From: Craig Topper <craig.topper@intel.com>
Date: Mon, 9 Sep 2019 17:48:05 +0000 (+0000)
Subject: [X86] Allow _MM_FROUND_CUR_DIRECTION and _MM_FROUND_NO_EXC to be used together on... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=a503e8648e12b2328f7fb931a08664857123e224;p=llvm

[X86] Allow _MM_FROUND_CUR_DIRECTION and _MM_FROUND_NO_EXC to be used together on instructions that only support SAE and not embedded rounding.

Current for SAE instructions we only allow _MM_FROUND_CUR_DIRECTION(bit 2) or _MM_FROUND_NO_EXC(bit 3) to be used as the immediate passed to the inrinsics. But these instructions don't perform rounding so _MM_FROUND_CUR_DIRECTION is just sort of a default placeholder when you don't want to suppress exceptions. Using _MM_FROUND_NO_EXC by itself is really bit equivalent to (_MM_FROUND_NO_EXC | _MM_FROUND_TO_NEAREST_INT) since _MM_FROUND_TO_NEAREST_INT is 0. Since we aren't rounding on these instructions we should also accept (_MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC) as equivalent to (_MM_FROUND_NO_EXC). icc allows this, but gcc does not.

Differential Revision: https://reviews.llvm.org/D67289

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@371430 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 0a8219214f4..2195f40c247 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -22706,8 +22706,16 @@ SDValue X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
     return false;
   };
   auto isRoundModeSAE = [](SDValue Rnd) {
-    if (auto *C = dyn_cast<ConstantSDNode>(Rnd))
-      return C->getAPIntValue() == X86::STATIC_ROUNDING::NO_EXC;
+    if (auto *C = dyn_cast<ConstantSDNode>(Rnd)) {
+      unsigned RC = C->getZExtValue();
+      if (RC & X86::STATIC_ROUNDING::NO_EXC) {
+        // Clear the NO_EXC bit and check remaining bits.
+        RC ^= X86::STATIC_ROUNDING::NO_EXC;
+        // As a convenience we allow no other bits or explicitly
+        // current direction.
+        return RC == 0 || RC == X86::STATIC_ROUNDING::CUR_DIRECTION;
+      }
+    }
 
     return false;
   };
diff --git a/test/CodeGen/X86/avx512-intrinsics.ll b/test/CodeGen/X86/avx512-intrinsics.ll
index 6f0aba31cf1..b2d6ce4dfc9 100644
--- a/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/test/CodeGen/X86/avx512-intrinsics.ll
@@ -755,7 +755,7 @@ define <8 x double> @test_getexp_round_pd_512(<8 x double> %a0) {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vgetexppd {sae}, %zmm0, %zmm0
 ; CHECK-NEXT:    ret{{[l|q]}}
-  %res = call <8 x double> @llvm.x86.avx512.mask.getexp.pd.512(<8 x double> %a0,  <8 x double> zeroinitializer, i8 -1, i32 8)
+  %res = call <8 x double> @llvm.x86.avx512.mask.getexp.pd.512(<8 x double> %a0,  <8 x double> zeroinitializer, i8 -1, i32 12)
   ret <8 x double> %res
 }
 declare <8 x double> @llvm.x86.avx512.mask.getexp.pd.512(<8 x double>, <8 x double>, i8, i32) nounwind readnone