[llvm-branch-commits] [llvm] release/18.x: [AArch64][GISEL] Consider fcmp true and fcmp false in cond code selection (#86972) (PR #91580)

Marc Auberer via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Thu May 9 03:10:12 PDT 2024


https://github.com/marcauberer created https://github.com/llvm/llvm-project/pull/91580

Backport c482fad2c1de367f8fef2b40361dec00523707f7

This cherry-pick includes some ref fixes due to the incompatibility between `main` and `release/18.x`.

>From 39f3ca9677a63ddc404e7ee13edace2bbf72fc93 Mon Sep 17 00:00:00 2001
From: Marc Auberer <marc.auberer at chillibits.com>
Date: Thu, 28 Mar 2024 23:08:38 +0100
Subject: [PATCH 1/2] [AArch64][GISEL] Consider fcmp true and fcmp false in
 cond code selection (#86972)

Fixes #86917

`FCMP_TRUE` and `FCMP_FALSE` were previously not considered and we ended
up in an llvm_unreachable assertion.
---
 .../AArch64/GISel/AArch64GlobalISelUtils.cpp  |   6 ++
 .../CodeGen/AArch64/GlobalISel/select.mir     |  20 ++++
 .../AArch64/neon-compare-instructions.ll      | 101 ++++++++++++++++++
 3 files changed, 127 insertions(+)

diff --git a/llvm/lib/Target/AArch64/GISel/AArch64GlobalISelUtils.cpp b/llvm/lib/Target/AArch64/GISel/AArch64GlobalISelUtils.cpp
index 92db89cc0915b..80fe4bcb8b58f 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64GlobalISelUtils.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64GlobalISelUtils.cpp
@@ -147,6 +147,12 @@ void AArch64GISelUtils::changeFCMPPredToAArch64CC(
   case CmpInst::FCMP_UNE:
     CondCode = AArch64CC::NE;
     break;
+  case CmpInst::FCMP_TRUE:
+    CondCode = AArch64CC::AL;
+    break;
+  case CmpInst::FCMP_FALSE:
+    CondCode = AArch64CC::NV;
+    break;
   }
 }
 
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/select.mir b/llvm/test/CodeGen/AArch64/GlobalISel/select.mir
index 60cddbf794bc7..ae78d4be0f88a 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/select.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/select.mir
@@ -183,6 +183,14 @@ registers:
   - { id: 5, class: gpr }
   - { id: 6, class: gpr }
   - { id: 7, class: gpr }
+  - { id: 8, class: fpr }
+  - { id: 9, class: gpr }
+  - { id: 10, class: fpr }
+  - { id: 11, class: gpr }
+  - { id: 12, class: gpr }
+  - { id: 13, class: gpr }
+  - { id: 14, class: gpr }
+  - { id: 15, class: gpr }
 
 # CHECK:  body:
 # CHECK:    nofpexcept FCMPSrr %0, %0, implicit-def $nzcv
@@ -209,6 +217,18 @@ body:             |
     %7(s32) = G_ANYEXT %5
     $w0 = COPY %7(s32)
 
+    %8(s32) = COPY $s0
+    %9(s32) = G_FCMP floatpred(true), %8, %8
+    %12(s8) = G_TRUNC %9(s32)
+    %14(s32) = G_ANYEXT %12
+    $w0 = COPY %14(s32)
+
+    %10(s64) = COPY $d0
+    %11(s32) = G_FCMP floatpred(false), %10, %10
+    %13(s8) = G_TRUNC %11(s32)
+    %15(s32) = G_ANYEXT %13
+    $w0 = COPY %15(s32)
+
 ...
 
 ---
diff --git a/llvm/test/CodeGen/AArch64/neon-compare-instructions.ll b/llvm/test/CodeGen/AArch64/neon-compare-instructions.ll
index 765c81e26e13c..c4c00f8e97942 100644
--- a/llvm/test/CodeGen/AArch64/neon-compare-instructions.ll
+++ b/llvm/test/CodeGen/AArch64/neon-compare-instructions.ll
@@ -2870,6 +2870,107 @@ define <2 x i64> @fcmune2xdouble(<2 x double> %A, <2 x double> %B) {
   ret <2 x i64> %tmp4
 }
 
