High bits of movmskp{s|d} and pmovmskb are known zero. rdar://10247336

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@141371 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index fc2d5ce64e8bcca6f456f0bec1b943de85c6267c..a351fa5ca86aa8adec7d483843cbebabc3f5fc46 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -12333,6 +12333,33 @@ void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
     KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
                                        Mask.getBitWidth() - 1);
     break;
+  case ISD::INTRINSIC_WO_CHAIN: {
+    unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
+    unsigned NumLoBits = 0;
+    switch (IntId) {
+    default: break;
+    case Intrinsic::x86_sse_movmsk_ps:
+    case Intrinsic::x86_avx_movmsk_ps_256:
+    case Intrinsic::x86_sse2_movmsk_pd:
+    case Intrinsic::x86_avx_movmsk_pd_256:
+    case Intrinsic::x86_mmx_pmovmskb:
+    case Intrinsic::x86_sse2_pmovmskb_128: {
+      // High bits of movmskp{s|d}, pmovmskb are known zero.
+      switch (IntId) {
+        case Intrinsic::x86_sse_movmsk_ps:      NumLoBits = 4; break;
+        case Intrinsic::x86_avx_movmsk_ps_256:  NumLoBits = 8; break;
+        case Intrinsic::x86_sse2_movmsk_pd:     NumLoBits = 2; break;
+        case Intrinsic::x86_avx_movmsk_pd_256:  NumLoBits = 4; break;
+        case Intrinsic::x86_mmx_pmovmskb:       NumLoBits = 8; break;
+        case Intrinsic::x86_sse2_pmovmskb_128:  NumLoBits = 16; break;
+      }
+      KnownZero = APInt::getHighBitsSet(Mask.getBitWidth(),
+                                        Mask.getBitWidth() - NumLoBits);
+      break;
+    }
+    }
+    break;
+  }
   }
 }
 

diff --git a/test/CodeGen/X86/2011-05-31-movmsk.ll b/test/CodeGen/X86/2011-05-31-movmsk.ll
deleted file mode 100644 (file)
index 2b54d5c..0000000
--- a/test/CodeGen/X86/2011-05-31-movmsk.ll
+++ /dev/null
@@ -1,79 +0,0 @@
-; RUN: llc -mcpu=core2 < %s | FileCheck %s
-; ModuleID = '<stdin>'
-target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
-target triple = "x86_64-apple-macosx10.6.6"
-
-%0 = type { double }
-%union.anon = type { float }
-
-define i32 @double_signbit(double %d1) nounwind uwtable readnone ssp {
-entry:
-  %__x.addr.i = alloca double, align 8
-  %__u.i = alloca %0, align 8
-  %0 = bitcast double* %__x.addr.i to i8*
-  %1 = bitcast %0* %__u.i to i8*
-  store double %d1, double* %__x.addr.i, align 8
-  %__f.i = getelementptr inbounds %0* %__u.i, i64 0, i32 0
-  store double %d1, double* %__f.i, align 8
-  %tmp = bitcast double %d1 to i64
-; CHECK-NOT: shr
-; CHECK: movmskpd
-; CHECK-NEXT: and
-  %tmp1 = lshr i64 %tmp, 63
-  %shr.i = trunc i64 %tmp1 to i32
-  ret i32 %shr.i
-}
-
-define i32 @double_add_signbit(double %d1, double %d2) nounwind uwtable readnone ssp {
-entry:
-  %__x.addr.i = alloca double, align 8
-  %__u.i = alloca %0, align 8
-  %add = fadd double %d1, %d2
-  %0 = bitcast double* %__x.addr.i to i8*
-  %1 = bitcast %0* %__u.i to i8*
-  store double %add, double* %__x.addr.i, align 8
-  %__f.i = getelementptr inbounds %0* %__u.i, i64 0, i32 0
-  store double %add, double* %__f.i, align 8
-  %tmp = bitcast double %add to i64
-; CHECK-NOT: shr
-; CHECK: movmskpd
-; CHECK-NEXT: and
-  %tmp1 = lshr i64 %tmp, 63
-  %shr.i = trunc i64 %tmp1 to i32
-  ret i32 %shr.i
-}
-
-define i32 @float_signbit(float %f1) nounwind uwtable readnone ssp {
-entry:
-  %__x.addr.i = alloca float, align 4
-  %__u.i = alloca %union.anon, align 4
-  %0 = bitcast float* %__x.addr.i to i8*
-  %1 = bitcast %union.anon* %__u.i to i8*
-  store float %f1, float* %__x.addr.i, align 4
-  %__f.i = getelementptr inbounds %union.anon* %__u.i, i64 0, i32 0
-  store float %f1, float* %__f.i, align 4
-  %2 = bitcast float %f1 to i32
-; CHECK-NOT: shr
-; CHECK: movmskps
-; CHECK-NEXT: and
-  %shr.i = lshr i32 %2, 31
-  ret i32 %shr.i
-}
-
-define i32 @float_add_signbit(float %f1, float %f2) nounwind uwtable readnone ssp {
-entry:
-  %__x.addr.i = alloca float, align 4
-  %__u.i = alloca %union.anon, align 4
-  %add = fadd float %f1, %f2
-  %0 = bitcast float* %__x.addr.i to i8*
-  %1 = bitcast %union.anon* %__u.i to i8*
-  store float %add, float* %__x.addr.i, align 4
-  %__f.i = getelementptr inbounds %union.anon* %__u.i, i64 0, i32 0
-  store float %add, float* %__f.i, align 4
-  %2 = bitcast float %add to i32
-; CHECK-NOT: shr
-; CHECK: movmskps
-; CHECK-NEXT: and
-  %shr.i = lshr i32 %2, 31
-  ret i32 %shr.i
-}

