From 8e7f6039fa059268f9355bbcd093b284ee656f82 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Thu, 30 Sep 2021 12:28:02 +0100
Subject: [PATCH] [X86] Atom SSE shift-by-variable take 2uops/3uops not 1uop

Based off the most recent llvm-exegesis captures (PR36895) and what Intel AoM / Agner / InstLatX64 reports as well.
---
 llvm/lib/Target/X86/X86ScheduleAtom.td        |  4 +--
 .../tools/llvm-mca/X86/Atom/resources-mmx.s   | 32 +++++++++----------
 .../tools/llvm-mca/X86/Atom/resources-sse2.s  | 32 +++++++++----------
 3 files changed, 34 insertions(+), 34 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ScheduleAtom.td b/llvm/lib/Target/X86/X86ScheduleAtom.td
index 1708226c6083..9f41b45140db 100644
--- a/llvm/lib/Target/X86/X86ScheduleAtom.td
+++ b/llvm/lib/Target/X86/X86ScheduleAtom.td
@@ -396,8 +396,8 @@ defm : X86WriteResPairUnsupported<WriteVecLogicZ>;
 defm : X86WriteResPairUnsupported<WriteVecTest>;
 defm : X86WriteResPairUnsupported<WriteVecTestY>;
 defm : X86WriteResPairUnsupported<WriteVecTestZ>;
-defm : AtomWriteResPair<WriteVecShift,     [AtomPort0,AtomPort1], [AtomPort0,AtomPort1], 2, 3, [1,1], [2,2]>;
-defm : AtomWriteResPair<WriteVecShiftX,    [AtomPort0,AtomPort1], [AtomPort0,AtomPort1], 2, 3, [1,1], [2,2]>;
+defm : AtomWriteResPair<WriteVecShift,     [AtomPort0,AtomPort1], [AtomPort0,AtomPort1], 2, 3, [1,1], [2,2], 2, 3>;
+defm : AtomWriteResPair<WriteVecShiftX,    [AtomPort0,AtomPort1], [AtomPort0,AtomPort1], 2, 3, [1,1], [2,2], 2, 3>;
 defm : X86WriteResPairUnsupported<WriteVecShiftY>;
 defm : X86WriteResPairUnsupported<WriteVecShiftZ>;
 defm : AtomWriteResPair<WriteVecShiftImm,   [AtomPort0],  [AtomPort0], 1, 1>;
diff --git a/llvm/test/tools/llvm-mca/X86/Atom/resources-mmx.s b/llvm/test/tools/llvm-mca/X86/Atom/resources-mmx.s
index bc929e14fc7f..bdefbe07b0bc 100644
--- a/llvm/test/tools/llvm-mca/X86/Atom/resources-mmx.s
+++ b/llvm/test/tools/llvm-mca/X86/Atom/resources-mmx.s
@@ -218,29 +218,29 @@ pxor        (%rax), %mm2
 # CHECK-NEXT:  1      1     0.50                        por	%mm0, %mm2
 # CHECK-NEXT:  1      1     1.00    *                   por	(%rax), %mm2
 # CHECK-NEXT:  1      1     1.00                        pslld	$1, %mm2
-# CHECK-NEXT:  1      2     1.00                        pslld	%mm0, %mm2
-# CHECK-NEXT:  1      3     2.00    *                   pslld	(%rax), %mm2
+# CHECK-NEXT:  2      2     1.00                        pslld	%mm0, %mm2
+# CHECK-NEXT:  3      3     2.00    *                   pslld	(%rax), %mm2
 # CHECK-NEXT:  1      1     1.00                        psllq	$1, %mm2
-# CHECK-NEXT:  1      2     1.00                        psllq	%mm0, %mm2
-# CHECK-NEXT:  1      3     2.00    *                   psllq	(%rax), %mm2
+# CHECK-NEXT:  2      2     1.00                        psllq	%mm0, %mm2
+# CHECK-NEXT:  3      3     2.00    *                   psllq	(%rax), %mm2
 # CHECK-NEXT:  1      1     1.00                        psllw	$1, %mm2
-# CHECK-NEXT:  1      2     1.00                        psllw	%mm0, %mm2
-# CHECK-NEXT:  1      3     2.00    *                   psllw	(%rax), %mm2
+# CHECK-NEXT:  2      2     1.00                        psllw	%mm0, %mm2
+# CHECK-NEXT:  3      3     2.00    *                   psllw	(%rax), %mm2
 # CHECK-NEXT:  1      1     1.00                        psrad	$1, %mm2
-# CHECK-NEXT:  1      2     1.00                        psrad	%mm0, %mm2
-# CHECK-NEXT:  1      3     2.00    *                   psrad	(%rax), %mm2
+# CHECK-NEXT:  2      2     1.00                        psrad	%mm0, %mm2
+# CHECK-NEXT:  3      3     2.00    *                   psrad	(%rax), %mm2
 # CHECK-NEXT:  1      1     1.00                        psraw	$1, %mm2
-# CHECK-NEXT:  1      2     1.00                        psraw	%mm0, %mm2
-# CHECK-NEXT:  1      3     2.00    *                   psraw	(%rax), %mm2
+# CHECK-NEXT:  2      2     1.00                        psraw	%mm0, %mm2
+# CHECK-NEXT:  3      3     2.00    *                   psraw	(%rax), %mm2
 # CHECK-NEXT:  1      1     1.00                        psrld	$1, %mm2
-# CHECK-NEXT:  1      2     1.00                        psrld	%mm0, %mm2
-# CHECK-NEXT:  1      3     2.00    *                   psrld	(%rax), %mm2
+# CHECK-NEXT:  2      2     1.00                        psrld	%mm0, %mm2
+# CHECK-NEXT:  3      3     2.00    *                   psrld	(%rax), %mm2
 # CHECK-NEXT:  1      1     1.00                        psrlq	$1, %mm2
-# CHECK-NEXT:  1      2     1.00                        psrlq	%mm0, %mm2
-# CHECK-NEXT:  1      3     2.00    *                   psrlq	(%rax), %mm2
