[X86][AVX512] Add masked MOVS[S|D] patterns

[guyblank]

Summary

The only mask bit used by masked scalar MOV is the lower bit, so no need to and with 1 before passing the mask. Add patterns to recognize that.

[craig.topper]

lib/Target/X86/X86InstrAVX512.td

4005

Does this extend past 80 columns?

[guyblank]

lib/Target/X86/X86InstrAVX512.td

4005

you're right, thanks!

[guyblank]

fit into 80 columns

[zvi]

LGTM

lib/Target/X86/X86InstrAVX512.td

 
 defm : avx512_load_scalar_lowering<"VMOVSSZ", avx512vl_f32_info,
                    (v16i1 (bitconvert (i16 (trunc (and GR32:$mask, (i32 1)))))), GR32>;
 defm : avx512_load_scalar_lowering_subreg<"VMOVSSZ", avx512vl_f32_info,
                    (v16i1 (bitconvert (i16 (and GR16:$mask, (i16 1))))), GR16, sub_16bit>;
 defm : avx512_load_scalar_lowering_subreg<"VMOVSDZ", avx512vl_f64_info,
                    (v8i1 (bitconvert (i8 (and GR8:$mask, (i8 1))))), GR8, sub_8bit>;
 
+def : Pat<(f32 (X86selects (scalar_to_vector (and GR8:$mask, (i8 1))),
+                           (f32 FR32X:$src1), (f32 FR32X:$src2))),
+          (COPY_TO_REGCLASS
+            (VMOVSSZrrk (COPY_TO_REGCLASS FR32X:$src2, VR128X),
+                        (COPY_TO_REGCLASS (i32 (INSERT_SUBREG (IMPLICIT_DEF),
+                          GR8:$mask, sub_8bit)), VK1WM),
+            (v4f32 (IMPLICIT_DEF)), FR32X:$src1), FR32X)>;
+
 def : Pat<(f32 (X86selects VK1WM:$mask, (f32 FR32X:$src1), (f32 FR32X:$src2))),
           (COPY_TO_REGCLASS (VMOVSSZrrk (COPY_TO_REGCLASS FR32X:$src2, VR128X),
            VK1WM:$mask, (v4f32 (IMPLICIT_DEF)), FR32X:$src1), FR32X)>;
 
+def : Pat<(f64 (X86selects (scalar_to_vector (and GR8:$mask, (i8 1))),
+                           (f64 FR64X:$src1), (f64 FR64X:$src2))),
+          (COPY_TO_REGCLASS
+            (VMOVSDZrrk (COPY_TO_REGCLASS FR64X:$src2, VR128X),
+                        (COPY_TO_REGCLASS (i32 (INSERT_SUBREG (IMPLICIT_DEF),
+                          GR8:$mask, sub_8bit)), VK1WM),
+            (v2f64 (IMPLICIT_DEF)), FR64X:$src1), FR64X)>;
+
 def : Pat<(f64 (X86selects VK1WM:$mask, (f64 FR64X:$src1), (f64 FR64X:$src2))),
           (COPY_TO_REGCLASS (VMOVSDZrrk (COPY_TO_REGCLASS FR64X:$src2, VR128X),
            VK1WM:$mask, (v2f64 (IMPLICIT_DEF)), FR64X:$src1), FR64X)>;
 
 def : Pat<(int_x86_avx512_mask_store_ss addr:$dst, VR128X:$src, GR8:$mask),
           (VMOVSSZmrk addr:$dst, (COPY_TO_REGCLASS (i32 (INSERT_SUBREG (IMPLICIT_DEF), GR8:$mask, sub_8bit)), VK1WM),
            (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
 

test/CodeGen/X86/avx512-load-store.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -O2 -mattr=avx512f -mtriple=x86_64-unknown | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK64
 ; RUN: llc < %s -O2 -mattr=avx512f -mtriple=i386-unknown | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK32
 
 define <4 x float> @test_mm_mask_move_ss(<4 x float> %__W, i8 zeroext %__U, <4 x float> %__A, <4 x float> %__B) local_unnamed_addr #0 {
 ; CHECK64-LABEL: test_mm_mask_move_ss:
 ; CHECK64:       # BB#0: # %entry
 ; CHECK64-NEXT:    kmovw %edi, %k1
 ; CHECK64-NEXT:    vmovss %xmm2, %xmm1, %xmm0 {%k1}
 ; CHECK64-NEXT:    retq
 ;
 ; CHECK32-LABEL: test_mm_mask_move_ss:
 ; CHECK32:       # BB#0: # %entry
 ; CHECK32-NEXT:    movb {{[0-9]+}}(%esp), %al
-; CHECK32-NEXT:    andb $1, %al
 ; CHECK32-NEXT:    kmovw %eax, %k1
 ; CHECK32-NEXT:    vmovss %xmm2, %xmm0, %xmm0 {%k1}
 ; CHECK32-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; CHECK32-NEXT:    retl
 entry:
   %0 = and i8 %__U, 1
   %tobool.i = icmp ne i8 %0, 0
   %__B.elt.i = extractelement <4 x float> %__B, i32 0
   %__W.elt.i = extractelement <4 x float> %__W, i32 0
   %vecext1.i = select i1 %tobool.i, float %__B.elt.i, float %__W.elt.i
   %vecins.i = insertelement <4 x float> %__A, float %vecext1.i, i32 0
   ret <4 x float> %vecins.i
 }
 
