SSE / SSE2 conversion intrinsics.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@27637 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 36ce4b0..e98d79e 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -479,11 +479,6 @@
}
// Conversion instructions
-def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (ops R32:$dst, FR32:$src),
- "cvtss2si {$src, $dst|$dst, $src}", []>;
-def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (ops R32:$dst, f32mem:$src),
- "cvtss2si {$src, $dst|$dst, $src}", []>;
-
def CVTTSS2SIrr: SSI<0x2C, MRMSrcReg, (ops R32:$dst, FR32:$src),
"cvttss2si {$src, $dst|$dst, $src}",
[(set R32:$dst, (fp_to_sint FR32:$src))]>;
@@ -514,6 +509,7 @@
def CVTSI2SDrm: SDI<0x2A, MRMSrcMem, (ops FR64:$dst, i32mem:$src),
"cvtsi2sd {$src, $dst|$dst, $src}",
[(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
+
// SSE2 instructions with XS prefix
def CVTSS2SDrr: I<0x5A, MRMSrcReg, (ops FR64:$dst, FR32:$src),
"cvtss2sd {$src, $dst|$dst, $src}",
@@ -524,7 +520,23 @@
[(set FR64:$dst, (fextend (loadf32 addr:$src)))]>, XS,
Requires<[HasSSE2]>;
-// Aliases to match intrinsics which expect XMM operand(s).
+// Match intrinsics which expect XMM operand(s).
+def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (ops R32:$dst, VR128:$src),
+ "cvtss2si {$src, $dst|$dst, $src}",
+ [(set R32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
+def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (ops R32:$dst, f32mem:$src),
+ "cvtss2si {$src, $dst|$dst, $src}",
+ [(set R32:$dst, (int_x86_sse_cvtss2si
+ (loadv4f32 addr:$src)))]>;
+
+// Aliases for intrinsics
+def Int_CVTTSS2SIrr: SSI<0x2C, MRMSrcReg, (ops R32:$dst, VR128:$src),
+ "cvttss2si {$src, $dst|$dst, $src}",
+ [(set R32:$dst, (int_x86_sse_cvttss2si VR128:$src))]>;
+def Int_CVTTSS2SIrm: SSI<0x2C, MRMSrcMem, (ops R32:$dst, f32mem:$src),
+ "cvttss2si {$src, $dst|$dst, $src}",
+ [(set R32:$dst, (int_x86_sse_cvttss2si
+ (loadv4f32 addr:$src)))]>;
def Int_CVTTSD2SIrr: SDI<0x2C, MRMSrcReg, (ops R32:$dst, VR128:$src),
"cvttsd2si {$src, $dst|$dst, $src}",
[(set R32:$dst, (int_x86_sse2_cvttsd2si VR128:$src))]>;
@@ -533,13 +545,18 @@
[(set R32:$dst, (int_x86_sse2_cvttsd2si
(loadv2f64 addr:$src)))]>;
-def CVTSD2SIrr: SDI<0x2D, MRMSrcReg, (ops R32:$dst, VR128:$src),
- "cvtsd2si {$src, $dst|$dst, $src}",
- [(set R32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
-def CVTSD2SIrm: SDI<0x2D, MRMSrcMem, (ops R32:$dst, f128mem:$src),
- "cvtsd2si {$src, $dst|$dst, $src}",
- [(set R32:$dst, (int_x86_sse2_cvtsd2si
- (loadv2f64 addr:$src)))]>;
+let isTwoAddress = 1 in {
+def Int_CVTSI2SSrr: SSI<0x2A, MRMSrcReg,
+ (ops VR128:$dst, VR128:$src1, R32:$src2),
+ "cvtsi2ss {$src2, $dst|$dst, $src2}",
+ [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
+ R32:$src2))]>;
+def Int_CVTSI2SSrm: SSI<0x2A, MRMSrcMem,
+ (ops VR128:$dst, VR128:$src1, i32mem:$src2),
+ "cvtsi2ss {$src2, $dst|$dst, $src2}",
+ [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
+ (loadi32 addr:$src2)))]>;
+}
// Comparison instructions
let isTwoAddress = 1 in {
@@ -816,16 +833,6 @@
MOVHLPS_shuffle_mask)))]>;
}
-// Conversion instructions
-def CVTPI2PSrr : PSI<0x2A, MRMSrcReg, (ops VR128:$dst, VR64:$src),
- "cvtpi2ps {$src, $dst|$dst, $src}", []>;
-def CVTPI2PSrm : PSI<0x2A, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
