From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Wed, 1 May 2019 11:17:11 +0000 (+0000)
Subject: [X86][SSE] Add scalar horizontal add/sub tests for element extractions from upper... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=d99b5efaf03111829450409435449ac512473cab;p=llvm

[X86][SSE] Add scalar horizontal add/sub tests for element extractions from upper lanes

As suggested on D61263

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@359671 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/test/CodeGen/X86/haddsub.ll b/test/CodeGen/X86/haddsub.ll
index e9a74a32558..b5cd15d5fc2 100644
--- a/test/CodeGen/X86/haddsub.ll
+++ b/test/CodeGen/X86/haddsub.ll
@@ -940,6 +940,29 @@ define float @extract_extract23_v8f32_fadd_f32(<8 x float> %x) {
   ret float %x01
 }
 
+define float @extract_extract67_v8f32_fadd_f32(<8 x float> %x) {
+; SSE3-LABEL: extract_extract67_v8f32_fadd_f32:
+; SSE3:       # %bb.0:
+; SSE3-NEXT:    movaps %xmm1, %xmm0
+; SSE3-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
+; SSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE3-NEXT:    addss %xmm1, %xmm0
+; SSE3-NEXT:    retq
+;
+; AVX-LABEL: extract_extract67_v8f32_fadd_f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX-NEXT:    vaddss %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+  %x0 = extractelement <8 x float> %x, i32 6
+  %x1 = extractelement <8 x float> %x, i32 7
+  %x01 = fadd float %x0, %x1
+  ret float %x01
+}
+
 define float @extract_extract01_v8f32_fadd_f32_commute(<8 x float> %x) {
 ; SSE3-SLOW-LABEL: extract_extract01_v8f32_fadd_f32_commute:
 ; SSE3-SLOW:       # %bb.0:
@@ -992,6 +1015,29 @@ define float @extract_extract23_v8f32_fadd_f32_commute(<8 x float> %x) {
   ret float %x01
 }
 
+define float @extract_extract67_v8f32_fadd_f32_commute(<8 x float> %x) {
+; SSE3-LABEL: extract_extract67_v8f32_fadd_f32_commute:
+; SSE3:       # %bb.0:
+; SSE3-NEXT:    movaps %xmm1, %xmm0
+; SSE3-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
+; SSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE3-NEXT:    addss %xmm1, %xmm0
+; SSE3-NEXT:    retq
+;
+; AVX-LABEL: extract_extract67_v8f32_fadd_f32_commute:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+  %x0 = extractelement <8 x float> %x, i32 6
+  %x1 = extractelement <8 x float> %x, i32 7
+  %x01 = fadd float %x1, %x0
+  ret float %x01
+}
+
 define double @extract_extract01_v4f64_fadd_f64(<4 x double> %x) {
 ; SSE3-SLOW-LABEL: extract_extract01_v4f64_fadd_f64:
 ; SSE3-SLOW:       # %bb.0:
@@ -1024,6 +1070,33 @@ define double @extract_extract01_v4f64_fadd_f64(<4 x double> %x) {
   ret double %x01
 }
 
+define double @extract_extract23_v4f64_fadd_f64(<4 x double> %x) {
+; SSE3-SLOW-LABEL: extract_extract23_v4f64_fadd_f64:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
+; SSE3-SLOW-NEXT:    addsd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
+;
+; SSE3-FAST-LABEL: extract_extract23_v4f64_fadd_f64:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    movapd %xmm1, %xmm0
+; SSE3-FAST-NEXT:    haddpd %xmm1, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-LABEL: extract_extract23_v4f64_fadd_f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+  %x0 = extractelement <4 x double> %x, i32 2
+  %x1 = extractelement <4 x double> %x, i32 3
+  %x01 = fadd double %x0, %x1
+  ret double %x01
+}
+
 define double @extract_extract01_v4f64_fadd_f64_commute(<4 x double> %x) {
 ; SSE3-SLOW-LABEL: extract_extract01_v4f64_fadd_f64_commute:
 ; SSE3-SLOW:       # %bb.0:
@@ -1056,6 +1129,33 @@ define double @extract_extract01_v4f64_fadd_f64_commute(<4 x double> %x) {
   ret double %x01
 }
 
