[X86] Use sse_load_f32/f64 in patterns for the memory forms of VRNDSCALESS/SD.

llvm-svn: 318009

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 96300b5..8d677be 100644 (file)
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -7672,9 +7672,8 @@ avx512_rndscale_scalar<bits<8> opc, string OpcodeStr, X86VectorVTInfo _> {
                          (ins _.RC:$src1, _.ScalarMemOp:$src2, i32u8imm:$src3),
                          OpcodeStr,
                          "$src3, $src2, $src1", "$src1, $src2, $src3",
-                         (_.VT (X86RndScales (_.VT _.RC:$src1),
-                          (_.VT (scalar_to_vector (_.ScalarLdFrag addr:$src2))),
-                          (i32 imm:$src3)))>;
+                         (_.VT (X86RndScales _.RC:$src1,
+                                _.ScalarIntMemCPat:$src2, (i32 imm:$src3)))>;
 
   let isCodeGenOnly = 1, hasSideEffects = 0 in {
     def r : I<opc, MRMSrcReg, (outs _.FRC:$dst),

diff --git a/llvm/test/CodeGen/X86/sse41-intrinsics-x86.ll b/llvm/test/CodeGen/X86/sse41-intrinsics-x86.ll
index 6c08671..eec4ef9 100644 (file)
--- a/llvm/test/CodeGen/X86/sse41-intrinsics-x86.ll
+++ b/llvm/test/CodeGen/X86/sse41-intrinsics-x86.ll
@@ -524,8 +524,7 @@ define <2 x double> @test_x86_sse41_round_sd_load(<2 x double> %a0, <2 x double>
 ; SKX-LABEL: test_x86_sse41_round_sd_load:
 ; SKX:       ## BB#0:
 ; SKX-NEXT:    movl {{[0-9]+}}(%esp), %eax ## encoding: [0x8b,0x44,0x24,0x04]
-; SKX-NEXT:    vmovapd (%eax), %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0x08]
-; SKX-NEXT:    vrndscalesd $7, %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf3,0xfd,0x08,0x0b,0xc1,0x07]
+; SKX-NEXT:    vrndscalesd $7, (%eax), %xmm0, %xmm0 ## encoding: [0x62,0xf3,0xfd,0x08,0x0b,0x00,0x07]
 ; SKX-NEXT:    retl ## encoding: [0xc3]
   %a1b = load <2 x double>, <2 x double>* %a1
   %res = call <2 x double> @llvm.x86.sse41.round.sd(<2 x double> %a0, <2 x double> %a1b, i32 7) ; <<2 x double>> [#uses=1]

diff --git a/llvm/test/CodeGen/X86/sse41-schedule.ll b/llvm/test/CodeGen/X86/sse41-schedule.ll
index a9bdf89..0eb3728 100644 (file)
--- a/llvm/test/CodeGen/X86/sse41-schedule.ll
+++ b/llvm/test/CodeGen/X86/sse41-schedule.ll
@@ -3152,8 +3152,7 @@ define <2 x double> @test_roundsd(<2 x double> %a0, <2 x double> %a1, <2 x doubl
 ; SKX-LABEL: test_roundsd:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vrndscalesd $7, %xmm1, %xmm0, %xmm1 # sched: [8:0.67]
-; SKX-NEXT:    vmovapd (%rdi), %xmm2 # sched: [6:0.50]
-; SKX-NEXT:    vrndscalesd $7, %xmm2, %xmm0, %xmm0 # sched: [8:0.67]
+; SKX-NEXT:    vrndscalesd $7, (%rdi), %xmm0, %xmm0 # sched: [14:0.67]
 ; SKX-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
@@ -3226,8 +3225,7 @@ define <4 x float> @test_roundss(<4 x float> %a0, <4 x float> %a1, <4 x float> *
 ; SKX-LABEL: test_roundss:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vrndscaless $7, %xmm1, %xmm0, %xmm1 # sched: [8:0.67]
-; SKX-NEXT:    vmovaps (%rdi), %xmm2 # sched: [6:0.50]
-; SKX-NEXT:    vrndscaless $7, %xmm2, %xmm0, %xmm0 # sched: [8:0.67]
+; SKX-NEXT:    vrndscaless $7, (%rdi), %xmm0, %xmm0 # sched: [14:0.67]
 ; SKX-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # sched: [4:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;