[X86-64] Add suport to raise PCMPGTrr instruction.

- Adresses failure exposed by issue #182. - Add suport to raise PCMPGTrm instruction. - Add tests to verify raising of newly added suport.
microsoft · Oct 3, 2022 · ba33acc · ba33acc
1 parent 4c9293f
commit ba33acc
Show file tree

Hide file tree

Showing 3 changed files with 38 additions and 3 deletions.
diff --git a/X86/X86AdditionalInstrInfo.cpp b/X86/X86AdditionalInstrInfo.cpp
@@ -1854,6 +1854,8 @@ static constexpr const_addl_instr_info::value_type MapData[] = {
     {X86::PCMPEQQrr, {0, SSE_COMPARE_RR}},
     {X86::PCMPEQWrm, {16, SSE_COMPARE_RM}},
     {X86::PCMPEQWrr, {0, SSE_COMPARE_RR}},
+    {X86::PCMPGTDrr, {0, SSE_COMPARE_RR}},
+    {X86::PCMPGTDrm, {0, SSE_COMPARE_RM}},
     {X86::PDEP32rm, {4, Unknown}},
     {X86::PDEP32rr, {0, Unknown}},
     {X86::PDEP64rm, {8, Unknown}},

diff --git a/X86/X86MachineInstructionRaiserSSE.cpp b/X86/X86MachineInstructionRaiserSSE.cpp
@@ -276,7 +276,9 @@ bool X86MachineInstructionRaiser::raiseSSECompareMachineInstr(
   case X86::PCMPEQDrm:
   case X86::PCMPEQDrr:
   case X86::PCMPEQQrm:
-  case X86::PCMPEQQrr: {
+  case X86::PCMPEQQrr:
+  case X86::PCMPGTDrr:
+  case X86::PCMPGTDrm: {
     // Compare a comparison of packed bytes/words/dwords/qwords
     // If a pair is equal, set the bits corresponding to 1, otherwise to 0
     LLVMContext &Ctx(MF.getFunction().getContext());
@@ -290,22 +292,32 @@ bool X86MachineInstructionRaiser::raiseSSECompareMachineInstr(
            "Expected operand types to be vector types of size 128");
 
     unsigned int ElementSizeInBits;
+    CmpInst::Predicate CmpPred = CmpInst::BAD_ICMP_PREDICATE;
     switch (MI.getOpcode()) {
     case X86::PCMPEQBrm:
     case X86::PCMPEQBrr:
       ElementSizeInBits = 8;
+      CmpPred = CmpInst::ICMP_EQ;
       break;
     case X86::PCMPEQWrm:
     case X86::PCMPEQWrr:
       ElementSizeInBits = 16;
+      CmpPred = CmpInst::ICMP_EQ;
       break;
     case X86::PCMPEQDrm:
     case X86::PCMPEQDrr:
       ElementSizeInBits = 32;
+      CmpPred = CmpInst::ICMP_EQ;
       break;
     case X86::PCMPEQQrm:
     case X86::PCMPEQQrr:
       ElementSizeInBits = 64;
+      CmpPred = CmpInst::ICMP_EQ;
+      break;
+    case X86::PCMPGTDrr:
+    case X86::PCMPGTDrm:
+      ElementSizeInBits = 32;
+      CmpPred = CmpInst::ICMP_SGT;
       break;
     default:
       llvm_unreachable("Unhandled pcmp instruction");
@@ -329,8 +341,8 @@ bool X86MachineInstructionRaiser::raiseSSECompareMachineInstr(
           ExtractElementInst::Create(CmpOpVal1, Index, "", RaisedBB);
       auto *CmpSegment2 =
           ExtractElementInst::Create(CmpOpVal2, Index, "", RaisedBB);
-      auto *CmpInst = new ICmpInst(*RaisedBB, CmpInst::ICMP_EQ, CmpSegment1,
-                                   CmpSegment2, "cmp_segment");
+      auto *CmpInst = new ICmpInst(*RaisedBB, CmpPred, CmpSegment1, CmpSegment2,
+                                   "cmp_segment");
       auto *SelectInstr =
           SelectInst::Create(CmpInst, BitmaskVal, ZeroVal, "segment", RaisedBB);
       Result =

diff --git a/test/asm_test/X86/raise-pcmp.s b/test/asm_test/X86/raise-pcmp.s
@@ -7,10 +7,12 @@
 // CHECK-NEXT: 0x0000ffffffffffffffffffffffffffff
 // CHECK-NEXT: 0x00000000ffffffffffffffffffffffff
 // CHECK-NEXT: 0x0000000000000000ffffffffffffffff
+// CHECK-NEXT: 0xffffffff000000000000000000000000
 // CHECK-NEXT: 0x00ffffffffffffffffffffffffffffff
 // CHECK-NEXT: 0x0000ffffffffffffffffffffffffffff
 // CHECK-NEXT: 0x00000000ffffffffffffffffffffffff
 // CHECK-NEXT: 0x0000000000000000ffffffffffffffff
+// CHECK-NEXT: 0xffffffff000000000000000000000000
 // CHECK-EMPTY
 
 .text
@@ -64,6 +66,16 @@ main:                                   # @main
     mov al, 0
     call printf
 
+    movdqa xmm0, [.L.val]
+    movdqa xmm1, [.L.val.1]
+    pcmpgtd xmm0, xmm1
+    movdqu [rsp], xmm0
+    mov rsi, QWORD PTR [rsp + 8]
+    mov rdx, QWORD PTR [rsp]
+    mov rdi, offset .L.str
+    mov al, 0
+    call printf
+
     # rm
     movdqa xmm0, [.L.val]
     pcmpeqb xmm0, [.L.val.1]
@@ -101,6 +113,15 @@ main:                                   # @main
     mov al, 0
     call printf
 
+    movdqa xmm0, [.L.val]
+    pcmpgtd xmm0, [.L.val.1]
+    movdqu [rsp], xmm0
+    mov rsi, QWORD PTR [rsp + 8]
+    mov rdx, QWORD PTR [rsp]
+    mov rdi, offset .L.str
+    mov al, 0
+    call printf
+
     add rsp, 16
     add rsp, 8
     xor rax, rax