABI AgcGetPacketSize(uint32_t* packet) {
-	const auto cmd_id = packet[0];
+	return Pm4::PacketSizeDw(packet[0]);
-	if ((cmd_id & 0x3fffff00u) == 0x3fff1000u) {
-		return 1;
-	}
-
-	return KYTY_PM4_LEN(cmd_id);
 }
 
 int KYTY_SYSV_ABI AgcSetPacketPredication(uint32_t* packet, uint32_t predication) {
@@ -3725,12 +3717,7 @@ int KYTY_SYSV_ABI AgcSetRangePredication(uint32_t* start, const volatile uint32_
 		const auto cmd_id = packet[0];
 		packet[0]         = (cmd_id & ~1u) | predication_bit;
 
-		auto size = KYTY_PM4_LEN(cmd_id);
+		packet_va += Pm4::PacketSizeDw(cmd_id) * sizeof(uint32_t);
-		if ((cmd_id & 0x3fffff00u) == 0x3fff1000u) {
-			size = 1;
-		}
-
-		packet_va += size * sizeof(uint32_t);
 		packet = reinterpret_cast<uint32_t*>(packet_va);
 	}
 

tests/ShaderRecompilerComputeTests.cpp

@@ -1,5 +1,6 @@
 #include "common/assert.h"
 #include "common/emulatorConfig.h"
+#include "common/file.h"
 #include "common/hostException.h"
 #include "common/logging/log.h"
 #include "common/subsystems.h"
@@ -41778,6 +41779,43 @@ void CheckPm4IndirectControlFlow(RenderContext &renderer) {
         static_cast<uint32_t>(address(target.data())),
         static_cast<uint32_t>(address(target.data()) >> 32u), control};
   };
+  // A one-word NOP must leave the following completion write reachable.
+  std::array<uint32_t, 7> nop_commands{};
+  CommandBufferLayout nop_dcb{nop_commands.data(), nop_commands.data() + nop_commands.size(),
+      nop_commands.data(), nop_commands.data() + nop_commands.size(), nullptr, nullptr, 0};
+  auto* nop_cb = reinterpret_cast<Gen5::CommandBuffer*>(&nop_dcb);
+  Require(name, "zero-sized NOP", Gen5::AgcCbNop(nop_cb, 0) == nullptr &&
+              nop_dcb.cursor_up == nop_commands.data() && nop_commands[0] == 0,
+          "zero-sized reservation changed the command buffer");
+  Require(name, "empty command list NOP", Gen5::AgcCbNop(nop_cb, 1) == nop_commands.data() &&
+              nop_dcb.cursor_up == nop_commands.data() + 1 && nop_commands[0] == 0xffff1000u &&
+              Gen5::AgcCbNopGetSize(1) == 4 && Gen5::AgcGetPacketSize(nop_commands.data()) == 1,
+          "native NOP(1) did not reserve one DWORD");
+  const auto completion = write(&selected, 7);
+  std::copy(completion.begin(), completion.end(), nop_commands.begin() + 1);
+  nop_commands.back() = 0xffff1000u;
+  for (uint32_t predication : {1u, 0u}) {
+    Gen5::AgcSetRangePredication(nop_commands.data(), nop_commands.data() + nop_commands.size(), predication);
+    Require(name, "NOP range predication", nop_commands[0] == (0xffff1000u | predication) &&
+                nop_commands.back() == (0xffff1000u | predication) &&
+                nop_commands[1] == (completion[0] | predication) && nop_commands[5] == 7,
+            "range predication skipped a packet or modified a payload");
+    Pm4Execution execution;
+    selected = 0;
+    Require(name, "NOP and completion execution",
+            processor.Process(execution, nop_commands) == Pm4ProcessResult::Complete && selected == 7,
+            "one-DWORD NOP prevented the completion write or overread the stream end");
+  }
+  Common::File dump;
+  Require(name, "NOP dump file", dump.CreateInMem(), "could not create PM4 dump");
+  Pm4::DumpPm4PacketStream(&dump, nop_commands.data(), 0, nop_commands.size());
+  std::array<uint32_t, 3> payload{0, 0x12345678u, 0x9abcdef0u};
+  CommandBufferLayout payload_dcb{payload.data(), payload.data() + payload.size(),
+      payload.data(), payload.data() + payload.size(), nullptr, nullptr, 0};
+  Require(name, "NOP payload reservation",
+          Gen5::AgcCbNop(reinterpret_cast<Gen5::CommandBuffer*>(&payload_dcb), 3) == payload.data() &&
+              payload == std::array<uint32_t, 3>{0xc0011000u, 0x12345678u, 0x9abcdef0u},
+          "NOP emitter overwrote the caller's reserved payload");
   const auto then_commands = write(&selected, 11);
   const auto else_commands = write(&selected, 33);
   const auto branch_suffix = write(&selected, 44);