[PATCH] [X86][SSE] Vector integer/float conversion memory folding (cvttps2dq / cvttpd2dq)

[Quentin Colombet]

Hi Simon,

Thanks for having split the patches.

See my comments inlined.

Cheers,
-Quentin

================
Comment at: lib/Target/X86/X86InstrInfo.cpp:936
@@ -931,5 +935,3 @@
     { X86::VCVTSS2SDrr,       X86::VCVTSS2SDrm,        0 },
     { X86::Int_VCVTSS2SDrr,   X86::Int_VCVTSS2SDrm,    0 },
     { X86::VRSQRTSSr,         X86::VRSQRTSSm,          0 },
----------------
While you are fixing this kind of issue, could you double check the opcode in there?
All the CVTs look suspicious to me.

================
Comment at: test/CodeGen/X86/avx1-stack-reload-folding.ll:22
@@ +21,3 @@
+  %3 = fptosi <64 x double> %1 to <64 x i32>
+  %4 = fptosi <64 x double> %2 to <64 x i32>
+  %5 = or <64 x i32> %3, %4
----------------
Could you trigger the transformation with something simpler (like load, cvt, store, with both addresses as argument)?
Maybe by using fast-isel?

http://reviews.llvm.org/D6001