[llvm] Optimize fptrunc(x)>=C1 --> x>=C2 (PR #99475)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,169 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=instcombine -S < %s | FileCheck %s
+
+
+define i1 @fcmp_trunc(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp oge double [[TMP0]], 0x4058FFFFF0000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_ult(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_ult(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp ult double [[TMP0]], 0x4068FFFFF0000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp ult float %trunc, 2.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_ole(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_ole(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp ole double [[TMP0]], 0x4072C00010000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp ole float %trunc, 3.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_ogt(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_ogt(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp ogt double [[TMP0]], 0x4079000010000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp ogt float %trunc, 4.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_zero(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_zero(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp oge double [[TMP0]], 0xB690000000000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp oge float %trunc, 0.000000
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_with_nnan(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_with_nnan(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp nnan oge double [[TMP0]], 0x4058FFFFF0000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp nnan oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_with_ninf(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_with_ninf(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp ninf oge double [[TMP0]], 0x4058FFFFF0000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp ninf oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_with_nsz(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_with_nsz(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp nsz oge double [[TMP0]], 0x4058FFFFF0000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp nsz oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_with_reassoc(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_with_reassoc(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp reassoc oge double [[TMP0]], 0x4058FFFFF0000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp reassoc oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_with_fast(double %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_with_fast(
+; CHECK-SAME: double [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp fast oge double [[TMP0]], 0x4058FFFFF0000000 
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc double %0 to float
+  %result = fcmp fast oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
+define <4 x i1> @fcmp_vec_trunc(<4 x double> %0) {
+; CHECK-LABEL: define <4 x i1> @fcmp_vec_trunc(
+; CHECK-SAME: <4 x double> [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp olt <4 x double> [[TMP0]], <double 0x3FEFFFFFF0000000, double 0x3FEFFFFFF0000000, double 0x3FEFFFFFF0000000, double 0x3FEFFFFFF0000000>
+; CHECK-NEXT:    ret <4 x i1> [[CMP]]
+;
+  %vec = fptrunc <4 x double> %0 to <4 x float>
+  %cmp = fcmp olt <4 x float> %vec, <float 1.0, float 1.0, float 1.0, float 1.0>
+  ret <4 x i1> %cmp
+}
+
+define <1 x i1> @fcmp_vec_trunc_scalar(<1 x double> %0) {
+; CHECK-LABEL: define <1 x i1> @fcmp_vec_trunc_scalar(
+; CHECK-SAME: <1 x double> [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp fast olt <1 x double> [[TMP0]], <double 0x3FEFFFFFF0000000>
+; CHECK-NEXT:    ret <1 x i1> [[CMP]]
+;
+  %vec = fptrunc <1 x double> %0 to <1 x float>
+  %cmp = fcmp fast olt <1 x float> %vec, <float 1.0>
+  ret <1 x i1> %cmp
+}
+
+define i1 @fcmp_trunc_fp128(fp128 %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_fp128(
+; CHECK-SAME: fp128 [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp fast oge fp128 [[TMP0]], 0xL000000000000000040058FFFFF000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc fp128 %0 to float
+  %result = fcmp fast oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_x86_fp80(x86_fp80 %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_x86_fp80(
+; CHECK-SAME: x86_fp80 [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp fast oge x86_fp80 [[TMP0]], 0xK4005C7FFFF8000000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc x86_fp80 %0 to float
+  %result = fcmp fast oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
+define i1 @fcmp_trunc_ppc_fp128(ppc_fp128 %0) {
+; CHECK-LABEL: define i1 @fcmp_trunc_ppc_fp128(
+; CHECK-SAME: ppc_fp128 [[TMP0:%.*]]) {
+; CHECK-NEXT:    [[RESULT:%.*]] = fcmp fast oge ppc_fp128 [[TMP0]], 0xM4058FFFFF0000000BD00000000000000
+; CHECK-NEXT:    ret i1 [[RESULT]]
+;
+  %trunc = fptrunc ppc_fp128 %0 to float
+  %result = fcmp fast oge float %trunc, 1.000000e+02
+  ret i1 %result
+}
+
----------------
arsenm wrote:

The set of tested constants seems too simple for the complexity of the loop testing constant validity. Should have negative tests for off by one bit in each direction. Also test with the edge case constants (inf, nan) and some denormal values? 

https://github.com/llvm/llvm-project/pull/99475