[NVPTX] Fix ISel patterns for i1 sint_to_fp (#110866)

NVPTX has ZeroOrNegativeOneBooleanContent, therefore we need to use -1 as the constant for i1 sint_to_fp operations in instruction selection.
llvm · Oct 8, 2024 · cc5ddae · cc5ddae
1 parent 5589096
commit cc5ddae
Show file tree

Hide file tree

Showing 2 changed files with 35 additions and 16 deletions.
diff --git a/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td b/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td
@@ -3126,11 +3126,12 @@ foreach ta = [v2f16, v2bf16, v2i16, v4i8, i32] in {
 
 // NOTE: pred->fp are currently sub-optimal due to an issue in TableGen where
 // we cannot specify floating-point literals in isel patterns.  Therefore, we
-// use an integer selp to select either 1 or 0 and then cvt to floating-point.
+// use an integer selp to select either 1 (or -1 in case of signed) or 0
+// and then cvt to floating-point.
 
 // sint -> f16
 def : Pat<(f16 (sint_to_fp Int1Regs:$a)),
-          (CVT_f16_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
+          (CVT_f16_s32 (SELP_s32ii -1, 0, Int1Regs:$a), CvtRN)>;
 def : Pat<(f16 (sint_to_fp Int16Regs:$a)),
           (CVT_f16_s16 Int16Regs:$a, CvtRN)>;
 def : Pat<(f16 (sint_to_fp Int32Regs:$a)),
@@ -3170,7 +3171,7 @@ def : Pat<(bf16 (uint_to_fp Int64Regs:$a)),
 
 // sint -> f32
 def : Pat<(f32 (sint_to_fp Int1Regs:$a)),
-          (CVT_f32_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
+          (CVT_f32_s32 (SELP_s32ii -1, 0, Int1Regs:$a), CvtRN)>;
 def : Pat<(f32 (sint_to_fp Int16Regs:$a)),
           (CVT_f32_s16 Int16Regs:$a, CvtRN)>;
 def : Pat<(f32 (sint_to_fp Int32Regs:$a)),
@@ -3190,7 +3191,7 @@ def : Pat<(f32 (uint_to_fp Int64Regs:$a)),
 
 // sint -> f64
 def : Pat<(f64 (sint_to_fp Int1Regs:$a)),
-          (CVT_f64_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
+          (CVT_f64_s32 (SELP_s32ii -1, 0, Int1Regs:$a), CvtRN)>;
 def : Pat<(f64 (sint_to_fp Int16Regs:$a)),
           (CVT_f64_s16 Int16Regs:$a, CvtRN)>;
 def : Pat<(f64 (sint_to_fp Int32Regs:$a)),

diff --git a/llvm/test/CodeGen/NVPTX/i1-int-to-fp.ll b/llvm/test/CodeGen/NVPTX/i1-int-to-fp.ll
@@ -2,37 +2,55 @@
 ; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 | %ptxas-verify %}
 
 ; CHECK-LABEL: foo
-; CHECK: setp
-; CHECK: selp
-; CHECK: cvt.rn.f32.u32
+; CHECK: setp.eq.b16 %[[P:p[0-9]+]], %{{.*}}, 1;
+; CHECK: selp.u32 %[[R:r[0-9]+]], 1, 0, %[[P]];
+; CHECK: cvt.rn.f32.u32 %f{{.*}}, %[[R]]
 define float @foo(i1 %a) {
   %ret = uitofp i1 %a to float
   ret float %ret
 }
 
 ; CHECK-LABEL: foo2
-; CHECK: setp
-; CHECK: selp
-; CHECK: cvt.rn.f32.s32
+; CHECK: setp.eq.b16 %[[P:p[0-9]+]], %{{.*}}, 1;
+; CHECK: selp.s32 %[[R:r[0-9]+]], -1, 0, %[[P]];
+; CHECK: cvt.rn.f32.s32 %f{{.*}}, %[[R]]
 define float @foo2(i1 %a) {
   %ret = sitofp i1 %a to float
   ret float %ret
 }
 
 ; CHECK-LABEL: foo3
-; CHECK: setp
-; CHECK: selp
-; CHECK: cvt.rn.f64.u32
+; CHECK: setp.eq.b16 %[[P:p[0-9]+]], %{{.*}}, 1;
+; CHECK: selp.u32 %[[R:r[0-9]+]], 1, 0, %[[P]];
+; CHECK: cvt.rn.f64.u32 %fd{{.*}}, %[[R]]
 define double @foo3(i1 %a) {
   %ret = uitofp i1 %a to double
   ret double %ret
 }
 
 ; CHECK-LABEL: foo4
-; CHECK: setp
-; CHECK: selp
-; CHECK: cvt.rn.f64.s32
+; CHECK: setp.eq.b16 %[[P:p[0-9]+]], %{{.*}}, 1;
+; CHECK: selp.s32 %[[R:r[0-9]+]], -1, 0, %[[P]];
+; CHECK: cvt.rn.f64.s32 %fd{{.*}}, %[[R]]
 define double @foo4(i1 %a) {
   %ret = sitofp i1 %a to double
   ret double %ret
 }
+
+; CHECK-LABEL: foo5
+; CHECK: setp.eq.b16 %[[P:p[0-9]+]], %{{.*}}, 1;
+; CHECK: selp.u32 %[[R:r[0-9]+]], 1, 0, %[[P]];
+; CHECK: cvt.rn.f16.u32 %{{.*}}, %[[R]]
+define half @foo5(i1 %a) {
+  %ret = uitofp i1 %a to half
+  ret half %ret
+}
+
+; CHECK-LABEL: foo6
+; CHECK: setp.eq.b16 %[[P:p[0-9]+]], %{{.*}}, 1;
+; CHECK: selp.s32 %[[R:r[0-9]+]], -1, 0, %[[P]];
+; CHECK: cvt.rn.f16.s32 %{{.*}}, %[[R]]
+define half @foo6(i1 %a) {
+  %ret = sitofp i1 %a to half
+  ret half %ret
+}