diff --git a/test/CodeGen/X86/movmsk.ll b/test/CodeGen/X86/movmsk.ll
new file mode 100644 (file)
index 0000000..2368548
--- /dev/null
+++ b/test/CodeGen/X86/movmsk.ll
@@ -0,0 +1,110 @@
+; RUN: llc -mcpu=core2 < %s | FileCheck %s
+; ModuleID = '<stdin>'
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-apple-macosx10.6.6"
+
+%0 = type { double }
+%union.anon = type { float }
+
+define i32 @double_signbit(double %d1) nounwind uwtable readnone ssp {
+entry:
+  %__x.addr.i = alloca double, align 8
+  %__u.i = alloca %0, align 8
+  %0 = bitcast double* %__x.addr.i to i8*
+  %1 = bitcast %0* %__u.i to i8*
+  store double %d1, double* %__x.addr.i, align 8
+  %__f.i = getelementptr inbounds %0* %__u.i, i64 0, i32 0
+  store double %d1, double* %__f.i, align 8
+  %tmp = bitcast double %d1 to i64
+; CHECK-NOT: shr
+; CHECK: movmskpd
+; CHECK-NEXT: and
+  %tmp1 = lshr i64 %tmp, 63
+  %shr.i = trunc i64 %tmp1 to i32
+  ret i32 %shr.i
+}
+
+define i32 @double_add_signbit(double %d1, double %d2) nounwind uwtable readnone ssp {
+entry:
+  %__x.addr.i = alloca double, align 8
+  %__u.i = alloca %0, align 8
+  %add = fadd double %d1, %d2
+  %0 = bitcast double* %__x.addr.i to i8*
+  %1 = bitcast %0* %__u.i to i8*
+  store double %add, double* %__x.addr.i, align 8
+  %__f.i = getelementptr inbounds %0* %__u.i, i64 0, i32 0
+  store double %add, double* %__f.i, align 8
+  %tmp = bitcast double %add to i64
+; CHECK-NOT: shr
+; CHECK: movmskpd
+; CHECK-NEXT: and
+  %tmp1 = lshr i64 %tmp, 63
+  %shr.i = trunc i64 %tmp1 to i32
+  ret i32 %shr.i
+}
+
+define i32 @float_signbit(float %f1) nounwind uwtable readnone ssp {
+entry:
+  %__x.addr.i = alloca float, align 4
+  %__u.i = alloca %union.anon, align 4
+  %0 = bitcast float* %__x.addr.i to i8*
+  %1 = bitcast %union.anon* %__u.i to i8*
+  store float %f1, float* %__x.addr.i, align 4
+  %__f.i = getelementptr inbounds %union.anon* %__u.i, i64 0, i32 0
+  store float %f1, float* %__f.i, align 4
+  %2 = bitcast float %f1 to i32
+; CHECK-NOT: shr
+; CHECK: movmskps
+; CHECK-NEXT: and
+  %shr.i = lshr i32 %2, 31
+  ret i32 %shr.i
+}
+
+define i32 @float_add_signbit(float %f1, float %f2) nounwind uwtable readnone ssp {
+entry:
+  %__x.addr.i = alloca float, align 4
+  %__u.i = alloca %union.anon, align 4
+  %add = fadd float %f1, %f2
+  %0 = bitcast float* %__x.addr.i to i8*
+  %1 = bitcast %union.anon* %__u.i to i8*
+  store float %add, float* %__x.addr.i, align 4
+  %__f.i = getelementptr inbounds %union.anon* %__u.i, i64 0, i32 0
+  store float %add, float* %__f.i, align 4
+  %2 = bitcast float %add to i32
+; CHECK-NOT: shr
+; CHECK: movmskps
+; CHECK-NEXT: and
+  %shr.i = lshr i32 %2, 31
+  ret i32 %shr.i
+}
+
+; rdar://10247336
+; movmskp{s|d} only set low 4/2 bits, high bits are known zero
+
+define i32 @t1(<4 x float> %x, i32* nocapture %indexTable) nounwind uwtable readonly ssp {
+entry:
+; CHECK: t1:
+; CHECK: movmskps
+; CHECK-NOT: movslq
+  %0 = tail call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %x) nounwind
+  %idxprom = sext i32 %0 to i64
+  %arrayidx = getelementptr inbounds i32* %indexTable, i64 %idxprom
+  %1 = load i32* %arrayidx, align 4
+  ret i32 %1
+}
+
+define i32 @t2(<4 x float> %x, i32* nocapture %indexTable) nounwind uwtable readonly ssp {
+entry:
+; CHECK: t2:
+; CHECK: movmskpd
+; CHECK-NOT: movslq
+  %0 = bitcast <4 x float> %x to <2 x double>
+  %1 = tail call i32 @llvm.x86.sse2.movmsk.pd(<2 x double> %0) nounwind
+  %idxprom = sext i32 %1 to i64
+  %arrayidx = getelementptr inbounds i32* %indexTable, i64 %idxprom
+  %2 = load i32* %arrayidx, align 4
+  ret i32 %2
+}
+
+declare i32 @llvm.x86.sse2.movmsk.pd(<2 x double>) nounwind readnone
+declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>) nounwind readnone