+define double @extract_extract23_v4f64_fadd_f64_commute(<4 x double> %x) {
+; SSE3-SLOW-LABEL: extract_extract23_v4f64_fadd_f64_commute:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
+; SSE3-SLOW-NEXT:    addsd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
+;
+; SSE3-FAST-LABEL: extract_extract23_v4f64_fadd_f64_commute:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    movapd %xmm1, %xmm0
+; SSE3-FAST-NEXT:    haddpd %xmm1, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-LABEL: extract_extract23_v4f64_fadd_f64_commute:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-NEXT:    vaddsd %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+  %x0 = extractelement <4 x double> %x, i32 2
+  %x1 = extractelement <4 x double> %x, i32 3
+  %x01 = fadd double %x1, %x0
+  ret double %x01
+}
+
 define float @extract_extract01_v8f32_fsub_f32(<8 x float> %x) {
 ; SSE3-SLOW-LABEL: extract_extract01_v8f32_fsub_f32:
 ; SSE3-SLOW:       # %bb.0:
@@ -1109,6 +1209,33 @@ define float @extract_extract23_v8f32_fsub_f32(<8 x float> %x) {
   ret float %x01
 }
 
+define float @extract_extract45_v8f32_fsub_f32(<8 x float> %x) {
+; SSE3-SLOW-LABEL: extract_extract45_v8f32_fsub_f32:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movaps %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm1[1,1,3,3]
+; SSE3-SLOW-NEXT:    subss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
+;
+; SSE3-FAST-LABEL: extract_extract45_v8f32_fsub_f32:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    movaps %xmm1, %xmm0
+; SSE3-FAST-NEXT:    hsubps %xmm1, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-LABEL: extract_extract45_v8f32_fsub_f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-NEXT:    vsubss %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+  %x0 = extractelement <8 x float> %x, i32 4
+  %x1 = extractelement <8 x float> %x, i32 5
+  %x01 = fsub float %x0, %x1
+  ret float %x01
+}
+
 ; Negative test...or get hoppy and negate?
 
 define float @extract_extract01_v8f32_fsub_f32_commute(<8 x float> %x) {
diff --git a/test/CodeGen/X86/phaddsub-extract.ll b/test/CodeGen/X86/phaddsub-extract.ll
index 606784e250d..17bc6f3b58f 100644
--- a/test/CodeGen/X86/phaddsub-extract.ll
+++ b/test/CodeGen/X86/phaddsub-extract.ll
@@ -496,6 +496,30 @@ define i32 @extract_extract23_v8i32_add_i32(<8 x i32> %x) {
   ret i32 %x01
 }
 
+define i32 @extract_extract67_v8i32_add_i32(<8 x i32> %x) {
+; SSE3-LABEL: extract_extract67_v8i32_add_i32:
+; SSE3:       # %bb.0:
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
+; SSE3-NEXT:    movd %xmm0, %ecx
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[3,1,2,3]
+; SSE3-NEXT:    movd %xmm0, %eax
+; SSE3-NEXT:    addl %ecx, %eax
+; SSE3-NEXT:    retq
+;
+; AVX-LABEL: extract_extract67_v8i32_add_i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vextractps $2, %xmm0, %ecx
+; AVX-NEXT:    vextractps $3, %xmm0, %eax
+; AVX-NEXT:    addl %ecx, %eax
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+  %x0 = extractelement <8 x i32> %x, i32 6
+  %x1 = extractelement <8 x i32> %x, i32 7
+  %x01 = add i32 %x0, %x1
+  ret i32 %x01
+}
+
 define i32 @extract_extract01_v8i32_add_i32_commute(<8 x i32> %x) {
 ; SSE3-SLOW-LABEL: extract_extract01_v8i32_add_i32_commute:
 ; SSE3-SLOW:       # %bb.0:
@@ -554,6 +578,30 @@ define i32 @extract_extract23_v8i32_add_i32_commute(<8 x i32> %x) {
   ret i32 %x01
 }
 
+define i32 @extract_extract67_v8i32_add_i32_commute(<8 x i32> %x) {
+; SSE3-LABEL: extract_extract67_v8i32_add_i32_commute:
+; SSE3:       # %bb.0:
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
+; SSE3-NEXT:    movd %xmm0, %ecx
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[3,1,2,3]
+; SSE3-NEXT:    movd %xmm0, %eax
+; SSE3-NEXT:    addl %ecx, %eax
+; SSE3-NEXT:    retq
+;
+; AVX-LABEL: extract_extract67_v8i32_add_i32_commute:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vextractps $2, %xmm0, %ecx
+; AVX-NEXT:    vextractps $3, %xmm0, %eax
+; AVX-NEXT:    addl %ecx, %eax
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+  %x0 = extractelement <8 x i32> %x, i32 6
+  %x1 = extractelement <8 x i32> %x, i32 7
+  %x01 = add i32 %x1, %x0
+  ret i32 %x01
+}
+
 define i16 @extract_extract01_v16i16_add_i16(<16 x i16> %x) {
 ; SSE3-SLOW-LABEL: extract_extract01_v16i16_add_i16:
 ; SSE3-SLOW:       # %bb.0:
@@ -615,6 +663,87 @@ define i16 @extract_extract23_v16i16_add_i16(<16 x i16> %x) {
   ret i16 %x01
 }
 