+define <2 x i32> @fcmal2xfloat(<2 x float> %A, <2 x float> %B) {
+; CHECK-SD-LABEL: fcmal2xfloat:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    movi v0.2d, #0xffffffffffffffff
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fcmal2xfloat:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    movi v0.2s, #1
+; CHECK-GI-NEXT:    shl v0.2s, v0.2s, #31
+; CHECK-GI-NEXT:    sshr v0.2s, v0.2s, #31
+; CHECK-GI-NEXT:    ret
+  %tmp3 = fcmp true <2 x float> %A, %B
+  %tmp4 = sext <2 x i1> %tmp3 to <2 x i32>
+  ret <2 x i32> %tmp4
+}
+
+define <4 x i32> @fcmal4xfloat(<4 x float> %A, <4 x float> %B) {
+; CHECK-SD-LABEL: fcmal4xfloat:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    movi v0.2d, #0xffffffffffffffff
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fcmal4xfloat:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    mov w8, #1 // =0x1
+; CHECK-GI-NEXT:    fmov s0, w8
+; CHECK-GI-NEXT:    mov v1.16b, v0.16b
+; CHECK-GI-NEXT:    mov v1.h[1], v0.h[0]
+; CHECK-GI-NEXT:    mov v0.h[1], v0.h[0]
+; CHECK-GI-NEXT:    ushll v1.4s, v1.4h, #0
+; CHECK-GI-NEXT:    ushll v0.4s, v0.4h, #0
+; CHECK-GI-NEXT:    mov v1.d[1], v0.d[0]
+; CHECK-GI-NEXT:    shl v0.4s, v1.4s, #31
+; CHECK-GI-NEXT:    sshr v0.4s, v0.4s, #31
+; CHECK-GI-NEXT:    ret
+  %tmp3 = fcmp true <4 x float> %A, %B
+  %tmp4 = sext <4 x i1> %tmp3 to <4 x i32>
+  ret <4 x i32> %tmp4
+}
+define <2 x i64> @fcmal2xdouble(<2 x double> %A, <2 x double> %B) {
+; CHECK-SD-LABEL: fcmal2xdouble:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    movi v0.2d, #0xffffffffffffffff
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fcmal2xdouble:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    adrp x8, .LCPI221_0
+; CHECK-GI-NEXT:    ldr q0, [x8, :lo12:.LCPI221_0]
+; CHECK-GI-NEXT:    shl v0.2d, v0.2d, #63
+; CHECK-GI-NEXT:    sshr v0.2d, v0.2d, #63
+; CHECK-GI-NEXT:    ret
+  %tmp3 = fcmp true <2 x double> %A, %B
+  %tmp4 = sext <2 x i1> %tmp3 to <2 x i64>
+  ret <2 x i64> %tmp4
+}
+
+define <2 x i32> @fcmnv2xfloat(<2 x float> %A, <2 x float> %B) {
+; CHECK-LABEL: fcmnv2xfloat:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    movi v0.2d, #0000000000000000
+; CHECK-NEXT:    ret
+  %tmp3 = fcmp false <2 x float> %A, %B
+  %tmp4 = sext <2 x i1> %tmp3 to <2 x i32>
+  ret <2 x i32> %tmp4
+}
+
+define <4 x i32> @fcmnv4xfloat(<4 x float> %A, <4 x float> %B) {
+; CHECK-SD-LABEL: fcmnv4xfloat:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    movi v0.2d, #0000000000000000
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fcmnv4xfloat:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    mov w8, #0 // =0x0
+; CHECK-GI-NEXT:    fmov s0, w8
+; CHECK-GI-NEXT:    mov v1.16b, v0.16b
+; CHECK-GI-NEXT:    mov v1.h[1], v0.h[0]
+; CHECK-GI-NEXT:    mov v0.h[1], v0.h[0]
+; CHECK-GI-NEXT:    ushll v1.4s, v1.4h, #0
+; CHECK-GI-NEXT:    ushll v0.4s, v0.4h, #0
+; CHECK-GI-NEXT:    mov v1.d[1], v0.d[0]
+; CHECK-GI-NEXT:    shl v0.4s, v1.4s, #31
+; CHECK-GI-NEXT:    sshr v0.4s, v0.4s, #31
+; CHECK-GI-NEXT:    ret
+  %tmp3 = fcmp false <4 x float> %A, %B
+  %tmp4 = sext <4 x i1> %tmp3 to <4 x i32>
+  ret <4 x i32> %tmp4
+}
+define <2 x i64> @fcmnv2xdouble(<2 x double> %A, <2 x double> %B) {
+; CHECK-LABEL: fcmnv2xdouble:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    movi v0.2d, #0000000000000000
+; CHECK-NEXT:    ret
+  %tmp3 = fcmp false <2 x double> %A, %B
+  %tmp4 = sext <2 x i1> %tmp3 to <2 x i64>
+  ret <2 x i64> %tmp4
+}
+
 define <2 x i32> @fcmoeqz2xfloat(<2 x float> %A) {
 ; CHECK-LABEL: fcmoeqz2xfloat:
 ; CHECK:       // %bb.0:

