vendor/llvm-project/llvmorg-15-init-15358-g53dc0f107877

1 files changed, 16 insertions, 1 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIModeRegister.cpp b/llvm/lib/Target/AMDGPU/SIModeRegister.cpp
index 24a8879b5684..a5816e2e8c73 100644
--- a/llvm/lib/Target/AMDGPU/SIModeRegister.cpp
+++ b/llvm/lib/Target/AMDGPU/SIModeRegister.cpp
@@ -17,6 +17,7 @@
 #include "GCNSubtarget.h"
 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
 #include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
 #include <queue>
 
 #define DEBUG_TYPE "si-mode-register"
@@ -162,7 +163,9 @@ FunctionPass *llvm::createSIModeRegisterPass() { return new SIModeRegister(); }
 // double precision setting.
 Status SIModeRegister::getInstructionMode(MachineInstr &MI,
                                           const SIInstrInfo *TII) {
-  if (TII->usesFPDPRounding(MI)) {
+  if (TII->usesFPDPRounding(MI) ||
+      MI.getOpcode() == AMDGPU::FPTRUNC_UPWARD_PSEUDO ||
+      MI.getOpcode() == AMDGPU::FPTRUNC_DOWNWARD_PSEUDO) {
     switch (MI.getOpcode()) {
     case AMDGPU::V_INTERP_P1LL_F16:
     case AMDGPU::V_INTERP_P1LV_F16:
@@ -170,6 +173,18 @@ Status SIModeRegister::getInstructionMode(MachineInstr &MI,
       // f16 interpolation instructions need double precision round to zero
       return Status(FP_ROUND_MODE_DP(3),
                     FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_ZERO));
+    case AMDGPU::FPTRUNC_UPWARD_PSEUDO: {
+      // Replacing the pseudo by a real instruction
+      MI.setDesc(TII->get(AMDGPU::V_CVT_F16_F32_e32));
+      return Status(FP_ROUND_MODE_DP(3),
+                    FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_INF));
+    }
+    case AMDGPU::FPTRUNC_DOWNWARD_PSEUDO: {
+      // Replacing the pseudo by a real instruction
+      MI.setDesc(TII->get(AMDGPU::V_CVT_F16_F32_e32));
+      return Status(FP_ROUND_MODE_DP(3),
+                    FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEGINF));
+    }
     default:
       return DefaultStatus;
     }