+define i16 @extract_extract89_v16i16_add_i16(<16 x i16> %x) {
+; SSE3-SLOW-LABEL: extract_extract89_v16i16_add_i16:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movd %xmm1, %ecx
+; SSE3-SLOW-NEXT:    pextrw $1, %xmm1, %eax
+; SSE3-SLOW-NEXT:    addl %ecx, %eax
+; SSE3-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
+; SSE3-SLOW-NEXT:    retq
+;
+; SSE3-FAST-LABEL: extract_extract89_v16i16_add_i16:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    phaddw %xmm1, %xmm1
+; SSE3-FAST-NEXT:    movd %xmm1, %eax
+; SSE3-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
+; SSE3-FAST-NEXT:    retq
+;
+; AVX1-SLOW-LABEL: extract_extract89_v16i16_add_i16:
+; AVX1-SLOW:       # %bb.0:
+; AVX1-SLOW-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-SLOW-NEXT:    vmovd %xmm0, %ecx
+; AVX1-SLOW-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX1-SLOW-NEXT:    addl %ecx, %eax
+; AVX1-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX1-SLOW-NEXT:    vzeroupper
+; AVX1-SLOW-NEXT:    retq
+;
+; AVX1-FAST-LABEL: extract_extract89_v16i16_add_i16:
+; AVX1-FAST:       # %bb.0:
+; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-FAST-NEXT:    vmovd %xmm0, %ecx
+; AVX1-FAST-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX1-FAST-NEXT:    addl %ecx, %eax
+; AVX1-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX1-FAST-NEXT:    vzeroupper
+; AVX1-FAST-NEXT:    retq
+;
+; AVX2-SLOW-LABEL: extract_extract89_v16i16_add_i16:
+; AVX2-SLOW:       # %bb.0:
+; AVX2-SLOW-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX2-SLOW-NEXT:    vmovd %xmm0, %ecx
+; AVX2-SLOW-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX2-SLOW-NEXT:    addl %ecx, %eax
+; AVX2-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX2-SLOW-NEXT:    vzeroupper
+; AVX2-SLOW-NEXT:    retq
+;
+; AVX2-FAST-LABEL: extract_extract89_v16i16_add_i16:
+; AVX2-FAST:       # %bb.0:
+; AVX2-FAST-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX2-FAST-NEXT:    vmovd %xmm0, %ecx
+; AVX2-FAST-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX2-FAST-NEXT:    addl %ecx, %eax
+; AVX2-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX2-FAST-NEXT:    vzeroupper
+; AVX2-FAST-NEXT:    retq
+;
+; AVX512-SLOW-LABEL: extract_extract89_v16i16_add_i16:
+; AVX512-SLOW:       # %bb.0:
+; AVX512-SLOW-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX512-SLOW-NEXT:    vmovd %xmm0, %ecx
+; AVX512-SLOW-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX512-SLOW-NEXT:    addl %ecx, %eax
+; AVX512-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX512-SLOW-NEXT:    vzeroupper
+; AVX512-SLOW-NEXT:    retq
+;
+; AVX512-FAST-LABEL: extract_extract89_v16i16_add_i16:
+; AVX512-FAST:       # %bb.0:
+; AVX512-FAST-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX512-FAST-NEXT:    vmovd %xmm0, %ecx
+; AVX512-FAST-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX512-FAST-NEXT:    addl %ecx, %eax
+; AVX512-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX512-FAST-NEXT:    vzeroupper
+; AVX512-FAST-NEXT:    retq
+  %x0 = extractelement <16 x i16> %x, i32 8
+  %x1 = extractelement <16 x i16> %x, i32 9
+  %x01 = add i16 %x0, %x1
+  ret i16 %x01
+}
+
 define i16 @extract_extract01_v16i16_add_i16_commute(<16 x i16> %x) {
 ; SSE3-SLOW-LABEL: extract_extract01_v16i16_add_i16_commute:
 ; SSE3-SLOW:       # %bb.0:
@@ -676,6 +805,87 @@ define i16 @extract_extract45_v16i16_add_i16_commute(<16 x i16> %x) {
   ret i16 %x01
 }
 