+# CHECK-NEXT:  2      2     1.00                        psrlq	%mm0, %mm2
+# CHECK-NEXT:  3      3     2.00    *                   psrlq	(%rax), %mm2
 # CHECK-NEXT:  1      1     1.00                        psrlw	$1, %mm2
-# CHECK-NEXT:  1      2     1.00                        psrlw	%mm0, %mm2
-# CHECK-NEXT:  1      3     2.00    *                   psrlw	(%rax), %mm2
+# CHECK-NEXT:  2      2     1.00                        psrlw	%mm0, %mm2
+# CHECK-NEXT:  3      3     2.00    *                   psrlw	(%rax), %mm2
 # CHECK-NEXT:  1      1     0.50                        psubb	%mm0, %mm2
 # CHECK-NEXT:  1      1     1.00    *                   psubb	(%rax), %mm2
 # CHECK-NEXT:  1      1     0.50                        psubd	%mm0, %mm2
diff --git a/llvm/test/tools/llvm-mca/X86/Atom/resources-sse2.s b/llvm/test/tools/llvm-mca/X86/Atom/resources-sse2.s
index 36b4c754e998..977c94f77b84 100644
--- a/llvm/test/tools/llvm-mca/X86/Atom/resources-sse2.s
+++ b/llvm/test/tools/llvm-mca/X86/Atom/resources-sse2.s
@@ -595,31 +595,31 @@ xorpd       (%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        pshuflw	$1, %xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00    *                   pshuflw	$1, (%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        pslld	$1, %xmm2
-# CHECK-NEXT:  1      2     1.00                        pslld	%xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00    *                   pslld	(%rax), %xmm2
+# CHECK-NEXT:  2      2     1.00                        pslld	%xmm0, %xmm2
+# CHECK-NEXT:  3      3     2.00    *                   pslld	(%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        pslldq	$1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psllq	$1, %xmm2
-# CHECK-NEXT:  1      2     1.00                        psllq	%xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00    *                   psllq	(%rax), %xmm2
+# CHECK-NEXT:  2      2     1.00                        psllq	%xmm0, %xmm2
+# CHECK-NEXT:  3      3     2.00    *                   psllq	(%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        psllw	$1, %xmm2
-# CHECK-NEXT:  1      2     1.00                        psllw	%xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00    *                   psllw	(%rax), %xmm2
+# CHECK-NEXT:  2      2     1.00                        psllw	%xmm0, %xmm2
+# CHECK-NEXT:  3      3     2.00    *                   psllw	(%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrad	$1, %xmm2
-# CHECK-NEXT:  1      2     1.00                        psrad	%xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00    *                   psrad	(%rax), %xmm2
+# CHECK-NEXT:  2      2     1.00                        psrad	%xmm0, %xmm2
+# CHECK-NEXT:  3      3     2.00    *                   psrad	(%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        psraw	$1, %xmm2
-# CHECK-NEXT:  1      2     1.00                        psraw	%xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00    *                   psraw	(%rax), %xmm2
+# CHECK-NEXT:  2      2     1.00                        psraw	%xmm0, %xmm2
+# CHECK-NEXT:  3      3     2.00    *                   psraw	(%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrld	$1, %xmm2
-# CHECK-NEXT:  1      2     1.00                        psrld	%xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00    *                   psrld	(%rax), %xmm2
+# CHECK-NEXT:  2      2     1.00                        psrld	%xmm0, %xmm2
+# CHECK-NEXT:  3      3     2.00    *                   psrld	(%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrldq	$1, %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrlq	$1, %xmm2
-# CHECK-NEXT:  1      2     1.00                        psrlq	%xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00    *                   psrlq	(%rax), %xmm2
+# CHECK-NEXT:  2      2     1.00                        psrlq	%xmm0, %xmm2
+# CHECK-NEXT:  3      3     2.00    *                   psrlq	(%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        psrlw	$1, %xmm2
-# CHECK-NEXT:  1      2     1.00                        psrlw	%xmm0, %xmm2
-# CHECK-NEXT:  1      3     2.00    *                   psrlw	(%rax), %xmm2
+# CHECK-NEXT:  2      2     1.00                        psrlw	%xmm0, %xmm2
+# CHECK-NEXT:  3      3     2.00    *                   psrlw	(%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        psubb	%xmm0, %xmm2
 # CHECK-NEXT:  1      1     1.00    *                   psubb	(%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        psubd	%xmm0, %xmm2