vml[as].f32 cause stalls in following advanced SIMD instructions. Avoid using
them for scalar floating point operations for now.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85697 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 822950c..d22ec35 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -2841,13 +2841,16 @@
def : N3VDsPat<fmul, VMULfd_sfp>;
// Vector Multiply-Accumulate/Subtract used for single-precision FP
-let neverHasSideEffects = 1 in
-def VMLAfd_sfp : N3VDMulOps<0, 0, 0b00, 0b1101, 1, IIC_VMACD, "vmla.f32", v2f32,fmul,fadd>;
-def : N3VDMulOpsPat<fmul, fadd, VMLAfd_sfp>;
+// vml[as].f32 can cause 4-8 cycle stalls in following ASIMD instructions, so
+// we want to avoid them for now. e.g., alternating vmla/vadd instructions.
-let neverHasSideEffects = 1 in
-def VMLSfd_sfp : N3VDMulOps<0, 0, 0b10, 0b1101, 1, IIC_VMACD, "vmls.f32", v2f32,fmul,fsub>;
-def : N3VDMulOpsPat<fmul, fsub, VMLSfd_sfp>;
+//let neverHasSideEffects = 1 in
+//def VMLAfd_sfp : N3VDMulOps<0, 0, 0b00, 0b1101, 1, IIC_VMACD, "vmla.f32", v2f32,fmul,fadd>;
+//def : N3VDMulOpsPat<fmul, fadd, VMLAfd_sfp>;
+
+//let neverHasSideEffects = 1 in
+//def VMLSfd_sfp : N3VDMulOps<0, 0, 0b10, 0b1101, 1, IIC_VMACD, "vmls.f32", v2f32,fmul,fsub>;
+//def : N3VDMulOpsPat<fmul, fsub, VMLSfd_sfp>;
// Vector Absolute used for single-precision FP
let neverHasSideEffects = 1 in