[APX] Enable APX-EVEX.ZU for SETcc instructions

src/coreclr/jit/codegen.h

@@ -1338,7 +1338,7 @@ class CodeGen final : public CodeGenInterface
     void inst_JMP(emitJumpKind jmp, BasicBlock* tgtBlock);
 #endif
 
-    void inst_SET(emitJumpKind condition, regNumber reg);
+    void inst_SET(emitJumpKind condition, regNumber reg, insOpts instOptions = INS_OPTS_NONE);
 
     void inst_RV(instruction ins, regNumber reg, var_types type, emitAttr size = EA_UNKNOWN);
 

src/coreclr/jit/codegenxarch.cpp

@@ -1807,19 +1807,29 @@ void CodeGen::inst_SETCC(GenCondition condition, var_types type, regNumber dstRe
     assert(varTypeIsIntegral(type));
     assert(genIsValidIntReg(dstReg) && isByteReg(dstReg));
 
-    const GenConditionDesc& desc = GenConditionDesc::Get(condition);
+    const GenConditionDesc& desc        = GenConditionDesc::Get(condition);
+    insOpts                 instOptions = INS_OPTS_NONE;
 
-    inst_SET(desc.jumpKind1, dstReg);
+    bool needsMovzx = !varTypeIsByte(type);
+    if (needsMovzx && compiler->canUseApxEvexEncoding() && JitConfig.EnableApxZU())
+    {
+        instOptions = INS_OPTS_EVEX_zu;
+        needsMovzx  = false;
+    }
+
+    inst_SET(desc.jumpKind1, dstReg, instOptions);
 
     if (desc.oper != GT_NONE)
     {
         BasicBlock* labelNext = genCreateTempLabel();
         inst_JMP((desc.oper == GT_OR) ? desc.jumpKind1 : emitter::emitReverseJumpKind(desc.jumpKind1), labelNext);
-        inst_SET(desc.jumpKind2, dstReg);
+        inst_SET(desc.jumpKind2, dstReg, instOptions);
         genDefineTempLabel(labelNext);
     }
 
-    if (!varTypeIsByte(type))
+    // we can apply EVEX.ZU to avoid this movzx.
+    // TODO-XArch-apx: evaluate setcc + movzx and xor + set
+    if (needsMovzx)
     {
         GetEmitter()->emitIns_Mov(INS_movzx, EA_1BYTE, dstReg, dstReg, /* canSkip */ false);
     }
@@ -9450,6 +9460,8 @@ void CodeGen::genAmd64EmitterUnitTestsApx()
 
     theEmitter->emitIns_Mov(INS_movd32, EA_4BYTE, REG_R16, REG_XMM0, false);
     theEmitter->emitIns_Mov(INS_movd32, EA_4BYTE, REG_R16, REG_XMM16, false);
+
+    theEmitter->emitIns_R(INS_seto_apx, EA_1BYTE, REG_R11, INS_OPTS_EVEX_zu);
 }
 
 void CodeGen::genAmd64EmitterUnitTestsAvx10v2()
@@ -10351,7 +10363,7 @@ void CodeGen::genPushCalleeSavedRegisters()
 #endif // DEBUG
 
 #ifdef TARGET_AMD64
-    if (compiler->canUseApxEncoding() && compiler->canUseEvexEncoding() && JitConfig.EnableApxPPX())
+    if (compiler->canUseApxEvexEncoding() && JitConfig.EnableApxPPX())
     {
         genPushCalleeSavedRegistersFromMaskAPX(rsPushRegs);
         return;
@@ -10477,7 +10489,7 @@ void CodeGen::genPopCalleeSavedRegisters(bool jmpEpilog)
         return;
     }
 
-    if (compiler->canUseApxEncoding() && compiler->canUseEvexEncoding() && JitConfig.EnableApxPPX())
+    if (compiler->canUseApxEvexEncoding() && JitConfig.EnableApxPPX())
     {
         regMaskTP      rsPopRegs = regSet.rsGetModifiedIntCalleeSavedRegsMask();
         const unsigned popCount  = genPopCalleeSavedRegistersFromMaskAPX(rsPopRegs);

src/coreclr/jit/compiler.h

@@ -9678,6 +9678,17 @@ class Compiler
         return compOpportunisticallyDependsOn(InstructionSet_APX);
     }
 
+    //------------------------------------------------------------------------
+    // canUseApxEvexEncoding - Answer the question: Are APX-EVEX encodings supported on this target.
+    //
+    // Returns:
+    //    `true` if APX-EVEX encoding is supported, `false` if not.
+    //
+    bool canUseApxEvexEncoding() const
+    {
+        return canUseApxEncoding() && canUseEvexEncoding();
+    }
+
 private:
     //------------------------------------------------------------------------
     // DoJitStressEvexEncoding- Answer the question: Do we force EVEX encoding.

src/coreclr/jit/emit.h

@@ -846,6 +846,8 @@ class emitter
 
 #define _idEvexbContext  (_idCustom6 << 1) | _idCustom5 /* Evex.b: embedded broadcast, rounding, SAE */
 #define _idEvexNdContext _idCustom5 /* bits used for the APX-EVEX.nd context for promoted legacy instructions */
+#define _idEvexZuContext _idCustom5 /* bits used for the APX-EVEX.zu context for promoted legacy instructions */
+
 #define _idEvexNfContext _idCustom6 /* bits used for the APX-EVEX.nf context for promoted legacy/vex instructions */
 
         // We repurpose 4 bits for the default flag value bits for ccmp instructions.
@@ -1793,15 +1795,28 @@ class emitter
 
         bool idIsEvexNdContextSet() const
         {
+            assert(IsApxNddCompatibleInstruction(_idIns));
             return _idEvexNdContext != 0;
         }
 
+        bool idIsEvexZuContextSet() const
+        {
+            assert(IsApxZuCompatibleInstruction(_idIns));
+            return (_idEvexZuContext != 0);
+        }
+
         void idSetEvexNdContext()
         {
             assert(!idIsEvexNdContextSet());
             _idEvexNdContext = 1;
         }
 
+        void idSetEvexZuContext()
+        {
+            assert(!idIsEvexZuContextSet());
+            _idEvexZuContext = 1;
+        }
+
         bool idIsEvexNfContextSet() const
         {
             return _idEvexNfContext != 0;