 define <4 x float> @test_mm_maskz_move_ss(i8 zeroext %__U, <4 x float> %__A, <4 x float> %__B) local_unnamed_addr #0 {
 ; CHECK64-LABEL: test_mm_maskz_move_ss:
 ; CHECK64:       # BB#0: # %entry
 ; CHECK64-NEXT:    kmovw %edi, %k1
 ; CHECK64-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} {z}
 ; CHECK64-NEXT:    retq
 ;
 ; CHECK32-LABEL: test_mm_maskz_move_ss:
 ; CHECK32:       # BB#0: # %entry
 ; CHECK32-NEXT:    movb {{[0-9]+}}(%esp), %al
-; CHECK32-NEXT:    andb $1, %al
 ; CHECK32-NEXT:    kmovw %eax, %k1
 ; CHECK32-NEXT:    vxorps %xmm2, %xmm2, %xmm2
 ; CHECK32-NEXT:    vmovss %xmm1, %xmm0, %xmm2 {%k1}
 ; CHECK32-NEXT:    vmovss {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3]
 ; CHECK32-NEXT:    retl
 entry:
   %0 = and i8 %__U, 1
   %tobool.i = icmp ne i8 %0, 0
   %vecext.i = extractelement <4 x float> %__B, i32 0
   %cond.i = select i1 %tobool.i, float %vecext.i, float 0.000000e+00
   %vecins.i = insertelement <4 x float> %__A, float %cond.i, i32 0
   ret <4 x float> %vecins.i
 }
 
 define <2 x double> @test_mm_mask_move_sd(<2 x double> %__W, i8 zeroext %__U, <2 x double> %__A, <2 x double> %__B) local_unnamed_addr #0 {
 ; CHECK64-LABEL: test_mm_mask_move_sd:
 ; CHECK64:       # BB#0: # %entry
 ; CHECK64-NEXT:    kmovw %edi, %k1
 ; CHECK64-NEXT:    vmovsd %xmm2, %xmm1, %xmm0 {%k1}
 ; CHECK64-NEXT:    retq
 ;
 ; CHECK32-LABEL: test_mm_mask_move_sd:
 ; CHECK32:       # BB#0: # %entry
 ; CHECK32-NEXT:    movb {{[0-9]+}}(%esp), %al
-; CHECK32-NEXT:    andb $1, %al
 ; CHECK32-NEXT:    kmovw %eax, %k1
 ; CHECK32-NEXT:    vmovsd %xmm2, %xmm0, %xmm0 {%k1}
 ; CHECK32-NEXT:    vmovsd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
 ; CHECK32-NEXT:    retl
 entry:
   %0 = and i8 %__U, 1
   %tobool.i = icmp ne i8 %0, 0
   %__B.elt.i = extractelement <2 x double> %__B, i32 0
   %__W.elt.i = extractelement <2 x double> %__W, i32 0
   %vecext1.i = select i1 %tobool.i, double %__B.elt.i, double %__W.elt.i
   %vecins.i = insertelement <2 x double> %__A, double %vecext1.i, i32 0
   ret <2 x double> %vecins.i
 }
 
 define <2 x double> @test_mm_maskz_move_sd(i8 zeroext %__U, <2 x double> %__A, <2 x double> %__B) local_unnamed_addr #0 {
 ; CHECK64-LABEL: test_mm_maskz_move_sd:
 ; CHECK64:       # BB#0: # %entry
 ; CHECK64-NEXT:    kmovw %edi, %k1
 ; CHECK64-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} {z}
 ; CHECK64-NEXT:    retq
 ;
 ; CHECK32-LABEL: test_mm_maskz_move_sd:
 ; CHECK32:       # BB#0: # %entry
 ; CHECK32-NEXT:    movb {{[0-9]+}}(%esp), %al
-; CHECK32-NEXT:    andb $1, %al
 ; CHECK32-NEXT:    kmovw %eax, %k1
 ; CHECK32-NEXT:    vxorpd %xmm2, %xmm2, %xmm2
 ; CHECK32-NEXT:    vmovsd %xmm1, %xmm0, %xmm2 {%k1}
 ; CHECK32-NEXT:    vmovsd {{.*#+}} xmm0 = xmm2[0],xmm0[1]
 ; CHECK32-NEXT:    retl
 entry:
   %0 = and i8 %__U, 1
   %tobool.i = icmp ne i8 %0, 0

test/CodeGen/X86/avx512-select.ll

 ; X86-NEXT:    leal {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    leal {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    cmovnel %eax, %ecx
 ; X86-NEXT:    fldl (%ecx)
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: pr30561_f64:
 ; X64:       # BB#0:
-; X64-NEXT:    andb $1, %dil
 ; X64-NEXT:    kmovw %edi, %k1
 ; X64-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1}
 ; X64-NEXT:    retq
   %cond = select i1 %c, double %a, double %b
   ret double %cond
 }
 
 define float @pr30561_f32(float %b, float %a, i1 %c) {
 ; X86-LABEL: pr30561_f32:
 ; X86:       # BB#0:
 ; X86-NEXT:    testb $1, {{[0-9]+}}(%esp)
 ; X86-NEXT:    leal {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    leal {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    cmovnel %eax, %ecx
 ; X86-NEXT:    flds (%ecx)
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: pr30561_f32:
 ; X64:       # BB#0:
-; X64-NEXT:    andb $1, %dil
 ; X64-NEXT:    kmovw %edi, %k1
 ; X64-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1}
 ; X64-NEXT:    retq
   %cond = select i1 %c, float %a, float %b
   ret float %cond
 }
 
 define <16 x i16> @pr31515(<16 x i1> %a, <16 x i1> %b, <16 x i16> %c) nounwind {