>From 66222e2555d730380d7b8cd31de9fa706058b0e5 Mon Sep 17 00:00:00 2001
From: Marc Auberer <marc.auberer at chillibits.com>
Date: Thu, 9 May 2024 12:06:59 +0200
Subject: [PATCH 2/2] Ref fixes

---
 .../AArch64/neon-compare-instructions.ll      | 24 +++++--------------
 1 file changed, 6 insertions(+), 18 deletions(-)

diff --git a/llvm/test/CodeGen/AArch64/neon-compare-instructions.ll b/llvm/test/CodeGen/AArch64/neon-compare-instructions.ll
index c4c00f8e97942..f398a7f8b8caa 100644
--- a/llvm/test/CodeGen/AArch64/neon-compare-instructions.ll
+++ b/llvm/test/CodeGen/AArch64/neon-compare-instructions.ll
@@ -2895,15 +2895,9 @@ define <4 x i32> @fcmal4xfloat(<4 x float> %A, <4 x float> %B) {
 ;
 ; CHECK-GI-LABEL: fcmal4xfloat:
 ; CHECK-GI:       // %bb.0:
-; CHECK-GI-NEXT:    mov w8, #1 // =0x1
-; CHECK-GI-NEXT:    fmov s0, w8
-; CHECK-GI-NEXT:    mov v1.16b, v0.16b
-; CHECK-GI-NEXT:    mov v1.h[1], v0.h[0]
-; CHECK-GI-NEXT:    mov v0.h[1], v0.h[0]
-; CHECK-GI-NEXT:    ushll v1.4s, v1.4h, #0
-; CHECK-GI-NEXT:    ushll v0.4s, v0.4h, #0
-; CHECK-GI-NEXT:    mov v1.d[1], v0.d[0]
-; CHECK-GI-NEXT:    shl v0.4s, v1.4s, #31
+; CHECK-GI-NEXT:    movi v0.2s, #1
+; CHECK-GI-NEXT:    mov v0.d[1], v0.d[0]
+; CHECK-GI-NEXT:    shl v0.4s, v0.4s, #31
 ; CHECK-GI-NEXT:    sshr v0.4s, v0.4s, #31
 ; CHECK-GI-NEXT:    ret
   %tmp3 = fcmp true <4 x float> %A, %B
@@ -2946,15 +2940,9 @@ define <4 x i32> @fcmnv4xfloat(<4 x float> %A, <4 x float> %B) {
 ;
 ; CHECK-GI-LABEL: fcmnv4xfloat:
 ; CHECK-GI:       // %bb.0:
-; CHECK-GI-NEXT:    mov w8, #0 // =0x0
-; CHECK-GI-NEXT:    fmov s0, w8
-; CHECK-GI-NEXT:    mov v1.16b, v0.16b
-; CHECK-GI-NEXT:    mov v1.h[1], v0.h[0]
-; CHECK-GI-NEXT:    mov v0.h[1], v0.h[0]
-; CHECK-GI-NEXT:    ushll v1.4s, v1.4h, #0
-; CHECK-GI-NEXT:    ushll v0.4s, v0.4h, #0
-; CHECK-GI-NEXT:    mov v1.d[1], v0.d[0]
-; CHECK-GI-NEXT:    shl v0.4s, v1.4s, #31
+; CHECK-GI-NEXT:    movi v0.2d, #0000000000000000
+; CHECK-GI-NEXT:    mov v0.d[1], v0.d[0]
+; CHECK-GI-NEXT:    shl v0.4s, v0.4s, #31
 ; CHECK-GI-NEXT:    sshr v0.4s, v0.4s, #31
 ; CHECK-GI-NEXT:    ret
   %tmp3 = fcmp false <4 x float> %A, %B



More information about the llvm-branch-commits mailing list