From 7f6eb639bd9563c53a6eb035d198f8d525ce3c0e Mon Sep 17 00:00:00 2001
From: Dale Johannesen <dalej@apple.com>
Date: Sat, 7 Aug 2010 00:33:42 +0000
Subject: [PATCH] Use sdmem and sse_load_f64 (etc.) for the vector form of
 CMPSD (etc.)  Matching a 128-bit memory operand is wrong, the instruction
 uses only 64 bits (same as ADDSD etc.)  8193553.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110491 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86InstrSSE.td        | 21 +++++++++++++--------
 test/CodeGen/X86/2010-08-06-cmpsd.ll | 27 +++++++++++++++++++++++++++
 2 files changed, 40 insertions(+), 8 deletions(-)
 create mode 100644 test/CodeGen/X86/2010-08-06-cmpsd.ll

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index ab5329d5eff..48de8899ebb 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -1163,31 +1163,36 @@ let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
                     "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}">, XD;
 }
 
-multiclass sse12_cmp_scalar_int<RegisterClass RC, X86MemOperand x86memop,
-                         Intrinsic Int, string asm> {
+multiclass sse12_cmp_scalar_int<RegisterClass RC, Operand memopr,
+                         ComplexPattern mem_cpat, Intrinsic Int, string asm> {
   def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
                       (ins VR128:$src1, VR128:$src, SSECC:$cc), asm,
                         [(set VR128:$dst, (Int VR128:$src1,
                                                VR128:$src, imm:$cc))]>;
   def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
-                      (ins VR128:$src1, f32mem:$src, SSECC:$cc), asm,
+                      (ins VR128:$src1, memopr:$src, SSECC:$cc), asm,
                         [(set VR128:$dst, (Int VR128:$src1,
-                                               (load addr:$src), imm:$cc))]>;
+                                              mem_cpat:$src, imm:$cc))]>;
 }
 
 // Aliases to match intrinsics which expect XMM operand(s).
+
 let isAsmParserOnly = 1 in {
-  defm Int_VCMPSS  : sse12_cmp_scalar_int<VR128, f32mem, int_x86_sse_cmp_ss,
+  defm Int_VCMPSS  : sse12_cmp_scalar_int<VR128, ssmem, sse_load_f32, 
+                       int_x86_sse_cmp_ss,
                        "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}">,
                        XS, VEX_4V;
-  defm Int_VCMPSD  : sse12_cmp_scalar_int<VR128, f64mem, int_x86_sse2_cmp_sd,
+  defm Int_VCMPSD  : sse12_cmp_scalar_int<VR128, sdmem, sse_load_f64,
+                       int_x86_sse2_cmp_sd,
                        "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}">,
                        XD, VEX_4V;
 }
 let Constraints = "$src1 = $dst" in {
-  defm Int_CMPSS  : sse12_cmp_scalar_int<VR128, f32mem, int_x86_sse_cmp_ss,
+  defm Int_CMPSS  : sse12_cmp_scalar_int<VR128, ssmem, sse_load_f32,
+                       int_x86_sse_cmp_ss,
                        "cmp${cc}ss\t{$src, $dst|$dst, $src}">, XS;
-  defm Int_CMPSD  : sse12_cmp_scalar_int<VR128, f64mem, int_x86_sse2_cmp_sd,
+  defm Int_CMPSD  : sse12_cmp_scalar_int<VR128, sdmem, sse_load_f64,
+                       int_x86_sse2_cmp_sd,
                        "cmp${cc}sd\t{$src, $dst|$dst, $src}">, XD;
 }
 
diff --git a/test/CodeGen/X86/2010-08-06-cmpsd.ll b/test/CodeGen/X86/2010-08-06-cmpsd.ll
new file mode 100644
index 00000000000..99b963f4537
--- /dev/null
+++ b/test/CodeGen/X86/2010-08-06-cmpsd.ll
@@ -0,0 +1,27 @@
+; RUN: llc < %s -mtriple=x86_64-applecl-darwin11 | FileCheck %s
+; 8193553
+
+define void @__math_kernel_Vectorized_wrapper(<4 x double> addrspace(1)* %a, <4 x double> addrspace(1)* %b, i64 addrspace(1)* %c, i64 addrspace(1)* %d) nounwind {
+entry.i:                                          ; preds = %entry.i, %loop
+; CHECK: math_kernel_Vectorized_wrapper
+; CHECK-NOT: cmpordsd (%rsi),
+  %0 = alloca i8
+  %1 = alloca i8
+  %2 = alloca i8
+  %tmp213.i = load <4 x double> addrspace(1)* %a ; <<4 x double>> [#uses=4]
+  %extract25.i = extractelement <4 x double> %tmp213.i, i32 1 ; <double> [#uses=1]
+  %tmp723.i = load <4 x double> addrspace(1)* %b ; <<4 x double>> [#uses=4]
+  %extract29.i = extractelement <4 x double> %tmp723.i, i32 1 ; <double> [#uses=1]
+  %tmp2.i26 = insertelement <2 x double> undef, double %extract25.i, i32 0 ; <<2 x double>> [#uses=1]
+  %tmp5.i27 = insertelement <2 x double> undef, double %extract29.i, i32 1 ; <<2 x double>> [#uses=1]
+  %cmpsd.i.i28 = call <2 x double> @llvm.x86.sse2.cmp.sd(<2 x double> %tmp2.i26, <2 x double> %tmp5.i27, i8 7) nounwind ; <<2 x double>> [#uses=1]
+  %3 = bitcast <2 x double> %cmpsd.i.i28 to <4 x i32> ; <<4 x i32>> [#uses=1]
+  %tmp12.i29 = extractelement <4 x i32> %3, i32 0 ; <i32> [#uses=1]
+  %and.i30 = and i32 %tmp12.i29, 1                ; <i32> [#uses=1]
+  %conv937.i36 = zext i32 %and.i30 to i64         ; <i64> [#uses=1]
+  store i64 %conv937.i36, i64 addrspace(1)* %d
+  ret void
+; CHECK: ret
+}
+
+declare <2 x double> @llvm.x86.sse2.cmp.sd(<2 x double>, <2 x double>, i8) nounwind readnone