[X86] Fix uops counts for SLM extract/extract-store instructions

Matches Intel AoM + Agner

diff --git a/llvm/lib/Target/X86/X86ScheduleSLM.td b/llvm/lib/Target/X86/X86ScheduleSLM.td
index 598e567..9025914 100644 (file)
--- a/llvm/lib/Target/X86/X86ScheduleSLM.td
+++ b/llvm/lib/Target/X86/X86ScheduleSLM.td
@@ -408,10 +408,12 @@ defm : X86WriteResPairUnsupported<WriteVPMOV256>;
 // Vector insert/extract operations.
 defm : SLMWriteResPair<WriteVecInsert, [SLM_FPC_RSV0],  1>;
 
-def  : WriteRes<WriteVecExtract, [SLM_FPC_RSV0]>;
+def  : WriteRes<WriteVecExtract, [SLM_FPC_RSV0]> {
+  let NumMicroOps = 2;
+}
 def  : WriteRes<WriteVecExtractSt, [SLM_FPC_RSV0, SLM_MEC_RSV]> {
   let Latency = 4;
-  let NumMicroOps = 2;
+  let NumMicroOps = 5;
   let ResourceCycles = [1, 2];
 }
 

diff --git a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse1.s b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse1.s
index e7554b9..a84a325 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse1.s
+++ b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse1.s
@@ -268,7 +268,7 @@ xorps       (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   pavgb  (%rax), %mm2
 # CHECK-NEXT:  1      1     0.50                        pavgw  %mm0, %mm2
 # CHECK-NEXT:  1      4     1.00    *                   pavgw  (%rax), %mm2
-# CHECK-NEXT:  1      1     1.00                        pextrw $1, %mm0, %ecx
+# CHECK-NEXT:  2      1     1.00                        pextrw $1, %mm0, %ecx
 # CHECK-NEXT:  1      1     1.00                        pinsrw $1, %eax, %mm2
 # CHECK-NEXT:  1      4     1.00    *                   pinsrw $1, (%rax), %mm2
 # CHECK-NEXT:  1      1     0.50                        pmaxsw %mm0, %mm2

diff --git a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s
index fcf9550..3c699a6 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s
+++ b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s
@@ -560,7 +560,7 @@ xorpd       (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   pcmpgtd        (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        pcmpgtw        %xmm0, %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   pcmpgtw        (%rax), %xmm2
-# CHECK-NEXT:  1      1     1.00                        pextrw $1, %xmm0, %ecx
+# CHECK-NEXT:  2      1     1.00                        pextrw $1, %xmm0, %ecx
 # CHECK-NEXT:  1      1     1.00                        pinsrw $1, %eax, %xmm0
 # CHECK-NEXT:  1      4     1.00    *                   pinsrw $1, (%rax), %xmm0
 # CHECK-NEXT:  1      5     2.00                        pmaddwd        %xmm0, %xmm2

diff --git a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s
index 3e4ece5..4b82cbc 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s
+++ b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s
@@ -167,8 +167,8 @@ roundss     $1, (%rax), %xmm2
 # CHECK-NEXT:  6      15    8.00    *                   dppd   $22, (%rax), %xmm2
 # CHECK-NEXT:  9      15    12.00                       dpps   $22, %xmm0, %xmm2
 # CHECK-NEXT:  10     18    12.00   *                   dpps   $22, (%rax), %xmm2
-# CHECK-NEXT:  1      1     1.00                        extractps      $1, %xmm0, %ecx
-# CHECK-NEXT:  2      4     2.00           *            extractps      $1, %xmm0, (%rax)
+# CHECK-NEXT:  2      1     1.00                        extractps      $1, %xmm0, %ecx
+# CHECK-NEXT:  5      4     2.00           *            extractps      $1, %xmm0, (%rax)
 # CHECK-NEXT:  1      1     1.00                        insertps       $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   insertps       $1, (%rax), %xmm2
 # CHECK-NEXT:  1      3     1.00    *                   movntdqa       (%rax), %xmm2
@@ -182,13 +182,13 @@ roundss     $1, (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   pblendw        $11, (%rax), %xmm2
 # CHECK-NEXT:  2      4     4.00                        pcmpeqq        %xmm0, %xmm2
 # CHECK-NEXT:  3      7     4.00    *                   pcmpeqq        (%rax), %xmm2
-# CHECK-NEXT:  1      1     1.00                        pextrb $1, %xmm0, %ecx
-# CHECK-NEXT:  2      4     2.00           *            pextrb $1, %xmm0, (%rax)
-# CHECK-NEXT:  1      1     1.00                        pextrd $1, %xmm0, %ecx
-# CHECK-NEXT:  2      4     2.00           *            pextrd $1, %xmm0, (%rax)
-# CHECK-NEXT:  1      1     1.00                        pextrq $1, %xmm0, %rcx
-# CHECK-NEXT:  2      4     2.00           *            pextrq $1, %xmm0, (%rax)
-# CHECK-NEXT:  2      4     2.00           *            pextrw $1, %xmm0, (%rax)
+# CHECK-NEXT:  2      1     1.00                        pextrb $1, %xmm0, %ecx
+# CHECK-NEXT:  5      4     2.00           *            pextrb $1, %xmm0, (%rax)
+# CHECK-NEXT:  2      1     1.00                        pextrd $1, %xmm0, %ecx
+# CHECK-NEXT:  5      4     2.00           *            pextrd $1, %xmm0, (%rax)
+# CHECK-NEXT:  2      1     1.00                        pextrq $1, %xmm0, %rcx
+# CHECK-NEXT:  5      4     2.00           *            pextrq $1, %xmm0, (%rax)
+# CHECK-NEXT:  5      4     2.00           *            pextrw $1, %xmm0, (%rax)
 # CHECK-NEXT:  1      4     1.00                        phminposuw     %xmm0, %xmm2
 # CHECK-NEXT:  1      7     1.00    *                   phminposuw     (%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        pinsrb $1, %eax, %xmm1