aco: use s_waitcnt_depctr to mitigate VMEMtoScalarWriteHazard

[mesa.git] / src / amd / compiler / aco_print_asm.cpp

diff --git a/src/amd/compiler/aco_print_asm.cpp b/src/amd/compiler/aco_print_asm.cpp
index 4b0ce26994be5d4d703a78f6b61f22a7f3c0ed1e..d4b0e0edbcf3e6acb2776e35846988bf8798bd7f 100644 (file)
--- a/src/amd/compiler/aco_print_asm.cpp
+++ b/src/amd/compiler/aco_print_asm.cpp
@@ -5,16 +5,97 @@
 #include "ac_llvm_util.h"
 
 #include <llvm/ADT/StringRef.h>
+#if LLVM_VERSION_MAJOR >= 11
+#include <llvm/MC/MCDisassembler/MCDisassembler.h>
+#endif
 
 namespace aco {
 
+/* LLVM disassembler only supports GFX8+, try to disassemble with CLRXdisasm
+ * for GFX6-GFX7 if found on the system, this is better than nothing.
+*/
+void print_asm_gfx6_gfx7(Program *program, std::vector<uint32_t>& binary,
+                         std::ostream& out)
+{
+   char path[] = "/tmp/fileXXXXXX";
+   char line[2048], command[128];
+   const char *gpu_type;
+   FILE *p;
+   int fd;
+
+   /* Dump the binary into a temporary file. */
+   fd = mkstemp(path);
+   if (fd < 0)
+      return;
+
+   for (uint32_t w : binary)
+   {
+      if (write(fd, &w, sizeof(w)) == -1)
+         goto fail;
+   }
+
+   /* Determine the GPU type for CLRXdisasm. Use the family for GFX6 chips
+    * because it doesn't allow to use gfx600 directly.
+    */
+   switch (program->chip_class) {
+   case GFX6:
+      switch (program->family) {
+      case CHIP_TAHITI:
+         gpu_type = "tahiti";
+         break;
+      case CHIP_PITCAIRN:
+         gpu_type = "pitcairn";
+         break;
+      case CHIP_VERDE:
+         gpu_type = "capeverde";
+         break;
+      case CHIP_OLAND:
+         gpu_type = "oland";
+         break;
+      case CHIP_HAINAN:
+         gpu_type = "hainan";
+         break;
+      default:
+         unreachable("Invalid GFX6 family!");
+      }
+      break;
+   case GFX7:
+      gpu_type = "gfx700";
+      break;
+   default:
+      unreachable("Invalid chip class!");
+   }
+
+   sprintf(command, "clrxdisasm --gpuType=%s -r %s", gpu_type, path);
+
+   p = popen(command, "r");
+   if (p) {
+      if (!fgets(line, sizeof(line), p)) {
+         out << "clrxdisasm not found\n";
+         pclose(p);
+         goto fail;
+      }
+
+      do {
+         out << line;
+      } while (fgets(line, sizeof(line), p));
+
+      pclose(p);
+   }
+
+fail:
+   close(fd);
+   unlink(path);
+}
+
 void print_asm(Program *program, std::vector<uint32_t>& binary,
                unsigned exec_size, std::ostream& out)
 {
    if (program->chip_class <= GFX7) {
-      out << "Disassembly for this GPU currently not supported." << std::endl;
+      print_asm_gfx6_gfx7(program, binary, out);
       return;
    }
+
    std::vector<bool> referenced_blocks(program->blocks.size());
    referenced_blocks[0] = true;
    for (Block& block : program->blocks) {
@@ -22,7 +103,11 @@ void print_asm(Program *program, std::vector<uint32_t>& binary,
          referenced_blocks[succ] = true;
    }
 
+   #if LLVM_VERSION_MAJOR >= 11
+   std::vector<llvm::SymbolInfoTy> symbols;
+   #else
    std::vector<std::tuple<uint64_t, llvm::StringRef, uint8_t>> symbols;
+   #endif
    std::vector<std::array<char,16>> block_names;
    block_names.reserve(program->blocks.size());
    for (Block& block : program->blocks) {
@@ -57,7 +142,7 @@ void print_asm(Program *program, std::vector<uint32_t>& binary,
 
       /* mask out src2 on v_writelane_b32 */
       if (((program->chip_class == GFX8 || program->chip_class == GFX9) && (binary[pos] & 0xffff8000) == 0xd28a0000) ||
-          (program->chip_class == GFX10 && (binary[pos] & 0xffff8000) == 0xd7610000)) {
+          (program->chip_class >= GFX10 && (binary[pos] & 0xffff8000) == 0xd7610000)) {
          binary[pos+1] = binary[pos+1] & 0xF803FFFF;
       }
 
@@ -67,8 +152,16 @@ void print_asm(Program *program, std::vector<uint32_t>& binary,
 
       size_t new_pos;
       const int align_width = 60;
-      if (!l && program->chip_class == GFX9 && ((binary[pos] & 0xffff8000) == 0xd1348000)) { /* not actually an invalid instruction */
-         out << std::left << std::setw(align_width) << std::setfill(' ') << "\tv_add_u32_e64 + clamp";
+      if (!l &&
+          ((program->chip_class >= GFX9 && (binary[pos] & 0xffff8000) == 0xd1348000) || /* v_add_u32_e64 + clamp */
+           (program->chip_class >= GFX10 && (binary[pos] & 0xffff8000) == 0xd7038000) || /* v_add_u16_e64 + clamp */
+           (program->chip_class <= GFX9 && (binary[pos] & 0xffff8000) == 0xd1268000)) /* v_add_u16_e64 + clamp */) {
+         out << std::left << std::setw(align_width) << std::setfill(' ') << "\tinteger addition + clamp";
+         bool has_literal = program->chip_class >= GFX10 &&
+                            (((binary[pos+1] & 0x1ff) == 0xff) || (((binary[pos+1] >> 9) & 0x1ff) == 0xff));
+         new_pos = pos + 2 + has_literal;
+      } else if (program->chip_class >= GFX10 && l == 4 && ((binary[pos] & 0xfe0001ff) == 0x020000f9)) {
+         out << std::left << std::setw(align_width) << std::setfill(' ') << "\tv_cndmask_b32 + sdwa";
          new_pos = pos + 2;
       } else if (!l) {
          out << std::left << std::setw(align_width) << std::setfill(' ') << "(invalid instruction)";