- "cvtpi2ps {$src, $dst|$dst, $src}", []>;
-def CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (ops VR128:$dst, VR64:$src),
- "cvtpi2pd {$src, $dst|$dst, $src}", []>;
-def CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (ops VR128:$dst, i64mem:$src),
- "cvtpi2pd {$src, $dst|$dst, $src}", []>;
-
// SSE2 instructions without OpSize prefix
def CVTDQ2PSrr : I<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
"cvtdq2ps {$src, $dst|$dst, $src}",
@@ -848,15 +855,6 @@
(bc_v4i32 (loadv2i64 addr:$src))))]>,
XS, Requires<[HasSSE2]>;
-def CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (ops VR64:$dst, VR128:$src),
- "cvtps2pi {$src, $dst|$dst, $src}", []>;
-def CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (ops VR64:$dst, f64mem:$src),
- "cvtps2pi {$src, $dst|$dst, $src}", []>;
-def CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (ops VR64:$dst, VR128:$src),
- "cvtpd2pi {$src, $dst|$dst, $src}", []>;
-def CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (ops VR64:$dst, f128mem:$src),
- "cvtpd2pi {$src, $dst|$dst, $src}", []>;
-
def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (ops VR128:$dst, VR128:$src),
"cvtps2dq {$src, $dst|$dst, $src}",
[(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
@@ -912,6 +910,52 @@
[(set VR128:$dst, (int_x86_sse2_cvtpd2ps
(loadv2f64 addr:$src)))]>;
+
+def CVTSD2SIrr: SDI<0x2D, MRMSrcReg, (ops R32:$dst, VR128:$src),
+ "cvtsd2si {$src, $dst|$dst, $src}",
+ [(set R32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
+def CVTSD2SIrm: SDI<0x2D, MRMSrcMem, (ops R32:$dst, f128mem:$src),
+ "cvtsd2si {$src, $dst|$dst, $src}",
+ [(set R32:$dst, (int_x86_sse2_cvtsd2si
+ (loadv2f64 addr:$src)))]>;
+
+// Match intrinsics which expect XMM operand(s).
+// Aliases for intrinsics
+let isTwoAddress = 1 in {
+def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
+ (ops VR128:$dst, VR128:$src1, R32:$src2),
+ "cvtsi2sd {$src2, $dst|$dst, $src2}",
+ [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
+ R32:$src2))]>;
+def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
+ (ops VR128:$dst, VR128:$src1, i32mem:$src2),
+ "cvtsi2sd {$src2, $dst|$dst, $src2}",
+ [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
+ (loadi32 addr:$src2)))]>;
+def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
+ (ops VR128:$dst, VR128:$src1, VR128:$src2),
+ "cvtsd2ss {$src2, $dst|$dst, $src2}",
+ [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
+ VR128:$src2))]>;
+def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
+ (ops VR128:$dst, VR128:$src1, f64mem:$src2),
+ "cvtsd2ss {$src2, $dst|$dst, $src2}",
+ [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
+ (loadv2f64 addr:$src2)))]>;
+def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
+ (ops VR128:$dst, VR128:$src1, VR128:$src2),
+ "cvtss2sd {$src2, $dst|$dst, $src2}",
+ [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
+ VR128:$src2))]>, XS,
+ Requires<[HasSSE2]>;
+def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
+ (ops VR128:$dst, VR128:$src1, f32mem:$src2),
+ "cvtss2sd {$src2, $dst|$dst, $src2}",
+ [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
+ (loadv4f32 addr:$src2)))]>, XS,
+ Requires<[HasSSE2]>;
+}
+
// Arithmetic
let isTwoAddress = 1 in {
let isCommutable = 1 in {