OpcodeDispatcher: optimize blsmsk flags

Signed-off-by: Alyssa Rosenzweig <[email protected]>
FEX-Emu · Jan 23, 2024 · 836835b · 836835b
1 parent 84313b4
commit 836835b
Show file tree

Hide file tree

Showing 3 changed files with 33 additions and 33 deletions.
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/Flags.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/Flags.cpp
@@ -1014,30 +1014,22 @@ void OpDispatchBuilder::CalculateFlags_BLSI(uint8_t SrcSize, OrderedNode *Src) {
 }
 
 void OpDispatchBuilder::CalculateFlags_BLSMSK(uint8_t SrcSize, OrderedNode *Result, OrderedNode *Src) {
-  // Now for the flags.
-
-  auto Zero = _Constant(0);
-  auto One = _Constant(1);
-
-  uint32_t FlagsMaskToZero =
-    (1U << X86State::RFLAG_ZF_RAW_LOC) |
-    (1U << X86State::RFLAG_OF_RAW_LOC);
-
-  ZeroMultipleFlags(FlagsMaskToZero);
-
   // PF/AF undefined
   _InvalidateFlags((1UL << X86State::RFLAG_PF_RAW_LOC) |
                    (1UL << X86State::RFLAG_AF_RAW_LOC));
 
+  // CF set according to the Src
+  auto Zero = _Constant(0);
+  auto One = _Constant(1);
   auto CFOp = _Select(IR::COND_EQ, Src, Zero, One, Zero);
-  SetRFLAG<X86State::RFLAG_CF_RAW_LOC>(CFOp);
 
-  auto SFOp = _Bfe(SizeToOpSize(SrcSize), 1, (SrcSize * 8) - 1, Result);
-  SetRFLAG<X86State::RFLAG_SF_RAW_LOC>(SFOp);
+  // The output of BLSMSK is always nonzero, so TST will clear Z (along with C
+  // and O) while setting S.
+  SetNZ_ZeroCV(SrcSize, Result);
+  SetRFLAG<X86State::RFLAG_CF_RAW_LOC>(CFOp);
 }
 
 void OpDispatchBuilder::CalculateFlags_BLSR(uint8_t SrcSize, OrderedNode *Result, OrderedNode *Src) {
-  // Now for flags.
   auto Zero = _Constant(0);
   auto One = _Constant(1);
 

diff --git a/unittests/InstructionCountCI/FlagM/VEX_map_group.json b/unittests/InstructionCountCI/FlagM/VEX_map_group.json
@@ -44,7 +44,7 @@
       ]
     },
     "blsmsk eax, ebx": {
-      "ExpectedInstructionCount": 11,
+      "ExpectedInstructionCount": 13,
       "Comment": [
         "Map group 17 0b010 32-bit"
       ],
@@ -53,17 +53,19 @@
         "sub x21, x20, #0x1 (1)",
         "eor x21, x21, x20",
         "mov w4, w21",
-        "mov w21, #0x50000000",
-        "mrs x22, nzcv",
-        "bic x21, x22, x21",
+        "mov w22, #0x50000000",
+        "mrs x23, nzcv",
+        "bic x22, x23, x22",
         "cmp x20, #0x0 (0)",
-        "cset x20, ne",
-        "msr nzcv, x21",
-        "rmif x20, #63, #nzCv"
+        "cset x20, eq",
+        "msr nzcv, x22",
+        "rmif x20, #63, #nzCv",
+        "lsr w20, w21, #31",
+        "rmif x20, #61, #Nzcv"
       ]
     },
     "blsmsk rax, rbx": {
-      "ExpectedInstructionCount": 9,
+      "ExpectedInstructionCount": 11,
       "Comment": [
         "Map group 17 0b010 64-bit"
       ],
@@ -74,9 +76,11 @@
         "mrs x21, nzcv",
         "bic x20, x21, x20",
         "cmp x7, #0x0 (0)",
-        "cset x21, ne",
+        "cset x21, eq",
         "msr nzcv, x20",
-        "rmif x21, #63, #nzCv"
+        "rmif x21, #63, #nzCv",
+        "lsr x20, x4, #63",
+        "rmif x20, #61, #Nzcv"
       ]
     },
     "blsi eax, ebx": {

diff --git a/unittests/InstructionCountCI/VEX_map_group.json b/unittests/InstructionCountCI/VEX_map_group.json
@@ -643,7 +643,7 @@
       ]
     },
     "blsmsk eax, ebx": {
-      "ExpectedInstructionCount": 13,
+      "ExpectedInstructionCount": 15,
       "Comment": [
         "Map group 17 0b010 32-bit"
       ],
@@ -652,19 +652,21 @@
         "sub x21, x20, #0x1 (1)",
         "eor x21, x21, x20",
         "mov w4, w21",
-        "mov w21, #0x50000000",
-        "mrs x22, nzcv",
-        "bic x21, x22, x21",
+        "mov w22, #0x50000000",
+        "mrs x23, nzcv",
+        "bic x22, x23, x22",
         "cmp x20, #0x0 (0)",
-        "cset x20, ne",
-        "mov w0, w21",
+        "cset x20, eq",
+        "mov w0, w22",
         "bfi w0, w20, #29, #1",
         "mov w20, w0",
+        "lsr w21, w21, #31",
+        "bfi w20, w21, #31, #1",
         "msr nzcv, x20"
       ]
     },
     "blsmsk rax, rbx": {
-      "ExpectedInstructionCount": 9,
+      "ExpectedInstructionCount": 11,
       "Comment": [
         "Map group 17 0b010 64-bit"
       ],
@@ -675,8 +677,10 @@
         "mrs x21, nzcv",
         "bic x20, x21, x20",
         "cmp x7, #0x0 (0)",
-        "cset x21, ne",
+        "cset x21, eq",
         "bfi w20, w21, #29, #1",
+        "lsr x21, x4, #63",
+        "bfi w20, w21, #31, #1",
         "msr nzcv, x20"
       ]
     },