+define i16 @extract_extract89_v16i16_add_i16_commute(<16 x i16> %x) {
+; SSE3-SLOW-LABEL: extract_extract89_v16i16_add_i16_commute:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movd %xmm1, %ecx
+; SSE3-SLOW-NEXT:    pextrw $1, %xmm1, %eax
+; SSE3-SLOW-NEXT:    addl %ecx, %eax
+; SSE3-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
+; SSE3-SLOW-NEXT:    retq
+;
+; SSE3-FAST-LABEL: extract_extract89_v16i16_add_i16_commute:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    phaddw %xmm1, %xmm1
+; SSE3-FAST-NEXT:    movd %xmm1, %eax
+; SSE3-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
+; SSE3-FAST-NEXT:    retq
+;
+; AVX1-SLOW-LABEL: extract_extract89_v16i16_add_i16_commute:
+; AVX1-SLOW:       # %bb.0:
+; AVX1-SLOW-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-SLOW-NEXT:    vmovd %xmm0, %ecx
+; AVX1-SLOW-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX1-SLOW-NEXT:    addl %ecx, %eax
+; AVX1-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX1-SLOW-NEXT:    vzeroupper
+; AVX1-SLOW-NEXT:    retq
+;
+; AVX1-FAST-LABEL: extract_extract89_v16i16_add_i16_commute:
+; AVX1-FAST:       # %bb.0:
+; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-FAST-NEXT:    vmovd %xmm0, %ecx
+; AVX1-FAST-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX1-FAST-NEXT:    addl %ecx, %eax
+; AVX1-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX1-FAST-NEXT:    vzeroupper
+; AVX1-FAST-NEXT:    retq
+;
+; AVX2-SLOW-LABEL: extract_extract89_v16i16_add_i16_commute:
+; AVX2-SLOW:       # %bb.0:
+; AVX2-SLOW-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX2-SLOW-NEXT:    vmovd %xmm0, %ecx
+; AVX2-SLOW-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX2-SLOW-NEXT:    addl %ecx, %eax
+; AVX2-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX2-SLOW-NEXT:    vzeroupper
+; AVX2-SLOW-NEXT:    retq
+;
+; AVX2-FAST-LABEL: extract_extract89_v16i16_add_i16_commute:
+; AVX2-FAST:       # %bb.0:
+; AVX2-FAST-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX2-FAST-NEXT:    vmovd %xmm0, %ecx
+; AVX2-FAST-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX2-FAST-NEXT:    addl %ecx, %eax
+; AVX2-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX2-FAST-NEXT:    vzeroupper
+; AVX2-FAST-NEXT:    retq
+;
+; AVX512-SLOW-LABEL: extract_extract89_v16i16_add_i16_commute:
+; AVX512-SLOW:       # %bb.0:
+; AVX512-SLOW-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX512-SLOW-NEXT:    vmovd %xmm0, %ecx
+; AVX512-SLOW-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX512-SLOW-NEXT:    addl %ecx, %eax
+; AVX512-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX512-SLOW-NEXT:    vzeroupper
+; AVX512-SLOW-NEXT:    retq
+;
+; AVX512-FAST-LABEL: extract_extract89_v16i16_add_i16_commute:
+; AVX512-FAST:       # %bb.0:
+; AVX512-FAST-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX512-FAST-NEXT:    vmovd %xmm0, %ecx
+; AVX512-FAST-NEXT:    vpextrw $1, %xmm0, %eax
+; AVX512-FAST-NEXT:    addl %ecx, %eax
+; AVX512-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX512-FAST-NEXT:    vzeroupper
+; AVX512-FAST-NEXT:    retq
+  %x0 = extractelement <16 x i16> %x, i32 8
+  %x1 = extractelement <16 x i16> %x, i32 9
+  %x01 = add i16 %x1, %x0
+  ret i16 %x01
+}
+
 define i32 @extract_extract01_v8i32_sub_i32(<8 x i32> %x) {
 ; SSE3-SLOW-LABEL: extract_extract01_v8i32_sub_i32:
 ; SSE3-SLOW:       # %bb.0:
@@ -734,6 +944,30 @@ define i32 @extract_extract23_v8i32_sub_i32(<8 x i32> %x) {
   ret i32 %x01
 }
 
+define i32 @extract_extract67_v8i32_sub_i32(<8 x i32> %x) {
+; SSE3-LABEL: extract_extract67_v8i32_sub_i32:
+; SSE3:       # %bb.0:
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
+; SSE3-NEXT:    movd %xmm0, %eax
+; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[3,1,2,3]
+; SSE3-NEXT:    movd %xmm0, %ecx
+; SSE3-NEXT:    subl %ecx, %eax
+; SSE3-NEXT:    retq
+;
+; AVX-LABEL: extract_extract67_v8i32_sub_i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vextractps $2, %xmm0, %eax
+; AVX-NEXT:    vextractps $3, %xmm0, %ecx
+; AVX-NEXT:    subl %ecx, %eax
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    retq
+  %x0 = extractelement <8 x i32> %x, i32 6
+  %x1 = extractelement <8 x i32> %x, i32 7
+  %x01 = sub i32 %x0, %x1
+  ret i32 %x01
+}
+
 ; Negative test...or get hoppy and negate?
 
 define i32 @extract_extract01_v8i32_sub_i32_commute(<8 x i32> %x) {