Switch to ymm after zmm in genZeroInitFrameUsingBlockInit (dotnet#115981)

Author: EgorBo

src/coreclr/jit/codegenxarch.cpp

@@ -11253,29 +11253,45 @@ void CodeGen::genZeroInitFrameUsingBlockInit(int untrLclHi, int untrLclLo, regNu
             //   movups  xmmword ptr [ebp/esp-OFFS], xmm4
             //   ...
             //   movups  xmmword ptr [ebp/esp-OFFS], xmm4
-            //   mov      qword ptr [ebp/esp-OFFS], rax
-            //
+
             // NOTE: it implicitly zeroes YMM4 and ZMM4 as well.
             emit->emitIns_SIMD_R_R_R(INS_xorps, EA_16BYTE, zeroSIMDReg, zeroSIMDReg, zeroSIMDReg, INS_OPTS_NONE);
 
-            int i = 0;
-            if (maxSimdSize > XMM_REGSIZE_BYTES)
+            assert((blkSize % XMM_REGSIZE_BYTES) == 0);
+
+            int regSize      = (int)compiler->roundDownSIMDSize(blkSize);
+            int lenRemaining = blkSize;
+            while (lenRemaining > 0)
             {
-                for (; i <= blkSize - maxSimdSize; i += maxSimdSize)
+                // Overlap with the previously zeroed memory if we can clear the remainder
+                // with just single store. Example: say we have 112 bytes to clear:
+                //
+                // Option 1 (no overlapping):
+                //   movups  zmmword ptr [+0]
+                //   movups  ymmword ptr [+64]
+                //   movups  xmmword ptr [+96]
+                //
+                // Option 2 (overlapping):
+                //   movups  zmmword ptr [+0]
+                //   movups  zmmword ptr [+48]
+                //
+                if ((regSize > lenRemaining) && !isPow2(lenRemaining))
                 {
-                    // We previously aligned data to 16 bytes which might not be aligned to maxSimdSize
-                    emit->emitIns_AR_R(simdUnalignedMovIns(), EA_ATTR(maxSimdSize), zeroSIMDReg, frameReg,
-                                       alignedLclLo + i);
+                    lenRemaining = regSize;
                 }
-                // Remainder will be handled by the xmm loop below
-            }
 
-            for (; i < blkSize; i += XMM_REGSIZE_BYTES)
-            {
-                emit->emitIns_AR_R(simdMov, EA_ATTR(XMM_REGSIZE_BYTES), zeroSIMDReg, frameReg, alignedLclLo + i);
-            }
+                // Use the largest SIMD register size that fits in the remaining length
+                regSize = (int)compiler->roundDownSIMDSize(lenRemaining);
+                assert(regSize >= XMM_REGSIZE_BYTES);
+
+                // frameReg is definitely not known to be 32B/64B aligned -> switch to unaligned movs
+                instruction ins    = regSize > XMM_REGSIZE_BYTES ? simdUnalignedMovIns() : simdMov;
+                const int   offset = blkSize - lenRemaining;
+                emit->emitIns_AR_R(ins, EA_ATTR(regSize), zeroSIMDReg, frameReg, alignedLclLo + offset);
 
-            assert(i == blkSize);
+                lenRemaining -= regSize;
+            }
+            assert(lenRemaining == 0);
         }
         else
         {